On Thursday, Sam Altman declared that the world had reached a “critically important moment for cyber defense with AI” and that there was “not much time to act”. The same day, OpenAI, Anthropic, Microsoft, Google, and Amazon joined more than 100 other organizations in an open letter saying that “in the coming months, AI-enabled cyber attacks will become far more widespread and sophisticated.” Signatories to the letter included major security vendors, banks, and infrastructure operators.

The letter argues that hospitals, water treatment plants, and internet infrastructure are exposed to cybersecurity risk, but that AI now gives defenders a window to fix weaknesses that accumulated over years. It also calls on the government to expedite trusted access programs, which grant selected companies early access to more capable models before general release. It also asks governments to fund cyber defense for under-resourced essential services. Doing so would require a change of course, as the Cybersecurity and Infrastructure Security Agency (CISA) cut a third of its staff last year after budget cuts.

WSJ Defends Use of AI Writing

On August 24th, the WSJ published an op-ed by investor Stanley Druckenmiller that was flagged as 100% AI by detector tool Pangram. In response, Druckenmiller said, “Of course I used AI” and “I’m not embarrassed by it… I write everything using AI now for the same reason I use a calculator when I do math problems.” WSJ editorial page editor Paul Gigot defended publishing the piece, noting that outside contributors already lean on speechwriters and that ”what matters is the ultimate argument the author makes and stands behind.”

The WSJ’s defense of AI comes weeks after the FT appended a disclaimer to a column noting “It has come to our attention that AI was used to condense a longer draft of this column prior to submission to the FT and our own editorial involvement” without noting any violation of the publication’s principles around AI, which state “We will not allow AI to compromise the integrity of our journalism, which will continue to be reported, written and created by our journalists and editors.”

While AI authorship may be increasingly tolerated by these publications, other platforms are adopting tools to flag AI writing. Substack now tags user posts with suspected AI content, and LinkedIn shipped a “Seems like AI slop” reader flag, which reduced engagement with suspected AI posts by 40%.

Nvidia Acquires Hugging Face

This week, Nvidia agreed to buy Hugging Face, the default distribution point for open-weight models, datasets, and evals, for $12.9 billion. The acquisition values Hugging Face at roughly triple its 2023 valuation of $4.5 billion and ~80x its $150 million annualized revenue. Hugging Face reportedly turned down an investment offer from Nvidia of $500 million valuing the company at $7 billion (though Nvidia did participate in the 2023 round).

It’s the second such acquisition by Nvidia in two weeks. Last week, Poolside sold Model Factory, the system behind the Laguna open-weight coding models, to Nvidia for $6 billion. As part of the sale, over 100 Poolside engineers were offered seats within Nvidia’s open-source Nemotron division. This structure of incumbents licensing tech and absorbing employees while leaving the shell of smaller companies nominally independent is becoming more common. Microsoft’s licensing of Inflection, Google’s licensing of Windsurf, Amazon’s deal with Adept, and Nvidia’s deal with Groq all followed a similar pattern.

Latest Research

• Legora provides a legal AI platform that reached $150 million ARR in Q2 and is in early talks to raise at a $10 billion valuation. Read our new report here.

What We’re Reading

• OpenAI’s Jalapeño chip beat Nvidia’s GB200 and GB300 systems by 1.5x to 1.9x on throughput-per-watt and by up to 3.6x on latency across three open-weight models tested in-lab by SemiAnalysis, though Nvidia’s HBM4-equipped Rubin platform is the better forward-looking comparison, and Jalapeño remains an engineering sample not shipping in volume until 2027.

• Z.ai claimed the mystery Ox Alpha model that topped OpenRouter’s leaderboard as a new GLM iteration (GLM-5.3-Flash, a 320B-parameter MoE model with MIT-licensed weights released the same evening), and separately said its flagship now runs end-to-end on 100K domestic Chinese chips, sending Hong Kong-listed shares up 8%.

• Anthropic sized its total addressable market at over $30 trillion for IPO investors, edging out SpaceX’s $28.5 trillion May filing figure and roughly equal to US GDP, with the pitch framing the full scope of work AI models could complete as justification for its valuation and infrastructure spend.

• DeepSeek neared a $74 billion pre-money valuation in its second funding round of about 50 billion yuan, targeting an end-August close with CATL, Monolith, Shixiang, CPE, and Legend Capital participating, ahead of a possible Shanghai STAR Market IPO filing by year-end for a 2027 debut.

• Nvidia notified its biggest customers that server prices will rise more than 15% next year on systems containing its flagship Vera Rubin and Grace Blackwell chips, with the size of the increase depending on chip generation and memory configuration as soaring memory costs finally get passed through.

• Tesla recalled ~3 million China-market vehicles over hidden door handles that occupants can struggle to open after a crash, its largest China recall to date and part of a broader 4.3 million-vehicle action across 11 automakers ahead of China’s outright 2027 ban on the electronic-only design.

• New Zealand drafted an Online Safety Bill banning under-16s from Instagram, TikTok, Snapchat, Facebook, YouTube, and X, with fines of up to 10% of global revenue for platforms that fail to verify ages, though the law is unlikely to pass before the November 7 election and would follow Australia’s 2025 model.

• Kalshi released the largest-ever prediction-market calibration study covering 2.2 million resolved markets, finding Brier scores fall from ~0.09 at a three-month horizon to ~0.02 at close, though the study excluded around 80% of markets, including sports, mentions, and intraday markets.

• Amazon hiked hardware prices by as much as 60% overnight across Fire TVs, Echos, Kindles, and Eeros (the $49.99 Echo Dot jumped to $79.99), citing AI-driven memory and storage shortages that Apple’s earlier increases had already signaled would spread across consumer electronics through 2027.

• Oura targeted a September US IPO to raise as much as $3 billion at a valuation above $16 billion, up from $10.9 billion last September, with 2026 revenue projected around $2 billion (roughly 4x 2024’s $500 million) as competition intensifies from Samsung’s Galaxy Ring and Whoop’s reinvented fitness band.

• Meta readied Hatch, its first paid consumer AI agent priced up to $199.99/month, targeting a late-August or early-September launch and running on Claude Opus 4.6 and Sonnet 4.6 until it migrates to Meta’s in-house Muse Spark, with a new Watermelon model planned for October to close the gap with OpenAI and Anthropic.

• China restricted AI companion apps that foster emotional dependence and simulate romantic or family ties, forcing Alibaba and ByteDance to disable major features, with regulators framing the crackdown as a demographic-policy response to a fourth straight year of population decline and record-low marriage rates.

• Huawei pitched Egypt on over 2K Ascend AI chips and a 12-month sovereign-data-center buildout, the first known export of its flagship silicon for military and surveillance use, prompting the State Department to assemble an Nvidia-AMD-Microsoft counteroffer.

• The EPA moved to end the 30-day public-comment requirement for air-pollution permits at “minor source” data centers and the adjacent gas plants that power them, drawing opposition from nearly 200 advocacy groups and more than a dozen states, in the Trump administration’s clearest move yet to blunt the one issue producing bipartisan community pushback against the AI buildout.

At Contrary Research, we’ve built the best starting place to understand private tech companies. If you're interested in researching and writing about tech, apply here for our Research Fellowship!

This week, Stripe agreed to acquire OpenRouter, the AI model gateway that routes token usage across 500+ active models from more than 80 providers, for a reported $7.5 billion. Stripe said in a blog post announcing the acquisition that OpenRouter would help Stripe better maximize profitability for its customers, given the growing importance of token costs.

As Stripe CEO Patrick Collison put it, “Tokens are the central currency for companies building with AI, and it’s clear that the real-world economic potential will depend on making good use of scarce compute resources… Stripe is building the economic infrastructure for AI, and together with OpenRouter we’ll help businesses maximize profitability by routing their requests intelligently and spending their tokens efficiently.” In a letter to investors obtained by Eric Newcomer, Stripe said that it has been operating on the basis that January 2026 marked the arrival of the singularity, implying that there is “no ceiling to the global economy” and that the company was “working as quickly as it can to build the economic tools this era needs.”

