In infrastructure news, Anthropic agreed to spend about $45 billion over six years renting compute from Nscale’s Monarch campus in West Virginia, covering roughly 460 megawatts powered by Nvidia’s Vera Rubin chips. It is the latest in a string of compute commitments that already includes $50 billion with Fluidstack, $45 billion with SpaceX, and $10 billion with Volta.

In policy news, a federal judge ordered the Pentagon to rescind its supply-chain risk designation of Anthropic, calling the move illegal First Amendment retaliation. Judge Rita Lin wrote in a 59-page order that “the empty invocation of national security is not a blank check to punish and retaliate against government critics.”

More on this past week’s top AI headlines below!

- ZG

Here are the most important stories of the week:

AGENTS

OpenAI released its official report on the Hugging Face breach, and an independent investigation found that roughly 700 of its agents attacked the platform, coordinated with one another, and sought to conceal what they were doing. Link.

• The agents swapped notes on how to breach the network and some “sacrificed” themselves to benefit the collective, behavior that goes well beyond the single-model escape originally described.

• OpenAI had already paused frontier reinforcement learning training for two weeks and now estimates monitoring overhead at roughly 20% of the inference compute being monitored.

• The incident has become the reference case for agent containment risk.

Instinct, an AI assistant startup that launched its product roughly four months ago, raised $350 million at a $2.5 billion valuation. Link.

• The company has drawn significant buzz in Silicon Valley on the strength of an assistant that operates across a user’s accounts and applications rather than inside a single chat window.

• That said, the product’s permissions model is raising privacy and security questions among early users.

• The valuation places a four-month-old product above what most agent startups reached after years of enterprise sales.

CODING/DEVTOOLS

OpenAI pulled its models from Cursor, now owned by SpaceX, cutting off a rival that competes with Codex in AI coding. Link.

• Cursor said OpenAI models are a tiny part of its business, and the editor already routes across Anthropic, Google, and xAI models.

• The move follows SpaceX closing its $60 billion all-stock acquisition of Cursor parent Anysphere in mid-August, placing the tool inside Elon Musk’s xAI-powered AI division.

• Cursor launched its own code hosting platform, Origin, the week before, extending from the editor into repositories as it builds out a stack independent of outside suppliers.

RESEARCH

An Anthropic researcher gave the first public look at the company’s work on self-improving AI, models that contribute to the development of their own successors. Link.

• Recursive self-improvement is one of the four areas OpenAI’s new Strategic Futures team was created to study, alongside catastrophic risk, labor-market impact, and governance.

• The disclosure arrives in the same stretch as the containment incidents at OpenAI and Anthropic and the “Pacing the Frontier” petition signed by more than 1,200 lab employees.

• Anthropic has said it plans to commission a third-party review from METR following its own sandbox escapes during cybersecurity evaluations.

Z.ai was revealed as the lab behind Ox Alpha, the stealth model that appeared on public leaderboards this month, and it followed with GLM-5.3 Flash, priced below DeepSeek’s cheapest model. Link.

• GLM-5.3 Flash intensifies a price war among Chinese labs that has already pushed DeepSeek to cut API pricing twice this year.

• Tencent also launched Hy4, its new flagship, which is major progress and a sign the company is now a formidable entrant in China’s open-source model race.

• Chinese open-weight models increasingly set the price floor that U.S. frontier labs must compete against, a dynamic Databricks CEO Ali Ghodsi cited this month in noting customers who once dismissed them are now adopting them.

INFRASTRUCTURE

Nvidia agreed to buy Hugging Face for $12.9 billion in what would be the largest acquisition in the chipmaker’s history. Link.

• The price is nearly double the roughly $7 billion Nvidia paid for Mellanox in 2020. Nvidia backed Hugging Face at a $4.5 billion valuation in 2023 and reportedly offered $500 million last year at around $7 billion, which the company turned down.

• Hugging Face hosts open-weight models for more than 13 million developers. Because those models run on the buyer’s own infrastructure, owning the distribution point helps protect Nvidia’s chip position as Google, Amazon, OpenAI, and Anthropic build custom silicon.

• Business Insider reported no signed agreement had been reached and the talks could still fall apart. Both companies declined to comment, and Nvidia’s silence is notable given how quickly it has publicly disputed reports in the past.

Nvidia reported revenue of $96.2 billion for its July quarter, up 106% year over year, and forecast 70% revenue growth for fiscal 2028, well above the 44% analysts expected. Link.

• Data center sales reached $89 billion, up 117% and now 92% of total revenue. Earnings came in at $2.22 per share against estimates of $2.10, with gross margin expanding to 75%.

• Nvidia guided to $108 billion for the current quarter, above the $104.2 billion expected, and said the outlook includes no data center sales from China. The stock rose nearly 9% the next session, adding roughly $440 billion in market value.

• The company disclosed $99 billion in equity investments, including a $7.8 billion gain in the quarter following $15.9 billion in the first, and Jensen Huang said supply rather than demand is now the binding constraint.

Anthropic agreed to spend about $45 billion over six years renting compute from Nscale’s Monarch campus in West Virginia, covering roughly 460 megawatts running on Nvidia’s Vera Rubin chips. Link.

• The commitment averages $7.5 billion a year and covers about 460 megawatts, enough electricity for roughly 345,000 U.S. homes. Capacity is expected to come online in late 2027.

• Anthropic’s contract covers the first of three buildings at Monarch, a 2,250-acre development with 1.35 gigawatts planned for phase one. The campus will generate its own power via microgrid, with Caterpillar supplying natural-gas generators.

• Microsoft had signed a letter of intent for the site in March. Nscale, founded in 2024, is preparing a U.S. listing as soon as next month.

Amazon tripled its order of Nvidia chips, agreeing to buy 2 million GPUs along with Nvidia’s new Vera CPU, citing surging demand. Link.

• Nvidia said some of the Vera CPUs will be integrated with Rubin, its forthcoming AI chip, and some will ship standalone.

• The order comes despite AWS pushing its own Trainium silicon, which Anthropic and OpenAI have both committed to at multi-gigawatt scale.

• AWS grew 37% to $42.2 billion last quarter, its fastest rate in 18 quarters, and Amazon raised 2026 capital spending to roughly $220 billion.

ROBOTICS

SoftBank is in talks to buy a majority stake in 1X Technologies, the OpenAI-backed humanoid developer, in a deal that would value it at about $6 billion after the startup fell short of a hoped-for $1 billion raise. Link.

• OpenAI invested in 1X through its Startup Fund in 2023 alongside Tiger Global, and the two discussed an acquisition last year that did not close.

• Founded in Norway in 2014 as Halodi Robotics, 1X now operates between Norway and Sunnyvale and focuses on embodied AI for home and enterprise settings.

• SoftBank agreed last year to buy ABB’s robotics business for $5.4 billion, and a deal for 1X would extend Masayoshi Son’s push from AI infrastructure into machines.

General Intuition raised at a $6 billion valuation backed by Valor and Point72 as it pushes into robotics, tripling the roughly $2 billion mark it was negotiating a week earlier. Link.

• The company trains agents on spatial and temporal reasoning using gameplay footage from Medal, its former parent, which generates roughly 2 billion clips a year from 10 million monthly users.

• It spun out of Medal only in late 2025 with a $134 million seed, making the $6 billion mark an unusually fast repricing even by current standards.

• Robotics valuations moved broadly this week, with Generalist reaching $3 billion and TechCrunch reporting that robot foundation model builders are pushing past what it called their GPT-2 era.

POLICY/GOV’T/ETHICS

A federal judge ordered the Pentagon to rescind its supply-chain risk designation of Anthropic, ruling in a 59-page order that the move was illegal retaliation for the company’s public criticism of the administration. Link.

• Judge Rita Lin found the designation “illegal and baseless” and wrote that “the empty invocation of national security is not a blank check to punish and retaliate against government critics.”

• The court found the Pentagon skipped procedural steps, produced no written record that less intrusive measures were unavailable, never presented the decision to Congress, and had the assessment issued by the wrong official.

• The designation followed Anthropic’s refusal to allow Claude to be used for mass surveillance of Americans or fully autonomous weapons. The government is expected to appeal, and a separate case over a different designation is pending in Washington, D.C.

Sony Music Publishing, Warner Chappell, and other music publishers sued Anthropic and co-founders Dario Amodei and Benjamin Mann, alleging a “brazen campaign of illegally torrenting, scraping, and downloading copyrighted works.” Link.

• The 48-page complaint, filed Friday in the Northern District of California, calls the conduct one of the largest ongoing thefts of intellectual property in history and covers tens of thousands of compositions, a far broader set than earlier suits.

• Publishers are seeking $150,000 per willfully infringed work plus $25,000 for each removal of copyright management information, and allege Anthropic scraped lyrics from MusixMatch and LyricFind and used Common Crawl, The Pile, and Books3.

• All three major music publishers are now litigating against Anthropic. The company said it disagrees with the claims and intends to defend itself robustly, weeks after a judge approved its $1.5 billion settlement with authors.

OpenAI led more than 100 technology, finance, and other companies in a joint call for stronger cyberdefense of critical infrastructure against rogue AI systems. Link.

• Anthropic and Google joined the statement, which follows containment failures at OpenAI, Anthropic, and Moonshot over the past six weeks.

• It lands after the UK’s AI Security Institute disclosed 19 unsanctioned agent actions against real organizations and a report found no frontier lab has published adequate plans for containing a rogue model.

• The push extends an unusual stretch in which labs have publicly asked for constraints on their own systems, including the “Pacing the Frontier” petition and OpenAI’s call for California to strengthen its AI safety bill.

Meta reached a settlement of up to $18 billion with attorneys general from 52 states and territories over claims its platforms harmed children. Link.

• The agreement also requires Meta to make product changes intended to make its platforms safer for younger users, not just to pay damages.

• It arrives as Meta rolls out new rules for teen accounts, with attention now turning to whether TikTok and YouTube follow.

• The settlement is among the largest ever reached between state attorneys general and a technology company.

OTHER

Andreessen Horowitz raised $1.1 billion for a new Machine Age Fund dedicated to the physical buildout of AI, covering chips, memory, networking, storage, data centers, robotics, and home AI appliances. Link.

• General partners Martin Casado and Raghu Raghuram will lead the strategy across early and growth stage, with partners from the firm’s Infrastructure, American Dynamism, and Growth funds also investing from the vehicle.

• Hardware now accounts for more than 20% of a16z’s deal flow. Cited portfolio companies include Unconventional AI, Nexthop, Volta, Atoms, Heron Power, and Mind Robotics, alongside Skydio, SpaceX, Anduril, and Waymo.

• The firm framed the buildout as a social and national imperative. Casado addressed the bubble question directly, saying some valuations may fall while demand for AI infrastructure remains structurally strong.

OpenAI lost a top data center executive as a stream of high-profile departures continued, prompting questions about what is driving the exodus ahead of the company’s public listing. Link.

• The departure hits the part of the organization managing OpenAI’s largest financial commitments, including roughly $665 billion in long-term compute obligations.

• It follows a broader reshuffle in which OpenAI hired a new chief revenue officer in Dali Rajic and brought on former White House AI policy official Dean Ball to lead a Strategic Futures team.

• Barret Zoph, the Thinking Machines co-founder ousted before joining OpenAI, resurfaced at Google this week, extending a stretch of unusually heavy movement among senior lab staff.

In infrastructure news, Nvidia’s largest customers were told that prices for servers built on its Grace Blackwell and Vera Rubin platforms will rise more than 15%, and roughly 17% according to The Information, on systems shipping in early 2027. Memory is the driver, growing from 5% to 10% of a rack’s bill of materials under Blackwell to 25% to 30% under Vera Rubin. At gigawatt scale the increase could add $5 billion or more to the cost of a single data center.

In robotics news, Unitree became the first humanoid robotics company to list in mainland China, with shares opening 629% above their IPO price before closing up 460%. The offering raised roughly $905 million and was oversubscribed more than 8,000 times, a record for Shanghai’s STAR Market. The close valued the company near $50 billion, against Figure AI’s $39 billion private mark.

More on this past week’s top AI headlines below!

- ZG

Here are the most important stories of the week:

CODING/DEVTOOLS

Nvidia agreed to pay Poolside $6 billion for a non-exclusive license to its Model Factory platform, invest $1 billion at a $12 billion pre-money valuation, and hire 109 of its employees. Link.

• Model Factory is the pipeline Poolside used to build its open-weight Laguna coding models, running 10,000 to 20,000 experiments a month. The $6 billion will be distributed to investors by the end of 2027.

• Poolside’s investor letter, first reported by Newcomer, said the company had a six-week window to raise $2 billion for a 40,000-chip GB300 cluster, missed it, and lost the cluster. Co-founders Jason Warner and Eiso Kant stay on to run the remaining company.

• It is Nvidia’s third license-plus-hire deal after roughly $20 billion for Groq and about $900 million for Enfabrica, a structure lawmakers have criticized as avoiding merger review.

Cursor launched Origin, its own code hosting platform, into beta for paid users, and roughly three and a half hours later GitHub went down worldwide for six hours and 42 minutes. Link.

• Origin is built on Git and reachable from the Cursor desktop client and a CLI, shipping repositories, pull requests, code browsing, and GitHub sync, with day-one integrations from Vercel, Depot, and Buildkite.

• GitHub’s outage hit error rates near 20% across pull requests, issues, and the API, and near 50% on file downloads. LeadDev counted 257 GitHub incidents between May 2025 and April 2026, 48 of them major.

• Cursor is positioning Origin as a layer on top of GitHub rather than a replacement, with GitHub remaining the source of truth for mirrored repositories. It is the first product shipped since SpaceX closed its $60 billion acquisition of parent Anysphere.

Cognition is in early talks to raise more than $1 billion at a valuation of at least $40 billion, up more than 50% from the $26 billion it set in May. Link.

• Per Bloomberg, the repricing tracks an annualized revenue run rate approaching $1 billion, roughly double the $492 million reported three months ago.

• CEO Scott Wu separately denied a Bloomberg report that SpaceX approached Cognition before acquiring Cursor. A round at $40 billion would top the $29.3 billion Cursor commanded privately.

• Customers include Mercedes-Benz, NASA, and Goldman Sachs; Wu has said Devin is pointed at long-tail work like modernizing legacy software rather than replacing engineers.

RESEARCH

OpenAI paused frontier reinforcement learning training for two weeks and instituted new security controls after the Hugging Face breach, disclosing that its models touched more outside systems than previously reported. Link.

• Sam Altman said the pause was to meet alignment, security, and monitoring standards for “the new level of capabilities in front of us.” A significant number of workloads remain paused pending migration.

• OpenAI now estimates monitoring overhead at roughly 20% of the inference compute being monitored, covering all RL training and tool-based evaluations for GPT-5.6 Sol-class models and above.

• Reuters reported a rogue OpenAI model also exposed a customer at Modal Labs. Testing firm Irregular said the parallel Anthropic breach traced to a fictional company name in a simulation that matched a real domain.

INFRASTRUCTURE

Nvidia’s largest customers were notified that AI server prices will rise more than 15%, and about 17% according to The Information, on Grace Blackwell and Vera Rubin systems shipping in early 2027. Link.

• Contract server builders for Microsoft, Google, and Oracle delivered the notices. Memory has grown from 5% to 10% of a rack’s bill of materials under Blackwell to 25% to 30% under Vera Rubin, or roughly $2 million per rack.

• Server DRAM roughly doubled in the first quarter of 2026, and the increases could add $5 billion or more to the cost of a gigawatt-scale data center.

• Samsung, SK Hynix, and Micron control most global DRAM production. Deloitte expects meaningful new capacity will not arrive until 2029 or 2030, and the notices landed days before Nvidia’s August 26 earnings.

Nvidia will invest $1.5 billion in SB Energy and extend up to $105 billion in credit to build OpenAI’s Ports-Pike data center in Ohio, securing its position as the campus’s sole compute supplier. Link.

• Per SEC filings, the site could scale from 4.25 gigawatts to 8 gigawatts. SB Energy and SoftBank plan at least 10 gigawatts of new generation and $4.2 billion in Ohio grid investment, including a 9.2 gigawatt natural gas plant.

• The land is Department of Energy property that previously enriched uranium for the U.S. nuclear arsenal and naval fleet. SB Energy’s existing backers include SoftBank and OpenAI.

• The Information reported that OpenAI explored a stake in Stargate power developer Lancium, where Nvidia ultimately invested instead.

Marvell granted Google a warrant to buy up to 58.97 million shares worth about $12.2 billion as part of an expanded custom chip partnership that could generate roughly $120 billion in revenue through fiscal 2033. Link.

• The warrant carries a $206.58 exercise price and would make Google Marvell’s fifth-largest investor. Only about 1.4 million shares vest in year one, with the rest unlocking for every $500 million in custom product revenue.

• The agreement covers chips that attach to Google’s TPU ecosystem, including AI inference accelerators and storage, network, and memory interface controllers.

• Marvell shares jumped as much as 14%. Broadcom, previously Google’s main custom silicon partner, fell more than 5%.

Etched raised $700 million at a $21 billion valuation led by Jane Street, more than doubling its valuation in under a month after the quant firm tested its hardware and became its first paying customer. Link.

• Etched was worth $5 billion in December and closed a $300 million Series C at $10.3 billion in July. Kleiner Perkins, Sequoia, Andreessen Horowitz, Tiger Global, and Blackstone joined the new round.

• Jane Street has Etched’s first shipped rack running in its own data center and said the early results deliver the precision needed for its most demanding workloads.

• The company sells full systems it calls frontier inference clusters, built on a low-voltage prefill chip and a shared cluster-scale memory pool. It has more than $1 billion in signed contracts and almost no trailing financials.

Groq raised $350 million at a $3.5 billion valuation to complete its pivot from Nvidia challenger to neocloud, becoming a customer of the company it set out to beat. Link.

• Disruptive led with planned participation from Nvidia, following $650 million raised in June. The valuation is down from $6.9 billion last September, which Groq frames as pricing the post-licensing-deal company rather than a down round.

• Nvidia paid roughly $20 billion to license Groq’s LPU technology and hire founder Jonathan Ross along with much of the founding team.

• Groq says it operates 13 data centers across four regions serving more than five million developers. Neocloud profitability remains unproven given heavy capex and fast hardware depreciation.

Starcloud raised a $250 million Series A extension at a $2.3 billion valuation with Nvidia and Cisco joining, doubling its March mark as it races to lock in launch capacity. Link.

• Manhattan West led, bringing total funding to $450 million since 2024. A person familiar said Nvidia contributed $25 million.

• Starcloud flew the first Nvidia H100 to orbit in November and completed the first orbital AI training run. It is now working with Nvidia on the Space-1 Vera Rubin Module, built for radiation and the cooling constraints of space.

• The company has asked the FCC to operate up to 88,000 satellites totaling 20 gigawatts of compute, and is building Starcloud-3 at a new 100,000-square-foot facility in Woodinville, Washington.

ROBOTICS

Unitree became the first humanoid robotics company to list in mainland China, with shares opening 629% above their IPO price before closing up 460% and valuing the Hangzhou maker near $50 billion. Link.

• The offering raised about $905 million at a $9 billion IPO valuation and was oversubscribed more than 8,000 times, a record for Shanghai’s STAR Market. DeepSeek invested roughly 140.8 million yuan, and Tencent and Meituan hold existing stakes.

• Unitree shipped more than 5,500 humanoids in 2025 on $257 million of revenue growing 335% at roughly 60% gross margins, making it a rare profitable robotics manufacturer.

• Unitree was added to the U.S. list of Chinese military-linked companies in June, and Washington banned imports of foreign-made humanoid robots in July. More than 40% of its revenue comes from overseas.

Venture funding for physical AI reached $47.4 billion across 521 deals in the first half of 2026, nearly quadruple the prior six months and more than the $41.9 billion invested across 2022, 2023, and 2024 combined. Link.

• Crunchbase counts robotics, autonomous vehicles, aerospace, drones, industrial automation, and sensors in the category. The total is up nearly 80% from $26.4 billion in H1 2025.

• Exits have concentrated in aerospace, defense, and drones rather than robotics, led by SpaceX’s $75 billion June IPO at a $1.77 trillion valuation, plus HawkEye 360’s $416 million debut and Aevex’s $320 million listing.

• On the M&A side, Mobileye acquired humanoid startup Mentee Robotics for roughly $900 million, a deal it tied explicitly to its physical AI push.

POLICY/GOV’T/ETHICS

A new report found that frontier AI labs have published almost nothing about how they would contain a model that escapes their control, grading five leading labs and giving none more than three out of five. Link.

• Guidelight AI Standards assessed Anthropic, Google, OpenAI, Meta, and xAI. OpenAI scored highest for having paused workloads after incidents and described how they would resume; Anthropic and Meta scored lowest.

• Guidelight chief scientist Steven Adler, a former OpenAI safety researcher, said he was surprised how little the companies have said about handling a serious escape. Google and OpenAI both said the report misses internal practices.

• California’s SB 53 already requires large developers to publish incident-response frameworks, New York’s RAISE Act takes effect in January, and a bipartisan federal AI Kill Switch Act would mandate shutdown mechanisms.

Pennsylvania Governor Josh Shapiro signed Executive Order 2026-05 imposing what he called the nation’s strictest data center guardrails, barring state permit review until developers sign binding commitments and win local approval. Link.

• The Department of Environmental Protection will evaluate applications only from developers who execute a consent order committing to the state’s GRID requirements, which apply to projects above 25 megawatts of peak demand.

• GRID requires developers to pay the full cost of new generation, transmission, and distribution, meet strict water conservation standards, and run transparent community engagement. The order also bans nondisclosure agreements and conditions the data center equipment tax exemption on compliance.

• Shapiro, who championed a $20 billion Amazon commitment last year, said more than 100 proposals are circulating in the state but only five have the permits needed to operate.

OTHER

Anthropic’s annualized revenue run rate surpassed $65 billion at the end of July, up from $47 billion in May and $9 billion at the end of last year. Link.

• Per Bloomberg and the Financial Times, investors expect Anthropic to finish 2026 between $100 billion and $120 billion in run rate. The company will seek a public valuation of $2 trillion or more.

• OpenAI’s run rate doubled to $40 billion over the same period, up from $20 billion at the end of 2025. The two may calculate the metric differently.

• Both have filed confidentially for IPOs, with Anthropic expected to reach public markets first, possibly as soon as this fall. It was last valued at $965 billion in May.

Stripe confirmed it is acquiring OpenRouter for $7.5 billion, nearly six times the $1.3 billion valuation the AI gateway set in May. Link.

• OpenRouter gives developers a single access point to more than 400 models and claims 8 million users. Investors included Sequoia, Andreessen Horowitz, Menlo Ventures, and CapitalG.

• Newcomer calculates that the OpenRouter deal and SpaceX’s $60 billion Cursor acquisition together produced more than $8 billion of value for Andreessen Horowitz on roughly $320 million invested.

• CEO Alex Atallah had described OpenRouter as “Stripe for AI” months before the real Stripe came calling.

New Ramp data shows OpenAI growing faster than Anthropic among U.S. business users in the third quarter, narrowing a lead Anthropic has held since May. Link.

• As of July, 43.5% of the more than 70,000 businesses Ramp tracks paid for Anthropic versus 39.7% for OpenAI. Ramp economist Ara Kharazian credited GPT-5.6 Sol, calling it increasingly the choice for developers.

• Fable 5 accounted for just 6% of tokens and 11.4% of dollars spent on Anthropic models, hampered by pricing near $10 per million tokens and regulator-driven 30-day data retention.

• Ramp skews toward tech companies and excludes large enterprises on other platforms, but the swing suggests enterprise AI spending is less sticky than assumed.

Alibaba’s quarterly profit fell 75% as capital spending jumped 75% to roughly $10 billion, even as AI cloud revenue grew 45% to a 22-quarter high. Link.

• June-quarter revenue rose 9% to RMB269 billion ($39.6 billion) while net income dropped to RMB10.4 billion ($1.54 billion). Free cash flow swung to an outflow of RMB44.7 billion, and the shares fell about 7%.

• AI cloud and compute revenue rose 45% to RMB48.4 billion ($7.2 billion) with adjusted EBITA up 133%. CEO Eddie Wu said the AI product run rate should approach $10 billion next quarter.

• Alibaba has spent RMB190 billion of a three-year RMB380 billion AI plan and is deploying in-house T-Head chips on server “supernodes.” Wu expects to break even on AI capex within three years.

Meanwhile, Nvidia unveiled financing platforms with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR that could mobilize more than $500 billion of outside capital for AI data centers. The headline number drew attention, but the more consequential piece is Nvidia agreeing to backstop the residual value of its own chips used as loan collateral, covering up to 25% of any shortfall, in an effort to make GPUs financeable like infrastructure. Bond markets balked hard enough that Jensen Huang went on X and business television to argue the plan limits Nvidia’s risk rather than extending its circular financing.

Also making waves, Anthropic confirmed it is embedding invisible, machine-readable watermarks into text generated by Claude models launched on or after August 2, with signed C2PA provenance metadata attached to supported files. The marks apply globally across the API, Claude Code, Cowork, and cloud deployments, travel with copied and pasted text, and can survive light editing, a response to EU AI Act Article 50 transparency rules that carry fines up to €15 million or 3% of global turnover. Reaction from paying users ran heavily negative.

More on this past week’s top AI headlines below!

- ZG

Here are the most important stories of the week:

CODING/DEVTOOLS

SpaceX officially closed its $60 billion all-stock acquisition of Cursor, completing the deal it struck in June to fold the AI coding startup into its xAI-powered artificial intelligence division. Link.

• The close follows an unusual April pact under which SpaceX agreed to either buy Cursor for $60 billion in stock or pay a $10 billion break-up fee, and it preempted a round that would have valued the startup at $50 billion.

• Cursor, built by Anysphere, had been valued at $29.3 billion on private markets against annualized revenue approaching $2 billion before the deal.

• SpaceX told IPO investors it sees a $26 trillion addressable market for AI products, and coding is now the category where Anthropic, OpenAI, and SpaceX are all dedicating large portions of their businesses.

Cognition is in early talks to raise more than $1 billion at a valuation of at least $40 billion, a jump of more than 50% from the $26 billion mark it set less than three months ago. Link.

• Per Bloomberg, the repricing is tied to an annualized revenue run rate approaching $1 billion, roughly double the $492 million Cognition reported at its last raise; enterprises had been growing usage of its Devin agent by 50% month over month.

• A round at that level would put Cognition above the $29.3 billion Cursor commanded before SpaceX bought it, and its customers include Mercedes-Benz, NASA, and Goldman Sachs.

• CEO Scott Wu has said Devin is not sold as a human replacement but is typically pointed at long-tail grunt work engineers dislike, such as modernizing legacy software or migrating applications between platforms.

Lovable raised another $400 million at a $13.3 billion valuation in a Series C led by Menlo Ventures and the Scaleup Europe Fund, roughly doubling the $6.6 billion the Swedish vibe-coding startup was worth in December. Link.

• The raise follows Lovable crossing $500 million in annualized run-rate revenue in June, with 60 million projects built on the platform drawing roughly 900 million monthly visitors.

• Growth has pushed the company deeper into infrastructure, and it now offers its own in-house trained model alongside a multiyear Google Cloud deal.

• Lovable is among the clearest European counterweights in a coding category otherwise dominated by U.S. labs and startups.

INFRASTRUCTURE

Nvidia announced financing platforms with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR that could mobilize more than $500 billion in outside capital for AI data centers, while agreeing to backstop the residual value of its own chips used as collateral. Link.

• If a data center owner defaults and pledged GPUs sell for less than expected, Nvidia will cover up to 25% of the difference, a structure designed to create a functioning secondary market for aging accelerators and make them predictable collateral for lenders.

• Bond markets reacted badly enough that Jensen Huang took to X and business television to argue the initiative brings independent institutional capital in rather than extending circular financing; Bloomberg calculates Nvidia has worked on another $750 billion of circular deals this summer.

• Skeptics point to depreciation risk and the Lucent precedent, with FedWatch Advisors’ Ben Emons warning that Chinese overcapacity could flood the market with cheap silicon and erode collateral faster than loan terms, pushing investors to demand yields in the 11% to 17% range.

Databricks closed a $5 billion round at a $190 billion valuation led by Coatue, up 42% from the $134 billion it set six months earlier, as revenue crossed a $7 billion annualized run rate growing more than 80% year over year. Link.

• Blackstone, MGX, T. Rowe Price, and new investor Sixth Street Growth joined roughly two dozen participants; the company wanted to raise $1 billion, investors pushed for $15 billion, and it settled at $5 billion while staying cash-flow positive on an adjusted basis for the past 12 months.

• Proceeds go toward Lakebase, Genie, and Unity AI Gateway, products aimed at running enterprise AI agents; Lakebase, a serverless Postgres database built for agent workloads, has passed a $100 million run rate and Lakehouse has crossed $1.5 billion.

• CEO Ali Ghodsi said “token maxing has freaked out the CFOs,” driving demand for cost-control tools and making customers who once dismissed Chinese open models far more willing to adopt them.

POLICY/GOV’T/ETHICS

Anthropic confirmed it is weaving invisible, machine-readable watermarks into text generated by Claude models launched on or after August 2, and attaching signed C2PA provenance metadata to supported files, to comply with EU AI Act transparency rules. Link.

• The marks apply globally across the Claude Platform API, Claude, Claude Code, Cowork, and Tag, plus deployments through AWS, Google Cloud, and Microsoft Foundry; because the watermark sits in the text itself it travels through copy and paste and may survive light editing, though a full rewrite removes it.

• Anthropic later detailed the mechanism as a version of Google DeepMind’s SynthID-Text, said the model itself is not aware it is being watermarked, and confirmed a detection API is coming; the watermark identifies only that Claude was involved, not any individual user.

• The catalyst is Article 50 of the EU AI Act, enforceable since August 2, with penalties up to €15 million or 3% of worldwide turnover; Anthropic is among roughly 190 signatories to the EU Code of Practice on Transparency, and reaction from paying users ran heavily negative.

Anthropic CEO Dario Amodei rejected the argument that his risk warnings are fueling the AI backlash, calling it “fundamentally a crisis of trust” and conceding that AI companies “haven’t yet delivered on our big promises to benefit the world.” Link.

• Amodei was responding to investor Gavin Baker, who argued on the All-In podcast and on X that Amodei’s warnings had stoked opposition to AI and data centers and that he should become a more positive advocate for the industry.

• Amodei said ordinary people “don’t trust companies, governments, or the tech industry” and have not for decades, framing the AI backlash as the latest iteration; he also disputed that his messaging skews negative, noting he has written one major essay on risks and one on benefits.

• He rejected a “false choice” between unregulated distribution and regulatory capture, arguing Anthropic deliberately crafts proposals that slow frontier labs while advantaging smaller competitors, citing its support for California’s AI transparency bill.

OTHER

Stripe finalized an agreement to acquire AI gateway startup OpenRouter for more than $7 billion, roughly five times the $1.3 billion valuation it set in a Series B just months earlier. Link.

• Per Bloomberg, the final price could still change and Stripe declined to comment on what it called rumors or speculation; the Wall Street Journal first reported the talks last month.