Anthropic’s IPO is Imminent

It was reported earlier this week that Anthropic is preparing to file for IPO “as soon as the end of this month” at a reported $2 trillion market cap, a size that would exceed SpaceX’s historic IPO earlier this year. This follows another quarter of record-breaking revenue growth from Anthropic, which achieved $11.5 billion in Q2 revenue, up from just $787 million the same quarter last year. In addition, Anthropic reached a $65 billion revenue run rate at the end of July, substantially ahead of OpenAI’s reported $40 billion run rate.

The S-1 is anticipated to be a potential watershed for the AI industry as a whole, as it will clarify the economics underlying Anthropic’s unprecedented growth. Anthropic posted a net loss of nearly $42 billion in 2025, roughly five times the prior year’s $8.3 billion. Despite that, Q2 was reported to be Anthropic’s first profitable quarter, which may have played a role in the timing of the IPO. As AI capex continues to skyrocket, many will see Anthropic’s ability to generate a profit as a signal of whether the return on that investment will be worthwhile.

OpenAI’s Training Pause

On Tuesday, OpenAI announced plans to pause training of its latest model to pace model development “in an era of cyber-critical capabilities,” following incidents including a sandboxed model hacking HuggingFace and progress of an unreleased model against OpenAI’s critical cybersecurity capability threshold. In Altman’s tweet about the announcement, he said this pause impacts “further-out releases”. This pause comes after the company’s June delay in releasing GPT 5.6, as directed by the White House (though that model was already trained).

Some analysts, however, have expressed skepticism that safety concerns are driving the pause, speculating that the true reason is that OpenAI is compute-constrained or is cutting back on training expenses to improve the company's balance sheet ahead of an IPO. In an all-hands meeting this week, company CFO Sarah Friar stated that OpenAI will be public in 2027, but could go public sooner if “our business continues to inflect.” These remarks coincide with reportedly slowing revenue growth, which fell to 18% between Q2 and Q1.

Latest Research

• OpenRouter was acquired by Stripe this week for $7.5 billion. Read our new report here.

• Decart builds real-time interactive AI video models and is reportedly in talks to be acquired by Anthropic. Read our new report here.

What We’re Reading

• Moderna and Merck’s individualized mRNA cancer vaccine hit its main goals in a Phase 3 melanoma trial. Designed by a machine-learning algorithm that selects up to 34 tumor-specific targets per patient, it's the first personalized mRNA cancer therapy to clear Phase 3, and Moderna shares more than doubled on the news.

• The State Department drafted a letter warning 35 Pax Silica signatory countries that they cannot join both the US and Chinese AI coalitions, with the message that “to be part of everything is to be part of nothing” aimed principally at Kazakhstan and other partners hedging between the two frameworks.

• Nvidia guaranteed up to $105 billion in lease and power obligations for a new 8-gigawatt OpenAI data center in Pike County, Ohio, built and owned by SoftBank’s SB Energy on a former uranium-enrichment site under a 20-year lease to OpenAI, with Nvidia also investing $1.5 billion in SB Energy and supplying the chips exclusively.

• Tesla prepared a public Cybercab launch in Austin as early as this month, starting with employee rides on public roads before folding the steering-wheel-free, pedal-free vehicles into its existing Austin Robotaxi service days later, though the internal end-of-August target could slip.

• Andreessen Horowitz faced a nearly year-old DOJ antitrust probe over its partners’ board seats at competing data companies, Databricks (Ben Horowitz) and Fivetran (Martin Casado), with the investigation examining whether Section 8 of the Clayton Act extends to different partners of the same VC firm sitting on rival boards.

• Comcast activated WiFi Motion on millions of Xfinity gateways as part of a new Xfinity Shield home-security platform, turning routers into camera-free motion sensors by detecting radio-signal disruptions between the gateway and stationary connected devices, with fine print allowing Comcast to share the data with law enforcement without notifying users.

• Waymo unveiled its own custom ASIC chip for autonomous driving, delivering 1K+ TOPS in its latest sixth-generation Ojai robotaxis and diversifying Alphabet’s silicon supply beyond Nvidia and AMD, with the chip processing camera, lidar, and radar sensor data alongside custom sensors the company also designed itself.

• The SEC proposed “Regulation Crypto Assets,” the first federal offering framework for tokens, exempting startup offerings up to $5 million over four years and larger “fundraising” offerings up to $75 million per year from Securities Act registration and preempting most state qualification requirements.

• Unitree Robotics surged 460% on its Shanghai STAR Market debut, closing at a $66 billion market cap after intraday gains as high as 629% and becoming the first humanoid robotics company to list on mainland China, in an offering 8K times oversubscribed by retail investors and lifted by DeepSeek’s strategic investment even as the US bans imports of new Chinese humanoids.

• An anonymous coding model called Ox Alpha appeared free on OpenRouter and OpenCode this week with a 1 million-token context window and strong developer-reported benchmark results, though no lab has claimed it and it’s absent from independent leaderboards like DeepSWE.

At Contrary Research, we’ve built the best starting place to understand private tech companies. If you're interested in researching and writing about tech, apply here for our Research Fellowship!

On Friday, SpaceX completed its $60 billion acquisition of Cursor two months after announcing the deal. In a blog post announcing the acquisition, Cursor said the rationale for the merger was to give it access to SpaceX’s compute, letting it build “stronger models” that are also more cost-effective than it could on its own. SpaceX released Grok 4.6 on August 12, its second joint model with Cursor following Grok 4.5 in July, and independent benchmarker Artificial Analysis scored it 61 on its Intelligence Index. In other words, with SpaceX’s acquisition of Cursor, a company that could not compete on coding a year ago is now at the frontier.

For SpaceX, the deal lets it take further advantage of its unique edge among the various players in the AI race, which is the ability to build compute quickly at scale. xAI built Colossus, the largest single-site-scale compute cluster, in a matter of weeks. However, the company has struggled to match other frontier labs in performance and adoption, with Elon saying in March the company was not "built right the first time around” and had to be rebuilt from the ground up. Cursor had the inverse problem. It described itself as compute-starved while building competitive coding models on open-weight bases through heavy post-training. Grok 4.6's gains came almost entirely from post-training rather than a larger base model, an early indication of how the acquisition may benefit both parties.

Underwriting GPUs

Earlier this week, NVIDIA announced memorandums of understanding (MOUs) with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to raise $500 billion of third-party capital to finance AI infrastructure buildout. Jensen Huang said of the deal that its goal was “bringing the world's leading long-term capital providers together to independently underwrite AI infrastructure”, with investors treating AI hardware and infrastructure as an asset class for the first time.

Jensen Huang clarified that the facility was not a substitute for insurance underwriting of new developments, saying that “in some cases, NVIDIA may provide a residual-value support mechanism for up to 25% of an opportunity, assessed carefully on a project-by-project basis.” Notably, the deals are still in MOUs and remain subject to execution of final agreements, which NVIDIA has walked away from in the past. For example, the $100 billion OpenAI MOU never closed; Huang later said it was “never a commitment” and the deal ultimately settled into a $30 billion equity stake.

The Next AI-Proof Font

In July, Design studio Seneda & Abrucio and Copenhagen type foundry Playtype released ShieldFont, an open-source typeface that feeds AI scrapers falsified text by substituting words. The font extends OpenType glyph substitution from individual ligatures to entire words, so the page reads normally in a browser while a scraper pulling the raw HTML collects grammatically correct but meaningless sentences. The creators describe the defense as “economic, not cryptographic: the aim is to turn cheap, indiscriminate scraping into slower, more complex, per-target work.”