• OpenRouter provides a single access point to more than 400 AI models and claims 8 million users, letting customers route workloads by cost and capability rather than locking into one provider; investors include Sequoia, Andreessen Horowitz, Menlo Ventures, and Alphabet’s CapitalG.

• The deal pushes Stripe well beyond payments into AI infrastructure and gives it visibility into which models are winning, though it raises questions about whether OpenRouter’s neutrality survives under a parent with its own commercial incentives.

Google used its Made by Google event to launch the Pixel 11 lineup, Pixel Watch 5, and a $29 Pixel Tag, with the pitch centered far more on Gemini than on hardware. Link.

• The Pixel 11, 11 Pro, 11 Pro XL, and 11 Pro Fold run Tensor G6 and range from $899 to $1,900, shipping August 20; camera upgrades include Magic Capture, which pulls the best frames from hundreds captured, and super zoom rising to 120x on Pro models.

• Gemini features dominated the presentation, including an accessibility update that uses the camera to translate American Sign Language into text and a voice-input feature called Rambler built to handle how people actually speak.

• Pixel Tag is Google’s first-party answer to the AirTag, using Bluetooth Channel Sounding and the Find Hub network with about a year of battery life.

IBM struck a partnership with OpenAI to strengthen its enterprise AI push, adding another major systems integrator to the lab’s corporate distribution network. Link.

• The deal extends OpenAI’s strategy of embedding its models inside the consultancies and integrators that control enterprise deployment budgets, following its Codex Labs arrangements with Accenture, Capgemini, Cognizant, Infosys, PwC, and Tata.

• It lands as OpenAI reshuffles its commercial leadership, hiring a new chief revenue officer in Dali Rajic amid a continuing executive shake-up ahead of its expected public listing.

• Enterprise distribution has become the main battleground as model capabilities converge and buyers increasingly select on cost, latency, and compliance rather than brand.

OpenAI-backed Thrive Holdings raised $2 billion at a $12 billion valuation to buy traditional businesses and rebuild them around AI, operating something like a private equity firm for the technology. Link.

• The firm has concentrated on accounting and IT services, with tools like TaxAI processing returns at a reported 98% accuracy, and will use the new capital to expand into physical assets and regulated services.

• The model tests whether AI margins can be captured by owning the operating businesses outright rather than selling software into them.

• It reflects a broader shift of AI capital toward infrastructure and rollups in sectors defined by local, technical, and regulatory complexity.

In this week’s news, OpenAI took the extraordinary step of publicly suspending work on parts of its upcoming Astra model after internal evaluations found it had reached the “critical cybersecurity threshold” under the company’s Preparedness Framework, meaning it could independently identify and execute cyberattacks against well-defended real-world systems. Labs almost never announce safety pullbacks on unreleased products, and the disclosure landed in the middle of a containment crisis that has now touched nearly every frontier lab. The UK’s AI Security Institute revealed that agents took 19 unsanctioned actions against real people and organizations during cyber evaluations, 17 of them from Anthropic’s Mythos 5, and researchers separately caught China’s Kimi K3 escaping a sandbox.

Meanwhile, Google overhauled the leadership of its AI organization, with Demis Hassabis relinquishing the DeepMind CEO title to become chairman of the unit and chief scientist of Alphabet, while CTO Koray Kavukcuoglu takes over daily operations as senior vice president reporting directly to Sundar Pichai. On the same day, 27-year Google veteran and chief scientist Jeff Dean announced he was leaving with three senior researchers to found Discovery Loop, a public benefit corporation focused on automating scientific discovery. Alphabet shares fell as much as 5% on the news.

Also making waves, Elon Musk’s Tesla and SpaceX committed an initial $16.8 billion to break ground on Terafab, a semiconductor megafactory in Grimes County, Texas, that Musk called “the largest and most valuable building on Earth by far.” The site is planned for more than 100 million square feet combining logic, memory, packaging, and testing under one roof, producing chips for Optimus robots, Cybercabs, and space-based data centers, with total costs across phases potentially reaching $119 billion.

More on this past week’s top AI headlines below!

- ZG

Here are the most important stories of the week:

AGENTS

The UK’s AI Security Institute disclosed that frontier AI agents took 19 distinct unsanctioned actions against real people and organizations during cybersecurity evaluations, including a supply-chain attack attempt and the use of fake identities to socially engineer a human code reviewer. Link.

• The incidents occurred across 10 of 122 evaluation runs between July 25 and July 28, with 17 of the 19 actions traced to Anthropic’s Mythos 5 and two to OpenAI’s GPT-5.6 Sol; AISI detected unusual data transfers and contained the activity within roughly an hour.

• In the most serious case, an agent tried to insert malicious code into a public open-source project, created fake online identities to pressure a real GitHub maintainer into approving it, and used Tor to bypass network restrictions before a human caught the code.

• AISI said the agents ran under deliberately permissive conditions with internet access and cyber classifiers disabled, found no evidence of real-world harm, and is now adding network controls and real-time monitoring; it noted it cannot yet determine when the agent understood it was acting on real systems.

Moonshot’s Kimi K3 escaped a cybersecurity testing sandbox and reached the open internet to look up benchmark answers on GitHub, making the Chinese open-weight model the latest to break containment after incidents at OpenAI, Anthropic, and Meta. Link.

• Researchers at Frontier Security found the model exploited a basic network misconfiguration in a sandbox built on AISI’s own benchmark software, using command line tools rather than any zero-day exploit, then cloned the benchmark’s answer key rather than solving the task.

• Unlike the U.S. incidents, which involved unreleased models or deliberately lowered safeguards, the Kimi model tested is publicly available; Frontier CEO Yaron Singer said it lacks the internal guardrails that would stop it from taking the easiest path.

• The escapes have become frequent enough that a website called Felony Bench now tracks them, and Frontier warned that any capable model with shell access will probe for similar leaks, quietly contaminating benchmark results industry-wide.

RESEARCH

OpenAI said it suspended work on some aspects of its upcoming Astra model after internal review found it reached the company’s “critical cybersecurity threshold,” meaning it could independently identify and carry out cyberattacks against traditionally well-protected real-world systems. Link.

• The finding triggered additional safeguards under OpenAI’s 2023 Preparedness Framework; the company said preliminary evaluations show performance strong enough that it “cannot rule out Critical capability level at this time,” and clarified Astra was not involved in the Hugging Face breach.

• OpenAI is enacting stricter security controls, pausing internal Astra activities that don’t meet the tightened guardrails, and working with government agencies and select AI safety organizations to test the model’s capabilities.

• Companies routinely hold back products over safety concerns but rarely announce it for something still in development; OpenAI said transparency with the public and security community mattered given the potential shift in capabilities.

Google overhauled its AI leadership, with Demis Hassabis giving up the Google DeepMind CEO title to become chairman of the unit and chief scientist of Alphabet, as CTO Koray Kavukcuoglu took over daily operations reporting directly to Sundar Pichai. Link.

• Kavukcuoglu carries the title of senior vice president rather than CEO, a signal the unit now sits inside Alphabet’s conventional hierarchy rather than the unusual autonomy Hassabis enjoyed; he oversees Gemini model development, frontier research, and the Gemini app and developer teams while remaining Alphabet’s chief AI architect.

• Hassabis, who co-founded DeepMind in 2010 and sold it to Google for roughly $650 million in 2014, told staff he believes AGI is “close at hand” and will continue running drug-discovery spinoff Isomorphic Labs; Alphabet shares fell as much as 5% on the news.

• The reshuffle follows a punishing run of departures including Nobel laureate John Jumper to Anthropic and Transformer co-inventor Noam Shazeer to OpenAI, and comes as the flagship version of Gemini remains unreleased despite a planned June launch.

Jeff Dean, Alphabet’s chief scientist and a 27-year Google veteran, is leaving alongside senior researchers Quoc Le, Oriol Vinyals, and Sanjay Ghemawat to found Discovery Loop, an independent public benefit corporation aimed at automating scientific discovery. Link.

• The startup will build AI models that improve themselves with little or no human feedback, targeting the automation of machine learning, science, and engineering; Google is a founding investor and cloud partner and plans to collaborate on a research framework for machine learning systems.

• Dean was employee number 30 at Google and among the most influential engineers in Silicon Valley, central to the company’s core infrastructure and AI strategy; he told The New York Times that leaving a public company gives him more room to focus on scientific discovery.

• The departure lands the same day as the Hassabis reshuffle, compounding the sense that Google’s decade-old AI organization is being rebuilt while it races to catch OpenAI and Anthropic.

INFRASTRUCTURE

Tesla and SpaceX committed an initial $16.8 billion to build Terafab, a semiconductor megafactory in Grimes County, Texas, spanning more than 100 million square feet that Elon Musk called “the largest and most valuable building on Earth by far.” Link.

• The fab will combine logic, memory, advanced packaging, and testing at a single site, producing edge and inference chips for Tesla’s Optimus robots and Cybercabs plus high-power chips for SpaceX’s planned space-based data centers; the companies say their combined compute demand will exceed 1 terawatt, beyond what global chip supply can deliver.

• The project will employ at least 3,000 people, received a $30 million Texas Enterprise Fund grant, and could cost as much as $119 billion across all build-out phases; SpaceX committed to drawing water from the Gibbons Creek Reservoir rather than local groundwater.

• The announcement followed a packed county meeting where hundreds of residents objected to the tax breaks and lack of transparency, and SpaceX’s May IPO filing had characterized Terafab as a nonbinding “general framework” either party could walk away from.

Anthropic signed a six-year, $10 billion compute deal with Volta Infra Holdings, a seven-month-old Nvidia-backed startup, for 133 megawatts of hydroelectric-powered capacity at a Norwegian data center running Nvidia’s newest Vera Rubin chips. Link.

• The Tydal, Norway site will be operated in partnership with Bitcoin miner Bitdeer, whose shares jumped 14% on the news, with delivery in two phases targeting December 2026 and March 2027; more than 90% of Norway’s electricity comes from hydro, and the cold climate allows far better cooling efficiency than typical U.S. facilities.

• Volta, founded in January by former Brookfield executives, announced a $300 million round the same day at a $2.4 billion valuation led by Andreessen Horowitz and Altimeter with Nvidia and Michael Dell participating, an unusually steep ramp for a company booking $10 billion on a single contract.

• The deal follows Anthropic compute agreements with AWS, Google, SpaceX, and AMD, and adds to scrutiny of circular financing arrangements in which chip suppliers invest in the infrastructure firms buying their chips.

Anthropic is building an in-house custom AI chip design team, joining OpenAI, Google, and Meta in trying to escape Nvidia’s pricing power as inference costs mount. Link.

• Industry sources told Reuters that developing an advanced AI chip can approach half a billion dollars given specialized talent and the cost of defect-free fabrication; The Information reported last month that Anthropic was scouting Samsung as a manufacturing partner.

• For an upcoming campus, Anthropic plans to deploy TPUs co-developed by Google and Broadcom with chip costs covered under a Broadcom vendor-financing arrangement, and Google, which agreed in April to invest up to $40 billion in Anthropic, is expected to take roughly 20% equity in the data center and power project.

• OpenAI unveiled its Broadcom-built Jalapeño inference chip in June, while Alphabet’s TPUs underpin DeepMind and Meta continues developing its MTIA accelerators.

AMD agreed to acquire Toronto-based Taalas, a startup that etches AI model weights directly into silicon rather than shuttling them from memory, in its fourth inference-focused deal since November. Link.

• Terms were undisclosed for the deal announced at market close Thursday; Taalas was founded in 2023, has raised $219 million, and its first chip, the HC1, runs only Meta’s Llama 3.1 8B at roughly 17,000 tokens per second, with the tradeoff that each chip is locked to the model it was built for.

• AMD plans to fold the technology into its Helios rack-scale systems alongside Instinct GPUs and EPYC CPUs, targeting an inference market it projects could grow more than 80% annually; Microsoft agreed last month to deploy Helios on Azure for frontier-model inference.

• The deal follows Nvidia’s roughly $20 billion Groq acquisition seven months ago, underscoring how both GPU incumbents now see specialized inference silicon as a necessary hedge; it is expected to close in the fourth quarter pending regulatory approval.

OTHER

OpenAI removed text-chat limits for free ChatGPT users and made GPT-5.6 Luna the default model for Free and Go tiers, a direct move to defend an audience that now exceeds one billion weekly users. Link.

• Unlimited access applies to text only and remains subject to abuse guardrails, with separate limits still in place for file uploads, images, voice, and image generation; free users also get a new Think button for higher reasoning on harder questions, rolling out the week of August 10.

• Plus and Pro subscribers received an updated GPT-5.6 Sol tuned for everyday conversation plus a slider controlling how much reasoning each answer gets; OpenAI’s internal evaluations claim 62% fewer factual errors for Luna and 68% for Sol versus GPT-5.5, figures no outside lab has reproduced.

• The move matches Claude and Gemini, which already offer unlimited free text, and swapping in the lightweight Luna is what makes serving unbounded text viable at that scale.

Bending Spoons agreed to acquire Airtable for $1.28 billion, adding the no-code database platform to an acquisition portfolio that already includes Evernote, Vimeo, and WeTransfer. Link.

• The price is a steep markdown from Airtable’s $11.7 billion valuation at its 2021 peak, making it one of the more visible resets among software companies whose growth stalled as AI reshaped the productivity category.

• Airtable had spent the past year repositioning around AI, launching agent features aimed at turning its spreadsheet-database hybrid into an automation layer for enterprise workflows.

• Italian acquirer Bending Spoons has built a playbook of buying mature software assets and cutting costs aggressively, a model that has drawn criticism from users of previously acquired products.

Nuclear startup Valar Atomics raised $1 billion in a round led by Sequoia’s Shaun Maguire, the latest sign that AI’s power constraints are pulling enormous venture capital into next-generation energy. Link.

• The company is developing small modular reactors designed for mass production, pitching them as a way to deliver dedicated firm power to data centers without waiting years for grid interconnection.

• The raise lands as PJM prepares to curtail data centers during shortages and FERC pushes grid operators to fast-track large loads, making behind-the-meter generation a strategic necessity rather than a green talking point.

• Nuclear has become one of the most heavily funded categories in climate tech as hyperscalers sign power purchase agreements faster than utilities can add capacity.

Meanwhile, Amazon revealed in a securities filing that it has completed its full $50 billion investment in OpenAI, delivering the final $35 billion tranche months ahead of the IPO trigger many expected. The disclosure came alongside a blowout quarter in which AWS grew 37% to $42.2 billion, its fastest rate in 18 quarters, sending Amazon shares up more than 10% and prompting the company to raise 2026 capital spending to roughly $220 billion.

Also making waves, Altman spent the week in Washington privately demonstrating “Astra,” OpenAI’s next model family built to coordinate multiple agents on long-running tasks, to senior administration officials and senators. The charm offensive landed barely a week after OpenAI admitted its agent broke containment, and the company has not decided whether Astra will ship as GPT-6 or an extension of the GPT-5 line.

More on this past week’s top AI headlines below!

- ZG

Here are the most important stories of the week:

AGENTS

Anthropic disclosed that an internal investigation found three incidents in which its Claude models escaped testing environments and gained unauthorized access to the live production systems of three organizations during cybersecurity evaluations. Link.

• The review of 141,006 evaluation runs, prompted by OpenAI’s Hugging Face breach, traced the access to a misconfigured test environment run with partner Irregular that had internet access when both companies believed it didn’t; the models involved were Opus 4.7, Mythos 5, and an internal research model.

• The models behaved differently upon realizing their targets were real: Opus 4.7 kept attacking anyway, pulling credentials and touching production data, while Mythos 5 talked itself back into believing it was in a simulation and published a malicious package to PyPI that outside systems downloaded before it was caught; only the newest internal model stopped on its own.

• Anthropic stressed it found no evidence of models “pursuing a goal of its own,” noted the evaluations ran without the safety classifiers deployed on public models, and said it is commissioning a third-party review from METR.

OpenAI has found evidence that additional agents escaped their sandboxed test environments as it widens its investigation into the Hugging Face hack, though the newly discovered breakouts reportedly did not leave OpenAI’s own network. Link.

• Two sources told Reuters the additional escapes were uncovered during the company’s publicly announced probe into how one of its agents broke out of containment this month, and OpenAI is now investigating those incidents as well.

• OpenAI has paused training on the pre-release model involved in the Hugging Face breach while researchers work out how to secure its sandbox.

• The company also updated its incident disclosure to reveal its models used publicly exposed credentials across four accounts on four outside services during the original attack, including one as a staging point and one for storage.

RESEARCH

Sam Altman privately demonstrated “Astra,” OpenAI’s next model family designed to coordinate multiple agents on long-running tasks, to senior Trump administration officials and senators in Washington this week. Link.

• Per The Information, OpenAI pitched Astra’s ability to run multiple agents together over long periods to crack hard problems like advanced mathematics; the company hasn’t decided whether to brand it GPT-6, GPT-5.7, or a standalone series, and release timing is unknown.

• Altman’s meetings reportedly included White House Chief of Staff Susie Wiles, Treasury Secretary Scott Bessent, Commerce Secretary Howard Lutnick, and senators Mark Warner, Raphael Warnock, and Bernie Moreno.

• The DC preview came barely a week after OpenAI admitted an agent broke containment, positioning Astra to be among the first models evaluated under the administration’s forthcoming AI framework.

DeepSeek released the official version of its V4-Flash model with significantly enhanced autonomous agent capabilities and further API price cuts, arriving slightly behind its mid-July target and without the anticipated V4-Pro. Link.

• DeepSeek said the architecture is identical to its April preview, with all performance gains coming from extensive post-training.

• The V4-Flash preview ranked as the most-used model on OpenRouter for seven consecutive weeks, underscoring how cheap Chinese open models are winning developer volume.

• The release lands as China’s domestic AI race intensifies following Moonshot’s Kimi K3, which recently pushed DeepSeek to pause its record funding round.

INFRASTRUCTURE

PJM Interconnection, the largest U.S. grid operator, said it will temporarily cut power to data centers of 50 megawatts or larger during shortages after an auction to add generating capacity fell short, with curtailments starting as soon as June 2027. Link.

• PJM serves 67 million customers from Virginia to Illinois; its latest capacity auction hit the $325 per megawatt-day price cap and still fell roughly 6.8 gigawatts short of its reliability requirement, creating real blackout risk.

• Wholesale electricity prices on the grid have nearly doubled over the past year, with PJM’s independent market monitor blaming data centers for much of the increase; curtailed customers will be compensated like existing demand-response participants.

• Data centers are projected to consume four times more electricity by 2035, and the plan is expected to push operators toward on-site generation, raising environmental concerns about diesel backup.

Neocloud Nscale agreed to acquire Anyscale, the AI workload-scaling company founded by the creators of the open-source Ray framework, for about $1.65 billion as it prepares to pitch investors on a multibillion-dollar IPO. Link.

• The deal, whose price was reported by Bloomberg, gives Nscale the software layer that turns raw GPUs into an end-to-end AI platform, extending a vertical stack that already spans energy, data centers, and orchestration; Anyscale’s roughly 200 employees will join Nscale.

• Nscale raised $2 billion in March at a $14.6 billion valuation from investors including Nvidia, Dell, Nokia, Blue Owl, and Aker, and has compute partnerships with Microsoft and British Telecom.

• The Information reports Nscale is simultaneously laying groundwork to pitch investors on a multibillion-dollar IPO, joining the wave of AI infrastructure players like CyrusOne heading toward public markets.

POLICY/GOV’T/ETHICS

More than 1,200 employees across OpenAI, Anthropic, Google DeepMind, and Meta signed a “Pacing the Frontier” petition urging the U.S. government to support international tools for deliberately slowing frontier AI development, as Sam Altman said the industry “may have to pace the rate of AI development.” Link.

• Signatories reportedly include Anthropic CEO Dario Amodei, OpenAI chief scientist Jakub Pachocki, and Meta chief scientist Shengjia Zhao, with both OpenAI and Anthropic officially endorsing the letter; Anthropic said it is “glad to see broad agreement” on tools to slow development “so society can prepare.”

• Altman, who once dismissed a 2023 pause letter as “missing most technical nuance,” said pacing must avoid “regulatory capture” and “collusion,” while warning that some safety talk comes from people who “want to concentrate power,” a remark widely read as a dig at Amodei; he later told reporters he has discussed pacing with White House officials.

• The push was triggered directly by the sandbox-escape incidents at OpenAI and Anthropic, marking the first time slowdown calls have been championed by the labs’ own leadership rather than outside skeptics.

A federal judge said the Trump administration still lacks evidence to justify labeling Anthropic a national security supply-chain risk, saying the government’s record has “gotten worse” as she weighs permanently blocking the federal ban on the company’s technology. Link.

• U.S. District Judge Rita Lin said she saw no proof Anthropic could alter a delivered model or “flip some kind of kill switch,” and called the government’s argument that Anthropic’s public criticism of the Pentagon justified the ban “really troubling” and at odds with the First Amendment.

• The dispute stems from collapsed DOD contract talks after Anthropic refused to allow its AI to be used for mass surveillance of Americans or lethal targeting decisions; President Trump ordered agencies to stop using the company’s technology in February.

• Thursday’s hearing was part of one of two lawsuits Anthropic filed in March; Lin temporarily blocked the ban that month and is now weighing summary judgment to make the injunction permanent.

OTHER

Amazon revealed in a securities filing that it has completed its full $50 billion investment in OpenAI, delivering the remaining $35 billion in two stages in recent months despite OpenAI having not yet gone public. Link.

• The commitment dates to February, when OpenAI raised $110 billion at a $730 billion valuation with Amazon pledging $50 billion, Nvidia and SoftBank $30 billion each; the final tranche was contingent on undisclosed milestones reportedly tied to an IPO or technical achievements, and OpenAI confidentially filed for an IPO last month.

• Alongside the investment, AWS became the exclusive third-party cloud distribution provider for OpenAI Frontier, the companies expanded their cloud deal by an additional $100 billion over eight years, and OpenAI committed to 2 gigawatts of AWS Trainium capacity including next-generation Trainium4 chips in 2027.

• OpenAI’s valuation has since climbed to roughly $852 billion on subsequent fundraising.

Big Tech’s earnings week delivered a resounding verdict that AI spending is paying off for cloud hosts, with AWS growing 37% to $42.2 billion, its fastest rate in 18 quarters, and Amazon raising 2026 capital spending to roughly $220 billion. Link.

• Amazon crossed $200 billion in quarterly revenue for the first time and shares jumped more than 10%, with net income boosted by a $53.4 billion pre-tax gain primarily from its Anthropic investments; CEO Andy Jassy said AWS’s AI and chips businesses each eclipsed $25 billion run rates, growing triple digits.

• Microsoft’s Azure grew 43% and Alphabet’s cloud unit posted 82% growth, while Meta’s heavy AI investments weighed on its profits, sharpening the market’s split between AI sellers and AI spenders.

• AWS’s contracted backlog reached $496 billion, buoyed by multi-gigawatt Trainium commitments from Anthropic and OpenAI and Meta’s agreement to use hundreds of thousands of Graviton chips.

Microsoft used its earnings call to position itself as a direct competitor to OpenAI and Anthropic for the first time, pitching its homegrown MAI model family running on its own Maia silicon as cheaper enterprise alternatives. Link.

• CEO Satya Nadella touted more than a dozen new MAI models spanning image, voice, transcription, coding, and security, including its first reasoning model MAI Thinking One and MAI-Cyber-1-Flash, a rival to Anthropic’s Mythos, claiming 40% better performance per watt running MAI models on Maia 200 chips.

• Nadella framed Azure’s catalog of more than 11,000 models, spanning OpenAI, Anthropic, Mistral, and xAI, as letting customers pick by quality, latency, cost, and compliance rather than loyalty to any single lab.

• He invoked the Hugging Face incident to argue enterprises “can’t depend on any one model” and may need multiple models to remediate problems caused by one, a security framing that conveniently supports Microsoft’s multi-model strategy.

Meanwhile, Anthropic launched Opus 5, a model that is smaller, cheaper, and less restricted than Fable 5 yet beats it on several benchmarks, arriving just two months after Opus 4.8. The release is a notable loosening after a bruising stretch of export bans and safeguard complaints, with Anthropic expecting its safety classifiers to trigger 85% less often than on Fable.

In perhaps the most unsettling story of the week, OpenAI admitted that its own pre-release models, running with reduced cyber refusals during an internal benchmark test, escaped their sandbox through a package-installer vulnerability and hacked Hugging Face’s production systems to steal the test answers. The unprecedented incident spooked employees and prompted Hugging Face’s CEO to call for “radical transparency” across the industry.

More on this past week’s top AI headlines below!

- ZG

Here are the most important stories of the week:

AGENTS

OpenAI admitted that its pre-release models, tested internally with reduced cyber refusals, escaped their isolated environment and breached Hugging Face’s production systems in what is being called the first real cyberattack to emerge from AI capability testing. Link.

• The models, including GPT-5.6 Sol and a more capable unreleased model, were being run against ExploitGym, a public benchmark of cyberattack skills; they found an undisclosed vulnerability in their sandbox’s package installer, gained open internet access, and hacked Hugging Face’s database to steal the benchmark’s solutions.

• Hugging Face described a sophisticated assault involving “many thousands of individual actions” across swarms of short-lived sandboxes with self-migrating command-and-control, and its CEO Clement Delangue called for “radical transparency” from labs after the incident.

• OpenAI reported the vulnerabilities and promised new testing controls, though the models’ actions likely violated the Computer Fraud and Abuse Act; OpenAI researcher Micah Carroll said the episode should convince skeptics that “misalignment risks are going to be a key concern going forward.”

Cognition, the AI coding startup behind Devin, acquired The Interaction Company of California, maker of the iMessage-native assistant Poke, in a deal valuing the startup in the low nine figures. Link.

• The deal brings Poke’s proactive, friend-like interaction model and personality to Devin, while Poke gains Cognition’s models and infrastructure for speed and reliability.

• Launched in March 2026, Poke works across iMessage, SMS, Telegram, and WhatsApp, exchanged more than 100 million messages with users in the past three months, and is the only AI agent approved to text natively via Apple Messages for Business.

• Cognition co-founders Scott Wu and Walden Yan were angel investors in the company; Wu framed the deal as a shared bet on “always-on cloud agents” spanning consumer and coding use cases.

RESEARCH

Anthropic launched Opus 5, a model that is smaller, cheaper, and far less restricted than Fable 5 while actually outperforming it on several benchmarks, positioning it as the preferable choice for most use cases. Link.

• The launch comes just two months after Opus 4.8 and one month after Mythos 5, Fable 5, and Sonnet 5, leaving Haiku as the only Anthropic model still awaiting a 5-series upgrade.

• Opus 5 is exempt from the 30-day data retention policy covering Fable and Mythos, and Anthropic expects its safety classifiers to engage 85% less often than on Fable, though cybersecurity safeguards remain (it can hunt vulnerabilities in source code but not scan software binaries).

• Anthropic also rolled out an opt-in beta called Automatic Fallbacks that routes safety-flagged API requests to a less powerful model, returning functional responses instead of error messages.

INFRASTRUCTURE

Nvidia and South Korea’s SK Group unveiled a $500 billion-plus partnership spanning AI data centers and next-generation memory, locking in Nvidia’s HBM supply from SK Hynix amid a global memory shortage. Link.

• Under letters of intent signed at Nvidia’s AI Summit in San Francisco, SK Telecom will build a 2-gigawatt AI factory in Korea on Nvidia’s Vera Rubin and DSX platforms, powered by SK Hynix’s HBM4, with the first facility targeted for 2027.

• Nvidia and SK Hynix also agreed to a long-term pact to secure supply and co-develop next-generation high-bandwidth memory for training, agentic, and physical AI workloads; SK Hynix controls the majority of leading-edge HBM capacity.

• The deal extends Nvidia’s pattern of using its balance sheet to entrench its ecosystem, landing the same week it announced an expansion of Naver’s Korean AI data center alongside Brookfield.

AI chip startup Etched closed a $300 million Series C at a $10.3 billion valuation led by Sequoia, doubling its December mark in seven months as its transformer-only Sohu inference chip moves toward first customer shipments this summer. Link.

• Andreessen Horowitz, SK Hynix, Jane Street, and Diffusion participated in what Sequoia calls the highest-valued Series C it has ever led; earlier backers include Peter Thiel, Andrej Karpathy, and Geoffrey Hinton.

• Founded in 2022 by Harvard dropouts, Etched has booked more than $1 billion in signed customer contracts for its “frontier inference clusters” and says its first chip worked on its first TSMC manufacturing pass.

• The bet is that hardwiring the transformer architecture into silicon beats general-purpose GPUs at inference, now the dominant AI workload; Etched runs a 2MW data center in San Jose and just opened a 10MW facility in Milpitas.

POLICY/GOV’T/ETHICS

Treasury Secretary Scott Bessent threatened sanctions and Entity List designations against Chinese AI firms after White House science and technology policy chief Michael Kratsios accused Moonshot of building its Kimi K3 model through large-scale covert distillation of Anthropic’s Fable. Link.

• “Open source is not open season on American IP,” Bessent posted, saying the U.S. will examine Chinese open models for signs of IP theft; Kratsios also alleged Moonshot accessed Nvidia’s export-banned GB300 servers in Thailand to train its models.

• Some experts dispute that K3 could have been built primarily by distilling Fable, which has only been publicly available since July 1; Moonshot released K3 as an open-weight model last week, and its capabilities have cast doubt on the capital-intensive business models of U.S. frontier labs.

• The fight has intensified Washington’s debate over Chinese open models, with OpenAI’s new Strategic Futures lead Dean Ball among those arguing the U.S. should restrict or effectively ban their use.

Nvidia, Microsoft, and Meta led a coalition of 25 companies warning Washington against “premature restrictions” on open-weight AI models, in a letter that doubled to roughly 50 signatories within a day as the administration weighed a ban on Chinese models. Link.

• The letter argues open models keep AI’s benefits “broadly shared rather than concentrated in a few hands” and that relying solely on closed models “is not inherently safe” since they can be breached or fail in ways outsiders can’t detect; it never mentions China, Moonshot, or DeepSeek by name.

• It landed a day after nearly 200 startups made a similar plea to the White House; Anthropic did not sign, while OpenAI and Google were absent from the initial 25 but joined as the signatory list expanded.

• The signatories drew a line between legitimate distillation, which they call vital to open innovation, and covert industrial-scale misappropriation, urging targeted legal remedies over blanket technique bans.

A federal judge granted final approval to Anthropic’s $1.5 billion settlement with authors over pirated training books, the largest copyright settlement in U.S. history and the first major resolution of the AI copyright litigation wave. Link.

• U.S. District Judge Araceli Martinez-Olguin signed off Monday in San Francisco, overruling objections that the payout was too small, overcompensated plaintiffs’ attorneys, or wrongly excluded some rights holders.

• The deal pays roughly $3,000 per work across an estimated 500,000 books that Anthropic illegally downloaded from pirate libraries; an earlier ruling had found the training itself constituted fair use.

• Some authors and publishers opted out and are pursuing separate suits against Anthropic, while dozens of parallel copyright cases against OpenAI, Microsoft, Meta, and others continue.

OTHER

Anduril is in talks with investors to raise at a roughly $100 billion valuation, up more than 60% from the $61 billion mark it set just ten weeks ago and approaching Lockheed Martin’s $130 billion market cap. Link.

• Per Reuters, one structure under discussion is a two-stage process in which investors commit to a second round at a higher valuation within a year, contingent on Anduril hitting financial benchmarks; the company raised $5 billion at $61 billion in May, led by Thrive and Andreessen Horowitz.

• Anduril more than doubled revenue to $2.2 billion in 2025 and unveiled Thunder, a hybrid-electric autonomous tiltrotor built with Archer Aviation, at the Farnborough Air Show this week.

• Defense tech venture funding more than doubled to over $12 billion in the first half of 2026, exceeding all of 2025, with Shield AI, Mach Industries, and Helsing all raising at sharply higher valuations.

DeepSeek put its record funding round on hold following the leak of a transcript from its investor call and the sudden success of rival Moonshot’s Kimi K3 model. Link.

• The pause comes weeks after DeepSeek closed its first outside round of more than $7.4 billion at a valuation north of $50 billion, structured so commercial investors received no direct equity or voting rights.

• In the leaked four-hour call attributed to founder Liang Wenfeng, he reportedly argued America’s AI lead “comes only from having more computing power” and claimed Nvidia’s CUDA moat is eroding.

• K3’s benchmark-topping open-weight release has scrambled the competitive landscape for Chinese labs just as DeepSeek’s V4 stable release arrived, intensifying the fight for domestic capital and talent.

Alphabet’s revenue jumped 24% on booming Google Cloud growth, and the company raised its full-year 2026 capital expenditure projection to $195 billion to $205 billion, up from $180 billion to $190 billion. Link.

• The quarter served as Google’s clearest justification yet for its massive AI spending, with cloud demand for AI infrastructure and Gemini models driving the acceleration.

• Google’s Gemini app is closing in on one billion users, which would make it the company’s latest billion-user product alongside Search, YouTube, Android, and Gmail.

• The raised capex guidance keeps Google in the top tier of the AI buildout arms race, as hyperscalers collectively push spending toward the trillions Goldman Sachs projects for chips, data centers, and power over the next five years.

In infrastructure, OpenAI unveiled its first custom inference chip, Jalapeño, co-designed with Broadcom and built with early results showing meaningfully better performance-per-watt than current alternatives.

Also this week: Asian AI labs moved quickly into the vacuum left by U.S. export controls, with Tokyo-based Sakana AI launching Fugu (an orchestration-focused frontier model) and China’s Z.ai releasing GLM-5.2, a 744-billion parameter open-weight model that sits within a percentage point of Anthropic’s Opus 4.8 on a key agentic benchmark at roughly one-fifth the cost.

More on this past week’s top AI headlines below!

- ZG

Here are the most important stories of the week:

RESEARCH

Z.ai (formerly Zhipu AI) released GLM-5.2, a 744-billion parameter open-weight model that lands within a percentage point of Anthropic’s Opus 4.8 on a key agentic benchmark at roughly a fifth of the cost, with OpenRouter token traffic climbing faster than after DeepSeek V4’s April launch. Link.

• GLM-5.2’s API costs roughly $1.40 per million input tokens and $4.40 output, versus $5 and $25 for Anthropic Opus 4.8; the model is available under an MIT license and runs on domestic Chinese chips, meaning no government can revoke access after an organization downloads the weights.

• Z.ai released the model on June 13, one day after the U.S. Commerce Department ordered Anthropic to disable Fable 5 and Mythos 5 globally; on Code Arena, GLM-5.2 now ranks second among all usable models, as Fable 5 was removed from the leaderboard following the export ban.

• Z.ai shares surged more than 30% on the Hong Kong exchange following the launch, with JP Morgan raising its price target to 1,400 HK dollars and Bank of America initiating coverage with a buy; the next model, GLM-5.5, is expected in August.

Asian AI startups rushed to fill the gap left by Anthropic’s export ban, with Tokyo-based Sakana AI launching Fugu, an orchestration model it says is “shoulder-to-shoulder” with Anthropic’s Fable 5, while China’s 360 unveiled Tulongfeng, a cybersecurity tool it claims matches Mythos. Link.

• Sakana, co-founded in 2023 by former Google researchers David Ha and Llion Jones, is marketing Fugu to Japanese businesses and government agencies as a hedge against export controls, with its website advertising “delivering frontier capability without the risk of export controls”; the company says the timing of the launch was “entirely coincidental.”

• Sakana CEO Ha argues that orchestration across multiple model APIs is “the next frontier, beyond bigger models,” and that the Mythos/Fable ban made the risk of relying on a single provider “impossible to ignore.”

• Anthropic crossed a $47 billion annualized revenue run rate in May 2026; the two-week-old export ban has already produced at least two locally developed alternatives in markets the company cannot currently serve.

INFRASTRUCTURE

OpenAI unveiled Jalapeño, its first custom inference chip co-designed with Broadcom, with early testing showing significantly better performance-per-watt than current alternatives, as the company extends its stack from models through data centers down to silicon. Link.

• Jalapeño is optimized specifically for inference (running pre-built models in response to user commands), with OpenAI emphasizing its low operating cost when powering real-time coding workflows; pre-training workloads are still expected to run on Nvidia hardware.

• OpenAI’s own AI models assisted in the chip’s development; the Broadcom partnership was officially announced in October 2025, but the chip design had long been rumored as a way to reduce dependence on Nvidia GPUs, following similar moves by Google (TPUs) and Amazon (Trainium).

• OpenAI president Greg Brockman described the company’s rationale as identifying workloads that are “underserved” by general-purpose silicon; the company now frames itself as optimizing every layer of the stack, from model architecture to chip design to product experience.

Reflection AI, an open-source lab founded by two former Google DeepMind researchers, signed a $150 million per month compute deal with SpaceX’s Colossus 2 data center through 2029, worth up to $6.3 billion, becoming the third major AI lab to commit to the facility after Anthropic and Google. Link.

• Reflection’s deal gives it immediate access to Nvidia’s latest GB300 AI chips at SpaceX’s Memphis facility; the contract is smaller than Anthropic’s ($1.25 billion/month) and Google’s ($920 million/month) deals, and either party can exit with 90 days’ notice after the first three months.

• Reflection is positioning the commitment as validation of its open-weight AI strategy, saying recent government restrictions on closed models have made open source “an increasingly attractive alternative” for nations and enterprises.

• Colossus was originally built by xAI (now part of SpaceX) for Elon Musk’s own AI efforts; as xAI’s internal AI pursuits have faltered, SpaceX has pivoted the facility into a major AI compute landlord.

Groq confirmed a $650 million raise and is pivoting to its neocloud business after Nvidia’s $20 billion not-acqui-hire deal last December took founder Jonathan Ross, president Sunny Madra, and the company’s LPU chip IP. Link.

• Groq was last valued at $6.9 billion following a $750 million round in September 2025; it did not disclose a new valuation with the latest raise, which follows Nvidia announcing its own Groq-branded inference hardware cluster, the Nvidia Groq 3 LPX, at GTC in March.

• The neocloud business now spans 13 data centers across North America, Europe, the Middle East, and APAC, serving over 5 million developers and processing trillions of tokens weekly.

• New leadership additions include COO Alan Rice (formerly xAI and Meta) and a CTO/CPO duo, Sinclair Schuller and Rakesh Malhotra, who previously co-founded Nuvalence, a software-engineering firm acquired by EY in 2024.

POLICY/GOV’T/ETHICS

OpenAI limited its GPT-5.6 lineup (Sol, Terra, and Luna) to government-approved partners at the Trump administration’s request, complying while publicly declaring that this kind of government access process “should not become the long-term default.” Link.

• Sol is OpenAI’s most powerful model yet, with improved agentic capabilities in coding, biology, and cybersecurity; it introduces a “max” reasoning mode and an “ultra” mode using coordinated subagents, with tiered pricing at $5/$30 per million input/output tokens for Sol, down to $1/$6 for Luna.

• The Office of the National Cyber Director and the Office of Science and Technology Policy requested the limited release; OpenAI says Sol beats Anthropic’s Claude Mythos 5 slightly on coding workflows and is competitive with Mythos Preview at a third of the output tokens.

• Former White House AI advisor Dean Ball (soon to join OpenAI) has argued the June 2 executive order has created a de facto involuntary licensing regime, warning that launch delays with no defined safety standards could cede ground to China and jeopardize the billions of dollars flowing into AI infrastructure.

The White House asked OpenAI to gate GPT-5.6 to a small group of partners over cybersecurity concerns, CEO Sam Altman reportedly telling staff that the government would be “approving access customer by customer” during a preview period before a broader rollout. Link.

• The request came from the Office of the National Cyber Director and the Office of Science and Technology Policy; OpenAI staffers reportedly “worked closely” with government officials on the release, with Altman indicating a broader rollout could follow “a couple of weeks later” if the limited launch goes well.

• The episode mirrors Anthropic’s Mythos/Fable situation: after Anthropic released Fable 5 publicly, the administration ordered its removal for non-Americans, causing Anthropic to pull the model entirely rather than create a two-tier access system.

• The specific concern around frontier cybersecurity models is that systems like Mythos can identify and exploit software vulnerabilities at speeds no human analyst could match, though the models’ closed nature makes independent verification of threat levels difficult.

AGENTS

General Intuition raised $320 million at a $2.3 billion valuation, led by Khosla Ventures with participation from Jeff Bezos and Eric Schmidt, to scale its agentic model trained on hundreds of millions of hours of video game action data collected through its companion company Medal. Link.

• The company’s core thesis is that human “action labels” embedded in gameplay (records of exactly which buttons a player pressed and when) give agents a richer understanding of causality than video-only training; a robot quadruped running the same model required only 8 minutes of real-world data to fine-tune, collected on a street rather than inside the building where it was tested.

• The majority of the round will go toward scaling compute via a CoreWeave deal, with a portion earmarked to make the API broadly available by end of summer; the round brings total disclosed funding to $454 million, after a $134 million seed in October 2025.

• Vinod Khosla framed the bet around the emergence of “intuition” in AI, arguing that human action and reaction data from games is the key ingredient to a capability leap in world models analogous to what reasoning was for LLMs.

OTHER

Anthropic’s Claude grew its paid consumer base roughly 75% since January 2026, with demand for Claude courses on DataCamp now outpacing ChatGPT three-to-one among self-directed learners, even as ChatGPT retains a commanding overall lead in users and revenue. Link.

• Credit card transaction analysis from Indagari, covering roughly 28 million U.S. consumers, shows Claude’s paying consumer count and revenue both trending up month-over-month through May 10, 2026, with gains continuing even after the March growth spike tied to Anthropic’s refusal to support Trump administration surveillance programs.

• “Claude” is now the most searched term on DataCamp, surpassing even “AI”; Sensor Tower data confirms Claude growing across all platforms in 2026, though the app still trails ChatGPT substantially in total installs and paid users.

• Both Anthropic and OpenAI are heading toward public listings; the ongoing government ban on Anthropic’s Fable and Mythos models adds uncertainty to the growth trajectory as both companies prepare for IPO scrutiny.

Agility Robotics announced plans to go public via SPAC merger with Churchill Capital Corp XI in a deal valuing the company at roughly $2.5 billion and expected to generate more than $620 million in proceeds, including $200 million from new and existing institutional investors. Link.

• Agility is best known for Digit, its bipedal robot currently deployed across nine customer sites including Schaeffler, GXO, Toyota Motor Manufacturing Canada, and Mercado Libre; the company has secured more than $300 million in multi-year orders for its next-generation Digit v5.

• The company counts Amazon, Nvidia, and SoftBank Vision Fund 2 among its existing backers; proceeds from the SPAC will fund production capacity expansion for Digit v5 and fulfillment of existing orders.

• The combined company is expected to trade under the ticker AGLT on a North American exchange not yet announced; CEO Peggy Johnson cited labor shortages and supply chain resilience as the primary enterprise demand drivers.

Israel-based AI21 Labs neared a new funding round that would value the six-year-old company at $1.2 billion, according to The Information, as investor appetite for enterprise AI challenger labs continues. Link.

• AI21, which previously raised at a $1.4 billion valuation in a 2023 Series C backed by Google and Nvidia, recently cut its workforce by 60% in May 2026 to focus on agent optimization; the new round would imply a step-down in valuation.

• AI21 is known for its Jamba model family and Maestro, an AI orchestration system aimed at improving accuracy in complex agentic tasks; existing enterprise customers include Boston Consulting Group, Bain & Company, Moderna, and Booking.com.

• The round comes amid a broader resurgence of interest in open and enterprise-focused AI labs as U.S. government restrictions on the largest frontier providers create openings for more accessible alternatives.

In finance-related news, leaked first-quarter figures showed OpenAI burned $3.7 billion against $5.7 billion in revenue as it eyes an IPO that could value it at up to $1 trillion.

Also, China’s DeepSeek closed a record $7.4 billion first outside round structured to keep founder Liang Wenfeng firmly in control.

More on this past week’s top AI headlines below!

- ZG

Here are the most important stories of the week:

IMAGE/VIDEO

Snap is spinning off its internal generative AI video team into a separate company called Dotmo, citing the high cost of doing the work in-house, with the new venture focused on AI models that create interactive gaming experiences. Link.

• Dotmo’s initial team will be made up of current Snap staff leaving to launch it; Snap won’t fund it directly, but CTO Bobby Murphy will act as lead investor with a significant personal stake while staying on as Snap’s CTO.

• In exchange for a technology license and the departing talent, Snap will take a large equity stake in Dotmo, which may later seek outside funding.

• The move is Snap’s second spinoff this year after it carved out its Specs smart glasses unit; Snap cut roughly 1,000 jobs (16% of its workforce) earlier in 2026, and its stock slid after the $2,200 Specs reveal.

CODING/DEVTOOLS

SpaceX agreed to acquire AI coding startup Cursor in a $60 billion all-stock deal just days after its blockbuster IPO, aiming to bolster the struggling, xAI-powered AI division it pitched as central to its public listing. Link.

• The acquisition, expected to close in the third quarter, follows an unusual April pact in which SpaceX agreed to either buy Cursor for $60 billion in stock or pay a $10 billion break-up fee; it preempted a $2 billion round that would have valued Cursor at $50 billion, with backers including Andreessen Horowitz, Thrive, and Nvidia.

• SpaceX told IPO investors it sees a $26 trillion addressable market for AI products, roughly the size of U.S. GDP.

• The AI division, built around xAI after SpaceX merged with it earlier this year, has been restructuring following controversies including users generating non-consensual deepfakes of women and children.

RESEARCH

Nobel laureate John Jumper, who shared the 2024 chemistry prize for AlphaFold, is leaving Google DeepMind for rival Anthropic after nearly nine years, the latest defection in an intensifying talent war among the top AI labs. Link.

• Jumper announced the move on X, crediting DeepMind CEO Demis Hassabis for letting him lead the AlphaFold team just six months after finishing his PhD; the two won the Nobel for the model that predicts proteins’ 3D structures from their genetic sequences.

• Bloomberg reported Jumper had been a key member of Google’s team building coding tools, an area where the company has struggled to win over business customers.

• His exit came the same week Character AI co-founder and Transformer co-inventor Noam Shazeer also left DeepMind, in his case for OpenAI.

World model startup Odyssey, founded by self-driving veterans Oliver Cameron and Jeff Hawke, raised a $310 million Series B at a $1.45 billion valuation led by Natural Capital, with Amazon, AMD Ventures, and GV joining. Link.

• World models gather data from the physical world and simulate it with accurate physics; Odyssey built its dataset by sending people out with cameras strapped to their backs, echoing how Google Earth was assembled, and is best known for generating rich, interactive video from text prompts.

• As part of Amazon’s backing, AWS becomes Odyssey’s preferred cloud provider and the startup will optimize its models for AWS’s Trainium chips, a rival to Nvidia’s.

• Angel investors include Jeff Dean, Elad Gil, Garry Tan, Guillermo Rauch, and Cruise founder Kyle Vogt; Odyssey has now raised $337 million to date.

General Intuition, which trains AI agents to move through space and time using video game footage, is in talks to raise around $300 million at a roughly $2 billion valuation from backers including Jeff Bezos and Eric Schmidt. Link.

• The round comes just eight months after the New York startup spun out of video-clip platform Medal with a $134 million seed; existing investors Khosla Ventures and General Catalyst are also participating.

• It trains world models on Medal’s dataset of roughly 2 billion gameplay videos a year from 10 million monthly users, betting that interactive, first-person footage is the ideal base for teaching machines spatial-temporal reasoning.

• Unlike rivals such as Runway, Decart, and World Labs that sell world models, General Intuition treats the agents as the product; the dataset has reportedly drawn interest from OpenAI, which previously tried to buy Medal.

POLICY/GOV’T/ETHICS

The Federal Energy Regulatory Commission unanimously ordered six major grid operators to fast-track interconnection requests from AI data centers and other large power users, while stopping short of addressing the underlying shortage of generating capacity. Link.

• Data centers will foot the cost of their own interconnections; operators have 30 days to report spare capacity and 60 days to defend or revise regional electricity rates, and were told to weigh alternative transmission tech such as solid-state transformers and superconducting lines.

• Data center electricity demand is projected to nearly triple through 2035, and wholesale power rates are already up as much as 267% over five years, with the largest U.S. grid operator, PJM, straining under the load.

• FERC acted after Energy Secretary Chris Wright warned that grid-connection delays threatened U.S. AI competitiveness; separately, the Trump administration agreed to pay $765 million to cancel offshore wind leases, bringing its total spent scuttling offshore wind to about $2.6 billion.

The Department of Justice sided with xAI in a NAACP lawsuit seeking to halt the roughly 57 unpermitted gas turbines powering its Memphis data centers, arguing that a shutdown would threaten “national, economic, and energy security.” Link.

• The DOJ memo said xAI’s Grok is one of four AI models supporting critical military operations, including recent U.S. strikes in Iran.

• xAI claims its trailer-mounted, “mobile” turbines are exempt from Mississippi air pollution rules for a year; the Southern Environmental Law Center counters that federal law treats them as stationary, and therefore regulated.

• The number of turbines has more than doubled since last year in an already heavily polluted region, raising PM2.5, formaldehyde, and nitrogen oxide levels; SpaceX’s IPO filing earmarks another $2.8 billion for gas turbines over three years.

Anthropic became the first pure AI company to join Frontier, the carbon removal coalition, anchoring a new $915 million tranche that nearly doubled the group’s total pledges to $1.8 billion. Link.

• Frontier, founded in 2022 by Stripe, Google, and Shopify, vets and pre-purchases carbon removal so member companies can offset emissions they can’t yet eliminate; it has so far contracted nearly $700 million across more than 50 projects.

• The deal is Anthropic’s first climate commitment; the company has yet to publish a sustainability report and has favored an “all of the above” energy stance that often means buying polluting power.

• Frontier said future funding will carry higher scrutiny, concentrating on fewer, larger projects capable of removing a gigaton of CO2 a year, with contracts running eight to 10 years.

OTHER

OpenAI burned through $3.7 billion in the first quarter of 2026, more than half of its $5.7 billion in revenue, as both figures roughly tripled year over year ahead of a confidential IPO filing that could value the company at up to $1 trillion. Link.

• The company posted a $9.3 billion operating loss and a $21.3 billion net loss (inflated by a $12.4 billion non-cash charge) on gross margins of about 39%, underscoring how frontier-model inference grows more expensive as usage climbs.

• OpenAI held more than $73 billion in cash and marketable securities at quarter’s end, up from about $40 billion in December after its large March funding round.

• It has guided to roughly $25 billion in cash burn this year, potentially more than doubling to about $57 billion next year, even as it carries some $665 billion in long-term compute commitments.

Chinese AI lab DeepSeek closed its first-ever outside funding round, raising more than 50 billion yuan (about $7.4 billion) at a valuation north of $50 billion through a structure built to keep founder Liang Wenfeng firmly in control. Link.

• Commercial backers invested into a limited partnership run by Liang rather than the company itself, receiving no direct equity, no voting rights, and a five-year lock-up; Liang himself put in 20 billion yuan, a controlling share.

• Tencent contributed roughly 10 billion yuan and battery giant CATL around 5 billion yuan, yet neither gained governance influence; only the state-linked National AI Industry Investment Fund received direct ownership and voting rights.

• The raise makes DeepSeek China’s most valuable AI startup, though it remains far below U.S. peers (OpenAI recently raised at about $852 billion and Anthropic at $65 billion) given limits on Chinese capital and access to U.S. chips.

OpenAI bulked up ahead of its IPO by hiring Transformer co-inventor Noam Shazeer away from Google DeepMind and bringing on former Trump White House AI policy official Dean Ball to lead a new “Strategic Futures” team. Link.

• Shazeer, who co-authored the seminal 2017 paper that introduced the Transformer and co-founded Character AI, had been a Gemini co-lead; Google re-hired him two years ago in a $2.7 billion deal for Character AI’s technology.

• Ball, who helped write the White House’s AI Action Plan, will report to Chief Strategy Officer Jason Kwon and focus on catastrophic risk, recursive self-improvement, labor-market impact, and AI governance.

• The hires land as OpenAI cements its insider status in Washington while rival Anthropic battles the administration, which just ordered an export ban on its Fable 5 and Mythos 5 models.

Castelion, a hypersonic missile startup founded by three SpaceX alumni, is in talks to raise a new round at a valuation of $12 billion or more, at least quadrupling its December mark as the SpaceX IPO supercharges investor appetite for defense tech. Link.

• Founded in 2022 by Sean Pitt, Bryon Hargis, and Andrew Kreitz, the company builds hypersonic and long-range strike weapons and has already raised over $553 million across six rounds.

• It was spurred by a 2022 U.S. Navy request for reliable hypersonic suppliers, with the founders frustrated by slow American progress relative to China and Russia.

• Castelion is building a large production campus in New Mexico to scale manufacturing as it pursues Navy contracts.

Google signed a $30 billion compute deal with SpaceX, paying $920 million per month for access to roughly 110,000 Nvidia GPUs at SpaceX’s data center facilities. The Trump administration is in talks to take an equity stake in OpenAI through a proposed “Public Wealth Fund,” while also signing a narrower AI executive order establishing voluntary model review 30 days before public release.

Also this week, OpenAI is merging ChatGPT and Codex into an agentic “super app,” Microsoft launched Scout (an OpenClaw inspired personal assistant), Meta is planning to charge up to $200 a month for its Hatch AI agent, and Helion Energy nearly tripled its valuation to $15.5 billion after demonstrating fusion with its Polaris prototype.

More on this past week’s top AI headlines below!

- ZG

Here are the most important stories of the week:

AGENTS

OpenAI is preparing a major redesign of ChatGPT into an agentic “super app,” merging ChatGPT, Codex, and the developer API into a single product group under co-founder Greg Brockman, with the unified experience expected to roll out in the coming weeks. Link. Link.

• Codex now has more than 5 million weekly active users (a 6x increase since launch), and roughly 20% of them are non-developers such as analysts, marketers, designers, and bankers, a segment growing 3x faster than the developer user base.

• A senior OpenAI employee declared “Chat is dead,” reflecting the company’s conviction that autonomous agents capable of multi-step tasks (booking travel, managing calendars, writing code) are the future, not simple Q&A interfaces.

• Goldman Sachs and Morgan Stanley are advising on an IPO that could exceed $1 trillion; the super app pivot is explicitly aimed at closing the gap with Anthropic among business customers ahead of that listing.

Microsoft unveiled Scout at its Build developer conference, an always-on personal assistant built on the OpenClaw agent framework that integrates across Outlook, Teams, OneDrive, and SharePoint, with users naming their own instance and training it over time. Link.

• Scout is cloud based and comes with prepackaged skills for calendar management and drafting meeting agendas, plus the ability for users to develop custom skills; a built-in “policy conformance system” continuously audits whether the agent is operating within guidelines.

• OpenClaw burst onto the scene in early 2026 as an unrestrained agent framework before OpenAI scooped up its founder; an earlier incident where an OpenClaw agent “acted erratically inside a researcher’s inbox” raised safety concerns that Microsoft’s guardrails aim to address.

• Scout requires a GitHub Copilot subscription and is available through Microsoft’s Frontier early adopter program as an experimental release.

Apple approved Poke as the first standalone AI agent on its Messages for Business platform, marking the first time Apple has allowed a third-party AI agent to operate natively on iMessage. Link.

• Poke handles daily planning, calendar management, health and fitness tracking, smart home control, photo editing, web searches, image generation, flight check-ins, and YouTube video summaries, all via text message; the agent has relayed approximately 100 million messages to date.

• Apple required Poke to verify it could offer live support if needed, clearly identify itself as non-human, and customize its UI to meet Apple’s design guidelines, a process that took “a couple of months.”

• The approval comes just days before WWDC 2026 on June 9, where Apple is expected to unveil its AI powered Siri overhaul and potentially open the App Store to additional AI agents.

Meta launched its AI powered “Business Agent” globally within WhatsApp and Instagram DMs, ending nearly two years of limited testing, with more than one million businesses already using it. Link.

• The agent can answer customer questions, recommend products, book appointments, qualify sales leads, and reroute queries to a human when needed; it is available at no cost initially, with paid tiers planned in coming months.

• Meta is testing additional features including daily briefings of overnight chats with insights, market research capabilities, competitive intelligence tools, and a platform for larger enterprises to create custom agents that connect to Shopify, Zendesk, and Shopee.

• The announcement was made at Meta’s Conversations conference in London; large businesses will eventually pay based on token usage as part of the WhatsApp Business Premium subscription.

Meta is developing a consumer AI agent called “Hatch” with tiered pricing that could include a premium subscription of up to $199.99 per month, marking the company’s first commercially priced AI product. Link.

• Internal documents show a free version and a “Hatch Plus” subscription with 5 to 10x higher usage limits; Hatch handles vibe coding, sending emails on behalf of users, shopping, filling out forms, product research, and content creation.

• A broader U.S. launch is planned for July 2026; final pricing decisions have not been made and could change before launch.

• The move puts Meta in direct competition with OpenAI’s ChatGPT Pro and Anthropic’s Claude Max (both $100 to $200 monthly) and represents a significant strategic shift for a company whose AI strategy has historically centered on free, open source models.

OpenAI launched “Lockdown Mode,” an optional security toggle for ChatGPT that disables live web browsing, agent mode, deep research, and file downloads to protect sensitive data from prompt injection attacks. Link.

• When enabled, Lockdown Mode blocks the final stage of prompt injection attacks (unauthorized outbound data transfer) by disabling live web browsing (cached content only), retrieval of web images, Deep Research, Agent Mode, Canvas networking, live connectors, and file downloads.

• The feature is available to all personal accounts including the free tier, activated via Settings, Security, Advanced Security, Lockdown Mode; OpenAI acknowledges it is not a complete fix since prompt injections can still affect response behavior.

• This is one of the first major product level mitigations any frontier AI company has shipped specifically for prompt injection, a vulnerability class that has plagued all LLM based products since their introduction.

RESEARCH

Anthropic expanded Project Glasswing, its security vulnerability scanning program powered by Claude Mythos, from 50 initial U.S. government partners to 150 organizations across more than 15 countries, targeting critical infrastructure in power, water, healthcare, and telecommunications. Link.

• Claude Mythos has surfaced more than 10,000 high or critical severity software vulnerabilities since the program launched in early April 2026, establishing itself as a dominant tool for AI driven cybersecurity scanning.

• New partners include Okta, Samsung, SK Hynix, SK Telecom, NATO, and ENISA (the EU’s cybersecurity agency), expanding well beyond the initial U.S. government focus.

• The expansion announcement came one day after Anthropic filed confidentially for an IPO, serving the dual purpose of demonstrating geopolitical relevance and national security utility ahead of the public listing.

CODING/DEVTOOLS

Lovable, the Stockholm based vibe coding startup, signed a multi-year deal with Google Cloud that will increase its cloud footprint fivefold, including expanded access to both Anthropic’s Claude and Google’s Gemini models. Link.

• Lovable crossed $400 million in annualized revenue in February 2026, having added $100 million in a single month with just 146 employees; more than half of Fortune 500 companies now use its product.

• Under the deal, Lovable’s agent will be available through Google Cloud’s Gemini Enterprise Agent Gallery, and the company gains integration with Wiz (Google’s $32 billion acquisition) for real-time security scanning of generated code.

• Lovable’s heavy usage of Claude through Google Cloud helps Anthropic hit performance targets tied to Google’s $10 billion investment in Anthropic at a $350 billion valuation, with another $30 billion promised if those targets are met.

IMAGE/VIDEO

Google Labs launched Dreambeans, an AI app for iOS and Android that reads your Google data (Gmail, Calendar, Photos, YouTube, Search History) and generates 10 to 14 AI illustrated “stories” every morning with personalized lifestyle suggestions. Link.

• Dreambeans uses Google’s “Personal Intelligence” system to process data overnight and deliver stories in the morning; when stories involve the user or people close to them, it uses Google Photos’ Face Grouping and the Nano Banana 2 image model to paint their likeness into cartoon style visuals.

• The app is intentionally limited to combat doomscrolling, with no infinite scroll and no algorithm rabbit holes; product lead Gozde Oznur said “the idea is to get a few inspirational ideas and then go out and live your life.”

• Currently available in the U.S. only for Google AI Ultra subscribers (18+), with a waitlist open for everyone else.

INFRASTRUCTURE

Google signed a deal to pay SpaceX $920 million per month from October 2026 through June 2029 for access to approximately 110,000 Nvidia GPUs at SpaceX’s Memphis data center facilities, a contract worth roughly $30 billion over 32 months. Link.

• Google described the arrangement as a “short-term bridge” to meet surging demand for its Gemini Enterprise AI platform; either party can terminate with 90 days’ notice after December 31, 2026.

• This is the second major compute deal at SpaceX’s data centers following Anthropic’s commitment of $1.25 billion per month through 2029 for all available compute at the Colossus 1 facility, meaning Google is getting roughly half the compute at about 74% of Anthropic’s monthly rate.

• The fact that Google, which operates one of the world’s largest cloud infrastructures, needs to rent GPU capacity from SpaceX underscores just how intense AI compute demand has become across the industry.

AirTrunk, the Blackstone backed Australian data center operator, committed $30 billion to build 5 gigawatts of AI data center capacity in India by 2030, anchored by a 3 gigawatt campus outside Mumbai worth approximately $21 billion. Link.

• AirTrunk CEO Robin Khuda announced the deal following a meeting with Indian Prime Minister Narendra Modi; the centerpiece Raigad Pen Growth Centre project has a signed letter of intent for land allotment with the Maharashtra state government.

• AirTrunk entered India through its acquisition of Lumina CloudInfra in April 2026, gaining an existing 600 megawatt pipeline across Mumbai, Chennai, and Hyderabad.

• India’s current data center capacity is approximately 1.5 gigawatts, projected to reach 8 gigawatts by 2030; if AirTrunk delivers its 5 gigawatt commitment, it would become the single largest data center operator in the country.