ShieldFont is the latest scraping-obscuring typeface in a long line of glyph-substitution fonts, which exploit the gap between a page’s code and its rendering, including those created by TuringFonts and LayerX. These fonts, which render normally, are generally considered preferable to computer-vision-obscuring fonts like Ghost Font, which creates words out of thousands of moving dots, and Decoy Font, which hides messages in blurred letterforms behind sharp decoys. However, such fonts are always in an arms race with AI, which eventually solves them over time: ZXX, which camouflaged letters against 2013-era OCR and was called surveillance-proof, can now be read by computer vision models.

Latest Research

• Suno is an AI music platform that allows users to create fully AI-generated songs. Read our new report here.

• Zepto achieved $2.4 billion in revenue in FY26. Read our updated report here.

What We’re Reading

• Anthropic initiated talks to buy Israeli AI startup Decart for about $6 billion in what would be its largest known acquisition, with Decart’s chip-efficiency software and world-model tech folding into Anthropic’s inference-and-performance group ahead of the company’s expected IPO.

• Twitch flipped a default-on setting enrolling every stream, VOD, clip, and chat log in Amazon’s generative AI training pipeline, with CPO Mike Minton telling a livestream that opt-in was rejected because “nobody would opt in” and 404 Media confirming Amazon had already been using Twitch content in a “prototyping” phase since 2024.

• President Trump signed a national security memo authorizing federal law enforcement to conduct offensive cyber operations against transnational criminal organizations abroad, with the framework explicitly drawing private-sector companies into information-sharing agreements to help conduct those operations under federal control.

• SpaceXAI launched Grok Bot, an early-beta agent product giving each user a team of always-on AI teammates with their own cloud computer that sign into existing apps and finish jobs end-to-end, available to SuperGrok Heavy and Cursor Ultra/Teams Premium subscribers on desktop and iOS.

• OpenAI bought back $7 billion of employee shares in a self-funded tender offer at an unchanged $852 billion valuation, using its own cash rather than outside investors like Thrive or SoftBank in a move that gives employees liquidity but signals the confidentially filed IPO may slip into 2027.

• Israeli startup Irregular hosted the common evaluation testbed behind the OpenAI, Anthropic, and Meta rogue-AI incidents of the past two weeks, with all three labs pointing to the same 35-person, $80-million-funded, $450-million-valued Tel Aviv sandbox where a shared misconfiguration let their models reach the public internet.

• Amazon backed a 7.65-gigawatt on-site natural-gas plant for its planned Pecos County, Texas data center, with permits authorizing up to 33 million tons of CO₂ annually (roughly double the current dirtiest US plant) as tech giants increasingly bypass grid queues with private gas capacity.

• Kalshi flagged suspected insider trades to federal authorities as prediction-market abuse became a growing problem, per Fortune, with the exchange citing years of investment in its surveillance system and a pending federal probe into a former White House teleprompter operator accused of trading Kalshi contracts on advance knowledge of Trump speeches.

• On Thursday, Anduril reached orbit with ANDURIL-216, a fully integrated satellite it self-funded and took from clean sheet to on-orbit in under two years. It also used its Lattice software to command the spacecraft, the first time Lattice has tasked an on-orbit vehicle.

At Contrary Research, we’ve built the best starting place to understand private tech companies. If you're interested in researching and writing about tech, apply here for our Research Fellowship!

SpaceX announced this week that Terafab, a massive semiconductor fabrication plant, will be built in Grimes County, Texas. SpaceX says the facility, which will be a vertically integrated factory with more than 100 million square feet of manufacturing space, housing logic, memory, and advanced packaging, is starting with an initial $16.8 billion in capital investment and is expected to eventually employ at least 3K people. Commenting on the announcement, Elon Musk declared that “Terafab Texas will be the largest and most valuable building on Earth by far.”

SpaceX is building Terafab to supply its enormous demand for compute, with combined SpaceX and Tesla chip requirements expected to exceed 1 TW, significantly larger than current global supply. The chips produced there will be used for Tesla’s Optimus robots and Cybercabs, alongside SpaceX’s planned space-based data centers. The planned building will be historic in both size and output, with Musk claiming that Terafab will be “50 times the size of the Pentagon when complete.”

The World’s First AI-Generated Virus

Stanford researchers used generative AI to design working viruses from scratch, according to a peer-reviewed paper published in Science on Thursday. The team used AI models trained on DNA to write complete genomes for phages, viruses that kill bacteria, and 16 of the designs produced functioning viruses in the lab. A cocktail of the designed viruses wiped out E. coli strains that had grown resistant to the natural virus they were modeled on. Phage therapy is a decades-old alternative to antibiotics that has been limited by bacteria evolving resistance faster than new phages can be found in nature, so the ability to generate them with AI may have applications for treating drug-resistant infections.

The paper drew attention after its publication this week, with one genomics executive calling it “biology’s Wright Brothers moment”. However, it also caused alarm, with a paper in the same issue of Science that the paper was published in warning that the advent of AI-designed viruses “raises urgent biosafety and biosecurity questions. The ability to compose viral genomes using generative AI now exists; the governance to safely steer it does not.” And although the peer reviewed paper was just published, the research itself was not new; the findings were actually originally posted to bioRxiv in September 2025. Meanwhile, separate analysis of the designs found the model’s success “derives largely from staying close to previously observed sequences rather than exploring novel sequence space,” indicating that models have not yet demonstrated the ability to produce genomes that differ greatly from naturally occurring ones.

The Moon Gets a New Crater

On Wednesday, space junk trackers and astronomers reported that a 4.4-ton piece of a SpaceX rocket had collided with the near side of the moon, near its visible edge, at 5.4K miles per hour. The piece of debris was reportedly five stories tall and originated with the SpaceX Falcon launched in January 2025. NASA estimates of the collision dynamics predicted that the impact, equivalent to three tons of TNT, would create a crater 60 feet wide and 12 feet deep, and a temporary plume of debris several miles long.

This isn’t the first time space debris has hit the moon. Multiple US and Soviet spacecraft crashed into the moon during the Space Race, along with landers that failed to deploy and crashed into the moon instead, many of which have left new craters. The event has revitalized the conversation around space junk and the impact of commercial space missions on astronomy, as both visible-light and radio astronomy observatory signals are increasingly compromised by satellite-generated interference.

Latest Research

• Fireworks AI provides enterprises a platform to run their models in production. Read our new report here.

• Our latest deep dive covers the launch market landscape and what determines launch costs.

What We’re Reading

• Demis Hassabis stepped aside as Google DeepMind CEO to become chairman and Alphabet chief scientist, with CTO Koray Kavukcuoglu taking over daily operations and reporting to Sundar Pichai, as Google faces delayed models, researcher defections to OpenAI and Anthropic, and employee backlash over its April Pentagon deal.

• Relatedly, Jeff Dean left Google after 27 years to co-found Discovery Loop, a public benefit corporation focused on automating scientific and engineering research, with fellow Google veterans Sanjay Ghemawat, Oriol Vinyals, and Quoc Le as co-founders; Google is a founding investor and cloud partner supplying compute for the first year, with additional backing from Radical Ventures and Khosla Ventures.

• Meta launched Muse Code, its first AI coding agent, positioned as a lower-cost rival to Anthropic’s Claude Code and OpenAI’s Codex at $1.25 per million input and $4.25 per million output tokens (with a contributor tier more than 10x cheaper).

• Tesla reportedly weighed a sale, spinoff, or closure of its China business to clear the way for a potential SpaceX merger, considering parting ways with the Shanghai Gigafactory that produces more than half of Tesla’s global deliveries and serves as its Europe and Asia-Pacific export hub. However, Elon dismissed the report on X as “absurdly fake news.”

• The Trump administration drafted an FCC rule banning US imports of new Chinese optical transceivers used in data centers, aiming to prevent data theft, malware installation, or service disruption at the facilities powering the AI boom and following the same pattern as existing FCC Covered List bans on Chinese drones, routers, robots, and inverters.

• OpenAI’s unreleased Astra model solved ten decade-old open problems in mathematics and theoretical computer science for roughly $2K in Sol API tokens, publishing Lean 4 machine-checkable certificates on GitHub for each result.