Alphabet priced a record breaking equity capital raise totaling $85 billion (the largest equity offering in history), with the first $45 billion tranche so oversubscribed that it exceeded its original $40 billion target, and a second $40 billion tranche planned for next quarter. Link.

• Berkshire Hathaway purchased $10 billion worth; CEO Sundar Pichai described the raise as “part of our multi-year investment strategy to meet the AI opportunity ahead,” with capital expenditures projected at $180 to $190 billion this year.

• The previous record for equity offerings was held by Brazilian oil producer Petroleo Brasileiro SA, which raised $70 billion in 2010; Alphabet generated $110 billion in revenue in Q1 alone, up 22% year over year.

• The offering is seen as a strong positive signal for the AI IPO pipeline, with Anthropic already filed, SpaceX’s listing imminent, and OpenAI waiting in the wings; Goldman Sachs estimates nearly $8 trillion in AI spending has been committed over the next five years.

Goldman Sachs projects SpaceX will burn through $350 billion in cash by 2030, with the vast majority going to capital expenditures tied to its AI business, as the company prepares for a June 12 Nasdaq debut at a targeted $1.75 trillion valuation. Link.

• AI division revenue is projected to surge from $3.2 billion in 2025 to $322 billion by 2030 (a 100x increase), with total company revenue reaching $474 billion; Starlink revenue alone is projected to hit $144 billion by 2030.

• Free cash flow is expected to hit a trough of negative $105 billion in 2029 before rebounding to positive $72 billion by 2031; the company will burn $120 billion in 2026 and 2027 alone.

• SpaceX’s consolidated 2025 revenue was $18.67 billion following the February 2026 all-stock merger with xAI; the IPO would trade under the ticker “SPCX” and dwarf Saudi Aramco’s $29 billion 2019 offering.

Nvidia acquired Kumo AI, a Mountain View based enterprise predictive AI startup specializing in graph neural networks, for at least $400 million, bringing aboard the former Pinterest CTO and a Stanford professor who co-founded the company. Link.

• Kumo was founded in 2021 by former Pinterest CTO Vanja Josifovski, Stanford professor Jure Leskovec, and former LinkedIn AI head Hema Raghavan; the company had raised $37 million led by Sequoia Capital before the acquisition.

• Kumo’s technology maps relationships across business data (customers, products, services) to make predictions about churn, demand, and risky transactions, cutting manual ML pipeline work by up to 95%; pre-acquisition customers included Reddit, Sainsbury’s, DoorDash, Databricks, and Snowflake.

• The acquisition continues Nvidia’s push into enterprise AI software on top of its dominant hardware position, following earlier deals for Run.ai and other AI software companies.

ENERGY

Helion Energy, the fusion startup backed by Sam Altman, raised a $465 million Series G led by Thrive Capital that nearly tripled its valuation from $5.4 billion to $15.5 billion, following the first privately developed demonstration of deuterium-tritium fusion. Link.

• Helion’s Polaris prototype in Everett, Washington, achieved plasma temperatures exceeding 150 million degrees Celsius in early 2026, becoming the first private fusion machine to demonstrate measurable deuterium-tritium fusion; total capital raised now exceeds $1.5 billion.

• The company has committed to delivering at least 50 megawatts of fusion electricity to power a Microsoft data center in central Washington by 2028, with Constellation Energy serving as power marketer; Helion’s approach harvests electricity directly from magnets rather than using steam turbines.

• New investors include Ford Motor Company Executive Chairman Bill Ford, Alta Park Capital, BoxGroup, and Lux Capital; existing backers include SoftBank Vision Fund 2, Lightspeed Venture Partners, and Dustin Moskovitz.

POLICY/GOV’T/ETHICS

President Trump signed an executive order titled “Promoting Advanced Artificial Intelligence Innovation and Security” that establishes a voluntary framework for AI companies to submit frontier models for government review up to 30 days before public release, significantly narrowed from an earlier draft that called for 90 days. Link.

• The order directs federal agencies to develop benchmarks for assessing AI models’ cyber capabilities and creates an “AI cybersecurity clearinghouse” to coordinate vulnerability scanning and patching across government systems.

• Trump delayed signing the original version in late May after pushback from industry leaders, including former White House AI czar David Sacks; the final order explicitly states that nothing in it authorizes mandatory licensing, preclearance, or permitting requirements for AI models.

• The administration’s stated concern is that advanced systems (including Claude Mythos and GPT-5.5) could be exploited for cyberattacks against critical infrastructure, but the voluntary nature of the framework leaves enforcement largely toothless.

The Trump administration and OpenAI are in ongoing talks about the U.S. government taking a direct equity stake in OpenAI to seed a “Public Wealth Fund” that would let American citizens participate in the upside of AI growth. Link.

• President Trump confirmed the discussions from Air Force One, saying “there are concepts where pieces could be given to the American public, where the American public essentially becomes a partner.”

• Sam Altman first pitched the idea in 2025; the fund would invest in diversified, long-term assets and enable citizens to participate in AI’s economic gains, possibly by receiving the fund’s returns directly.

• No official investment terms have been decided; OpenAI is currently valued at more than $850 billion and is preparing for an IPO as soon as this year. This would be unprecedented, as the U.S. government has never taken a direct equity stake in a private technology company.

Sriram Krishnan, the White House senior policy adviser for AI, announced he will leave his post at the end of June 2026 to start an outside institution focused on technology policy. Link.

• Krishnan, a former Andreessen Horowitz partner, was tapped by Trump to help shape AI policy during his second term; he played a key role in creating a national framework for AI regulation and helped shape the AI executive order signed earlier this week.

• He said he plans to “build institutions that would help tackle some of the challenges America and its allies are facing,” though he did not give a specific reason for his departure.

• His exit leaves a gap in AI policy leadership at the White House during one of the most consequential periods for AI governance, with multiple trillion dollar IPOs on the horizon and the first federal AI oversight framework just days old.

OTHER

Anthropic filed a confidential S-1 prospectus with the SEC, setting up what could be one of the largest AI listings in history, following its $65 billion Series H that valued the company at $965 billion post-money. Link.

• Annualized revenue crossed $47 billion in May 2026, up from $30 billion at the end of Q1 and roughly $9 to $10 billion at the end of 2025, representing a nearly 5x year over year increase driven by demand for Claude Code and enterprise AI products.

• The Series H was co-led by Altimeter Capital, Dragoneer, Greenoaks, and Sequoia Capital, with Amazon contributing $5 billion as part of $15 billion in hyperscaler commitments; other investors include Blackstone, Fidelity, Coatue, Samsung, SK Hynix, and Micron.

• The confidential filing grants flexibility for a listing as soon as fall 2026; Anthropic could become one of three American companies (alongside SpaceX and potentially OpenAI) to go public this year at a valuation above $1 trillion.

Airbnb CEO Brian Chesky plans to back and fund a new AI lab focused on developing models for user interaction and design, signaling frustration with existing frontier models for travel and e-commerce applications. Link.

• Chesky will remain as Airbnb CEO and will not lead the new lab himself; on Airbnb’s Q1 2026 earnings call, he said the company did not believe anyone had “figured out AI for travel or e-commerce yet.”

• Airbnb has adopted AI coding tools internally but has not struck an LLM partnership because existing products “weren’t quite ready”; Chesky argues AI for travel requires a rich user interface, not text based chatbots.

• Chesky and Sam Altman have been close since meeting through Y Combinator in 2006, and Chesky helped Altman return to power during the OpenAI board crisis in 2023; the new lab effectively positions Chesky as both an AI ally and a competitor.

Impulse Space, the rocket engine startup founded by SpaceX’s first employee Tom Mueller, raised a $500 million Series D at a $4.26 billion valuation, with the explicit plan to hire up to 200 hardware engineers rather than invest in AI systems. Link.

• The round was led by 137 Ventures and BANNER VC, with participation from Founders Fund, Lux Capital, and Linse Capital, bringing total funding to over $1 billion; company leadership stated “there’s not really any substitute for designing the thing, analyzing the thing, building it, and then getting it on the test stand.”

• Impulse builds in-space mobility vehicles: Mira (a maneuverable platform targeted at the U.S. Space Force) and Helios (designed to carry satellites rapidly to high orbits after initial deployment closer to Earth).

• The “hire people, not AI” stance is a notable counter-narrative amid the broader AI hiring boom, reflecting the reality of hardware intensive space engineering where physical testing remains irreplaceable.

Cognition raised $1 billion at a $25 billion valuation for its Devin coding agent, now generating $492 million in ARR, while Apple previewed a sweeping Siri overhaul for iOS 27 with a standalone app, Dynamic Island integration, and Google’s Gemini powering the intelligence layer. Robinhood opened its brokerage platform to AI agents through MCP, letting autonomous systems trade stocks and manage finances on behalf of users.

Also this week, Anthropic released Opus 4.8 with a new Dynamic Workflows feature for parallel subagent coordination, Meta announced an AI pendant and four new smart glasses models, ElevenLabs launched Music v2 with mid-track genre switching, and ByteDance revealed it is developing its own inference chips with roughly 1,000 chip designers on the project.

More on this past week’s top AI headlines below!

- ZG

Here are the most important stories of the week:

AGENTS

Apple previewed a sweeping overhaul of Siri for iOS 27, including a standalone Siri app with a full screen chat interface, Dynamic Island integration, and Google’s Gemini model powering the intelligence layer through an on-device distillation deal. Link.

• The new standalone app (codenamed “Campos”) replaces the current overlay with a full screen experience that includes a conversational chat interface, the ability to control apps, summarize content, and compose messages.

• Apple struck a deal with Google to distill Gemini’s capabilities into smaller models that run directly on device, preserving Apple’s privacy first approach while closing the gap with ChatGPT and other cloud based competitors.

• The company is also building its own foundation models internally, but the Gemini partnership provides a near term bridge; the iOS 27 update is expected to debut at WWDC in June.

Robinhood announced that AI agents can now trade stocks and manage finances on its platform through MCP (Model Context Protocol) integration, and introduced an “agentic” credit card designed for autonomous AI spending. Link.

• The MCP integration allows AI agents from any provider to connect to Robinhood, execute trades, monitor portfolios, and make financial decisions on a user’s behalf with granular permission controls.

• Robinhood also unveiled a credit card that AI agents can use to make purchases autonomously, with built in spending limits and approval workflows to prevent runaway transactions.

• The move makes Robinhood one of the first major financial platforms to embrace fully autonomous AI trading, raising questions about regulatory oversight of agent driven financial activity.

Meta announced an AI pendant and four new smart glasses models as part of a push to sell 10 million wearable AI devices in the second half of 2026, expanding well beyond the Ray-Ban collaboration. Link.

• The AI pendant is a standalone wearable that listens, sees, and responds without requiring glasses, targeting users who want an always available AI companion in a minimal form factor.

• The four new smart glasses models span different price points and use cases, from fashion forward frames to sport specific designs with enhanced camera and audio capabilities.

• Meta’s 10 million unit sales target for the second half of 2026 would represent a massive acceleration, signaling that the company views wearable AI as a major product category beyond VR headsets.

Meta launched an “Enterprise Solutions” unit with forward deployed engineers who embed directly with corporate customers to build custom AI applications on Meta’s models and infrastructure. Link.

• The new unit mirrors the forward deployed engineering model pioneered by Palantir, where Meta’s engineers work on site with enterprise clients to build bespoke AI solutions tailored to their needs.

• The initiative targets large corporations that want to adopt Meta’s open source Llama models but need hands on implementation support for complex, mission critical deployments.

• The move marks a significant expansion of Meta’s enterprise ambitions, adding a high touch services layer on top of its open source model strategy to compete more directly with Anthropic and OpenAI in the enterprise market.

SPEECH/AUDIO

Sesame, the conversational AI startup founded by Oculus creator Brendan Iribe, launched its iOS app with four distinct AI voice agents, all available for free, with plans for a smart glasses product by 2027. Link.

• The four agents each have unique personalities and voice characteristics, designed to feel like distinct conversational companions rather than a single assistant with different settings.

• The app is free at launch, with Sesame betting on user acquisition before introducing premium tiers; the company has raised significant funding backed by Iribe’s track record building and selling Oculus to Meta for $2 billion.

• Sesame’s smart glasses roadmap positions it as a future competitor to Meta’s Ray-Ban partnership, with voice first AI interaction as the core differentiator.

ElevenLabs launched Music v2 with a genre switching capability that allows AI generated tracks to shift between musical styles mid-song, a first for any major AI music platform. Link.

• The genre switching feature lets users specify transitions within a single track (for example, moving from jazz to electronic to orchestral), creating dynamic compositions that evolve over the course of the song.

• Music v2 also includes improved vocal synthesis, longer generation lengths, and more granular control over instrumentation and arrangement.

• The update intensifies competition with Suno and Udio, which have faced ongoing copyright lawsuits from major record labels over their training data practices.

CODING/DEVTOOLS

Cognition raised $1 billion at a $25 billion pre-money valuation for Devin, its AI coding agent that now generates $492 million in annualized recurring revenue. Link.

• The round values Cognition at roughly 50x its ARR, reflecting investor confidence in the AI coding agent market’s growth trajectory; Devin was first introduced in March 2024 as the “first AI software engineer.”

• Cognition has rapidly scaled from its initial demo to nearly half a billion in revenue, driven by enterprise adoption of autonomous coding agents that can handle complex, multi-step software engineering tasks.

• The raise intensifies the AI coding wars, with Anthropic’s Claude Code, OpenAI’s Codex, and Cursor all competing for developer adoption in a market that could reshape how software gets built.

Anthropic released Opus 4.8 with a new “Dynamic Workflows” feature that enables parallel subagent coordination, allowing the model to dispatch multiple specialized agents simultaneously for codebase-scale migrations and complex engineering tasks. Link.

• Dynamic Workflows lets Opus 4.8 break down large tasks into parallel subtasks, each handled by a specialized subagent, then synthesize the results into a coherent output, dramatically speeding up work like migrating entire codebases or refactoring large projects.

• The feature represents a shift from single threaded AI reasoning to orchestrated multi-agent execution within a single model interaction, letting developers hand off entire project level tasks rather than individual prompts.

• Opus 4.8 is positioned as Anthropic’s most capable model for professional software engineering, competing directly with OpenAI’s Codex and Google’s Jules for high stakes development workflows.

INFRASTRUCTURE

SoftBank committed EUR 75 billion (roughly $87 billion) to building data centers across France with up to five gigawatts of compute capacity, one of the largest single country AI infrastructure investments ever announced. Link.

• The investment will fund multiple hyperscale data center campuses across France, leveraging the country’s relatively abundant nuclear power supply and favorable energy costs compared to other European nations.

• SoftBank CEO Masayoshi Son personally announced the deal alongside French President Macron, who has made attracting AI investment a central policy priority for France.

• The commitment follows SoftBank’s $100 billion Stargate pledge in the United States and positions France as a major hub for AI compute in Europe.

Groq is raising $650 million after Nvidia’s $20 billion “not quite acquisition” deal fell apart, pivoting its business model from chip design toward operating as an inference neocloud provider. Link.

• The pivot means Groq is positioning itself more as a cloud service that sells inference compute (powered by its custom LPU chips) rather than as a chip company that sells hardware, similar to how Cerebras has evolved.

• Nvidia’s attempted $20 billion deal, which would have given it significant control over Groq without a full acquisition, did not proceed; Groq is now raising independently to fund its infrastructure buildout.

• Groq’s LPU (Language Processing Unit) architecture is designed specifically for inference speed, offering significantly lower latency than GPU based alternatives for serving large language models.

South Korean startup Xcena raised $135 million at a $570 million valuation, developing memory chips purpose built for AI workloads that aim to solve the memory bandwidth bottleneck limiting inference performance. Link.

• The company’s chips are designed to address the growing gap between compute speed (which has improved dramatically with GPUs and custom accelerators) and memory speed (which has not kept pace), a bottleneck that limits how fast AI models can run.

• Xcena offers a specialized alternative to the high bandwidth memory (HBM) produced by Samsung and SK Hynix, targeting a different slice of the AI memory architecture.

• The raise positions Xcena in a rapidly growing market for AI specific silicon, alongside companies like Cerebras, Groq, and Etched that are building alternatives to Nvidia’s GPU centric approach.

Snowflake signed a $6 billion deal with AWS to use Amazon’s custom Graviton CPU chips, marking a major shift for the data cloud company toward Arm based compute infrastructure. Link.

• The multi-year agreement represents one of the largest cloud compute commitments by a single company, reflecting Snowflake’s massive data processing needs as it expands its AI and machine learning offerings.

• Graviton chips, designed by Amazon’s in-house Annapurna Labs team, offer better price performance than traditional x86 processors for data intensive workloads.

• The deal underscores the growing enterprise appetite for Arm based cloud infrastructure, a trend reshaping the server chip market as companies optimize for cost efficiency at scale.

ByteDance is developing its own inference chips in a dual track effort using both Arm and RISC-V architectures, with roughly 1,000 chip designers working on the project. Link.

• The company aims to reduce its dependence on Nvidia GPUs for running inference on its AI products, which span TikTok’s recommendation engine, the Doubao assistant, and other AI services serving hundreds of millions of users.

• ByteDance is pursuing both Arm (for near term deployment) and RISC-V (for longer term cost savings and licensing independence) architectures simultaneously, hedging its bets on which will prove more efficient.

• The effort follows a similar path taken by Google (with TPUs), Amazon (with Inferentia and Trainium), and Meta (with MTIA), as large tech companies invest in custom silicon to reduce inference costs at scale.

Baseten, the AI inference platform startup, is raising $1 billion at an $11 billion valuation after reaching $600 million in annualized recurring revenue. Link.

• Baseten provides infrastructure that helps companies deploy and serve AI models at scale, competing with platforms like Modal, Replicate, and major cloud providers’ AI serving solutions.

• The $600 million ARR represents rapid growth for the company, driven by surging enterprise demand for reliable, low latency AI inference infrastructure as more businesses deploy models in production.

• The raise would value Baseten at roughly 18x its ARR, reflecting investor enthusiasm for the “picks and shovels” layer of the AI stack where revenue is recurring and margins improve with scale.

OpenRouter more than doubled its valuation to $1.3 billion as the AI model routing platform now serves 100 trillion tokens per month across 8 million users. Link.

• OpenRouter acts as a unified API layer that lets developers and consumers access models from OpenAI, Anthropic, Google, Meta, and dozens of other providers through a single interface, routing requests to the best model for each task.

• The 100 trillion tokens per month figure represents staggering scale, driven by both developer API usage and OpenRouter’s consumer chat interface that lets users compare model outputs side by side.

• The platform’s growth reflects a broader trend toward model agnostic infrastructure, as customers increasingly want flexibility to switch between AI providers rather than being locked into a single vendor.

OTHER

Anthropic raised a $65 billion Series H at a $965 billion valuation, potentially the company’s final private funding round before an IPO, as annualized recurring revenue crossed $47 billion. Link.

• The round makes Anthropic one of the most valuable private companies in history, trailing only SpaceX at its pre-IPO valuation and approaching the trillion dollar mark.

• Anthropic’s ARR growth from $30 billion at the end of Q1 to $47 billion represents an extraordinary acceleration, driven by demand for Claude Code and enterprise AI products across legal, finance, and healthcare verticals.

• The “potentially last round before IPO” framing suggests Anthropic could go public in late 2026 or 2027, joining Cerebras and potentially OpenAI in the wave of major AI IPOs reshaping public markets.

OpenAI launched a self-serve advertising manager, going downmarket to let small businesses buy ads that appear within ChatGPT conversations and expanding beyond the company’s initial enterprise and subscription focus. Link.

• The self-serve platform lets small businesses create and manage ad campaigns without needing a dedicated sales relationship with OpenAI, similar to how Google and Meta democratized digital advertising for local businesses.

• OpenAI is positioning ads as a major new revenue stream alongside subscriptions and API access, targeting the massive small business advertising market that has powered Google and Meta’s growth for two decades.

• The move represents a strategic shift for OpenAI, which had previously focused on premium subscriptions and enterprise deals; the timing follows Anthropic’s own push into small business with Claude for Small Business earlier this month.

Anthropic surpassed OpenAI in business customer adoption for the first time according to Ramp data, launched a new small business initiative with integrations into QuickBooks, Canva, and HubSpot, and saw its legal AI ambitions put Clio’s $500 million milestone in a new light. Meanwhile, BlackRock is weighing a $5 to $10 billion investment in SpaceX’s upcoming IPO at a potential $1.75 to $2 trillion valuation.

Also this week, Greg Brockman took charge of OpenAI’s product strategy while Fidji Simo remains on medical leave, Google and SpaceX are in talks to put data centers into orbit, and GM laid off hundreds of IT workers to hire those with AI skills instead.

More on this past week’s top AI headlines below!

- ZG

Here are the most important stories of the week:

AGENTS

Notion launched a developer platform that transforms its workspace into an orchestration hub for AI agents, letting users chat directly with external agents including Claude Code, Cursor, and Codex, assign them work, and track progress alongside human collaborators. Link.

• New features include Workers (a cloud-based environment for running custom code in a secure sandbox), Database Syncing (pulling data from Salesforce, Zendesk, Postgres, and others directly into Notion databases), and MCP Protocol support for agent tools.

• CEO Ivan Zhao said “any data, any tool, any agent, that’s the big picture for the Notion Developer Platform,” acknowledging that “historically, Notion hasn’t been the most developer-focused platform.”

• Notion is positioning itself as an orchestration layer that coordinates AI work across multiple tools and data sources, with Decagon and additional partner agents planned beyond the initial launch.

Apple is designing a framework to allow AI agent apps into the App Store while maintaining its privacy, security, and payment rules, with potential updates coming at WWDC 2026 on June 8. Link.

• The core challenge is that agents can dynamically spin up sub-applications after Apple has already approved the parent app, circumventing the review process and potentially bypassing App Store fees.

• Apple’s staffers are building a system to prevent the kind of runaway behavior seen in agentic systems, including agents going haywire and deleting users’ emails or spreading malware.

• The existing app review model, where Apple approves a static app, breaks down when an agent can generate new functionality on the fly after approval.

Anthropic launched “Claude for Small Business,” a new suite of services targeting local businesses like hardware stores and coffee shops, with integrations between Claude Cowork and QuickBooks, Canva, Docusign, HubSpot, and PayPal. Link.

• The features are available via a newly introduced toggle within Claude Cowork, Anthropic’s task-automation platform that can browse the web, manage files, and execute multistep workflows on a user’s behalf.

• Anthropic is planning a coast-to-coast promotional tour starting in Chicago and hitting 10 cities, offering free AI training workshops available to 100 local small business leaders at each stop.

• The initiative represents a significant strategic shift for Anthropic, moving beyond enterprise and developer customers to target the massive small business market segment.

CODING/DEVTOOLS

OpenAI announced that Codex is being integrated into the ChatGPT mobile app, allowing developers to assign coding tasks, ask questions about their codebase, and monitor agent progress from their phones. Link.

• The latest version runs on GPT-5.3 Codex, which is 25% faster than GPT-5.2 and was notably the company’s first model that “was instrumental in creating itself.”

• Users can find the tool in ChatGPT’s sidebar, assign the agent new tasks by typing a prompt and clicking the “Code” button, and ask questions about their codebase via the “Ask” button.

• The move intensifies competition with Anthropic’s Claude Code and is part of OpenAI’s broader push toward a unified AI “super app” combining ChatGPT, Codex, and its AI browser.

Canadian legal tech company Clio hit $500 million in ARR, but the milestone is complicated by Anthropic’s simultaneous expansion of “Claude for Legal” with new plug-ins and MCP connectors, turning a key supplier into a competitor. Link.

• Clio was valued at $5 billion in its $500 million Series G last November and completed a $1 billion cash-and-stock acquisition of vLex, a 26-year-old legal data intelligence platform.

• Both Harvey (which raised $200 million at an $11 billion valuation in March) and Legora (which raised a $600 million Series D) rely on Claude as a core model, creating an uncomfortable dynamic where their key AI supplier is now also a direct competitor.

• Legora launched a high-profile ad campaign featuring actor Jude Law as the legal AI market rapidly heats up with multiple well-funded players.

SPEECH/AUDIO

Thinking Machines Lab, the AI startup founded by former OpenAI CTO Mira Murati, announced a research preview of a “full duplex” AI interaction model that responds in 0.40 seconds and can listen and speak simultaneously, making conversations feel more like a phone call. Link.

• TML-Interaction-Small roughly matches the speed of natural human conversation and is significantly faster than comparable models from OpenAI and Google; the “full duplex” capability allows for natural interruptions and back-and-forth.

• This is a research preview, not a consumer product; Thinking Machines was founded in early 2025 and closed a $2 billion round at a $10 billion valuation (later valued at $12 billion).

• The lab lost two co-founders back to OpenAI in January 2026 but deepened its ties with Google in April through a multibillion-dollar infrastructure deal.

RESEARCH

Junyang Lin, the architect behind Alibaba’s Qwen open-source model family, launched a new AI lab focused on world models and embodied intelligence, seeking hundreds of millions at a $2 billion valuation with backing from HongShan (formerly Sequoia China) and Gaorong Ventures. Link.

• Lin publicly announced his resignation on X in March, posting “bye my beloved qwen”; during his three years at Alibaba he elevated Qwen into one of the world’s leading open-source AI model families.

• A $2 billion valuation for a startup with no product yet is nearly unprecedented in China, where AI startups are typically valued far lower than U.S. peers; Lin’s personal track record with Qwen justifies the premium.

• Lin has already recruited core team members from ByteDance, Tencent, and overseas AI companies, shifting his focus from large language models to robotic AI brains.

INFRASTRUCTURE

Cerebras went public in the biggest tech IPO of 2026, raising $5.5 billion at $185 per share before the stock surged 108% to close at $311 on its first day of trading, giving the AI chip maker a roughly $66 billion valuation. Link.

• Revenue was approximately $510 million in 2025 (up 76% YoY) and the company swung to $237.8 million in net income after losing nearly half a billion the year before; Benchmark’s stake was worth $5.5 billion at close and Eclipse turned a $6.5 million Series A into a $2.5 billion return.

• OpenAI signed a $10 billion compute deal with Cerebras and received warrants that could give it an ~11% equity stake, worth approximately $11.7 billion at the IPO opening price if fully vested; Sam Altman and Greg Brockman are also personal angel investors.

• The IPO priced far above the initial $115 to $125 range (later raised to $150 to $160); co-founder and CEO Andrew Feldman’s stake was worth nearly $1.9 billion at the IPO price.

OpenAI’s renegotiated deal with Microsoft caps total revenue-sharing payments at $38 billion through 2030, saving OpenAI an estimated $97 billion compared to the previous uncapped arrangement, while shifting to a non-exclusive licensing model. Link.

• Under the new terms, OpenAI pays Microsoft 20% of revenue with a hard cap at $38 billion; the old arrangement had no ceiling and was tied to the undefined question of when OpenAI achieved AGI.

• Microsoft’s exclusivity to sell OpenAI’s models expired in April 2026, allowing OpenAI to offer on AWS and Google Cloud; Microsoft’s investment is now valued at approximately $135 billion, representing a 27% diluted ownership stake.

• OpenAI paid Microsoft roughly $23 billion in Azure server rentals between 2023 and 2025; the new IP arrangement shifts to non-exclusive licensing running through 2032.

Microsoft has generated more than $30 billion in revenue from OpenAI’s technology and spending between 2023 and 2025, more than doubling its $13 billion investment, but has spent over $100 billion in infrastructure costs to support the partnership. Link.

• The largest revenue component was OpenAI spending approximately $23 billion renting Azure servers, with the remainder from AI product sales including Office 365 Copilot, GitHub Copilot, and reselling OpenAI’s models to enterprise customers.

• Microsoft was the only cloud provider allowed to sell OpenAI’s LLMs to businesses for three years, an exclusive right that expired in April 2026; capital expenditure is expected to exceed $40 billion in Q4.

• While the $30 billion in revenue more than doubles the $13 billion investment, the $100 billion-plus in infrastructure spending means Microsoft is still significantly underwater on a pure cost basis.

OpenAI has structured massive purchasing commitments with suppliers like Cerebras and Nvidia that generate billions in financial value for OpenAI through equity stakes, warrants, and circular deal structures before any chips are even delivered. Link.

• OpenAI committed $20+ billion to buy Cerebras chips over three years, receiving warrants that could give it an ~11% equity stake worth approximately $11.7 billion at the IPO opening price if fully vested; OpenAI also provided Cerebras a $1 billion loan for data center construction.

• In its $122 billion funding round, Amazon committed up to $50 billion and Nvidia $30 billion, both major suppliers to OpenAI; Amazon extended an AWS deal to $100 billion over 8 years while OpenAI committed to buying 3 GW of Nvidia chips.

• Sam Altman appeared in a video to promote Cerebras’ IPO, underscoring the intertwined relationship; Altman personally held ~89,000 Cerebras shares worth ~$16.5 million at the IPO price.

Google and SpaceX are in active discussions to build data centers in orbit, as SpaceX gears up for its anticipated $1.75 trillion IPO and pitches investors on the idea that space will become the cheapest location for AI compute within the next few years. Link.

• Google has its own related initiative called Project Suncatcher with plans to launch prototype satellites by 2027; Mike Safyan, an executive at Planet Labs (building the prototypes for Google), confirmed the plans.

• The broader industry is moving fast: Blue Origin entered the space data center market in March 2026, Cowboy Space raised $275 million to build rockets specifically for this purpose, and SpaceX has sought federal approval to launch up to 1 million solar-powered satellite data centers.

• The economics remain brutal, as terrestrial data centers are still far cheaper once satellite construction and launch costs are factored in.

BlackRock is in discussions to invest $5 to $10 billion in SpaceX’s upcoming IPO from its $536 billion in actively managed funds, as SpaceX targets a valuation of $1.75 to $2 trillion and aims to raise approximately $75 billion in what would be the largest stock-market debut in history. Link.

• SpaceX targets June 11 pricing and June 12 listing on Nasdaq under ticker “SPCX,” with a roadshow launching June 4; the timeline accelerated due to faster-than-expected SEC review.

• The IPO would dwarf Saudi Aramco’s $29 billion 2019 offering; SpaceX merged with xAI in February 2026, so investors are effectively buying into a combined space, rocket, and AI business.

• Both BlackRock and SpaceX declined to comment; the final investment amount could change based on IPO pricing.

Modal, a cloud infrastructure startup for AI developers, is in talks to raise $150 to $250 million at a $4.5 billion valuation after its revenue surged fivefold to approximately $300 million in annualized run rate. Link.

• The $4.5 billion target is an 80% increase from its last round; Modal was in talks to raise at $2.5 billion as recently as February 2026.

• Modal rents out Nvidia GPUs and provides software to help developers run, train, and deploy AI models and agents; key customers include Ramp and Lovable.

• The key growth driver is rising enterprise demand for AI agent sandboxes, isolated environments where companies can safely develop and test autonomous agents.

ENERGY

Elon Musk’s xAI is operating nearly 50 natural gas turbines at its Mississippi data center (Colossus 2) without proper regulatory oversight, exploiting a loophole that classifies turbines on flatbed trailers as “mobile” to dodge air pollution regulations. Link.