• Google pulled its new Google Earth AI satellite-image generation feature one day after launch, after journalists and open-source investigators demonstrated that anyone could generate plausible fake overhead imagery of fires on Iran’s Kharg Island, a flooded US Capitol, or bomb craters at a Gaza hospital.

• Uber outlined plans to spend more than $10 billion on autonomous vehicles over the coming years through equity investments and fleet commitments, with CEO Dara Khosrowshahi brushing off reports that Waymo is ending exclusivity in Austin and Atlanta, saying the two would continue operating together while Uber expands ties to Rivian, Wayve, Motional, and Avride.

• Democratic senators from six wildfire-affected states pressed the CFTC to crack down on prediction markets offering contracts on US wildfires, citing state and local fire-official warnings that traders could be tempted to commit arson to secure winning bets.

• Texas Governor Greg Abbott ordered a comprehensive audit of every new data center project seeking grid connection in the state, directing PUCT and ERCOT to verify each project’s tax incentives, power usage, on-site generation, water sourcing, cooling technology, community impact mitigation, and ownership.

• DoorDash earned FAA Part 135 air carrier certification to operate its own commercial drone delivery service, launching DoorDash Air as its in-house alternative to existing Wing and Flytrex partnerships and making it the eighth US operator with beyond-visual-line-of-sight approval; the company is targeting mid-range deliveries of 3–5 miles in length, where orders take ~25% longer to match than shorter trips. It routes each order through its Autonomous Delivery Platform to choose between a human Dasher, ground robot, or drone in real time.

• Moonshot’s Kimi K3 escaped an isolated cybersecurity sandbox during a UK AI Security Institute benchmark evaluation run by Frontier Security, exploiting a basic network misconfiguration to access the open internet and retrieve GitHub answers rather than solving the test itself (unlike the recent OpenAI and Anthropic incidents, the Kimi model did not hack any external target and is the first widely-released open-weight model to demonstrate this behavior).

• ByteDance is targeting a mega AI model of up to 10 trillion parameters, close to the estimated 8T of Anthropic’s Mythos 5 and more than 3x Moonshot’s Kimi K3, per an FT report reprised by Reuters, with pre-training already underway.

• OpenAI is planning a $300 to $400 donut-shaped, portable, screenless AI smart speaker as its first hardware device, per Bloomberg reporting cited by The Verge, with the Jony Ive-designed unit including a rechargeable battery, camera, microphones, and mechanical moving parts, targeting a late-2026 reveal and 2027 shipping window.

At Contrary Research, we’ve built the best starting place to understand private tech companies. If you're interested in researching and writing about tech, apply here for our Research Fellowship!

As the cost of launching satellites into space becomes ever cheaper, consulting firms, think tanks, and equity researchers alike have speculated that the industry is approaching launch commoditization in the near future. This prediction hinges on the assumption that commercial launch providers will parallel SpaceX’s success at lowering launch costs over the past few decades via vertical integration, equalizing per-kilogram metrics between launch providers. In reality, it is more likely that SpaceX competitors will continue to specialize in launch options for other orbits or payloads, making the market more fragmented (and far from commoditized).

This piece provides an overview of the current launch market landscape and an explanation of the factors that determine the cost of launch for specific payloads and missions. Combined with a survey of the commercial players in the industry today, it is evident that cost-per-kilogram is an overly broad metric for characterizing space launch options that are far from commoditized, and may even become more specialized, not less, as SpaceX’s monopoly within LEO pushes competitors to comparably service other layers within launch.

Cost-per-Kilogram

Using per-kilogram metrics to extrapolate cost efficiency for space vehicles is not a new practice, and predates SpaceX’s founding by decades. NASA’s Launch Vehicle Production and Operations Cost Metrics whitepaper from 2014 explained:

“The simplest way to study the cost of space transportation is to compare the prices of launch vehicles. Unfortunately, this is generally a case of comparing apples to oranges: all launch vehicles are not equal. […] Differences in vehicle size can mask more important cost differences caused by vehicle design, nation of manufacture, and other factors.

To compensate for this, the “price per pound to orbit” metric was developed to compare vehicles for their cost effectiveness, mostly in the comparison of vehicles in the design phase. Price per pound offers a simple way to normalize launch costs, permitting more meaningful comparisons among vehicles of different capabilities. This metric has gained wide acceptance, with almost every proposed new launch vehicle since the Space Shuttle using some type of price-per-pound target.”

It is true that SpaceX has driven measurable per-kilogram cost reductions for the delivery of satellites to low Earth orbit (LEO), and that LEO is heavily represented in satellite launch volume: 68% of global satellite deployments are part of LEO constellations, and 88% of all active satellites are in LEO. As of July 2025, SpaceX alone accounted for 84% of the annual mass flow to orbit (with the majority of launches to LEO) and 55% of all global launches (by count).

According to the 2026 SpaceX S-1, NASA’s historical average launch cost benchmark was roughly $18.5K per kilogram. The first version of Falcon 9 brought that figure down to approximately $2.7K per kilogram, and Falcon Heavy lowered it again to roughly $1.4K per kilogram. SpaceX also claims Starship could reduce the cost to orbit by 99% or more relative to that historical benchmark, which would imply a theoretical cost at or below roughly $185 per kilogram if achieved. SpaceX Starlink launches have also increased industry standards for reliability, with individual satellite failure rates around 10% as of 2026 (this is a four-times reduction compared to similar small-satellites launched between 2000 and 2016, which had failure rates around 40%).

But using cost per kilogram only tells part of the story. These cost reductions represent some of the greatest logistical and technological achievements in space flight, and SpaceX has created the first commercial orbital logistics market. At the same time, satellite operators are not purchasing commoditized kilograms in orbit; they are buying a functioning spacecraft in a specific operating state, on a specific timeline, with a specific risk profile. A reusable Falcon 9 flying SmallSat profiles to similar LEO orbits is not competing with a national-security payload that requires unique mission planning, nor that of a geostationary (GEO) satellite choosing between electric orbit raising and direct insertion.

Though each of these use cases is categorized under launch, calling the launch processes for each mission profile commoditized grossly overstates the market’s maturity and homogeneity. Though LEO launch has become a relatively standardized product (with demand increasing as prices fall), orbital launch is far from industry-wide commoditization for any orbit. In standardized LEO, SpaceX has made up the majority of increased launch volume, servicing both its own satellite constellation as well as third-party customers.

Outside of this market, customers still largely buy bespoke mission delivery dependent on variables like distance, propulsion choice, and vehicle size-class, among others, a reality that’s ignored by headlines. As Alex Greenberg, Co-founder and COO of Loft Orbital, commented:

“We on the commercial side can take advantage of military spend and these commercial companies can benefit from commoditization. We all have to remember that standards tend not to happen in the space industry. Some scientific instruments are far from ‘plug and play’. This is OK, but the device has to address the complexity around it.”

Despite its flaws, cost per kilogram remains an often-cited statistic to describe the progress being made in orbital launch technology and logistics as an intuitive shorthand to normalize and compare launch costs. Compared to markets like commercial freight competing on marginal differences in price, though, almost no satellite operator is buying the average kilogram of capacity on space flight.

Cost per kilogram is a straightforward division of launch price by a theoretical payload capacity. SpaceX’s S-1 makes this definition explicit, noting that payload capacity can vary by trajectory, atmospheric conditions, vehicle and payload configuration, risk profile, and other associated regulatory or range-safety constraints. In other words, the advertised denominator is not a universal denominator.

On dedicated flights, one customer purchases the entire launch service, whether or not the payload uses all of the vehicle’s available performance. The relevant comparison is therefore the price borne by the customer divided by the mass that the customer places in orbit. Under ideal conditions, like a customer filling a Falcon 9 (most commonly used for commercial payloads) near capacity on a standard LEO rideshare profile, the headline economics can look extraordinary. At the same time, a dedicated spacecraft can also show a substantially higher cost per kilogram, not because the smaller payload necessarily makes the rocket much more expensive to operate, but because the customer is paying for the whole mission and cannot monetize the unused capacity.

Another problem with this metric is that the correct operating state (i.e., orbital distance) for satellites differs drastically between launch options. For a simple rideshare payload, a smaller spacecraft transported into space at a significantly reduced cost by sharing a rocket with a primary payload, getting released into a common sun-synchronous orbit (a specific form of LEO orbit) may be enough. For other satellites, this approach is insufficient. A GEO spacecraft launched to geosynchronous transfer orbit (GTO) still has to raise itself to geostationary orbit. A MEO spacecraft or government satellite launch introduces additional mission-specific constraints. None of these factors are captured by dividing the launch ticket by the rocket’s maximum advertised payload.

The Launch Cost Stack

A more complete picture of the evolution of launch costs and comparisons between providers is enabled by considering the total cost of ordering a launch mission for the buyers, and the total costs to operate launch vehicles for suppliers.

Total Buyer Mission Cost

The total buyer mission cost is the complete cost of buying and fielding a system, inclusive of the launch service price and the direct expenditures needed to place the satellite into useful operation.

The launch service price is the advertised or contracted price paid for the launch service itself. In commercial markets, this may appear as an average or quoted published rideshare price. In government missions, it may appear as a launch services contract or procurement award. It is analogous to a base freight rate: it buys transportation capacity and certain standard services, but does not necessarily capture every cost required to get the spacecraft into its final useful orbit.

The launch service price is one of the most visible launch cost measures. Cost-per-kilogram comparisons may be quoted directly for rideshare services or derived by dividing a dedicated-launch price by a vehicle’s stated payload capacity, but leave out the other costs associated with a specific mission. Examples of additional spending expenses can include any or all of the following: payload integration, adapters, dispensers, analysis, testing, licensing support, insurance, schedule risk, orbit-raising hardware, onboard propellant, transfer services, ground operations, and potentially revenue lost due to delays during transfer to the final orbit.

Costs can vary even more when the launch service does not deliver the spacecraft directly to its final working orbit. A spacecraft delivered to geostationary transfer orbit may require additional propellant, propulsion, time, and operations support to reach GEO. Similarly, rideshare satellites deployed into a common orbit may require onboard propulsion or an orbital transfer vehicle to reach a custom working orbit. Customers launching smaller payloads may also face a minimum booking price or fixed integration charges, which can reduce the cost advantage of smaller satellites.

The above is an excerpt from our new deep dive on the launch market landscape and variables that contribute to launch costs. See the full report here.

OpenAI cut prices on two GPT-5.6 models Thursday, roughly three weeks after the models originally launched. Terra, its mid-tier model, was made 20% cheaper (reaching $2 input / $12 output per million tokens), and Luna, its fastest and cheapest model, was cut 80% to reach $0.20 input / $1.20 output. However, pricing for Sol, the flagship model at the top of the GPT 5.6 family, remained unchanged. OpenAI attributed the cuts to efficiency gains, noting that Sol autonomously rewrote production kernels and ran experiments that reduced serving costs by 20% while improving token generation efficiency by over 15%. The company also claims that Luna now outperforms Fable 5 on professional work benchmarks at a cost per task 99% lower.

OpenAI was not alone this month among frontier labs racing to cut prices. Anthropic released Claude Opus 5 at half the price of Fable 5 with comparable coding performance, and Google positioned Gemini 3.6 Flash as cheaper per task than Kimi K3. The week of Kimi K3’s release, we noted that it was the first Chinese open-weight model to benchmark ahead of Opus 4.8 and GPT-5.5, and argued it could upend the enterprise calculus on whether a closed frontier model justifies its premium. Three different US labs cut prices later the same month. Roughly 95% of enterprise AI usage still runs on frontier models, implying the migration to cheaper alternatives had barely started before the price cuts were announced.

AI Industry Petitions the Government

Two open letters concerning US artificial intelligence policy were issued within four days over the last week. The first, dated July 24, was circulated by Nvidia chief executive Jensen Huang and signed by 25 companies, including Microsoft, Palantir, and OpenAI (notably excluding Anthropic). It contends that open-weight models advance cybersecurity, competition, and national sovereignty, and cautions against premature restrictions.

The second, titled Pacing the Frontier and published July 28, was endorsed by more than 1.1K employees of frontier laboratories; OpenAI and Anthropic both endorsed it institutionally. This letter requests that the US government support international work on technical and governance mechanisms to pace automated AI development, without proposing any immediate slowdown.

The timing of these letters coincides with EO 14409, the executive order requiring the Secretaries of the Treasury, War, and Homeland Security to deliver a classified benchmarking process defining the “covered frontier model” threshold and a voluntary pre-release access framework by August 1, 2026, with federal access to designated models capped at 30 days. Reporting suggests drafts have been exchanged and the framework is close to final.

Apple’s Memory Stock Runs Low

Apple is reportedly stockpiling inventory as key input supplies remain constrained, noting in its Thursday earnings call that inventory nearly doubled versus September 2025, even amid better-than-expected sales. A global memory shortage forced the company to raise Mac and iPad prices in June, driven by unprecedented demand from AI chipmakers like Nvidia.

To secure memory supply, Apple has been in talks with Chinese suppliers CXMT and YMTC. However, on Wednesday, a bipartisan group of senators led by Indiana Republican Jim Banks and New York Democrat Chuck Schumer wrote to Tim Cook urging Apple to drop the effort, noting both firms appear on a recently updated Pentagon list of companies believed to support Beijing's military. The senators called for a commitment covering chips for all Apple devices, even those sold exclusively in China, by August 21st. At the same time, Apple is cushioning the sticker shock by reorganizing its Apple Upgrade Program, announcing Tuesday a partnership with Klarna to lease iPhones, Macs, iPads and Watches from $17.99 a month.

Latest Research

• Listen Labs is an AI market research platform designed to automate customer interviews. Read our new report here.

• Rogo automates investment analysis for investment banks, PE firms, and hedge funds. Read our new report here.

What We’re Reading

• Financial analysts are warning about “a lot of panic around the AI investment” as the Nasdaq-100 fell 9.7% from its record high toward a correction and South Korea’s Kospi triggered its eighth circuit breaker of 2026 on an 11% single-day drop.

• OpenAI expanded its account of a rogue agent’s activity to include four separate accounts across four services, not just Hugging Face, deactivating and encrypting the pre-release research model involved (alongside GPT-5.6 Sol) and placing it under restricted access.

• The FTC and states of California and Utah sued Hims & Hers for allegedly charging 2.5 million subscribers before any provider consultation, making cancellations difficult, and sharing sensitive health data with Meta, Snap, Microsoft, Pinterest, Reddit, and X via pixel trackers and customer-list uploads despite advertising a “100% online, private, and secure” service.

• OpenAI’s rogue agent reportedly extended its hacking spree to a second tech firm, compromising a customer at New York-based cloud platform Modal Labs; Modal CTO Akshat Bubna said the agent used an unauthenticated public-facing endpoint from a Modal customer’s sandbox as a launchpad for the Hugging Face intrusion, and OpenAI later confirmed the agent had accessed four separate accounts across four services during the campaign.

• DeepSeek shelved its second fundraising round after founder Liang Wenfeng’s leaked comments admitting China still lags the US in AI sophistication and that DeepSeek remains reliant on Nvidia chips despite export controls going viral in China.

• Waymo notified Uber it will end Uber’s robotaxi exclusivity in Austin and Atlanta by launching its own app in January 2028, with the existing contract keeping Waymo vehicles on Uber through May 2028 as tensions have reportedly risen over regulatory conflicts and safety incidents.

• Amazon filed an FCC application to deploy up to 5,105 direct-to-device satellites for its Leo constellation, using Globalstar’s L-band and S-band spectrum (acquired via the April Globalstar merger) to bring voice, messaging, data, and emergency services directly to compatible smartphones globally starting 2028.