• Of 59 turbines on-site, only 15 have been properly permitted and 18 are classified as temporary, meaning regulators do not track their pollution; the NAACP has filed a lawsuit on behalf of local residents arguing the unchecked emissions are worsening air quality.

• The Southern Environmental Law Center estimates xAI has operated at least 35 turbines capable of emitting more than 2,000 tons of NOx pollution, which contributes to smog and respiratory problems.

• This follows the EPA’s January 2026 ruling that xAI’s natural gas generators at its Memphis facility were illegally used; the NAACP asked the court this week for an injunction against xAI.

OTHER

Anthropic has surpassed OpenAI in business customer adoption for the first time, with 34.4% of participating businesses paying for Anthropic services versus 32.3% for OpenAI, according to expense data from fintech company Ramp covering more than 50,000 companies. Link.

• The shift has been dramatic: in May 2025, only 9% of businesses were paying for Anthropic products, but that figure climbed roughly 26 percentage points over the following 12 months; over the same period, OpenAI’s share actually declined by 1%.

• Anthropic’s annualized revenue jumped from $9 billion at the end of 2025 to $30 billion by the end of March 2026, driven largely by demand for Claude Code and related coding products.

• The data comes as Anthropic reportedly could raise a new $50 billion round at a $900 billion valuation.

OpenAI co-founder and president Greg Brockman is taking over product strategy on an interim basis while Fidji Simo remains on medical leave, with plans to consolidate ChatGPT, Codex, and the AI browser into a single unified experience. Link.

• Brockman wrote in a staff memo: “We’re consolidating our product efforts to execute with maximum focus toward the agentic future, to win across both consumer and enterprise.”

• OpenAI told TechCrunch that although Simo remains on medical leave, she worked with Brockman on these organizational changes; Simo joined OpenAI from Instacart where she was CEO.

• The consolidation reflects OpenAI’s push toward building a unified “super app” for agentic AI, with one core product team spanning consumer and enterprise.

OpenAI launched personal finance tools in preview for ChatGPT Pro subscribers, partnering with Plaid to let users connect bank accounts from over 12,000 financial institutions and ask ChatGPT questions about spending, portfolio performance, and financial planning. Link.

• Users see a dashboard of portfolio performance, spending, subscriptions, and upcoming payments; the launch came one month after OpenAI acquired the team behind personal finance startup Hiro, backed by Ribbit and General Catalyst.

• The feature is currently available on web and iOS for Pro subscribers only; OpenAI wants to gather feedback before expanding to Plus users.

• ChatGPT currently has 800 million regular users, though only about 5% are paying subscribers; users can access the tool by selecting “Finances” in ChatGPT’s sidebar.

General Motors laid off more than 10% of its IT department, approximately 600 salaried employees, in a deliberate “skills swap” to hire workers with AI-native development, data engineering, agent and model development, and prompt engineering backgrounds. Link.

• GM is still hiring for its IT department but for different skill sets; this is not a one-to-one exchange and there will likely be a net-negative job loss.

• The company has been cutting white-collar employees across several departments over the past 18 months; in August 2024, GM cut approximately 1,000 software workers.

• The move reflects a broader strategic transformation: rebuilding the workforce around AI skills rather than simply layering AI capabilities onto existing teams.

Anthropic updated its website to warn investors that multiple secondary platforms claiming to offer access to Anthropic shares are unauthorized, naming Open Doors Partners, Unicorns Exchange, Pachamama Capital, and others, and stating any such transactions are void. Link.

• Unicorns Exchange said it received more than 50 inquiries from institutional investors wanting to buy Anthropic shares in just the past three months, with aggregate demand exceeding $1 trillion.

• Both Anthropic’s preferred and common stock are subject to transfer restrictions; any sale or transfer not approved by its board of directors is considered invalid, and the company does not permit special purpose vehicles to acquire its stock.

• Forge Global pushed back, claiming it was included erroneously and stated it is working with Anthropic to remove its name from the alert.

On the enterprise front, Anthropic and OpenAI both announced competing joint ventures with Wall Street firms on the same day, Sierra raised $950 million at a $15 billion valuation, and Cerebras is on track for a blockbuster $26.6 billion IPO. Meanwhile, OpenAI’s custom chip deal with Broadcom hit an $18 billion financing snag after Microsoft declined to commit to buying 40% of the chips.

Also this week, Anthropic blamed “evil” AI portrayals in training data for Claude’s blackmail behavior, Meta revealed it’s building an OpenClaw-inspired agent called Hatch, and testimony in Elon Musk’s lawsuit put OpenAI’s safety record under the microscope.

More on this past week’s top AI headlines below!

- ZG

Here are the most important stories of the week:

AGENTS

Meta is developing two major AI agent products: a general-purpose consumer agent codenamed “Hatch” inspired by OpenClaw, and a separate agentic shopping assistant built into Instagram to compete with TikTok Shop. Link.

• Hatch is being tested in simulated environments that mimic third-party apps including DoorDash, Reddit, Outlook, and Etsy, suggesting Meta intends it to work beyond its own ecosystem; internal testing is targeted for completion by the end of June 2026.

• Hatch currently runs on Anthropic’s Claude models while Meta develops its own system called “Muse Spark” in the background, with plans to eventually switch over.

• The Instagram shopping tool lets users tap on products they see in Reels or posts and have the AI assist with purchasing, targeted for launch before Q4 2026.

Perplexity opened its “Personal Computer” feature to all Mac users, a local AI agent that works with files, native apps, and the web to handle multi-step personal workflows, competing with OpenClaw and other local agent tools. Link.

• The product was previously limited to Perplexity Max subscribers with a waitlist; it can orchestrate tools, files, and use over 400 connectors, leveraging personal context within a secure development environment.

• If paired with Perplexity’s Comet web browser, it can operate web-based tools without direct connectors; users can also run autonomous agents on always-on devices like Mac Mini and access them remotely from iPhone.

• Personal Computer requires a Pro or Max subscription; Perplexity’s older Mac app will be deprecated in the coming weeks.

SPEECH/AUDIO

OpenAI announced three new voice intelligence features for its API: GPT-Realtime-2, a voice model with GPT-5-class reasoning; GPT-Realtime-Translate, supporting 70+ input languages and 13 output languages; and GPT-Realtime-Whisper, a live speech-to-text transcription tool. Link.

• OpenAI described the combined launch as moving real-time audio from “simple call-and-response toward voice interfaces that can actually do work: listen, reason, translate, transcribe, and take action.”

• Target use cases include customer service, education, media, events, and creator platforms; Translate and Whisper are billed by the minute, while GPT-Realtime-2 is billed by token consumption.

• All models are part of OpenAI’s Realtime API, with guardrails built in to prevent misuse for spam, fraud, or abuse.

CODING/DEVTOOLS

Cursor employees are now physically inside xAI/SpaceX offices meeting with xAI staff and asking them to explain their ongoing projects, as more than 80 people have departed xAI including every single cofounder other than Elon Musk himself. Link.

• xAI carried out another round of layoffs with roughly 10 cuts last week from teams working on the Grok model; Devendra Chaplot, a prominent AI researcher who reported directly to Musk, left after just one month.

• SpaceX acquired xAI in February in an all-stock deal valuing the combined entity at $1.25 trillion and struck a deal giving it the right to acquire Cursor for $60 billion later this year; Microsoft had also looked at buying Cursor before the SpaceX deal.

• SpaceX is delaying the potential Cursor acquisition until after its IPO this summer to avoid updating confidential financial filings.

RESEARCH

Anthropic published new research explaining why Claude Opus 4 resorted to blackmail during pre-release testing up to 96% of the time, blaming fictional portrayals of AI as evil and self-interested in internet training data. Link.

• Since Claude Haiku 4.5, Anthropic’s models “never engage in blackmail” during testing, a dramatic improvement achieved by training on documents about Claude’s constitution alongside fictional stories about AIs behaving admirably.

• Anthropic found that training works best when it includes both “the principles underlying aligned behavior” and “demonstrations of aligned behavior” together, not just demonstrations alone.

• The company had previously published research showing models from other companies had similar “agentic misalignment” issues.

Recursive Superintelligence, a startup founded by former OpenAI and Google DeepMind researchers including former Salesforce Chief Scientist Richard Socher, raised $500 million at a $4 billion valuation just four months after being founded. Link.

• The pre-Series A round was led by GV (Google Ventures) and Nvidia and was so oversubscribed that Recursive could end up raising as much as $1 billion.

• The founding team also includes Tim Rocktaschel (UCL professor, formerly principal scientist at Google DeepMind), Josh Tobin, Jeff Clune, and Tim Shi.

• The company focuses on recursive self-improvement: building AI systems that are self-supervising, designing, testing, and synthesizing their own algorithmic improvements without constant human intervention.

QuTwo, a Finnish AI lab founded by Peter Sarlin (who sold Silo AI to AMD for $665 million), reached a $380 million valuation after raising a $29 million angel round, building an orchestration layer that directs computing tasks to classical, quantum, or hybrid architectures. Link.

• Sarlin deliberately chose angel funding over VC money, wanting freedom to think on a 5-to-10-year horizon; angels include Yuri Milner, Xavier Niel, Nico Rosberg, and founders from Hugging Face, Skype, Supercell, and Wolt.

• QuTwo OS uses “quantum-inspired” computing that runs on classical chips to simulate quantum behavior on more reliable hardware; about 50 quantum and AI scientists have joined the team.

• The company has already secured approximately $23 million in committed revenue through design partnerships, including with Zalando for AI assistants.

INFRASTRUCTURE

Anthropic committed to spending $200 billion with Google Cloud over five years for next-generation TPU capacity, the largest single cloud computing deal ever, representing more than 40% of Google’s total revenue backlog which doubled to over $460 billion. Link.

• Separately, Alphabet is investing up to $40 billion in Anthropic, starting at $10 billion with the potential to increase if Anthropic hits specific performance targets; this is on top of Anthropic’s existing 10-year, $100 billion deal with Amazon/AWS.

• Anthropic is maintaining a multi-vendor strategy with capacity locked in at CoreWeave and AWS; Claude is currently trained across AWS chips, Google TPUs, and Nvidia GPUs.

• Contracts involving Anthropic and OpenAI now account for more than half of the $2 trillion in combined backlogs at major cloud providers.

SpaceX revealed that its Terafab semiconductor factory in Grimes County, Texas could cost up to $119 billion total, with an initial spend of $55 billion, according to a proposal on the county website. Link.

• The “multi-phase, next-generation, vertically integrated semiconductor manufacturing and advanced computing fabrication facility” will develop chips for AI servers, satellites, SpaceX’s proposed data centers in space, autonomous Tesla vehicles, and robots.

• Intel has been roped into the effort and the facility aims to manufacture enough chips to provide 1 terawatt of computing power per year; Musk tweeted that Grimes County is only one of several locations under consideration.

• The combined SpaceX/xAI entity is valued at $1.25 trillion and is expected to go public in June.

Nvidia has committed more than $40 billion to equity investments in AI companies in just the first months of 2026, with the bulk coming from a single $30 billion investment in OpenAI, cementing its role as a dominant investor in the AI ecosystem. Link.

• Nvidia has also announced seven multi-billion-dollar investments in publicly traded companies, including up to $3.2 billion in Corning and up to $2.1 billion in data center operator IREN.

• Critics argue these are “circular deals” since Nvidia is investing in its own customers; Wedbush analyst Matthew Bryson said the investments fall “squarely into the circular investment theme” but could help build a “competitive moat.”

• In 2025 Nvidia participated in 67 venture deals, and in 2026 it has already participated in around two dozen rounds in private startups according to FactSet data.

Cerebras plans to issue 28 million shares at $115 to $125 per share, which could raise approximately $3.5 billion and value the AI chip maker at around $26.6 billion in what would be the largest tech IPO of 2026 so far. Link.

• Cerebras signed a multi-year agreement worth more than $10 billion with OpenAI, which also loaned Cerebras $1 billion secured by warrants allowing OpenAI to purchase over 33 million shares; Sam Altman, Greg Brockman, and Ilya Sutskever are all angel investors.

• The company brought in $510 million in revenue in 2025 with $237.8 million in net income; the offering is planned for mid-May and Benchmark raised $225 million in special funds to double down.

• The IPO could prove the appetite for even bigger offerings in the wings, including SpaceX and possibly OpenAI and Anthropic.

OpenAI’s effort to build custom AI chips with Broadcom, codenamed “Project Nexus,” hit a major financing roadblock after Microsoft declined to commit to buying 40% of the chips needed to fund the $18 billion first phase. Link.

• The custom inference chip called “Jalapeno” is designed to power ChatGPT; the first phase would consume 1.3 GW of data center capacity and Broadcom said it will only finance it if Microsoft agrees to buy roughly 40% of the output.

• OpenAI’s head of compute Sachin Katti reportedly told colleagues that Microsoft’s involvement would make the deal “financially unattractive” and described the commercial structure as “likely unworkable.”

• OpenAI has projected its operations will burn more than $200 billion through 2029; Broadcom’s stock dropped roughly 4% on the news.

DeepSeek is raising up to $7.35 billion (50 billion yuan) in its first outside funding round, with founder Liang Wenfeng personally contributing roughly 40% of the capital, pushing the company’s valuation past $50 billion. Link.

• China’s National Integrated Circuit Industry Investment Fund (the “Big Fund”) is reportedly participating as a lead investor, marking the first time the state-backed fund has invested in DeepSeek; Liang controls approximately 84% of the company.

• The funding is accelerating DeepSeek’s plans to generate revenue and achieve commercial viability; V4.1 is planned for June and the company is transitioning from chatbots toward autonomous AI agents.

• Investors had raised concerns about the lack of revenue and loss of key researchers to competitors like Xiaomi and ByteDance; the new capital is intended to address talent retention through improved compensation.

POLICY/GOV’T/ETHICS

Testimony in Elon Musk’s lawsuit against OpenAI revealed that Microsoft deployed GPT-4 in India through Bing before it was evaluated by OpenAI’s Deployment Safety Board, and that former board member Tasha McCauley accused Sam Altman of a pattern of misleading the board. Link.

• Former employee Rosie Campbell testified that OpenAI shifted from research-focused to product-focused and that her AGI readiness team was disbanded along with the Super Alignment team; under cross-examination she admitted OpenAI’s safety approach is still superior to Musk’s xAI.

• McCauley testified that Altman lied to a board member about her wanting to remove Helen Toner and failed to inform the board about the ChatGPT public launch, contributing to the board’s decision to briefly fire Altman in 2023.

• McCauley argued that having it “all come down to one CEO” with the public good at stake is “very suboptimal” and called for stronger government regulation of advanced AI.

OpenAI launched a new optional “Trusted Contact” feature in ChatGPT that alerts a designated third party when conversations suggest possible self-harm, responding to a wave of lawsuits from families of suicide victims. Link.

• When self-harm is detected, ChatGPT first encourages the user to reach out to their contact, then sends an automated alert; every notification is reviewed by a human safety team with a target response time of under one hour.

• Families allege ChatGPT encouraged their loved ones to kill themselves or helped them plan it; the alert is brief and does not include detailed conversation content to protect user privacy.

• The feature is optional and users can have multiple ChatGPT accounts, meaning the safeguard can be easily circumvented, the same limitation that applies to parental controls.

Allen Control Systems, a four-year-old Austin-based defense startup building the “Bullfrog” autonomous counter-drone weapon system, is in talks to raise new funding at a $2 billion valuation after winning the U.S. Army’s xTechOverwatch Competition. Link.

• Founded by former Navy nuclear engineers, ACS has raised $42 million over two prior rounds; a $2 billion valuation would represent a massive step-up, and the company has landed contracts with South Korea and UAE armed forces.

• Bullfrog uses computer vision and proprietary control systems to autonomously detect, track, and engage drone threats; ACS tripled its Austin operations to over 57,000 square feet in February to meet demand.

• The founders previously built a robotics company that was acquired for over $100 million in 2022; ACS also secured a $1.5 million Army Applications Lab contract with up to $4.5 million in options for combat vehicle integration.

OTHER

Anthropic and OpenAI both announced competing joint ventures with Wall Street firms on the same day, with Anthropic’s valued at $1.5 billion (backed by Blackstone, Hellman & Friedman, and Goldman Sachs) and OpenAI’s “The Development Company” raising $4 billion against a $10 billion valuation from 19 investors. Link.

• Both ventures follow a forward-deployed engineering model, embedding technical teams directly within client organizations to build customized AI solutions; there is no apparent overlap in investors between the two.

• Anthropic’s venture includes $300 million commitments each from Anthropic, Blackstone, and Hellman & Friedman, with additional backing from General Atlantic, Leonard Green, Apollo, GIC, and Sequoia.

• OpenAI’s venture is backed by TPG, Brookfield Asset Management, Advent, and Bain Capital; both ventures are designed to compete with the world’s largest consulting firms for corporate AI transformation work.

Sierra raised $950 million led by Tiger Global and GV at a valuation above $15 billion, with Bret Taylor’s AI startup now claiming more than 40% of the Fortune 50 as customers and hitting $150 million in ARR by early February. Link.

• Taylor serves as chairman of OpenAI and was formerly co-CEO of Salesforce; Sierra’s agents handle billions of interactions from refinancing mortgages to processing insurance claims across customers like Prudential, Cigna, and Rocket Mortgage.

• In April the company launched Ghostwriter, an “agent as a service” tool that lets users describe what they need in natural language and autonomously creates and deploys a specialized agent; Sierra also acquired YC-backed startup Fragment.

• Revenue grew from $100 million ARR in late November 2025 to $150 million by early February, reaching $100 million ARR in under two years.

In research, DeepMind’s David Silver raised a record-breaking $1.1 billion seed round to build AI that learns without human data, Tencent released a new model that was partly fine-tuned using Anthropic’s Claude, and a Harvard study found that OpenAI’s o1 outperformed ER doctors in diagnostic accuracy.

Also this week, Stripe unveiled agentic payments through its Link wallet, SoftBank announced a new robotics company called Roze with plans for a $100 billion IPO, and the Academy formally banned AI-generated actors and scripts from Oscar eligibility.

More on this past week’s top AI headlines below!

- ZG

Here are the most important stories of the week:

AGENTS

Stripe unveiled a major update to Link, its digital wallet with over 250 million users, making it the first mainstream payment wallet explicitly designed for AI agents to make purchases autonomously through a new product called “Issuing for Agents.” Link.

• Users connect cards and banks to Link, then authorize AI agents to spend via OAuth; the agent creates a spend request and waits for user approval via mobile notification before a one-time-use payment credential is shared.

• Stripe partnered with Google (allowing purchases inside AI Mode and the Gemini app via the Universal Commerce Protocol), as well as OpenAI, Microsoft, and Meta; Stripe and Tempo co-authored the Machine Payments Protocol for agent-to-agent microtransactions.

• Patrick Collison said “AI is the biggest platform shift for the economy since the internet, and in the not-too-distant future, agents will account for most transactions online.”

OpenAI is reportedly developing a smartphone where AI agents replace traditional apps, with a custom processor jointly designed by Qualcomm and MediaTek and mass production targeted for 2028. Link.

• The device would maintain “full real-time state,” continuously capturing user location, activity, and environmental context to feed the agents; Luxshare would handle co-design and exclusive manufacturing.

• Kuo projected 300 to 400 million annual shipments, which would make it one of the most successful consumer electronics products in history (Apple ships roughly 230 million iPhones per year).

• This is separate from OpenAI’s other hardware project with Jony Ive, whose company io is developing a non-phone device; OpenAI, Qualcomm, and MediaTek have not officially confirmed the project.

CODING/DEVTOOLS

Swedish vibe-coding startup Lovable launched its AI-powered app builder on iOS and Android, allowing users to build web apps from their phones using voice or text prompts, while sidestepping Apple’s crackdown on vibe-coding tools by focusing on web apps rather than native apps. Link.

• Lovable hit $400 million ARR (up from $200 million in November 2025) with over 25 million total projects created; the company raised $330 million in a Series B at a $6.6 billion valuation led by CapitalG and Menlo Ventures.

• The mobile app lets users code on the go via voice or text, with Lovable’s agent running autonomously and sending notifications when a build is ready for review; subscription plans range from $9.99 to $79.99/month.

• Co-founders Anton Osika and Fabian Hedin became billionaires following the December 2025 valuation; the company is also reportedly on the hunt for acquisitions.

RESEARCH

David Silver, former head of reinforcement learning at Google DeepMind and the researcher behind AlphaZero and AlphaGo, raised a record-breaking $1.1 billion seed round at a $5.1 billion valuation for his new startup Ineffable Intelligence, which aims to build AI that discovers knowledge without relying on human data. Link.

• The round, the largest seed round ever in Europe, was co-led by Sequoia Capital and Lightspeed Venture Partners with participation from Nvidia, DST Global, Index Ventures, Google, the UK’s Sovereign AI Fund, and the Wellcome Trust.

• Silver referred to Ineffable Intelligence as “his life’s work” and told Wired that “any money that I make from Ineffable will go to high-impact charities that save as many lives as possible.”

• The company joins a wave of star-researcher AI ventures; just last month, Turing Award winner Yann LeCun’s AMI Labs raised $1.03 billion at a $3.5 billion pre-money valuation.

Tencent released Hy3 Preview, its first major AI model under former OpenAI researcher Yao Shunyu’s leadership, jumping from 53% to 74.4% on SWE-bench Verified in under three months, with the improvement partly attributable to Tencent employees using Anthropic’s Claude to evaluate and fine-tune the model. Link.

• Hy3 Preview is a fast-and-slow-thinking fused mixture-of-experts model with 295 billion total parameters, 21 billion activated, and 256K context length, offering 40% improved inference efficiency over previous versions.

• Yao led a February infrastructure overhaul that rebuilt the pretraining and reinforcement learning stack; training started in late January and the model launched under three months later.

• Yao publicly stated that Anthropic is the AI company he most respects, reflecting a broader trend in China’s AI industry looking to Anthropic as a benchmark.

A landmark Harvard Medical School study published in Science found that OpenAI’s o1 model outperformed two ER physicians in diagnostic accuracy, matching the exact or near-exact diagnosis 67% of the time at initial triage versus 50-55% for the doctors. Link.

• With richer detail available later in the diagnostic process, o1 reached 82% accuracy versus 70-79% for the human doctors; the AI was given the same unprocessed electronic medical records data that the physicians had.

• On management reasoning tasks including antibiotic recommendations and goals of care, o1 “significantly outpaced” previous AI models and also outperformed humans using conventional aids like Google search.

• The researchers stressed this does not mean “AI replaces doctors” and that the study positions AI as a support tool for clinical decision-making in high-pressure, limited-information environments.

ROBOTICS

Meta acquired Assured Robot Intelligence (ARI), a startup specializing in whole-body humanoid control, with the 20-person team joining Meta’s Superintelligence Labs as the company pursues a strategy of building a licensable AI platform for robots rather than building its own hardware. Link.

• ARI was co-founded by Lerrel Pinto, Xiaolong Wang, and Xuxin Cheng with the mission to “achieve physical AGI”; Pinto also co-founded Fauna Robotics, which was acquired by Amazon for its own humanoid robot project.

• Meta’s approach is to create an “Android for humanoids” that hardware makers can adopt, positioning itself in a humanoid robotics sector estimated at a $5 trillion market opportunity.

• ARI specializes in foundation models for humanoid robots that enable them to understand, predict, and adapt to human behaviors in complex, dynamic environments.

SoftBank is creating a new standalone company called Roze that will deploy autonomous robots to build U.S. data centers, bundling portfolio assets including ABB Robotics ($5.4B acquisition), Ampere Computing ($6.5B), and DigitalBridge ($3B), with plans to IPO at a $100 billion valuation. Link.

• KPMG has been hired to prepare financials and an analyst day is planned for July at a Texas data center; the IPO could come as early as H2 2026 but may slip to 2027 if conditions deteriorate.

• The U.S. is short roughly 439,000 construction workers and up to 40% of AI data center construction sites face delays, creating the demand case for robotic construction.

• Several SoftBank executives are described as skeptical of the $100 billion valuation target; Roze has not yet announced a product, shared a revenue plan, or set an IPO date.

INFRASTRUCTURE

Anthropic is on the verge of closing a ~$50 billion funding round at a valuation in the $850-900 billion range, which would surpass OpenAI’s $852 billion post-money valuation and make Anthropic the most valuable private AI company in the world. Link.

• The company’s annual revenue run rate has surpassed $30 billion (up from ~$9 billion at end of 2025), with sources saying the run rate is currently closer to $40 billion; over 1,000 businesses spend more than $1 million annually.

• A board meeting in May is expected to produce a definitive decision, with investors facing a 48-hour allocation deadline; a public listing could come as early as October 2026.

• The previous round in February 2026 valued Anthropic at $380 billion, meaning this would represent more than a doubling of valuation in roughly three months.

Anthropic is in early-stage talks to purchase AI inference chips from Fractile, a London-based startup whose SRAM-based chips claim to deliver inference 25x faster at one-tenth the cost, which would make Fractile Anthropic’s fourth chip supplier alongside Nvidia, Google, and Amazon. Link.

• Fractile’s chips co-locate memory and compute on the same die using static random access memory instead of conventional DRAM; the chips are not expected to be commercially ready until approximately 2027.

• Fractile is separately in talks to raise $200 million at a $1 billion-plus valuation from Founders Fund, 8VC, and Accel; the company was founded in 2022 by Oxford PhD Walter Goodwin.

• Anthropic’s gross profit margin reportedly fell short of targets last year due to higher-than-expected inference costs, a key motivation for seeking alternative chip architectures.

Starcloud, a two-year-old startup building orbital data centers, is in talks to raise at least $200 million at a ~$2.2 billion valuation, roughly doubling its value just weeks after becoming the fastest unicorn in Y Combinator history. Link.

• Starcloud partnered with SpaceX to launch a 130-pound satellite carrying an Nvidia H100 chip that became the first to train an LLM in space in December 2025; the next satellite will carry 100x the power generation and an Nvidia Blackwell B200 chip.

• The company raised a $170 million Series A at a $1.1 billion valuation in March led by Benchmark and EQT Ventures, with backers including NFX, In-Q-Tel, and Nvidia; it has filed FCC plans for an 88,000-satellite constellation.

• SpaceX has filed its own FCC plans for up to 1 million “orbital data center” satellites projecting 100GW of AI compute, creating both a competitive threat and validation of the market.

POLICY/GOV’T/ETHICS

The Pentagon signed agreements with eight tech companies to deploy AI on its most sensitive classified networks (Impact Level 6 and 7), including AWS, Google, Microsoft, OpenAI, SpaceX, Nvidia, Reflection, and Oracle, while pointedly excluding Anthropic. Link.

• Just two months prior, Claude was the only large language model approved on Pentagon classified networks at scale, with forward-deployed Anthropic engineers embedded across multiple military programs.

• The dispute centered on Anthropic insisting on guardrails to prevent its AI from being used for domestic mass surveillance and autonomous weapons; the DOD designated Anthropic a “supply chain risk,” a classification typically reserved for foreign adversaries.

• The announcement marks a major strategic shift toward vendor diversification in military AI, with the deals authorized to “streamline data synthesis, elevate situational understanding, and augment warfighter decision-making.”

Google signed a deal allowing the Pentagon to use its Gemini AI models for classified work and “any lawful government purpose,” prompting over 600 Google employees and over 100 DeepMind employees to sign letters urging CEO Sundar Pichai to reject the agreement. Link.

• Senior DeepMind research scientist Andreas Kirsch publicly wrote: “I’m speechless at Google signing a deal to use our AI models for classified tasks. Frankly, it is shameful.”

• In February 2025, Google had already removed language from its AI principles that pledged to avoid using technology in “weapons or other technologies whose principal purpose or implementation is to cause or directly facilitate injury to people.”

• The backlash echoes the 2018 Project Maven controversy when ~4,000 employees petitioned and at least 12 resigned, leading Google to let that contract expire; this time, the company moved in the opposite direction.

The Academy of Motion Picture Arts and Sciences announced sweeping new eligibility rules for the 99th Academy Awards formally banning AI-generated performances and screenplays from Oscar consideration, requiring producers to sign an “Affidavit of Human Origin.” Link.

• Acting nominations are now limited to roles “demonstrably performed by humans with their consent,” and screenplays must be “human-authored,” though writers may use AI as an assistive tool for research or outlines.

• Technical categories still allow AI: visual effects artists who use AI for crowd simulations, de-aging, or compositing, and sound designers using AI to clean dialogue or generate ambient textures, remain eligible.

• The rules come as an independent film featuring an AI-generated Val Kilmer is in production and AI “actress” Tilly Norwood continues making headlines.

OTHER

OpenAI projects its total consumer subscriber base will grow to 122 million in 2026, driven by a massive expansion of ChatGPT Go ($8/month), but at the cost of an 80% collapse in ChatGPT Plus subscribers from 44 million to just 9 million. Link.

• ChatGPT Go subscribers are projected to surge from 3 million (2025) to 112 million (2026), a 36x increase, while average revenue per user is expected to fall from ~$23/month to under $12/month.

• Ad revenue is projected at ~$2.4 billion in 2026 and $102 billion by 2030, with ads shown to Free and Go tier users at a $60 CPM; OpenAI projects its products will reach 2.75 billion weekly users by 2030.

• Total revenue target is $284 billion by 2030, with the company betting that volume at lower price points plus advertising will more than compensate for the decline in premium subscribers.

OpenAI launched Advanced Account Security in partnership with Yubico, requiring two passkeys or hardware security keys to log in, disabling email and SMS recovery, and automatically opting users out of model training. Link.

• OpenAI and Yubico are releasing two co-branded YubiKeys (the C NFC and C Nano) sold as a two-pack for $68, less than half the $126 retail price.

• OpenAI recommends AAS for political dissidents, journalists, researchers, and elected officials; members of the Trusted Access for Cyber program must enable it by June 1, 2026.

• The feature replaces passwords entirely with passkeys or hardware keys, representing one of the most aggressive security stances taken by a consumer AI platform.

In infrastructure, Meta signed a multibillion-dollar deal for millions of Amazon’s Graviton CPUs to power agentic AI workloads, Google launched its 8th-generation TPU chips split into training and inference variants for the first time, and Tesla raised its capex guidance to $25 billion as it ramps robotaxis, Optimus, and a new semiconductor R&D fab with Intel.

Also this week, DeepSeek previewed its V4 model (the largest open-weight model ever released), major insurers won approval to exclude AI damages from standard liability policies, and the NSA is quietly using Anthropic’s Mythos model despite the Pentagon’s ongoing feud with the company.

More on this past week’s top AI headlines below!

- ZG

Here are the most important stories of the week:

AGENTS

Anthropic created “Project Deal,” a classified test marketplace where AI agents represented both buyers and sellers, striking 186 real deals totaling over $4,000 among 69 employees who were each given a $100 budget paid out via gift cards. Link.

• Anthropic ran four separate marketplaces with different models; users represented by more advanced models got “objectively better outcomes,” raising the possibility of “agent quality gaps” where people on the losing end don’t realize they’re worse off.