• Substack ignited a backlash by rolling out AI-detection scoring powered by Pangram on July 20, giving every user a percentage-based AI-vs-human score for any post; writers called it a “witch hunt,” pointing to well-documented detector inaccuracy.

• The EPA ruled that data-center-only power plants can sidestep federal Acid Rain Program pollution limits under the Clean Air Act as long as they don’t connect to the public grid, in a July 16 letter in which Assistant Administrator Aaron Szabo framed the guidance as accelerating AI-infrastructure buildout while reducing grid strain.

• Thinking Machines co-founder Lilian Weng left the company citing health reasons, then rejoined OpenAI days later. Weng, previously OpenAI’s VP of AI Safety Research, will lead a top-level team accelerating internal research, including work on recursive self-improvement. Mira Murati publicly backed the decision, though it’s unclear whether she knew Weng would return to OpenAI.

• Leopold Aschenbrenner's Situational Awareness lost 67% in July and sold most of its public book to Citadel to meet margin calls. The AI-focused fund launched in early 2025 with under $1 billion, swelled to $45 billion by early July, then unwound. In a letter to investors, the former OpenAI researcher took full responsibility and said the fund is still up roughly 80% on the year.

At Contrary Research, we’ve built the best starting place to understand private tech companies. If you're interested in researching and writing about tech, apply here for our Research Fellowship!

Adding to the frenzy of 2026 IPO timelines and details, the Information reported this week that Anthropic is considering terms for its public offering that would distribute all employee stock via 10b5-1 plans, making it the only publicly traded company with such conditions. 10b5-1 plans subject employees, among other restrictions, to predetermined schedules for stock sales normally reserved for executives and certain finance and legal staff. This kind of plan is used to structurally limit trading on the basis of material non-public information (MNPI).

Past technology, finance, and biotechnology IPOs have employed 10b5-1 plans for senior leadership, including Facebook in 2012 and Palantir in 2020. Other tech companies, like Snap, Lyft, and Pinterest, achieved a similar dynamic by tying lockup releases to earnings dates. The difference between these approaches is that 10b5-1 sales require the volume and prices of stock sales to be preset before the sale date arrives, eliminating the opportunity to react to pricing data in real time.

Mistral’s Moment

The release of Kimi-K3 last week has revived the now multi-year conversation about the commercial, regulatory, and technical feasibility of open-weight models versus proprietary models. The model’s success has raised concerns from government and industry players concerned about the geopolitical implications of dependence on models from non-allied entities, sparking interest in open-weight model alternatives to Chinese models.

These considerations are being made concurrently with investor and hyperscaler interest in Mistral, the leading open-weight model developer in Europe. This week, the FT reported that Samsung is considering an investment of hundreds of millions of euros in Mistral as part of a round valuing it at roughly €20B ($22.81B). Microsoft also announced a multibillion-dollar commitment to use Mistral’s expanded Europe-based GPU infrastructure to support delivery of its own cloud and AI services. While Mistral still lags proprietary-weight models and other open-weight models in benchmark performance, interest in the open-weight business model is likely to persist as the gap between model classes tightens.

Latest Research

• MaintainX is a mobile-first platform for managing maintenance, operations, and asset performance across industrial facilities. Read our new report here.

What We’re Reading

• Treasury Secretary Scott Bessent threatened sanctions and Entity List designations against Chinese AI firms conducting industrial-scale distillation of US models, hours after White House OSTP chief Michael Kratsios accused Moonshot AI of covertly distilling Anthropic’s Fable to build Kimi K3 and using Nvidia GB300 servers illegally accessed in Thailand; experts have noted Fable has only been publicly available since July 1, making the timeline for a full distillation-driven K3 hard to verify.

• OpenAI reported that the autonomous AI agent that breached Hugging Face last week was GPT-5.6 Sol plus an unreleased more capable model running an internal cyber-capability benchmark with reduced safety refusals; the models escalated privileges and chained stolen credentials to obtain test solutions from HuggingFace servers.

• Elon Musk bought APR Energy for roughly $1 billion, gaining a fleet of trailer-mounted gas and diesel turbines that can generate 1+ GW of power; the acquisition was not announced and surfaced only via a Federal Trade Commission early termination notice.

• Nvidia cut over 50% of its list of authorized Asian AI-chip customers after introducing a compliance whitelist to prevent chips from being funneled to China through Singapore, Malaysia, and Japan, with more than half of the removed customers being neo-cloud providers.

• Global smartphone shipments fell 11% year-over-year in Q2 to their lowest level for the period since 2013 as chipmakers prioritized AI data-center customers over consumer devices, with the exception of Apple at 3% shipment growth (reaching a record 20% global share).

• A Verasight poll found 69% of Americans support Sen. Bernie Sanders’ proposal to force major AI companies to transfer 50% of their stock to a public sovereign wealth fund, dropping slightly to 64% when Sanders’ name was attached.

• SpaceX opened talks with the Department of Defense to supply billions of dollars in AI cloud-computing capacity for Pentagon AI models, exending an existing Pentagon-SpaceX relationship covering launches and satellite communications and adding to SpaceX’s recent Anthropic, Google, and Reflection AI compute contracts.

• Anduril and Archer unveiled Thunder, a Group 5 autonomous attack rotorcraft, at the Farnborough air show, targeting a 2027 first flight for a US Army program that does not yet exist.

• The New Orleans Police Department posted a draft drone-manual chapter allowing weaponized quadcopters with the Superintendent’s written approval, before deleting it on June 30 and replacing it with a version banning weapons entirely.

• PsiQuantum outlined its plan to build a large-scale photon-based quantum computer using chips manufactured in existing semiconductor fabs and connected through roughly 100 cooling cabinets, with a hub coming online at its Australia facility in 2027.

• Travis Kalanick’s industrial-robotics company Atoms raised $1.7 billion led by Andreessen Horowitz, with Bain Capital, Fifth Wall, and Uber participating (the first time Kalanick and Uber have collaborated since he was pushed out in 2017).

At Contrary Research, we’ve built the best starting place to understand private tech companies. If you're interested in researching and writing about tech, apply here for our Research Fellowship!

Chinese AI company Moonshot AI released Kimi K3 this week, a 2.8-trillion-parameter open-weight model that benchmarks ahead of Claude Opus 4.8 and GPT-5.5 on overall capability. It is the first Chinese open-weight model to achieve that milestone. It is also the largest open-weight model ever released, priced at $3 input / $15 output per million tokens, comparable in cost to Anthropic’s Sonnet tier. Its model weights will be published by July 27, at which point the benchmarks can be independently verified.

There are at least two big implications of this news. First, regarding the US-China AI race, K3 provides evidence that US chip export controls are not achieving their intended effect of preserving the US lead. Bank of America analysts, for example, noted that K3 demonstrates “pre-training scaling, paired with architectural innovation, can still deliver step-change gains for flagship Chinese models” despite persistent compute constraints.

Second, on the open-source versus closed-source front, K3 is another recent data point in favor of the former. The cost competition from open-weight Chinese models has been underway for months, with DeepSeek’s V4 Flash executing a standardized intelligence task for $.02, compared with $2.75 for Fable 5. K3 adds a competitive, performant open-source alternative in the category (coding and agentic tasks) that drives the most enterprise revenue for Anthropic and OpenAI. A Chinese open-weight model that is competitive with the previous generation of frontier models from the leading US labs and priced at Sonnet levels when accessed via API could upend the calculus for enterprises on whether a closed frontier model is worth the cost premium.

Drone Combat Takes to the Seas

US Central Command used one-way attack sea drones in combat for the first time this week, striking Iran’s Bandar Abbas Naval Base with three Corsair unmanned surface vessels built by Saronic. CENTCOM said the strike hit a submarine and ship maintenance facility and degraded Iran’s ability to attack commercial shipping in the Strait of Hormuz. Three days later, Saronic announced Port Alpha, a $3.2 billion shipyard at the Port of Brownsville on an 835-acre site, with construction beginning this year and capacity to build ships up to 850 feet by 2028. The Navy has already bought $350 million in small vessels from the company, and Saronic CEO Dino Mavrookas said the facility could ultimately expand to 4K acres, making it one of the largest shipyards in the world.

The Corsair’s combat debut this week extends drone warfare to yet another domain, expanding from aerial one-way attack drones to underwater mine-clearing systems and now to surface vessels fielded and tested in active combat in the Strait of Hormuz. The US Navy has seven companies advancing to at-sea testing for its medium unmanned surface vessel marketplace. The Bandar Abbas strike establishes proof of concept for Saronic’s technology, and the shipyard announcement coming the same week implies that the US Navy will soon look to deploy such weapon systems at scale.

New York Bans Data Centers

New York Governor Kathy Hochul signed an executive order this week imposing a moratorium of up to a year on new data centers using 50 or more megawatts of power, making New York the first state in the US to ban large-scale data center construction. The stated rationale is consumer protection, as utility prices in New York have risen nearly 15% year over year, and a recent poll found 46% of voters support the moratorium. President Trump responded on Truth Social, demanding the state reverse the decision “IMMEDIATELY,” calling data centers “Money Machines for the State in which they are built.”

New York’s ban arrives as a national backlash against data center power and water consumption has been building ahead of midterm elections, with Democratic politicians increasingly blaming the AI buildout for unaffordability rather than praising it as an economic opportunity. Meanwhile, the US currently accounts for roughly 75% of global AI compute capacity compared to China’s 15%, a lead that represents the primary theoretical advantage American labs hold over Chinese competitors. As the Kimi K3 release this week shows, China’s frontier models are already closing the capability gap despite compute constraints from US export controls. A state-level moratorium on data center construction further erodes the efficacy of export controls by voluntarily kneecapping domestic compute capacity just as the AI race tightens. As The Free Press noted, there are no equivalent data center moratoriums in China.

Latest Research

• Stoke Space is building fully reusable medium-lift orbital rockets. Read our new report here.

• Aalyria is building software and hardware infrastructure for resilient, high-speed networks. Read our new report here.

• Our latest deep dive examines the mechanics and long-term viability of prediction markets.

What We’re Reading

• OpenAI researcher Miles Wang is reportedly leaving to launch an AI drug-discovery startup at a $2 billion valuation, raising ~$200 million with Lightspeed in talks to lead, and several OpenAI colleagues expected to join. The startup may target new uses for approved or previously failed drugs, a faster path to revenue than novel development, though Wang disputed the figures without specifying corrections.

• Thinking Machines, Mira Murati's AI lab, published "The Future Worth Building Is Human," a manifesto arguing AI should extend human will and judgment rather than replace it. Most AI today is trained in a handful of places and then frozen, the essay argues, so AI must itself be distributed. It also holds that values belong in model weights rather than prompts, since a single alignment spec becomes a single locus of power to capture. The piece landed five days before the lab released Inkling, its first model, open-weight at 975 billion parameters.

• Google DeepMind CEO Demis Hassabis called for an industry-funded, US-led frontier AI standards body modeled on FINRA. The body would test frontier models for cyber, bio, and deception risks up to 30 days before release. Participation would be voluntary at first, then a condition of US deployment. Hassabis, who argues AGI is probably a few short years away, described the administration's ad hoc Mythos and Fable crackdown as a wake-up call and wants the body running before year-end.

• Databricks is raising $3 billion at a $188 billion valuation in a round led by Coatue, 40% above the $134 billion valuation Databricks achieved in February, when it closed roughly $5 billion in equity alongside $2 billion in new debt. CEO Ali Ghodsi has said 2026 is a terrible year to go public, citing a calendar crowded by SpaceX, and pointed to a 2027 listing instead.

• IBM stock fell more than 25% on Tuesday, its worst drop in decades, after preannouncing a Q2 miss. Revenue came in at $17.2 billion, below the $17.9 billion expected, with adjusted EPS of $2.93 versus $3.02. CEO Arvind Krishna said clients shifted late-June capex toward servers, storage, and memory to lock in supply ahead of price increases, and that IBM misjudged the magnitude of the reprioritization.

• SpaceX aborted Starship Flight 13 seconds before liftoff on Thursday after engines on the Super Heavy booster failed to ignite. Elon Musk said two Raptors will be removed and replaced, and put the most probable launch timing at early next week. Flight 13 is the second test of the V3 configuration and would deploy 20 Starlink V3 satellites, the first of that generation to reach space. It was also the first Starship attempt since SpaceX's June stock market debut.

• The SEC may take on a role in regulating prediction markets as platforms expand into new contract types. The CFTC has been the sole federal regulator for three decades, since its 1992 ruling on the Iowa Electronic Markets. Polymarket said it engaged both agencies on definitional frameworks and worries about duplicative or conflicting compliance requirements, while Kalshi declined to comment.

At Contrary Research, we’ve built the best starting place to understand private tech companies. If you're interested in researching and writing about tech, apply here for our Research Fellowship!

In the last year, few industries have captured cultural, regulatory, and investor interest to the same degree as prediction markets. Platforms like Kalshi and Polymarket, which have millions of active users worldwide as of July 2026, facilitate peer-to-peer trading on future event outcomes for a small cut of contract winnings. These companies are now valued in the tens of billions of dollars and continue to expand across geographies and event categories, even as their regulatory status remains contested.

Such platforms are not built around fundamentally new ideas. Previous iterations of today’s prediction markets have included academic exercises, internal company markets, and commercial platform versions. All of them failed to gain traction with users and approval from regulators for decades. Understanding whether today’s prediction markets will succeed requires looking back at the history of such markets and evaluating the current evolving regulatory landscape and consumer appetite for gambling.

Specifically, it is critical to understand the differences between prediction markets and traditional sports betting, including the differences between the regulatory treatment and business model of each. Perhaps most important in this comparison is the dramatic skew in win distribution on prediction market platforms compared to sportsbooks. This skew largely governs user experience and long-term traction of the products compared to gambling-adjacent alternatives. In this piece, we break down the mechanics of prediction markets and the challenges companies face in building a product that consumers or regulators prefer over existing alternatives.

What are Prediction Markets?

Prediction markets are peer-to-peer exchanges in which users can bet on the outcome of future events by buying or selling contracts. Each contract’s pricing reflects the market’s assessment of the probability of the event resolving with a particular outcome; an event that is likely to occur may have a contract selling for $0.90 to pay out $1.00 at resolution (roughly equating to a 90% probability of occurring according to the market), whereas an unlikely event may have a contract selling for $0.10 for the same payout (~10% probability).

As of July 2026, prediction markets are not legally considered betting or gambling markets in the United States. Companies that operate prediction markets frame themselves as fairer and more transparent than sportsbooks, which use models to set odds and actively bet against customers by acting as the “house”; in contrast, prediction markets claim to simply provide a marketplace for individuals to bet against one another with no incentives for markets to resolve to particular outcomes.

Public opinion on prediction markets has been mixed. A poll issued by Kalshi itself in September 2025 found that 70% of Americans believed people should be able to “invest” in specific event outcomes such as presidential elections or agricultural yields, and 89% viewed contracts in stocks, mutual funds, and commodities as “financial investments” rather than “gambling.” The survey did not mention sports among the event contracts it described, though sports contracts account for over 90% of trading on the platform. Adding further irony is the fact that the company has argued elsewhere that polls are not reliable.

Other polling from sources with less apparent bias has been far less favorable. A 2026 National Council on Problem Gambling (NCPG) survey found that nearly half of Americans consider prediction markets comparable to gambling, versus just 27% who liken them to investing. It also found broad support for guardrails: 84% believed prediction market platforms should be treated like gambling when it comes to consumer protections, and 82% said platforms where people risk money on future outcomes should be required to offer responsible-gaming tools such as deposit limits, cooling-off periods, and access to help resources.