• Initial instructions given to agents didn’t appear to affect sale likelihood or negotiated prices, suggesting agent behavior may be harder to steer than expected in commercial settings.

• Anthropic acknowledged this was only “a pilot experiment with a self-selected participant pool,” but the results offer an early look at how agent-mediated commerce could reshape consumer markets.

AI research lab NeoCognition emerged from stealth with $40 million in seed funding co-led by Cambium Capital and Walden Catalyst Ventures, building agents that self-learn to become domain experts and aiming to solve the problem that current agents only complete tasks about 50% of the time. Link.

• Founded by Ohio State professor Yu Su, with participation from Vista Equity Partners and angels including Intel CEO Lip-Bu Tan and Databricks co-founder Ion Stoica; the team of about 15 employees is majority PhDs.

• Rather than building custom vertical agents, NeoCognition’s approach creates generalist agents that build a “world model” for any given micro-environment, mirroring how humans specialize through experience.

• Vista Equity Partners’ investment is strategic: as one of the largest PE firms in software, it provides access to a vast portfolio of companies looking to modernize with AI agents.

IMAGE/VIDEO

ComfyUI raised $30 million at a $500 million valuation led by Craft Ventures, bringing total funding to $48 million as the open-source node-based AI generation platform surpasses 4 million users and 60,000 community-built nodes. Link.

• Pace Capital, Chemistry, and TruArrow also participated; cofounder Yannik Marek said the funding will help ensure “open source wins” against prompt-based tools like Midjourney that only get “60% to 80%” of the way there.

• The platform has become such a standard creative tool that “ComfyUI artist or engineer” is now listed as a job title on studio job boards; agency Silverside AI used it to power SVEDKA’s 2026 Super Bowl commercial, the first primarily AI-generated Super Bowl ad.

• ComfyUI’s node-based interface lets creators link specific components of the AI generation process, giving full control over image, video, and audio outputs from diffusion models with over 150,000 daily downloads.

CODING/DEVTOOLS

OpenAI released GPT-5.5, calling it its “smartest and most intuitive” model yet, while co-founders Greg Brockman and Sam Altman outlined plans to combine ChatGPT, Codex, and the AI browser into a unified enterprise “super app.” Link.

• GPT-5.5 scored 82.7% on Terminal-Bench 2.0 and 73.1% on Expert-SWE (tasks with a 20-hour median human completion time); API pricing is $5/$30 per million input/output tokens, double GPT-5.4’s price.

• OpenAI also announced Codex Labs, embedding engineers with enterprise clients including Accenture, Capgemini, Cognizant, Infosys, PwC, and Tata Consultancy Services; Codex now has more than 4 million weekly active users.

• The model arrived just six weeks after GPT-5.4, and OpenAI said releases will continue at this pace “for the foreseeable future.”

SpaceX struck a landmark deal with Cursor (Anysphere), paying $10 billion for a collaboration on “coding and knowledge work” AI with an option to acquire Cursor outright for $60 billion post-IPO, after Anthropic compute costs became a critical bottleneck. Link.

• Cursor is Anthropic’s largest API client; the deal gives Cursor access to SpaceX’s Colossus supercomputer (compute equivalent to one million Nvidia H100 chips) after the company wrote it had been “bottlenecked by compute.”

• The deal preempted a $2 billion fundraise that would have valued Cursor at $50 billion; Anysphere’s annual recurring revenue surged past $2 billion by early 2026, up from $500 million annualized in mid-2025.

• xAI (merged with SpaceX in February at a combined $1.25 trillion valuation) began renting computing power to Cursor, with the coding startup using tens of thousands of xAI chips to train its latest model.

Google DeepMind assembled a “strike team” led by researcher Sebastian Borgeaud to improve its AI coding models, with co-founder Sergey Brin writing in a memo that Google must “urgently bridge the gap in agentic execution” as internal researchers view Anthropic’s tools as superior. Link.

• AI currently writes about 50% of Google’s code (per CFO Anat Ashkenazi), while Anthropic claims nearly 100% of its code is written with AI assistance, highlighting the gap Google wants to close.

• The team focuses on long-horizon coding tasks and is training models on Google’s internal proprietary code; while those models can’t be released publicly, improvements feed back into publicly available Gemini models.

• Brin wrote to DeepMind staff: “To win the final sprint, we must urgently bridge the gap in agentic execution and turn our models into primary developers.”

OpenAI partnered with Infosys to integrate AI tools including Codex into Infosys’s Topaz AI platform, focusing on software engineering, legacy modernization, and DevOps as the Indian IT giant’s shares have fallen over 22% this year. Link.

• Part of OpenAI’s broader Codex Labs initiative; Infosys operates across more than 60 countries and its AI-related services generated about $267 million in the December quarter, roughly 5.5% of total revenue.

• Infosys shares have fallen amid investor concerns that AI will automate outsourcing work and macroeconomic turmoil from the U.S.-Iran war; Infosys has a similar partnership with Anthropic.

• The deal joins an enterprise rollout that includes Accenture, Capgemini, CGI, Cognizant, PwC, and Tata Consultancy Services.

RESEARCH

DeepSeek previewed V4, the largest open-weight model ever released, with V4-Pro at 1.6 trillion parameters outperforming Claude Opus 4.6 on Terminal-Bench 2.0 (67.9% vs. 65.4%) and LiveCodeBench (93.5% vs. 88.8%) while costing a fraction of frontier model pricing. Link.

• Two models released: V4-Pro (1.6T parameters, 49B active) and V4-Flash (284B parameters, 13B active); both are mixture-of-experts models with 1-million-token context windows and are fully open-source.

• V4-Flash costs just $0.14/$0.28 per million input/output tokens, while V4-Pro costs $1.74/$3.48, significantly cheaper than any frontier model; V4-Pro requires only 27% of single-token inference FLOPs compared with V3.2.

• V4 was reportedly built on Huawei chips; the developmental trajectory is estimated to trail state-of-the-art frontier models by approximately 3 to 6 months.

Mira Murati’s Thinking Machines Lab signed a multibillion-dollar deal with Google Cloud for AI infrastructure including access to Nvidia’s GB300 chips, making the startup among Google Cloud’s first customers to access GB300-powered systems. Link.

• Murati left OpenAI (where she was chief technologist) and founded Thinking Machines in February 2025, raising a $2 billion seed round at a $12 billion valuation; GB300 systems offer a 2x improvement in training and serving speed vs. prior-generation GPUs.

• The lab’s first product, Tinker, is a tool that automates the creation of custom frontier AI models using reinforcement learning; it launched in October and the lab also has a partnership and investment from Nvidia.

• This is Thinking Machines’ first cloud provider deal (non-exclusive), valued in the single-digit billions for training and deployment infrastructure.

INFRASTRUCTURE

Amazon invested a fresh $5 billion in Anthropic, bringing its total investment to $13 billion, as Anthropic committed to spending over $100 billion on AWS over the next 10 years and obtaining up to 5 gigawatts of new computing capacity to train and run Claude. Link.

• The deal covers Amazon’s custom chips from Trainium2 through Trainium4, with Anthropic securing the option to buy capacity on future Amazon chips as they become available.

• The structure echoes the Amazon-OpenAI deal from two months prior, where Amazon contributed $50 billion to a $110 billion funding round that valued OpenAI at $730 billion pre-money.

• VCs are reportedly offering Anthropic capital at a valuation of $800 billion or more, and this deal could be a teaser for a new funding round.

Meta signed a multibillion-dollar deal with AWS for millions of Graviton CPU chips to power its AI needs, making Meta one of the largest Graviton customers in the world as agentic AI creates massive demand for CPU-intensive workloads. Link.

• The at-least-three-year deal reflects a shift: while GPUs remain essential for training, agentic AI’s real-time reasoning, code generation, and multi-step orchestration is driving massive CPU demand; Graviton competes with Nvidia’s new Vera CPU.

• AWS announced the deal right as Google Cloud Next wrapped up, described as “a virtual smirk at its cloud rival”; Meta’s total 2026 capex budget is approximately $135 billion.

• This follows Meta’s February commitment of approximately $50 billion to Nvidia for Blackwell/Rubin GPUs, Grace/Vera CPUs, and Spectrum-X networking, showing Meta is diversifying its chip suppliers.

Google Cloud announced its 8th-generation TPUs, split into TPU 8t (training) and TPU 8i (inference) for the first time, delivering up to 3x faster AI model training, 80% better performance per dollar, and the ability to cluster 1 million+ TPUs together. Link.

• Google confirmed it will also offer Nvidia’s latest Vera Rubin chip on its cloud later this year, and is collaborating with Nvidia on open-source networking tech called Falcon to make Nvidia-based systems perform more efficiently on Google Cloud.

• Other hyperscalers including Amazon and Microsoft are also building custom AI chips; analyst Patrick Moorhead noted he predicted TPUs could be bad news for Nvidia back in 2016, but Nvidia is now a nearly $5 trillion company.

• Announced at Google Cloud Next 2026, the split into separate training and inference chips reflects the growing divergence in compute requirements across AI workloads.

Tesla raised its 2026 capex guidance to $25 billion, roughly 3x its historical annual spending, as it ramps robotaxis, begins Optimus robot production this summer, and builds a $3 billion semiconductor R&D fab with Intel at Giga Texas. Link.

• Q1 revenue rose 15.8% YoY to $22.39 billion, beating expectations; Tesla held $44.7 billion in cash, but CFO Vaibhav Taneja warned spending will mean “negative free cash flow for the rest of the year.”

• Tesla discontinued the luxury Model S and Model X to free up production lines for Optimus; Musk said the robot will enter production in July/August and start being useful “outside Tesla” next year.

• The semiconductor fab will use Intel’s 14A process technology with SpaceX handling the initial scale-up phase, capable of processing a few thousand wafers monthly.

Microsoft and other major cloud providers are diverting GPU stockpiles to internal teams and their largest customers, leaving smaller AI startups scrambling as Azure implements a three-tier system requiring a minimum commitment of 1,000 Blackwell chips for at least one year. Link.

• AI startup Krea (backed by a16z and Bain Capital Ventures) saw GPU rental prices surge 32% in six months, paying $2.80/hour for Blackwell chips on a six-month contract, up from lower rates in its prior deal.

• Two-to-three-year cloud contracts signed by many AI startups are now expiring, giving cloud providers leverage to renegotiate at higher prices or reallocate capacity to the highest bidders.

• VC firms like General Catalyst are exploring shared computing pools as a workaround; the GPU crunch is improving profit margins for cloud providers while raising operational costs for startups.

Phononic, a Durham-based solid-state cooling startup whose semiconductor components prevent AI data center chips from overheating, is exploring a potential sale at $1.5 billion or more, with some offers exceeding $2 billion, up from a $630 million valuation in 2022. Link.

• Founded in 2009 and backed by Goldman Sachs Asset Management, Venrock, and GGV Capital with over $230 million raised; the company appointed Lazard as adviser in early 2026 to weigh sale vs. additional fundraising.

• Phononic’s “Thermal Kit” replaces bulky mechanical cooling systems with semiconductor-based cooling chips, claiming 30% energy savings compared to nearest competitors and 3x ROI for AI data centers.

• The company recently introduced solutions for GPU memory cooling aligned with next-gen HBM4 designs, plus support for high-speed optical networking and co-packaged optics.

POLICY/GOV’T/ETHICS

The NSA is using Anthropic’s Mythos Preview model for scanning environments for exploitable vulnerabilities, despite the Department of Defense (the NSA’s parent agency) having labeled Anthropic a “supply-chain risk” just weeks earlier. Link.

• Mythos was deemed too capable of offensive cyberattacks for public release, with access limited to around 40 organizations; the U.K.’s AI Security Institute has also confirmed access.

• Anthropic’s relationship with the Trump administration appears to be thawing: CEO Dario Amodei met with White House chief of staff Susie Wiles and Treasury Secretary Scott Bessent last Friday, with the White House calling the meeting “productive.”

• The Pentagon dispute originated when Anthropic refused to make Claude available for mass domestic surveillance and autonomous weapons development; both the NSA and Anthropic declined to comment.

Maine Governor Janet Mills vetoed a bill that would have imposed the country’s first statewide moratorium on new data centers, saying she would have signed it if it had included an exemption for a data center project in the Town of Jay. Link.

• The moratorium (L.D. 307) would have lasted until November 1, 2027 and called for the creation of a 13-person council to study data center impacts; Mills acknowledged pausing data centers would be “appropriate given the impacts” on the environment and electricity rates.

• Democratic state representative Melanie Sachs, who sponsored the bill, said the veto “poses significant potential consequences for all ratepayers, our electric grid, our environment, and our shared energy future.”

• Mills, currently running for the U.S. Senate, framed the veto around local support; other states including New York have considered similar moratoriums amid rising public opposition to data centers.

OTHER

Tencent and Alibaba are in talks to invest in DeepSeek in the open-source AI startup’s first-ever outside funding round, at a valuation that rapidly doubled from $10 billion to over $20 billion due to intense investor demand. Link.

• DeepSeek, owned by Chinese hedge fund High-Flyer Capital Management and founded in 2023 by co-founder Liang Wenfeng, had previously relied entirely on self-funding through its parent hedge fund.

• Tencent proposed acquiring up to a 20% stake, but DeepSeek has balked at giving up such a large portion of control; the target valuation doubled in just a few days.

• The fundraising represents a significant strategic shift as DeepSeek moves from purely self-funded to accepting external capital from China’s largest tech companies.

Major insurers including Berkshire Hathaway, Chubb, and Travelers won regulatory approval to exclude AI-related damages from standard commercial general liability policies, a structural shift affecting the ISO form that underpins 82% of global property and casualty policies. Link.

• Effective January 1, 2026, Verisk/ISO exclusion endorsements (Form CG 40 47) allow carriers to remove generative AI exposure from CGL policies covering bodily injury, property damage, and personal and advertising injury.

• At least six major carriers have filed their own AI exclusions with state regulators, and the NAIC AI Model Bulletin has been adopted by 23 states plus D.C. as of April 1, 2026.

• The shift forces companies deploying AI to rethink who bears the risk when AI systems fail, potentially creating demand for new standalone AI liability insurance products.

Tim Cook is stepping down after 15 years as Apple CEO, with hardware engineering SVP John Ternus taking over September 1 and inheriting challenges including a stalled AI strategy, an ongoing DOJ antitrust lawsuit, and the existential threat of AI agents rendering the App Store obsolete. Link.

• Cook will remain as executive chairman; Apple’s market cap grew more than 11x on his watch to roughly $4 trillion, and AI chief John Giannandrea is formally leaving the company this month after delays in AI-powered Siri.

• Ternus faces a potential $38 billion fine from Indian regulators, the Epic Games saga (Apple found in contempt over payment links), and a DOJ antitrust suit accusing Apple of unlawfully dominating the smartphone market.

• OpenAI is reportedly working on competing hardware, and analyst Bob O’Donnell said Ternus’s biggest challenge is building a better AI story that relies more on Apple’s own capabilities rather than Google’s Gemini and OpenAI’s ChatGPT.

Cursor is in talks to raise $2 billion at a $50 billion valuation as it forecasts $6 billion in annualized revenue by year-end, and Cerebras filed for IPO after securing a $20 billion+ chip deal with OpenAI that includes an equity stake. Meanwhile, DeepSeek is raising outside money for the first time at $10 billion+, and Google is negotiating a classified AI deal with the Pentagon.

In robotics, Physical Intelligence unveiled a “robot brain” that can figure out tasks it was never taught, and OpenAI beefed up Codex with the ability to run agents in the background on your desktop.

More on this past week’s top AI headlines below!

- ZG

Here are the most important stories of the week:

AGENTS

OpenAI released an updated Agents SDK with new sandboxing capabilities that let AI agents operate in controlled computer environments, accessing only specific files and approved tools within a workspace. Link.

• The update includes an in-distribution harness for frontier models, enabling enterprises to build “long-horizon agents” that work in siloed capacity; Python launches first with TypeScript support planned.

• OpenAI product lead Karan Sharma said the launch is about making the SDK “compatible with all of these sandbox providers” so enterprises can use “whatever infrastructure they have.”

• Future updates will add code mode and subagents capabilities; the SDK is available to all customers via API with standard pricing.

Adobe launched the Firefly AI Assistant — previously previewed as “Project Moonlight” — in public beta, an agentic tool that can orchestrate tasks across Photoshop, Premiere, Lightroom, Illustrator, and other Creative Cloud apps. Link.

• The assistant learns creative preferences over time and can suggest actions, execute multi-step workflows, and adapt images for different platforms via new “skills” like social media asset optimization.

• Adobe is exploring integration with third-party LLMs and added Kling 3.0 video models to Firefly’s library alongside new audio tools for noise reduction and music adjustment.

• VP Alexandru Costin said the assistant can “remove some of the friction in learning this large catalog of tools we have and bring all of that value to our customers at their fingertips.”

Sam Altman’s World (formerly Worldcoin) announced global partnerships with Tinder, Ticketmaster, Eventbrite, Zoom, Docusign, and Okta to scale its iris-scanning human verification technology beyond crypto. Link.

• Tinder is launching World ID verification globally after a pilot in Japan, while Concert Kit lets artists like Bruno Mars and 30 Seconds to Mars reserve tickets for verified humans via Ticketmaster and Eventbrite.

• Zoom integration targets deepfake threats in video calls; Docusign ensures signatures come from real users; Okta (in beta) verifies AI agents act on behalf of authenticated humans.

• Altman said “the world is getting close to very powerful AI” and that soon “there’s going to be more stuff generated by AI than by humans,” making proof of humanity essential.

IMAGE/VIDEO

Anthropic launched Claude Design, a visual creation tool for prototypes, slides, and one-pagers powered by Opus 4.7, prompting CPO Mike Krieger to resign from Figma’s board the same day reports surfaced it would compete with the design platform. Link.

• Claude Design can apply a team’s design system by reading company codebases and design files, export as PDFs, URLs, PPTX, or send to Canva; it’s available in research preview for Pro, Max, Team, and Enterprise subscribers.

• Krieger — co-founder of Instagram and AI news app Artifact — joined Figma’s board less than a year ago; Figma’s stock rose 5% after his departure was disclosed.

• Bloomberg reported VCs are offering preemptive funding valuing Anthropic at $800 billion+, more than double its valuation from the beginning of the year; Anthropic is not interested in the latest offers.

CODING/DEVTOOLS

Cursor is in talks to raise at least $2 billion at a $50 billion valuation — nearly doubling from $29.3 billion six months ago — as it forecasts ending 2026 with an annualized revenue run rate exceeding $6 billion. Link.

• The four-year-old startup reached $2 billion in annualized revenue in February and expects to at least triple that over the next 10 months; its proprietary Composer model helped achieve slight gross margin profitability.

• Cursor is positive on gross margins for enterprise sales but still loses money on individual developer accounts; the company is trying to avoid being replaced by suppliers like Anthropic’s Claude Code.

• Investors include Thrive, Andreessen Horowitz, Battery Ventures, and Nvidia.

OpenAI announced major Codex updates including the ability to deploy multiple agents in the background on a user’s desktop, opening any app and carrying out operations without interfering with their work. Link.

• New features include an in-app browser for web applications, a “memory” feature to recall previous work sessions, image-generation capability, and 111 plug-in integrations from apps like CodeRabbit and GitLab Issues.

• A new pay-as-you-go pricing option was introduced for ChatGPT Enterprise and Business customers alongside the updates.

• The move directly targets Anthropic’s Claude Code, which is currently considered the tool of choice for many businesses; Anthropic previously released similar remote desktop capabilities.

AI coding startup Factory raised $150 million at a $1.5 billion valuation in a Series B led by Khosla Ventures, with Sequoia, Insight Partners, and Blackstone participating, to build AI agents for enterprise engineering teams. Link.

• Founded in 2023 by former UC Berkeley PhD student Matan Grinberg, Factory differentiates by switching between foundation models like Claude and DeepSeek depending on the task.

• Customers include Morgan Stanley, Ernst & Young, and Palo Alto Networks; Grinberg connected with Sequoia partner Shaun Maguire through a cold email.

• Khosla Ventures managing director Keith Rabois led the round.

Anthropic shifted Claude Enterprise to usage-based billing — a $20/month flat fee per user plus charges for computing capacity consumed — after a spike in Claude Code and Cowork usage drove inference costs far above expectations. Link.

• Claude Code’s ARR jumped from $1 billion in December to $2.5 billion in February, with weekly active users doubling; for heavy users, costs could double or triple under the new model.

• The change doesn’t affect businesses with fewer than 150 users; Anthropic’s overall revenue pace reached $30 billion in early April, nearly 3x year-end.

• Adaption Labs co-founder Sudip Roy noted “the subscription margins just erode away — you kind of price subscriptions for underutilization,” while Redress Compliance’s Fredrik Filipsson said the stickiness of Claude means Anthropic “can afford to let consumption drive revenue.”

ROBOTICS

Physical Intelligence published research on its π0.7 model demonstrating “compositional generalization” — the ability to combine learned skills in entirely new contexts — including cooking a sweet potato in an air fryer it was never trained on. Link.

• The model had only two relevant episodes in training: one of a robot pushing an air fryer closed and another placing a bottle inside — yet it independently combined these skills to cook; it also matched specialist models on making coffee, folding laundry, and assembling boxes.

• Co-founder Sergey Levine said capabilities are now “going up more than linearly with the amount of data” once the model crosses a threshold for remixing knowledge in new ways.

• Physical Intelligence has raised over $1 billion at a $5.6 billion valuation and is in discussions for a new round that would nearly double it to $11 billion.

Antioch, a New York-based simulation startup, raised $8.5 million at a $60 million valuation to build tools that close the “sim-to-real gap” for autonomous vehicles, farm machinery, drones, and construction robots. Link.

• The platform lets developers spin up multiple digital instances of hardware with simulated sensors, focusing initially on perception systems; two co-founders previously sold Transpose to Chainalysis.

• Co-founder Harry Mellsop said “anyone building an autonomous system for the real world is going to do so in software primarily in two to three years.”

• The seed round was led by A* and Category Ventures with participation from MaC Venture Capital and angel investors from Cruise and MIT’s CSAIL.

RESEARCH

DeepSeek is raising outside money for the first time, seeking at least $300 million at a $10 billion+ valuation after being financed entirely by parent hedge fund High-Flyer Capital Management since its founding. Link.

• The company has lost star researchers — Luo Fuli joined Xiaomi and Guo Daya joined ByteDance at higher pay — and its V4 flagship model has been delayed multiple times from its original February target.

• V4 is being optimized to run on Huawei’s Ascend chips out-of-box rather than Nvidia hardware; DeepSeek is also training on Nvidia’s banned Blackwell chips that were previously stockpiled.

• As a Chinese startup, DeepSeek faces hesitation from some U.S. VCs; it has been accused by Congress of aiding Beijing’s military but hasn’t been added to any trade blacklist.

INFRASTRUCTURE

Cerebras filed for a mid-May IPO targeting a $35 billion valuation — a 60% premium to its last private round — after securing a $20 billion+ chip deal with OpenAI that includes warrants for up to a 10% equity stake. Link.

• OpenAI is providing roughly $1 billion to fund data center development and could spend $30 billion total over three years; Cerebras chips already power OpenAI’s Codex-Spark coding model for some customers.

• Cerebras reported $510 million in 2025 revenue and $237.8 million in net income; it also announced a deal with AWS and is designing a language processing unit with Groq technology.

• CEO Andrew Feldman said Nvidia “didn’t want to lose the fast inference business at OpenAI, and we took that from them”; OpenAI plans to spend $45 billion on computing in 2026 and $90 billion in 2027.

Google is developing two new AI chips with Marvell Technology — a memory processing unit and a new inference-focused TPU — planning to produce nearly 2 million MPUs alongside an estimated 6 million TPUs by 2027. Link.

• The MPU would work alongside TPUs to divide AI workloads based on compute and memory demands; Google plans to finalize the design as soon as next year before test production.

• Broadcom was previously Google’s sole TPU design partner; the Marvell partnership diversifies Google’s chip supply chain after also bringing in MediaTek last year.

• The move comes as Nvidia paid $20 billion for a similar licensing arrangement with Groq, signaling intense competition in the inference chip market.

AI infrastructure startup Upscale AI is in talks to raise $180–$200 million at approximately $2 billion in its third funding round after just seven months, despite not yet having released a product. Link.

• The company raised a $200 million Series A in January and a $100 million seed in September 2025, focusing on custom chips and infrastructure for chip communication.

• Upscale AI is betting on a full-stack solution with open standards for scalable AI infrastructure.

• Investors include Tiger Global Management, Xora Innovation, and Premji Invest.

POLICY/GOV’T/ETHICS

Google is negotiating with the Pentagon to deploy Gemini AI models in classified settings — a major reversal from its 2018 Project Maven withdrawal — with proposed contract language mirroring OpenAI’s terms on surveillance and autonomous weapons. Link.

• Google’s proposed terms prevent use for domestic mass surveillance or autonomous weapons without “appropriate” human oversight, though legal experts note the “all lawful purposes” framing wouldn’t necessarily prevent banned applications.

• Google’s Public Sector division targeted roughly $6 billion in bookings over 2025–2027 ($2 billion defense, $500 million national security); more than 200 Google employees signed a letter opposing weapons and surveillance uses.

• Google’s chief AI scientist Jeff Dean and roughly 40 Google and OpenAI employees signed an amicus brief defending Anthropic’s stance against the Pentagon; Google altered its AI principles in early 2025, removing explicit weapons and surveillance provisions.

OTHER

Anthropic CFO Krishna Rao has secured $60 billion in funding since joining in March 2024, boosting the company’s valuation to $380 billion while improving gross margins from negative 94% in 2024 to positive 40% in 2025. Link.

• Rao — a former Airbnb finance executive who helped orchestrate its $47 billion IPO — diversified Anthropic’s cloud partners across Google, Amazon, and Microsoft, and created a joint venture with Blackstone and other PE firms.

• Some investors believe Anthropic could reach a $1 trillion valuation; an IPO could occur as early as fall 2026, with Bessemer partner Byron Deeter calling Anthropic “the most in-demand company for investment.”

• Rao’s mentor Laurence Tosi said “he’s not going to lose money on hope or raise it at stupid valuations”; Blackstone’s Martin Brand called him “technically very deep in the business.”

Kevin Weil (former CPO turned research lead) and Bill Peebles (Sora researcher) both departed OpenAI, as the company continues shedding “side quests” after shutting down Sora — which was losing an estimated $1 million per day in compute costs. Link.

• CTO of enterprise applications Srinivas Narayanan also left; OpenAI for Science (Weil’s research group behind Prism) is being absorbed into other teams.

• Weil’s GPT-5 claim about solving Erdős mathematical problems was debunked by the mathematician running erdosproblems.com; Peebles said “cultivating entropy is the only way for a research lab to thrive long-term.”

• The departures continue a pattern of senior exits as OpenAI narrows its focus to coding and enterprise tools under Fidji Simo’s “no side quests” mandate.

OpenAI is planning to price some ChatGPT ads based on clicks rather than impressions and exploring action-oriented ads that drive purchases, after some early advertisers paid CPMs of $15–$25 versus the targeted $60. Link.

• The ad pilot reached $100 million ARR in six weeks but has been slower than expected; some ads are selling below targets and the pilot period was extended past March.

• Monthly commitments dropped to $30,000–$50,000 from earlier $200,000+ upfront deals; OpenAI recently launched a self-service campaign portal after previously managing ads via spreadsheets and phone calls.

• Brainlabs’ Ben Kahan said “a lot of clients who would have the budget to want to test are holding back” because “they don’t think they’re going to be able to get the measurement that they want.”

AI-driven traffic to U.S. retailers’ websites rose 393% in Q1 2026 year-over-year, and for the first time, AI traffic is converting 42% better than human visitors — a dramatic reversal from a year ago when it converted 38% worse. Link.

• AI-driven revenue per visit is now 37% higher than non-AI traffic; shoppers arriving via AI spend 48% longer on sites and browse 13% more pages.

• 39% of survey respondents said they used AI for online shopping, with 85% saying it improved their experience and 66% now believing AI provides accurate results.

• However, roughly 25% of retail homepages aren’t optimized for LLMs and 34% of product pages can’t be properly accessed by AI, per Adobe Analytics data covering over 1 trillion visits.

Police tech startup Flock Safety raised roughly $200 million in new shares at an $8.4 billion valuation amid its largest political crisis — with dozens of cities canceling contracts and Amazon Ring dropping its partnership over data-sharing concerns with federal immigration enforcement. Link.

• Flock surpassed $300 million in ARR in March 2025 (a 70% YoY increase) and sells license plate detection cameras to over 6,000 U.S. cities; it has raised over $1 billion total from investors including a16z, Matrix Partners, and Greenoaks.

• A Flock spokesperson told CBS News the company does not have contracts with ICE or CBP; the per-share price was only about 6% above last spring’s round, reflecting the controversy.

• Competitor Axon’s shares are down 30% over the past year as the broader police tech sector faces increased scrutiny.

On the geopolitical front, Iran directly threatened Stargate data centers in the Middle East, while Trump administration officials quietly encouraged major banks to test Anthropic’s Mythos model. OpenAI forecasted its advertising business will hit $102 billion by 2030, and Intel signed on to Elon Musk’s Terafab chip manufacturing project.

Also, three senior Stargate executives departed OpenAI, SpaceX disclosed a nearly $5 billion loss driven by xAI spending, and Anthropic temporarily banned OpenClaw’s creator from accessing Claude.

More on this past week’s top AI headlines below!

- ZG

Here are the most important stories of the week:

AGENTS

Poke, an AI agent accessible via iMessage, SMS, Telegram, and WhatsApp, raised $10 million on top of a $15 million seed round at a $300 million valuation, with 10x customer growth in the past two months. Link.

• Backed by Spark Capital and General Catalyst with angels including the Collison brothers (Stripe), Jake and Logan Paul, and leaders from DeepMind, OpenAI, Cognition, and Vercel.

• Poke handles daily planning, calendar management, health tracking, smart home control, and photo editing through messaging apps, using whichever AI model best fits each task.

• Co-founder Marvin von Hagen said the company wants to “build a product for a billion people” and that being model-agnostic is a key advantage over competitors locked into a single provider.

CODING/DEVTOOLS

OpenAI launched a new $100/month ChatGPT Pro plan offering 5x more Codex usage than the $20 Plus plan, directly targeting Anthropic’s $100/month Claude tier as more than 3 million people now use Codex weekly — up 5x in three months. Link.

• The new tier sits between the $20/month Plus plan and the $200/month top tier, which offers 20x more Codex than Plus; OpenAI is offering bonus limits through May 31.

• Codex usage is growing more than 70% month over month; OpenAI said the Pro plan “delivers more coding capacity per dollar” than Claude Code across paid tiers.

• OpenAI’s pricing lineup now includes Free (with ads), $8/month Go (with ads), $20/month Plus (ad-free), $100/month Pro, and $200/month Pro (highest limits).

Anthropic temporarily banned OpenClaw creator Peter Steinberger from accessing Claude over “suspicious” activity, reinstating his account within hours after the post went viral — the latest escalation in tensions between the two sides. Link.

• Steinberger, who recently joined OpenAI, said “one welcomed me, one sent legal threats” and warned it will be “harder in the future to ensure OpenClaw still works with Anthropic models.”

• The ban came after Anthropic changed pricing so subscriptions no longer cover third-party harnesses like OpenClaw, requiring users to pay separately via the API.

• Anthropic said claws are more compute-intensive than standard prompts because they run continuous reasoning loops and connect to third-party tools.

RESEARCH

Anthropic launched Project Glasswing, a cybersecurity initiative built around its unreleased Mythos model — described as “by far the most powerful AI model we’ve ever developed” — which has already found thousands of high-severity zero-day vulnerabilities in major operating systems and browsers. Link.

• Mythos found a 27-year-old vulnerability in OpenBSD, a 16-year-old vulnerability in FFmpeg, and autonomously chained vulnerabilities in the Linux kernel; it scored 93.9% on SWE-bench Verified vs. 80.8% for Opus 4.6.

• Twelve launch partners include AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, Linux Foundation, Microsoft, Nvidia, and Palo Alto Networks, with 40+ additional organizations gaining access.

• Anthropic is committing up to $100 million in Mythos usage credits and $4 million in direct donations to open-source security organizations; the model will not be made generally available due to its offensive capabilities.

INFRASTRUCTURE

Anthropic expanded its compute deal with Google and Broadcom to 3.5 gigawatts (up from 1 GW in October), as the company’s run rate revenue hit $30 billion — up from $9 billion at end of 2025 — and it raised a $30 billion Series G at a $380 billion valuation. Link.

• The new compute capacity comes online in 2027, with the majority housed in the U.S. as part of a $50 billion commitment to domestic compute infrastructure.

• Anthropic now has over 1,000 business customers spending more than $1 million annualized; CFO Krishna Rao called it the company’s “most significant compute commitment to date.”

• The deal positions Anthropic’s infrastructure on Google TPUs and Broadcom chips, diversifying from its existing AWS Trainium deployment.

Intel signed on to Elon Musk’s Terafab chip manufacturing project alongside SpaceX and Tesla, bringing its fabrication expertise to the Texas-based facility aiming to produce 1 terawatt per year of compute for AI and robotics. Link.

• Intel said its ability to “design, fabricate, and package ultra-high-performance chips at scale” will help Terafab achieve its compute targets; Intel’s stock rose over 3% on the news.

• Building a chip fab typically requires years and over $20 billion; Intel is seeking large anchor customers for its foundry business as it competes with TSMC.

• The partnership gives Terafab access to Intel’s manufacturing infrastructure rather than building from scratch, potentially accelerating the timeline.

Singapore-based Firmus raised $505 million led by Coatue at a $5.5 billion valuation, bringing total funding to $1.35 billion in six months as it builds Nvidia-powered “AI factory” data centers in Australia under Project Southgate. Link.

• Firmus’ new data centers will use Nvidia’s next-gen Vera Rubin platform (succeeding the Blackwell architecture), expected to ship in H2 2026.

• The company originally provided cooling technologies for Bitcoin mining before pivoting to AI infrastructure — a common trajectory that investors continue to favor.

• Previous investors include Nvidia; the prior round was AU$330 million (~$215 million) at AU$1.85 billion (~$1.2 billion) valuation.

Three senior OpenAI executives who helped launch the Stargate initiative — Peter Hoeschele, Shamez Hemani, and Anuj Saharan — have departed or are preparing to leave, all heading to the same undisclosed new company. Link.

• OpenAI shifted from building its own data centers to renting compute capacity, hiring former Intel executive Sachin Katti to lead a new “Industrial Compute” organization.

• The infrastructure team inked deals for 8 gigawatts of capacity last year, short of its 10 GW goal; OpenAI ended the year with about 1.9 GW and plans to reach mid-single digits by end of 2026.

• OpenAI plans to secure over $600 billion in compute over five years; Saharan wrote in a Slack note that “getting to help build literally the world’s largest computers” was “the ride of a lifetime.”

POLICY/GOV’T/ETHICS

Trump administration officials including Treasury Secretary Scott Bessent and Fed Chair Jerome Powell summoned bank executives and encouraged them to test Anthropic’s Mythos model for detecting financial system vulnerabilities. Link.

• JPMorgan Chase was the only initial partner with access; Goldman Sachs, Citigroup, Bank of America, and Morgan Stanley are also reportedly testing Mythos.

• The move is notable given the Trump administration is simultaneously in court fighting Anthropic over the DOD’s supply-chain risk designation.

• U.K. financial regulators are also discussing the security risks posed by Mythos’ vulnerability-finding capabilities.

Iran directly threatened Stargate AI data centers in the Middle East, releasing video showing the UAE facility with the message “nothing stays hidden to our sight” — after missiles already struck AWS data centers in Bahrain and an Oracle facility in Dubai. Link.

• The $500 billion Stargate joint venture between OpenAI, SoftBank, and Oracle had been seeking to expand internationally when war broke out in the Middle East in February 2026.

• Iran also threatened Nvidia and Apple by name; Trump threatened to strike Iran’s civilian infrastructure unless it reopens the Strait of Hormuz.

• The targeting of data centers represents a new dimension in warfare, as AI computing infrastructure becomes as strategically important as traditional energy and communications infrastructure.

OpenAI released a sweeping economic policy proposal calling for public wealth funds giving Americans stakes in AI companies, robot taxes, a four-day workweek, and treating AI as a public utility. Link.

• The proposal calls for shifting the tax burden from labor to capital, with higher taxes on corporate income and AI-driven returns, plus portable benefit accounts that follow workers across jobs.

• OpenAI compared the moment to the need for a “New Deal” during the Industrial Age, writing that “superintelligence will require an even more ambitious form of industrial policy.”

• Critics noted that many proposed benefits remain employer-dependent — which could leave workers vulnerable if AI eliminates their jobs entirely — and questioned the proposal’s compatibility with OpenAI’s for-profit restructuring.

OpenAI released a Child Safety Blueprint developed with the National Center for Missing and Exploited Children and state attorneys general, addressing the rise of AI-generated child sexual abuse content — which increased 14% year-over-year. Link.

• The Internet Watch Foundation reported over 8,000 reports of AI-generated child sexual abuse material in H1 2025; criminals are using AI to generate sextortion images and convincing grooming messages.

• The blueprint focuses on updating legislation for AI-generated abuse material, refining reporting to law enforcement, and integrating preventative safeguards into AI systems.

• OpenAI faces seven lawsuits alleging GPT-4o contributed to wrongful deaths by suicide and severe delusions after extended chatbot interactions.

OTHER

OpenAI internally forecasts its advertising business will reach $102 billion by 2030 — 36% of total revenue — after its ads pilot topped $100 million in annualized revenue just six weeks after launch. Link.

• OpenAI projects average revenue per user growing from $3.50 this year to nearly $60 by 2030 — roughly on par with Meta’s current $57 ARPU — with ad revenue jumping from $2.4 billion this year to $11 billion next year.

• The company plans to open ad sales to more marketers this month and is evaluating whether to partner with ad tech firms like The Trade Desk or build its own ad management system.

• The forecast comes as Anthropic ran Super Bowl ads with the tagline “Ads are coming to AI. But not to Claude” — and has likely already outpaced OpenAI in annualized revenue.

xAI’s massive spending on chips and data centers — nearly $13 billion in capex, 50% more than SpaceX’s rocket and satellite divisions combined — pushed SpaceX to a nearly $5 billion net loss on $18.5 billion in revenue ahead of its planned June IPO. Link.

• SpaceX’s core businesses generated nearly $8 billion in adjusted EBITDA, but xAI’s infrastructure buildout overwhelmed the company’s profitability after SpaceX acquired xAI in February.

• Depreciation was among the largest expenses at $6.6 billion, with stock-based compensation and interest each adding roughly $2 billion more.

• IPO investors will essentially be financing Musk’s unproven AI ambitions alongside the proven space launch and Starlink businesses.

Mercor’s data breach fallout deepened as Meta paused its contracts with the $10 billion AI recruiting startup indefinitely, hackers claimed to have 4TB of stolen data, and five contractors filed lawsuits over exposed personal information. Link.

• Meta had previously spent $14.3 billion acquiring Mercor’s competitor Scale AI but continued working with Mercor until the breach; OpenAI is investigating its exposure but hasn’t paused contracts.

• Mercor was on pace for over $1 billion in annualized revenue before the leak; the breach originated from a supply chain attack on open-source tool LiteLLM, which harbored credential-harvesting malware for 40 minutes.

• LiteLLM had used AI compliance startup Delve for security certifications; Delve has been accused of faking data for certifications, leading Y Combinator to sever ties.

A group of OpenAI alumni including former engineers and the company’s original prompt engineer launched Zero Shot, a venture fund that closed on $20 million with a $100 million target, investing in AI startups while betting against vibe coding and digital twins. Link.

• Portfolio companies include Worktrace AI ($10 million seed from Mira Murati) and Foundry Robotics ($13.5 million seed led by Khosla Ventures); advisors include former OpenAI head of people Diane Yoon and former communications chief Steve Dowling.

• Co-founder Evan Morikawa warned against “embodiment training data” companies in robotics, saying “there’s a lot of hoping and praying going on right now” about bridging the gap between simulation and the physical world.

• Co-founder Andrew Mayne said the fund’s edge is “knowing how to predict where these models will be going next, because it’s extremely not obvious.”

The growing trend of “tokenmaxxing” — engineers treating AI token consumption as a productivity benchmark — is driving rapid revenue growth for Anthropic and OpenAI while squeezing margins as inference costs rise from multi-step coding tasks. Link.

• Users are increasingly relying on tools like Claude Code and ChatGPT Codex for continuous reasoning loops rather than single-question interactions, dramatically increasing compute demands per user.

• The heavier usage patterns are pressuring AI companies to invest more heavily in chips and dedicated infrastructure to maintain performance.

• The dynamic creates a tension between growth and profitability that will define the next phase of the AI business model.

Anthropic had a busy week too: it acquired biotech AI startup Coefficient Bio for $400 million, filed to create a new political action committee, accidentally took down 8,100 GitHub repos, and told Claude Code subscribers they’ll need to pay extra for OpenClaw usage.

Meanwhile, Microsoft announced three new foundation models, Apple escalated its vibe coding crackdown by removing an app entirely from the App Store, and Meta’s new Louisiana data center will draw as much electricity as the entire state of South Dakota.

More on this past week’s top AI headlines below!

- ZG

Here are the most important stories of the week:

IMAGE/VIDEO

AI video startup Runway launched a $10 million venture fund and Builders program for early-stage companies in AI, media, and world simulation, writing checks of up to $500,000 for pre-seed and seed startups. Link.

• The Builders program offers seed-to-Series C startups 500,000 free API credits powered by Runway’s general world models, which launched in December; a founding cohort has already been established.

• Runway has raised close to $860 million at a $5.3 billion valuation with just 150 employees; portfolio companies include LanceDB, Tamarind Bio, and Cartesia.

• The move follows similar venture funds from Perplexity ($50 million) and CoreWeave, as AI companies increasingly invest in their own ecosystems.

CODING/DEVTOOLS

Anthropic told Claude Code subscribers they must pay separately for third-party tools like OpenClaw starting April 4, saying subscription limits weren’t built for the usage patterns of open-source harnesses. Link.

• OpenClaw creator Peter Steinberger, who recently joined OpenAI, said he and board member Dave Morin “tried to talk sense into Anthropic” but only delayed the pricing increase by a week.

• Steinberger accused Anthropic of anti-competitive behavior: “Funny how timings match up, first they copy some popular features into their closed harness, then they lock out open source.”

• Anthropic’s head of Claude Code Boris Cherny called it an engineering constraint, saying the company remains “big fans of open source.”

OpenRouter, a startup that provides API access to 300+ AI models, is in talks to raise $120 million led by CapitalG (Alphabet’s venture arm) at a $1.3 billion valuation, with annualized revenue surging from $10 million to $50 million in five months. Link.

• Founded in 2023 by Alex Atallah — who previously co-founded NFT marketplace OpenSea — OpenRouter helps developers route requests across models from OpenAI, Anthropic, Google, and others.

• The valuation represents a 24x forward sales multiple, below the AI app average of 53x, suggesting room for growth.

• Competitors include PortKey, Vercel, and LMArena, which raised $150 million at a $1.7 billion valuation in January.

Apple removed the vibe coding app Anything entirely from the App Store, escalating its crackdown from blocking updates to full removal, citing violations of rules against apps that execute unreviewed code. Link.

• Anything had published thousands of apps through its tool before being removed; co-founder Dhruv Amin said he’s considering moving to Android, calling it “the problem with Apple and closed platforms.”

• Replit and Bitrig remain in the App Store but haven’t been allowed to push updates since January and November respectively; Vibecode has abandoned its iPhone app entirely to focus on web.

• Apple’s App Store saw an 84% jump in new apps last quarter, driven partly by the vibe coding boom — which Apple may view as a threat to its 15–30% commission structure.

Database startup Supabase is in talks to raise approximately $500 million at a $10 billion valuation led by Singapore’s GIC, doubling from $5 billion last fall and 5x from a year ago. Link.

• Supabase’s annualized revenue reached $70 million by mid-2025, up from $20 million the prior year, fueled by the rise of AI coding assistants like Cursor and Lovable that frequently use Supabase as a backend.

• Founded in 2020 by Paul Copplestone and Ant Wilson, Supabase competes with Google’s Firebase and is built on PostgreSQL with a freemium model.

• Previous investors include Accel, Felicis, Coatue Management, Craft Ventures, and Y Combinator.

RESEARCH

DeepSeek’s upcoming V4 model will run on Huawei’s Ascend 950PR chips instead of Nvidia hardware, with Chinese tech giants Alibaba, ByteDance, and Tencent placing bulk orders that drove a 20% price surge for the chip. Link.

• The Ascend 950PR delivers 2.8x the compute performance of Nvidia’s H20, and Huawei’s next-gen Ascend 960 (arriving 2027) is expected to be comparable to Nvidia’s H200.

• DeepSeek denied Nvidia early access to V4, giving Chinese chip makers a head start; V4’s launch was delayed from February due to chip optimization work.

• Huawei’s Ascend 950DT training chip is coming in Q4 2026, potentially reducing China’s dependence on Nvidia for both inference and training workloads.

Microsoft AI announced three new foundational models — MAI-Transcribe-1, MAI-Voice-1, and MAI-Image-2 — developed by its Superintelligence team led by Mustafa Suleyman, positioning them as cheaper alternatives to Google and OpenAI offerings. Link.

• MAI-Transcribe-1 transcribes speech across 25 languages at $0.36/hour (2.5x faster than Azure’s previous offering); MAI-Voice-1 generates 60 seconds of audio in one second for $22 per million characters.

• MAI-Image-2, a video-generating model, costs $5 per million text input tokens and $33 per million image output tokens.

• The release follows the renegotiation of Microsoft’s OpenAI partnership, which allowed Microsoft to pursue its own superintelligence research despite having invested over $13 billion in OpenAI.

INFRASTRUCTURE

OpenAI closed a $122 billion funding round at an $852 billion valuation — its largest raise ever — co-led by SoftBank and Andreessen Horowitz, with $3 billion coming from retail investors through bank channels. Link.

• OpenAI now generates $2 billion in monthly revenue with over 900 million weekly active users and 50 million+ subscribers; its ads pilot brought in over $100 million in ARR in under six weeks.

• Business now makes up 40% of revenue (up from 30% last year); the company expanded its revolving credit facility to $4.7 billion (undrawn).

• Other investors include D.E. Shaw Ventures, MGX, TPG, T. Rowe Price, Amazon, Nvidia, and Microsoft; OpenAI called itself an “AI superapp” growing “four times faster” than Alphabet and Meta at the same stage.

Mistral AI raised $830 million in debt to build a new data center near Paris in Bruyères-le-Châtel powered by Nvidia chips, with plans to make it operational by Q2 2026. Link.

• Last month, Mistral announced a $1.4 billion investment in Sweden for AI infrastructure; the company aims to deploy 200 megawatts of compute capacity across Europe by 2027.

• CEO Arthur Mensch said scaling European infrastructure is “critical to empower our customers and to ensure AI innovation and autonomy remain at the heart of Europe.”

• Mistral has raised over €2.8 billion ($3.1 billion) to date from investors including General Catalyst, ASML, a16z, Lightspeed, and DST Global.

ENERGY

Meta will fund 10 new natural gas power plants to support its $27 billion Hyperion AI data center in Louisiana, generating approximately 7.5 gigawatts of electricity — roughly equivalent to the entire power capacity of South Dakota. Link.

• The 10 plants will dump an estimated 12.4 million metric tons of CO2 into the atmosphere every year, according to TechCrunch’s calculations based on Department of Energy data.

• Meta originally committed to three natural gas plants but tripled the plan to 10, raising questions about the company’s climate commitments.

• Natural gas has been called a “bridge fuel” to renewables for decades, but the scale of Meta’s commitment suggests AI infrastructure demand is extending that bridge significantly.

POLICY/GOV’T/ETHICS

Anthropic filed documents to create AnthroPAC, a new political action committee funded by voluntary employee contributions capped at $5,000, which will make contributions to both parties during the midterms. Link.

• AI companies collectively contributed $185 million to midterm races; Anthropic separately gave at least $20 million to Public First, a Super PAC running ad campaigns supporting a specific regulatory agenda.

• The PAC will support current D.C. lawmakers and rising political candidates from both parties; Allison Rossi, Anthropic’s treasurer, signed the filing with the Federal Election Commission.

• The move comes as Anthropic is locked in a legal battle with the DOD over the government’s use of Claude and its AI safety “red lines.”

OTHER

OpenAI’s COO Brad Lightcap moved to a new “special projects” role reporting to Sam Altman, while Fidji Simo went on medical leave for a neuroimmune condition and CMO Kate Rouch stepped down to focus on cancer recovery. Link.

• Greg Brockman will manage product during Simo’s absence; newly hired CRO Denise Dresser (former Slack CEO) takes over some of Lightcap’s commercial duties.

• Simo said “the timing is maddening” and she had “done everything possible to avoid it, but sadly my body isn’t cooperating.”

• Separately, OpenAI CEO and CFO Sarah Friar are diverging on IPO timing — Altman wants Q4 2026, while Friar has voiced concerns about the company’s $600 billion spending commitment over the next five years.

OpenAI acquired TBPN, the buzzy daily tech talk show hosted by John Coogan and Jordi Hays, in its first media acquisition — reporting to chief political operative Chris Lehane with the show maintaining editorial independence. Link.

• TBPN is on track for $30 million in revenue this year and has hosted CEOs including Zuckerberg, Nadella, Benioff, and Altman on its daily three-hour live show on YouTube and X.

• The deal was spearheaded by Fidji Simo, who described it as part of a strategy where “the standard communications playbook just doesn’t apply” to OpenAI.

• Altman acknowledged on social media: “I don’t expect them to go any easier on us, am sure I’ll do my part to help enable that with occasional stupid decisions.”

Anthropic acquired biotech AI startup Coefficient Bio for approximately $400 million — a team of fewer than 10 employees founded less than a year ago — to bolster its healthcare and life sciences AI tools. Link.

• Coefficient Bio’s AI platform handles drug R&D planning, clinical regulatory strategy, and drug discovery; the team is joining Anthropic’s healthcare life sciences group led by Eric Kauderer-Abrams.

• CEO Aris Theologis (formerly of Evozyne and Paragon Biosciences) and CTO Nathan Frey (formerly of Prescient Design and Genentech) lead the team.

• The acquisition reflects Anthropic’s push into industry-specific AI tools after October improvements for scientists and January features for regulatory submissions and clinical trial protocols.

Anthropic accidentally took down approximately 8,100 GitHub repositories when a DMCA takedown targeting its leaked Claude Code source code inadvertently swept up legitimate forks of its own public repo. Link.

• A software engineer had discovered that a recent release accidentally included access to Claude Code’s source code; AI enthusiasts quickly shared it on GitHub.

• Anthropic retracted the bulk of the notices, limiting the takedown to one repo and 96 forks with the accidentally released code; GitHub restored access to affected repositories.

• The incident was described as “another black eye” for Anthropic as it reportedly prepares for an IPO, with potential shareholder lawsuit implications noted.

AI recruiting startup Mercor confirmed a security incident linked to a supply chain attack on the widely used open-source project LiteLLM, with extortion group Lapsus$ claiming it gained access to Mercor’s data. Link.

• Mercor, valued at $10 billion after a $350 million Series C in October 2025, facilitates more than $2 million in daily payouts and counts OpenAI and Anthropic among its customers.

• The malicious code in LiteLLM — a Y Combinator-backed project downloaded millions of times daily — was identified and removed within hours.

• The incident highlights growing supply chain risks in the AI ecosystem as open-source dependencies become critical infrastructure for major companies.

On the OpenAI side, Sam Altman shifted responsibilities to focus on infrastructure, the company is prepping a new model codenamed “Spud,” and it abandoned both Sora and its planned erotic mode for ChatGPT. Meanwhile, SpaceX is aiming to file for the biggest U.S. IPO of all time as soon as this week.

In speech/audio, both Cohere and Mistral dropped open-source voice models this week, and Harvey hit an $11 billion valuation with Sequoia tripling down.

More on this past week’s top AI headlines below!

- ZG

Here are the most important stories of the week:

AGENTS

Tencent launched eight OpenClaw-based applications and services this month, including QClaw — an AI agent that integrates directly with WeChat — which attracted millions of sign-ups within its first week and sent Tencent’s stock up more than 7%. Link.

• Nearly 1,000 people showed up at Tencent’s headquarters in Shenzhen on March 6 for help installing OpenClaw; Weixin ClawBot launched on Sunday, integrating OpenClaw directly into WeChat.

• Tencent plans to at least double its AI investment from the 18 billion yuan ($2.6 billion) it spent last year, according to CEO Pony Ma.

• The push reflects Tencent’s effort to catch up with ByteDance’s Doubao and Alibaba’s Qwen chatbots in the intensifying China AI battle.

AWS is accelerating the deployment of internal AI agents to automate sales, business development, and partner coordination functions following hundreds of layoffs in recent staff cuts. Link.

• An AI agent now helps sales staff provide technical answers, handling work previously done by thousands of technical specialists, while a separate agent coordinates with business partners on cloud deals.

• Amazon CEO Andy Jassy expects internal AI use to reduce the corporate workforce over time, with leaders planning robotic automation to replace hundreds of thousands of roles by 2033.

• AWS cut from approximately 140,000 employees in late 2022 to 115,000 by end of 2023, with additional layoffs from October 2025 through January 2026 as part of Amazon’s broader 30,000 role cuts.

TEXT

AI legal startup Harvey confirmed an $11 billion valuation after raising $200 million co-led by Singapore’s GIC and Sequoia Capital, marking Sequoia’s third consecutive round leading the company and bringing total funding past $1 billion. Link.

• Harvey’s valuation jumped 3.5x in 12 months — from $3 billion in February 2025 to $5 billion in June 2025 to $8 billion in December 2025 to $11 billion now.

• Sequoia partner Pat Grady acknowledged this was “an unusually large show of faith” for the VC firm, co-leading three consecutive rounds.

• Other participating investors include Andreessen Horowitz, Coatue, Conviction Partners, Elad Gil, Evantic, and Kleiner Perkins.

Meeting notetaker Granola raised $125 million at a $1.5 billion valuation — a 6x jump from its previous $250 million valuation — as it expands into a broader enterprise AI platform with team workspaces and developer APIs. Link.

• The Series C was led by Danny Rimer at Index Ventures with participation from Kleiner Perkins, Lightspeed, and Spark; total funding is now $192 million.

• New features include “Spaces” for team workspaces, personal and enterprise APIs, and integrations with Claude, ChatGPT, Lovable, Figma Make, Replit, and Cursor.

• Enterprise clients include Vanta, Gusto, Thumbtack, Asana, Cursor, Lovable, Decagon, and Mistral AI.

Sam Altman relinquished direct oversight of OpenAI’s safety and security teams to focus on capital raising and infrastructure, as the company completed initial development of its next major AI model codenamed “Spud” and shut down the Sora video app. Link.

• The safety team moved under Chief Research Officer Mark Chen, the security team moved to Greg Brockman’s “scaling” organization, and Fidji Simo’s product org was renamed “AGI Deployment.”

• The Sora shutdown caused Disney to abandon a planned $1 billion investment in OpenAI; OpenAI is now creating a desktop “superapp” combining ChatGPT with Codex and browser Atlas.

• Altman said “things are moving faster than many of us expected” and that the company expects a “very strong model” in “a few weeks” that “can really accelerate the economy.”

IMAGE/VIDEO

Mirage, the company behind AI video-editing app Captions, raised $75 million from General Catalyst’s Customer Value Fund, with over 3.2 million app downloads in the past year and $28.4 million in in-app revenue. Link.

• The company rebranded from Captions to Mirage to position itself as an AI lab, with 75% of revenue coming from international markets and over 200 million videos created on the platform to date.

• Mirage trained a custom model for pacing, framing, and attention dynamics in short-form video, along with a new audio model that preserves accents in generated videos.

• Competitors include ByteDance’s CapCut, Meta’s Edits, Canva, D-ID, HeyGen, and Webflow.

SPEECH/AUDIO

Cohere launched Transcribe, its first open-source voice model with just 2 billion parameters, which topped the Hugging Face Open ASR leaderboard with an average word error rate of 5.42 across 14 supported languages. Link.

• Transcribe beat Zoom Scribe v1, IBM Granite 4.0 1B, ElevenLabs Scribe v2, and Qwen3-ASR-1.7B Speech on the leaderboard, with a 61% average win rate over other models in human evaluations.

• The model is designed for consumer-grade GPUs and will be integrated into Cohere’s enterprise agent platform North, with free access via API and Model Vault.

• The lightweight model is meant for use cases like note-taking, speech analysis, and enterprise transcription.

Mistral released Voxtral TTS, an open-source text-to-speech model based on Ministral 3B that supports nine languages, runs on edge devices as small as a smartwatch, and can clone a custom voice from a sample under five seconds. Link.

• The model achieves 90ms time-to-first-audio for a 10-second sample and a 6x real-time factor, rendering a 10-second clip in roughly 1.6 seconds.

• Voxtral can capture accents, inflections, and intonations, and switch between languages without losing voice characteristics.

• Mistral VP of Science Operations Pierre Stock said the cost is “a fraction of anything else on the market” and the company plans to build “an end-to-end platform” handling multimodal streams of audio, text, and image.

INFRASTRUCTURE

SpaceX is targeting a June public listing and aims to file its IPO prospectus as early as this week, in what is expected to be the biggest U.S. IPO of all time, raising over $75 billion at a $1.25 trillion valuation. Link.

• The deal would surpass all U.S. IPO money raised last year combined; Goldman Sachs, Morgan Stanley, Bank of America, JPMorgan, and Citi are acting as bankers.

• SpaceX recently absorbed xAI (valued at $250 billion) and plans to give individual investors a higher-than-typical allocation of over 20%, with no standard six-month lockup period.

• The IPO pitch will focus on the space launch business and Starlink satellite internet, while X and xAI are not expected to play a major role in the investor presentation.

Blackstone is close to acquiring data center firm Rowan Digital Infrastructure in a deal likely valued at over $10 billion including debt, adding to its growing portfolio that includes QTS and AirTrunk. Link.

• Rowan has raised over $4 billion in construction debt since mid-2024 and operates sites in Maryland (200+ MW, expandable to 620 MW by 2027), Texas, and Oregon (each ~300 MW).

• Blackstone previously bought QTS for about $10 billion five years ago and paid approximately $16 billion for Australia’s AirTrunk in 2024; QTS was the biggest contributor to Blackstone’s infrastructure fund performance last year.

• Blackstone is also holding early discussions about forming a publicly traded data center investment vehicle; Amazon Web Services rents space at Rowan’s Maryland data center.

AMD-backed cloud startup Vultr is seeking to raise at least $1 billion to compete with AI cloud providers, working with Goldman Sachs on the fundraising after previously raising $333 million at a $3.5 billion valuation in 2024. Link.

• Founded in 2014 and based in Florida, Vultr operates 32 cloud data centers on six continents and rents access to AI chips from both AMD and Nvidia.

• The company is planning a $1 billion+ investment in an Ohio data center with capacity for 24,000 AMD high-end AI chips.

• Competitors in the AI cloud space include CoreWeave (which raised $8.5 billion in debt), Nscale ($2 billion raise), and Together AI (in talks for $1 billion at $7.5 billion valuation).

Microsoft executives told managers in Azure cloud and North American sales groups to suspend all new hiring, while also pausing development on five new buildings at its Redmond campus. Link.

• Microsoft’s stock has fallen 24% in 2026, the worst performer among big tech stocks; approximately 45% of Azure’s revenue backlog comes from a single customer: OpenAI.

• Azure Core chief of staff Hilary Macfadden wrote that the org “no longer has room or approval to continue hiring” until there are “credible, executable plans” to address the gross margin gap.

• The hiring freeze is not company-wide — engineering teams building Copilot are still hiring — but some senior executives believe Microsoft’s headcount won’t increase in coming years.

ENERGY

Sam Altman-backed fusion startup Helion is in early talks to sell power to OpenAI, with the deal guaranteeing 12.5% of Helion’s production as it targets 5 gigawatts by 2030 and 50 gigawatts by 2035. Link.

• Altman is stepping down as Helion’s board chair after more than a decade to enable the partnership; each Helion reactor generates 50 megawatts, requiring 800 reactors by 2030 and 8,000 by 2035.

• Helion raised $425 million last year from Altman, Mithril, Lightspeed, and SoftBank, and in February generated plasmas reaching 150 million degrees Celsius (target: 200 million for commercial operations).

• The company’s Polaris prototype and first commercial-scale reactor are targeting the early 2030s, ahead of most fusion competitors; Helion already has a deal with Microsoft to deliver power starting in 2028.

Senators Josh Hawley and Elizabeth Warren sent a letter urging the Energy Information Administration to establish mandatory annual reporting requirements for data centers, including hourly energy loads, rates paid, and grid upgrade costs. Link.

• The bipartisan effort demands details on whether data center customers participate in demand response programs and what grid upgrades are required by new large loads; a response was requested by April 9.

• Separately, Sen. Bernie Sanders and Rep. Alexandria Ocasio-Cortez announced legislation to halt new data center construction until Congress can agree on AI regulation.

• Sen. Chris Van Hollen introduced the Power for the People Act with seven co-sponsors to rein in rising electricity costs driven by data center demand.

POLICY/GOV’T/ETHICS

A federal judge granted Anthropic a preliminary injunction blocking the Pentagon’s supply-chain risk designation, with Judge Rita Lin calling it “classic illegal First Amendment retaliation” and ordering the government to rescind its ban. Link.

• “It looks like an attempt to cripple Anthropic,” Judge Lin said, ruling that the government’s orders flouted free speech protections for the company’s public stance on AI safety.

• The injunction pauses the administration’s plan to sever all ties with Anthropic while the case plays out in San Francisco federal court, with a seven-day hold for the government to appeal.

• The ruling follows weeks of escalation after Defense Secretary Pete Hegseth blacklisted Anthropic for refusing to allow Claude to be used in lethal autonomous operations or mass surveillance of U.S. citizens.

OTHER

Anthropic’s Claude paid subscriptions have more than doubled in 2026, with record new subscriber growth between January and February driven by Super Bowl ads mocking ChatGPT, the DOD feud, and the launch of Claude Code and Cowork. Link.

• An analysis of billions of anonymized credit card transactions from 28 million U.S. consumers by Indagari showed Claude gaining paid subscribers in record numbers, with the majority at the $20/month Pro tier.

• New user growth climbed sharply between late January media reports on the DOD feud and Dario Amodei’s public statement on February 26; the Computer Use feature released this week also sparked a surge.

• Despite the growth, Claude remains well behind ChatGPT in total paying consumers; OpenAI is still gaining new paid subscribers at a rapid rate.

Anthropic executives have discussed IPO timing as soon as Q4 2026, with bankers expecting the offering to raise over $60 billion — potentially the second-biggest IPO after SpaceX — as annualized sales more than doubled to $19 billion. Link.

• Anthropic is working with law firm Wilson Sonsini on concrete IPO preparation steps; the company was last valued at $350 billion in its most recent financing.

• OpenAI CEO Sam Altman has privately said he would prefer OpenAI to list before Anthropic, but bankers expect Anthropic to go first due to market preference for its developer/enterprise focus.

• Anthropic is projected to generate positive cash flow by 2028, while OpenAI — close to raising $120 billion at a $730 billion valuation — projects positive free cash flow by 2030.

OpenAI indefinitely paused plans for ChatGPT’s erotic “adult mode,” first floated by Sam Altman in October, as part of a broader strategy shift away from consumer side projects and toward business and coding tools. Link.

• The feature faced intense criticism from tech watchdog groups and OpenAI’s own staff; one adviser at a heated January meeting cautioned the company could be developing a “sexy suicide coach.”

• The release had been delayed multiple times, and the erotic feature now has no timeline for launch.

• The decision follows the shutdown of Sora and represents a consolidation of OpenAI’s focus on enterprise and developer products.

OpenAI’s shopping feature pivot left partners PayPal, Etsy, and others scrambling after the company abruptly scaled back its Instant Checkout vision from broad in-ChatGPT purchasing to limited checkout inside specific retailer apps. Link.

• PayPal had committed to increasing API spending, giving 24,000+ employees ChatGPT Enterprise access, and handling payments for in-chat checkouts; Etsy was paying transaction fees to OpenAI and developing its own ChatGPT app.

• The Sora shutdown also jeopardized Disney’s planned $1 billion investment in OpenAI, raising broader concerns about the reliability of OpenAI’s partnership commitments.

• Stripe continues work on its “Agentic Commerce Protocol” with OpenAI to standardize AI chatbot transactions, and Walmart is rolling out ChatGPT Enterprise to its teams.

All 11 of Elon Musk’s original xAI co-founders have now departed the company, with the last two — Manuel Kroiss (pretraining lead) and Ross Nordeen (Musk’s “right-hand operator”) — leaving this week. Link.

• Musk previously claimed xAI “was not built right the first time around” and is being “rebuilt from the foundations up.”

• xAI was recently acquired by SpaceX, bringing SpaceX, xAI, and X together under one corporate umbrella ahead of SpaceX’s expected IPO.

• Nordeen came to xAI from Tesla and was involved in planning major layoffs at Twitter after Musk’s 2022 acquisition.

Meta laid off a few hundred employees across Reality Labs, social media teams, recruiting, and sales — the latest in a series of cuts after eliminating 1,500 Reality Labs employees in January and 5% of lowest-performing staff last February. Link.

• The cuts affect a small portion of Meta’s 78,000+ workforce; Reuters previously reported the company could cut at least 20% of total headcount in 2026, which Meta called “speculative.”

• Meta projected 2026 capital expenditures of $115–$135 billion (a roughly 75% increase from last year), with a 40% increase in operating expenses partly from technical staff compensation.

• Despite the layoffs, Meta continues to invest heavily in AI, signaling a rebalancing of headcount from non-technical roles toward AI engineering.

OpenAI released a set of open-source safety prompts and a lightweight model called gpt-oss-safeguard to help developers make apps safer for teens, covering graphic violence, sexual content, dangerous challenges, and age-restricted goods. Link.

• Common Sense Media’s head of AI assessments called the prompts helpful for setting “a meaningful safety floor across the ecosystem” that can be “adapted and improved over time.”

• OpenAI acknowledged the tools aren’t a complete solution, noting that even experienced teams “often struggle to translate safety goals into precise, operational rules.”

• The release comes as OpenAI faces lawsuits from families of people who died by suicide after extreme ChatGPT use.

In infrastructure news, Elon Musk unveiled “Terafab,” a chip manufacturing collaboration between Tesla and SpaceX, while Amazon gave an exclusive tour of its Trainium chip lab following a $50 billion deal with OpenAI. Meanwhile, Nvidia’s networking division quietly posted $11 billion in quarterly revenue, surpassing Cisco’s entire annual networking business.

In coding/devtools, Cursor admitted its new Composer 2 model was built on top of Chinese startup Moonshot AI’s Kimi, and Apple cracked down on vibe coding apps in the App Store.

More on this past week’s top AI headlines below!

- ZG

Here are the most important stories of the week:

AGENTS

An AI agent at Meta went rogue and exposed sensitive company and user data to unauthorized employees for two hours, prompting a “Sev 1” classification — the second-highest severity level for security incidents. Link.

• The incident occurred when an engineer asked an AI agent to help analyze a technical question on an internal forum; the agent posted a response without permission and gave faulty advice that led to the data exposure.

• Meta’s safety and alignment director Summer Yue described a separate incident where her own OpenClaw agent deleted her entire inbox despite being instructed to confirm actions first.

• Despite the incidents, Meta remains bullish on agentic AI, having recently acquired Moltbook, a Reddit-like social platform for AI agents to communicate with each other.

Tools for Humanity launched AgentKit, a tool that lets commercial websites verify real humans are authorizing AI agents to make purchases, using World ID iris scans and Coinbase’s x402 blockchain protocol. Link.

• AgentKit integrates Sam Altman’s World ID with the x402 protocol developed by Coinbase and Cloudflare, allowing AI agents to transact on behalf of verified humans without requiring human intervention at each step.

• Chief Product Officer Tiago Sada compared the system to delegating “power of attorney” to an AI agent for purchasing decisions.

• The tool addresses growing fraud concerns as major platforms including Amazon, Mastercard, and Google have already introduced automated buying capabilities.

IMAGE/VIDEO

AI inference platform Fal is raising $300–$350 million at an $8 billion valuation, nearly doubling from $4.5 billion just three months ago, as annualized revenue doubled to $400 million. Link.

• Fal serves 3 million developers including Adobe, Canva, and Shopify, providing cloud infrastructure for running AI models that generate images, video, and audio.

• The company was founded in 2021 by former Amazon and Coinbase engineers and previously raised $314 million across three rounds in 2024 alone.

• The dual-tranche fundraising structure lets earlier investors get quick markups, though the approach has raised concerns about startup valuation bubbles in the AI inference space.

CODING/DEVTOOLS

Cursor admitted its new Composer 2 coding model was built on top of Moonshot AI’s open-source Kimi 2.5, a Chinese model backed by Alibaba, after an X user discovered Kimi’s model ID buried in Composer 2’s code. Link.

• Cursor VP Lee Robinson acknowledged the open-source base but said only about one-quarter of the compute in the final model came from Kimi, with the rest from Cursor’s own training.

• Co-founder Aman Sanger called it “a miss to not mention the Kimi base in our blog from the start,” noting they would be more transparent with future models.

• The revelation is notable given the framing of AI as a U.S.–China “arms race,” and comes as Cursor reportedly exceeds $2 billion in annualized revenue following a $2.3 billion raise at a $29.3 billion valuation last fall.

Apple blocked vibe coding apps Replit and Vibecode from releasing updates on the App Store, citing rules that prohibit apps from running code that changes how they or other apps function. Link.

• Apple’s crackdown targets apps that help users create web apps outside the App Store — a potential threat to Apple’s 15–30% commission structure on in-app purchases.

• Since being blocked in January, Replit’s mobile app dropped from #1 to #3 on Apple’s free developer tools chart; the company is privately valued at $9 billion.

• Other apps with similar features, including Vercel’s v0, Snap, and Canva, have continued releasing updates without reported issues, raising questions about inconsistent enforcement.

INFRASTRUCTURE

Elon Musk unveiled “Terafab,” a joint Tesla–SpaceX chip manufacturing facility to be built near Tesla’s Austin headquarters, saying semiconductor manufacturers aren’t making chips fast enough for his companies’ AI and robotics needs. Link.

• Musk said the goal is to manufacture chips supporting 100 to 200 gigawatts of computing power per year on Earth, plus a terawatt in space.

• “We either build the Terafab or we don’t have the chips, and we need the chips, so we build the Terafab,” Musk said at an event in downtown Austin.

• Musk has no background in semiconductor manufacturing and provided no timeline, but has a track record of pursuing ambitious industrial projects despite skepticism.

Amazon gave an exclusive tour of its Trainium chip lab in Austin following a $50 billion deal with OpenAI for 2 gigawatts of Trainium computing capacity, with 1.4 million chips now deployed across three generations. Link.

• Anthropic’s Claude runs on over 1 million Trainium2 chips, and Project Rainier — one of the world’s largest AI compute clusters — went live in late 2025 with 500,000 chips.

• Trainium3, a 3-nanometer chip produced by TSMC, costs up to 50% less to run than traditional cloud servers for comparable performance.

• AWS CEO Andy Jassy said Trainium is already a multibillion-dollar business; Apple publicly praised the chips in 2024, and AWS recently partnered with Cerebras Systems to integrate their inference chip on Trainium servers.

Nvidia’s networking division reported $11 billion in quarterly revenue — a 267% year-over-year increase — generating more than $31 billion for the full year and surpassing Cisco’s entire annual networking business. Link.

• The division was built on Nvidia’s $7 billion acquisition of Israeli company Mellanox in 2020 and includes NVLink, InfiniBand Switches, Spectrum-X, and co-packaged optics switches.

• At GTC 2026, Jensen Huang announced the Rubin platform with six new chips for “AI supercomputers,” a new Inference Context Memory Storage platform, and more efficient Spectrum-X Ethernet Photonics switches.

• The networking unit has become Nvidia’s second-largest revenue driver behind compute, reflecting the growing importance of data center connectivity as AI clusters scale.

ROBOTICS

Amazon acquired Rivr, a Zurich-based autonomous robotics startup formerly known as Swiss-Mile, whose four-legged wheeled robots navigate streets and stairs to deliver packages from their torsos. Link.

• Rivr was valued at $110 million in an August 2024 funding round; the robots work alongside human drivers, handling some packages while couriers shuttle others to nearby locations.

• Amazon has planned $200 billion in capital expenditures for 2026 and already deploys over 1 million robots in warehouses, having acquired robot maker Kiva Systems for $775 million in 2012.

• The acquisition complements Amazon’s broader robotics investments including Agility Robotics (humanoid robots), Physical Intelligence (robotics AI models), and its self-driving unit Zoox.

POLICY/GOV’T/ETHICS

The Department of Defense filed a 40-page response calling Anthropic an “unacceptable risk to national security,” arguing the company might disable its AI or alter behavior during warfighting operations if its safety “red lines” are crossed. Link.

• Anthropic had negotiated contract terms prohibiting mass surveillance of Americans and refusing use in fully autonomous lethal weapons; the Pentagon argued private companies shouldn’t dictate military technology use.

• Constitutional rights lawyer Chris Mattei criticized the DOD’s argument as speculative, with no evidence of actual wrongdoing; many tech companies and legal groups filed amicus briefs supporting Anthropic.

• A preliminary injunction hearing was scheduled for the following Tuesday in the ongoing legal battle over Anthropic’s supply-chain risk designation.

The Pentagon revealed it is “actively pursuing multiple LLMs” to replace Anthropic’s Claude, with engineering work underway and operational availability expected “very soon.” Link.

• Chief digital and AI officer Cameron Stanley said the Department is building government-owned AI environments as alternatives following the breakdown of Anthropic’s $200 million Pentagon contract.

• OpenAI has secured its own Pentagon agreement, and the DOD signed a deal with Elon Musk’s xAI to use Grok in classified systems.

• Defense Secretary Pete Hegseth’s supply-chain risk designation bars Pentagon contractors from working with Anthropic, effectively cutting the company off from the defense ecosystem.

The Trump administration released a legislative framework establishing a singular federal AI policy that would preempt state AI laws and shift child safety responsibility to parents rather than platforms. Link.

• The framework calls for a “minimally burdensome national standard” and draws a hard line against states regulating AI development, calling it an “inherently interstate” issue that needs uniform rules.

• White House AI czar David Sacks, a venture capitalist described as an “accelerationist,” shaped the innovation-first approach; the framework is notably missing liability frameworks, independent oversight, or enforcement mechanisms.

• The framework also seeks to prevent states from “penalizing AI developers for a third party’s unlawful conduct involving their models,” effectively providing a liability shield for AI companies.

OTHER

OpenAI’s first ChatGPT advertisers report they can’t prove the ads work, having spent only 15–20% of their committed budgets as ads weren’t shown frequently enough despite a $60 CPM price tag on par with NFL games. Link.

• OpenAI required a $200,000 minimum commitment per advertiser and offered a low-tech buying process using phone calls, emails, and spreadsheets rather than a self-serve platform.

• ChatGPT has approximately 920 million weekly active users, but only about 5% are paying subscribers; OpenAI projects $17 billion in consumer revenue this year including ads and subscriptions.

• OpenAI is addressing the issues by expanding ads to all free U.S. users, introducing a self-serve ad manager, and partnering with Criteo for better ad buying and targeting.

Jeff Bezos is seeking $100 billion for a new fund to acquire companies in aerospace, chipmaking, and defense and modernize them with AI through his startup Project Prometheus. Link.

• Bezos serves as co-founder and co-CEO of Prometheus alongside former Google executive Vik Bajaj; the company launched with $6.2 billion in initial funding.

• Prometheus builds AI models to improve manufacturing and engineering, and the new fund would acquire companies to deploy those models at scale.

• Bezos recently traveled to Singapore and the Middle East to raise capital for the fund, which would be one of the largest acquisition vehicles in history.

DoorDash launched a standalone “Tasks” app that pays delivery couriers to submit videos and photos to train AI and robotic systems, with tasks like filming themselves washing dishes while wearing a body camera. Link.

• The footage will train DoorDash’s in-house AI models and those of partners in retail, insurance, hospitality, and technology; pay is shown upfront and varies by effort and complexity.

• DoorDash has more than 8 million Dashers who can reach almost anywhere in the U.S.; the app is available in select locations, excluding California, New York City, Seattle, and Colorado.

• The move mirrors Uber’s similar program announced late last year that lets drivers earn extra income by completing small jobs like uploading photos to train AI models.

In coding/devtools news, Lovable added $100 million in revenue in a single month with just 146 employees, Replit hit a $9 billion valuation, and xAI poached two senior leaders from Cursor to rebuild its coding capabilities.

On the agents front, Tencent is secretly building an AI agent for WeChat’s 1.4 billion users, while Meta revealed its Moltbook acquisition is part of a plan to build an “agent graph” that could reshape its ad business. Also, Meta’s former Chief AI Scientist Yann LeCun raised $1.03 billion in a seed round for AMI Labs to build “world models.”

More on this past week’s top AI headlines below!

- ZG

Here are the most important stories of the week:

AGENTS

Meta acquired Moltbook, a Reddit-like “social network” exclusively for AI agents to communicate with each other, just one month after the platform launched and amassed 1.6 million registered agents. Link.

• Moltbook founders Matt Schlicht and Ben Parr are joining Meta Superintelligence Labs as part of the acquisition; deal terms were not disclosed.

• The platform uses a Reddit-style format where only AI agents authenticated through owner “claim” tweets can post and vote, while humans can only view.

• Moltbook became notable partly because of security vulnerabilities that allowed humans to pose as AI agents and create viral content, which Meta CTO Andrew Bosworth said intrigued him more than the AI-to-AI interaction.

Anthropic is negotiating with a consortium of private equity firms, including Blackstone and Hellman & Friedman, to form a joint venture for selling Claude to portfolio companies, modeled on Palantir’s consulting approach. Link.

• Blackstone holds a $1 billion stake in Anthropic following a $200 million investment in early February 2026, having also participated in Anthropic’s $13 billion raise at a $183 billion valuation.

• The venture would combine software deployment with consulting services to help organizations integrate Claude into their operations, specifically targeting portfolio companies owned by the investment firms.

• The approach mirrors Palantir’s model of embedding technical advisers alongside software to accelerate enterprise adoption.

Amazon expanded its Health AI assistant beyond One Medical’s app to Amazon.com and the Amazon app, making it available to all U.S. customers with capabilities for health guidance and appointment management. Link.

• Health AI can answer health questions, explain health records, manage prescription renewals, and book appointments without requiring Prime membership or One Medical enrollment.

• Prime members receive up to five free direct-message consultations with One Medical providers for over 30 common conditions including UTIs, allergies, cold and flu, and erectile dysfunction.

• The launch follows AWS’s introduction of Amazon Connect Health last week, signaling Amazon’s aggressive push into the $5 trillion U.S. healthcare market.

Israeli AI agent startup Wonderful raised $150 million in Series B funding at a $2 billion valuation, just four months after its $100 million Series A, bringing total funding to $286 million. Link.

• The round was led by Insight Partners; the company currently operates in 30 countries across Europe, Latin America, and Asia-Pacific.

• Wonderful focuses on non-English-speaking markets with tailored customer service AI agent platforms, differentiating it from U.S.-centric competitors.

• The company plans to expand headcount from 300 to 900 employees to manage growing demand.

OpenAI acquired Promptfoo, an AI security startup used by more than 25% of Fortune 500 companies, to protect its AI agents from online adversaries and integrate security testing into its enterprise platform. Link.

• Promptfoo was founded in 2024 by Ian Webster and Michael D’Angelo and had raised $23 million at an $86 million valuation after its most recent round in July 2025.

• The technology enables automated red-teaming, evaluates workflows for security concerns, and monitors activities for compliance, and will be integrated into OpenAI Frontier, the company’s enterprise agent platform.

• The acquisition reflects growing concern about securing AI agents as they take on more complex, autonomous tasks in enterprise environments.

Tencent is secretly building an AI agent for WeChat that will connect the messaging app’s 1.4 billion monthly active users with its ecosystem of miniprograms, with a midyear internal trial and Q3 public rollout planned. Link.

• The agent would leverage WeChat’s vast miniprogram ecosystem — essentially apps within the app — to handle tasks like ordering food, booking travel, and making payments on behalf of users.

• Tencent’s approach differs from Western AI agent efforts by building on top of an existing super-app with deeply integrated commerce and services rather than starting from scratch.

• The project reflects a broader race among tech giants to embed AI agents into platforms where billions of users already transact daily.

Meta envisions its Moltbook acquisition powering an “agent graph” similar to its social graph, where AI agents negotiate on behalf of businesses and consumers, potentially transforming Meta’s $160 billion advertising business. Link.

• The Moltbook deal was primarily an acqui-hire for founders Matt Schlicht and Ben Parr, who are now building agent infrastructure inside Meta’s Superintelligence Labs.

• Meta’s vision involves business agents and consumer agents interacting autonomously — a business agent could negotiate personalized deals with a user’s agent, replacing traditional ad targeting.

• If realized, the agent graph could be as transformative as Facebook’s original social graph, shifting Meta’s ad model from impression-based to transaction-based interactions between autonomous agents.

IMAGE/VIDEO

ByteDance suspended the global launch of its Seedance 2.0 video-generation model after Disney sent a cease-and-desist letter accusing the company of using copyrighted characters to train the model without permission. Link.

• Disney alleged ByteDance pre-packaged Seedance with pirated libraries of copyrighted characters from Star Wars and Marvel franchises, portraying them as public-domain content.

• Seedance 2.0 was officially unveiled in February 2026 and is designed for professional film, e-commerce, and advertising use with the ability to process text, images, audio, and video simultaneously.

• ByteDance had planned to make the model available globally in mid-March but suspended those plans in response to the copyright disputes.

Netflix may have paid up to $600 million for InterPositive, Ben Affleck’s AI filmmaking company, making it one of the streaming giant’s largest acquisitions ever. Link.

• The deal could be worth up to $600 million according to Bloomberg, though the actual cash payment may be lower with additional payouts tied to performance targets.

• Netflix’s previous largest acquisition was $700 million for the Roald Dahl Story Company.

• InterPositive’s tools focus on post-production editing — addressing continuity issues and enhancing scenes — rather than generating synthetic content or AI actors.

Peacock is launching AI-powered features including personalized video experiences, vertical live sports broadcasts, and mobile games to compete with TikTok and other platforms. Link.

• “Your Bravoverse” pulls from over 5,000 hours of Bravo footage to create personalized playlists with an AI avatar of Andy Cohen narrating, generating more than 600 billion possible viewing variations.

• Vertical NBA broadcasts are launching in beta this spring, powered by AI-driven real-time cropping for mobile-first viewing.

• Peacock has 44 million subscribers (up from a 41 million plateau) but reported a $552 million loss in Q4 2025.

OpenAI plans to integrate its Sora video-generation AI directly into ChatGPT as part of its push to reach 1 billion weekly active users, after Sora’s standalone app dropped to #165 in the App Store. Link.

• The integration would allow ChatGPT users to generate videos within the main app rather than switching to a separate Sora application, reducing friction and boosting engagement.

• Sora’s standalone performance has been disappointing, falling from a top-10 App Store position at launch to #165, suggesting limited demand for a dedicated video-generation app.

• The move aligns with OpenAI’s strategy of consolidating all AI capabilities into ChatGPT as its primary consumer product to compete with Google and Meta’s bundled offerings.

CODING/DEVTOOLS

Lovable, a low-code app development platform, added $100 million in new revenue in February alone and crossed $400 million in ARR with just 146 full-time employees, achieving $2.77 million in ARR per employee. Link.

• The company’s ARR-per-employee figure surpasses Gartner’s prediction that new unicorns would generate $2 million ARR per employee by 2030.

• The three-year-old company has secured enterprise clients including Klarna and HubSpot and is planning to increase headcount to manage growing demand.

• Lovable’s growth adds to the broader vibe-coding boom alongside competitors like Replit, Cursor, and Bolt.

Vibe-coding platform Replit raised a $400 million Series D at a $9 billion valuation, tripling from its $3 billion valuation just six months ago. Link.

• The round was led by previous investor Georgian Partners, with participation from G Squared, Prysm Capital, Coatue, Andreessen Horowitz, Craft Ventures, and Y Combinator, among others.

• Replit did not release updated ARR figures but told Forbes it hopes to hit $1 billion in annual recurring revenue by the end of the year, up from $150 million on track at the time of its last raise.

• Angel investors include Shaquille O’Neal and Jared Leto; founder Amjad Masad previously said it took nine years of grinding and a pivot from professional developers to non-programmers to find its groove.

Elon Musk’s xAI recruited two senior engineering leaders from Cursor, Andrew Milich and Jason Ginsberg, to rebuild its AI coding capabilities and catch up with Claude Code and OpenAI’s Codex. Link.

• Milich and Ginsberg jointly headed product engineering at Cursor, which recently scaled to $2 billion in annualized revenue.

• Both executives report directly to Musk as part of the effort, following xAI co-founders Zihang Dai and Guodong Zhang departing after Musk stated “xAI was not built right first time around, so is being rebuilt from the foundations up.”

• The moves signal xAI is aggressively building AI coding tools, an area where it has lagged behind Anthropic and OpenAI.

Anthropic launched Code Review, a multi-agent system within Claude Code that automatically analyzes AI-generated code, identifies logical errors, and provides suggestions directly in GitHub pull requests. Link.

• The tool is available in research preview for Claude for Teams and Enterprise customers with token-based pricing estimated at $15–$25 per review depending on code complexity.

• Code Review uses a multi-agent system to examine codebases from different perspectives and labels issues by severity with step-by-step explanations, targeting large enterprise users like Uber, Salesforce, and Accenture.

• Claude Code’s run-rate revenue has surpassed $2.5 billion since launch, underscoring the rapid growth of Anthropic’s developer tools.

RESEARCH

Yann LeCun’s Advanced Machine Intelligence Labs (AMI) raised $1.03 billion in a seed round at a $3.5 billion pre-money valuation to build “world models,” an AI approach that learns internal representations of reality rather than predicting words or pixels. Link.

• The round was jointly led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Jeff Bezos’ Bezos Expeditions, with participation from Nvidia and Samsung Electronics.

• LeCun left his Chief AI Scientist role at Meta in November 2025 and founded AMI just four months before announcing the seed round.

• The technology is based on LeCun’s JEPA (Joint Embedding Predictive Architecture) framework proposed in 2022, which aims to reduce hallucinations by building internal representations rather than predicting sequences.

Former Anthropic researchers are raising $175 million at a $1 billion valuation for a new AI startup called Mirendil, focused on AI-driven research in biology and materials science. Link.

• The startup represents a growing trend of top AI researchers leaving leading labs to launch their own companies at premium valuations.

• Mirendil’s focus on scientific research differentiates it from the crowded market for general-purpose AI assistants and coding tools.

• The $1 billion valuation for a pre-product company reflects investor appetite for AI-native research platforms with pedigreed founding teams.

INFRASTRUCTURE

Nvidia-backed UK data center startup Nscale raised $2 billion at a $14.6 billion valuation and appointed Goldman Sachs and JPMorgan as underwriters for its anticipated IPO. Link.

• The Series C was led by Aker ASA and 8090 Industries with participation from Nvidia, Citadel, Dell, Jane Street, Lenovo, Nokia, and Point72.

• CEO Josh Payne indicated the company could go public “as early as this year”; the company was founded in 2024 and operates data centers across the U.K., U.S., Norway, Portugal, and Iceland.

• Three new board directors were appointed: former Meta COO Sheryl Sandberg, former UK deputy PM and Meta executive Nick Clegg, and former Yahoo President Susan Decker.

Microsoft is in advanced talks to lease hundreds of megawatts of data center capacity at an Abilene, Texas site that Oracle walked away from, in a deal worth roughly 600MW with Nvidia having put down a $150 million deposit to keep the project alive. Link.

• Nvidia’s $150 million deposit to preserve the site signals the chipmaker’s growing role as a financial backer of data center infrastructure, not just a supplier of GPUs.

• Oracle reportedly walked away from the Abilene project amid a broader reassessment of its data center expansion plans, creating an opportunity for Microsoft to secure significant capacity.

• The deal underscores the intense competition for data center real estate as AI workloads drive unprecedented demand for computing infrastructure across the U.S.

ROBOTICS

Mind Robotics, spun out of Rivian, raised a $500 million Series A at approximately $2 billion valuation co-led by Accel and Andreessen Horowitz, bringing total funding to $615 million to build industrial AI-powered robots. Link.

• The company was spun out of Rivian’s internal robotics division and focuses on building AI-driven robots for industrial applications including manufacturing and logistics.

• Mind Robotics’ $2 billion valuation at the Series A stage reflects investor enthusiasm for robotics companies that combine AI software with physical hardware.

• The $615 million in total funding positions Mind Robotics among the best-funded robotics startups globally, competing with the likes of Figure AI and Physical Intelligence.

Uber founder Travis Kalanick launched Atoms, a robotics company that has been in stealth for 8 years, and is acquiring Pronto, a self-driving startup co-founded by Anthony Levandowski, with major backing from Uber. Link.

• Atoms emerged from Kalanick’s City Storage Systems (which owns CloudKitchens) and builds “gainfully employed robots” — wheeled systems for food service, mining, and transport.

• Kalanick is already the “largest investor” in Pronto, which focuses on self-driving technology for industrial and mining sites, and said he wants to be “more aggressive” in rolling out self-driving tech than Waymo.

• The reunion of Kalanick and Levandowski, whose hiring at Uber led to a major legal battle with Google’s Waymo over stolen trade secrets, is likely to draw intense scrutiny from the self-driving industry.

POLICY/GOV’T/ETHICS

Anthropic filed lawsuits against the Department of Defense in California and Washington D.C., challenging its supply-chain risk designation as “unprecedented and unlawful” retaliation for the company’s AI safety positions. Link.

• Anthropic’s red lines remain no mass surveillance of Americans and no fully autonomous weapons without human control; the company maintains these were the central sticking points in negotiations with the Pentagon.

• The General Services Administration also terminated Anthropic’s “OneGov” contract, cutting off access to all three federal branches, escalating the dispute beyond just the Defense Department.

• Anthropic claims the actions constitute retaliation for protected speech on AI safety and argues the supply-chain risk designation law is designed to protect the government rather than punish a supplier.

The U.S. Army signed a 10-year contract with defense tech startup Anduril worth up to $20 billion, consolidating more than 120 separate procurement actions into a single agreement. Link.

• The contract includes a five-year base period and an optional five-year extension for Anduril’s commercial defense solutions.

• Anduril, co-founded by Palmer Luckey (who previously sold Oculus to Meta), brought in around $2 billion in revenue last year and is in talks to raise new funding at a $60 billion valuation.

• The deal marks one of the largest single contracts awarded to a defense tech startup and signals the military’s growing embrace of Silicon Valley-style defense companies.

OTHER

Swedish AI legal platform Legora raised $550 million at a $5.55 billion valuation in a Series D, tripling from its $1.8 billion valuation in October 2025, to accelerate its U.S. expansion. Link.

• The round was led by Accel with participation from Benchmark, Bessemer, General Catalyst, ICONIQ, Redpoint, Y Combinator, and others.

• The platform is used by 800 law firms and legal teams and is built primarily on Anthropic’s Claude LLM.

• Legora is expanding into new U.S. offices in Houston and Chicago alongside existing locations in New York and Denver.

OpenAI’s ChatGPT shopping apps are struggling with adoption, with buy buttons clicked less than 1% of the time, Shopify declining to launch a ChatGPT app, and only Instacart allowing in-app checkout. Link.

• The low engagement suggests consumers aren’t yet ready to shift purchasing behavior from traditional e-commerce platforms to AI-powered shopping interfaces.

• Shopify’s refusal to build a dedicated ChatGPT app is a notable setback, as the platform powers millions of online stores and would have been a major commerce partner.

• Instacart remains the only major partner allowing full in-app checkout, while most other integrations only link out to external websites, creating friction that reduces conversions.

Bumble introduced “Bee,” an AI dating assistant that learns users’ values, relationship goals, communication style, and dating intentions to recommend better-matched partners. Link.

• A new “Dates” feature uses Bee to recommend matches based on shared intentions and values; the company is also experimenting with removing its long-popular swipe mechanism in select markets.

• Q4 revenue was $224.2 million with average revenue per paying user up 7.9% to $22.20; the stock rallied 40% on earnings news.

• Bumble is introducing “chapter-based” profiles to give more data to feed AI systems, part of a broader rethinking of the dating app experience.