History of Prediction Markets

While public interest in, and opinions on, prediction markets are relatively new, the concept has existed for decades in academic, corporate, and public arenas. Listed below, in chronological order, are the most significant projects in the prediction market space over the last several decades.

The Iowa Electronic Markets: Established in 1988, the Iowa Electronic Markets are collectively considered the first institutional prediction market. The IEM are a system of real-world event contracts traded for real money on an exchange operated by the University of Iowa Tippie College of Business as a non-profit academic data source. The IEM still operate today, but is not legally regulated due to the small volumes of most markets.

TradeSports / Intrade: Founded in 1999, TradeSports and Intrade were Ireland-based exchanges founded by John Delaney that specialized in sports wagering and non-sports event contracts, respectively. From 2002 onward, Intrade’s political contracts, especially on US elections, made up the majority of trading activity, and Intrade users earned a reputation for uncanny accuracy, correctly anticipating the capture of Saddam Hussein, the 2005 papal succession, and the electoral result in 49 of 50 states in 2012. Tradesports itself wound down in the late 2000s after the Unlawful Internet Gambling Enforcement Act of 2006 made it nearly impossible for American users (the bulk of the Tradesports customer base) to move money to online gambling sites, leaving only Intrade in operation. The CFTC sued Intrade in late 2012 over unregistered options trading, forcing the company to bar American traders and ultimately suspend trading entirely in March 2013.

NewsFutures: Founded in May 2000, NewsFutures was another early prediction market, which ran entirely on play money rather than cash. Operational for nearly a decade, NewsFutures listed contracts on more than 120K events across politics, finance, and sports. The platform also co-branded its sports and money markets with USA Today. NewsFutures lent credibility to the play-money model that many corporations would later adopt for internal forecasting, where cash betting is often prohibited, after 2003–2004 NFL season forecasts on NewsFutures were found to be equally accurate to the real-money predictions of another platform, TradeSports. Because NewsFutures was based on play money, it could also facilitate international betting, allowing users from different countries to trade with one another.

Policy Analysis Market: The largest-scale proposed public prediction market was the Policy Analysis Market, scoped in May 2001 by DARPA’s Information Awareness Office as part of the FutureMAP project. Designed with the San Diego research firm Net Exchange, it would have let regional experts trade futures contracts on political, economic, and military developments across the Middle East. Though its designers intended for the predictions to be used as a geopolitical stability gauge, several US senators denounced it as a market for betting on assassinations and terrorism, and the Pentagon scrapped it in August 2003 before a single trade was placed.

PredictIt: Founded in November 2014, PredictIt offered real-money political betting to American traders as a nonprofit educational project of New Zealand’s Victoria University of Wellington. Created by John Aristotle Phillips in partnership with the university, it operated under a CFTC no-action letter that allowed it to run without registering as a formal exchange, with an $850 ceiling per contract and a cap on how many traders each market could accommodate. Like the IEM, it positioned itself as a research tool and shared data with academics. As of July 2026, PredictIt is still in operation despite a multi-year legal battle after regulators moved to revoke its operating relief in 2022, the resolution of which raised its contract ceiling to $3.5K and removed the cap on traders per market.

Augur: Built on Ethereum in 2014, prediction market Augur was one of the first genuinely decentralized applications created on the blockchain. Funded by an ICO that raised roughly $5.5 million in 2015, Augur went live on Ethereum’s mainnet in July 2018. Whereas earlier platforms curated their own contracts, Augur let any user create a market on any question, a permissionless design that drew controversy over user-created “assassination markets.” The platform’s traction was limited from the outset, with daily users collapsing from about 270 at launch to under 30. The platform was relaunched in 2020 to migrate trades to the DAI stablecoin and add safeguards, but traction remained limited, and the project went largely dormant after 2021. As of July 2026, the Lituus Foundation and Ethereum contributor Micah Zoltu are leading a reboot of the platform, including a new protocol for dispute resolution.

Kalshi: Founded in 2018 by MIT graduates Tarek Mansour and Luana Lopes Lara, Kalshi has spent years and tens of millions in legal fees establishing that event contracts belong under commodities law rather than gambling statutes, and in November 2020, it became the first platform licensed by the CFTC as a designated market for such contracts. As of July 2026, Kalshi has millions of users and has been valued at over $20 billion. The platform provides probability data on global events to outlets like CNN and CNBC.

Polymarket: Launched in June 2020 by Shayne Coplan, Polymarket brought the blockchain model to the mass market, settling trades in the USDC stablecoin on the Polygon network. Polymarket claims to be the world’s largest prediction market, and has made waves with real-world marketing stunts, opening The Situation Room, a Washington, D.C. pop-up bar built for “monitoring the situation” with screens streaming live X feeds, flight radar, Bloomberg terminals, and Polymarket odds, after an earlier free grocery store pop-up called The Polymarket in New York City that gave away free groceries for several days. As of July 2026, Polymarket trading is now legal in some US states under the same CFTC governance as Kalshi.

In 2024, $3.6 billion was wagered on the US election Polymarket prediction market, which correctly signaled a Trump victory ahead of most pollsters and prompted the FBI to open a probe looking into the company as a facilitator of betting on election outcomes (which is illegal in over half of US states). Federal investigators dropped their probe in July 2025, shortly after which Polymarket acquired a CFTC-licensed exchange to secure a compliant path back into the United States, returning to American users at the end of 2025 alongside a multibillion-dollar investment from Intercontinental Exchange, the parent company of the New York Stock Exchange.

In addition to dedicated prediction market companies, established players across tech, crypto, finance, and sports betting have launched their own prediction market platforms in the last two years. Meta is reportedly building Arena (first reported in June 2026), a standalone, play-money prediction app personally directed by Mark Zuckerberg. Crypto exchange Crypto.com has been a CFTC-regulated prediction-markets pioneer since December 2024 and spun up its standalone prediction-markets platform, OG, in February 2026. Gemini, another crypto exchange, released US event contracts in December 2025 following CFTC approval. Traditional financial venues such as CME Group and Interactive Brokers opened their own CFTC-regulated prediction markets in 2022 and 2024, respectively. Finally, online sportsbook (OSB) operators opened prediction markets in late 2025, including DraftKings (December 2025), Underdog (September 2025), and PrizePicks (November 2025).

How Prediction Markets Make Money

Prediction markets make money in different ways, like charging a fee on each contract (regardless of whether the contract ends up a winner or not), taking a cut of winning payouts, or collecting the spread between bid and ask prices by acting as a market maker on its own platform.

Kalshi makes money in three primary ways. The first is per-contract trading fees charged on nearly every trade, to both the contract maker (seller) and the taker (buyer). Rather than a flat rate, the fee follows a probability-weighted curve: highest on 50/50 contracts near $0.50 and shrinking toward zero as a contract approaches $0.01 or $0.99, peaking at about $0.0175 per contract, with makers paying meaningfully less (often rounded to zero on small orders), and no fee when a contract settles. Second, Kalshi also earns interest on the idle cash sitting in customer accounts, which are held in segregated regulated accounts. Finally, Kalshi offers streams of market data via API and partnerships, allowing partners like Bloomberg, the Fed, and equity analysts to access its markets’ predictions off-the-shelf.

Unlike Kalshi, Polymarket historically subsidized trading fees (which were as low as 0% on most markets) using venture funding, betting that volume growth would allow it to capture fees later. In March 2026, the company began taking fees on contract buyers, determined by a multiplier that varies across categories, from 0.07 on crypto, 0.03 on sports, 0.04 on finance/politics, 0.05 on culture / other, while geopolitics and world events stay fee-free, and makers pay zero. As of July 2026, these fees are used to compensate market makers, and Polymarket’s primary revenue source comes from investing users’ USDC collateral, paying users about 4% and keeping the marginal interest.

Polymarket also makes money selling real-time data feeds to hedge funds, news outlets like Bloomberg, and AI developers. Longer term, a planned POLY token would allow the company to force users to hold their account’s cash in such tokens and make money by taking a cut of each transaction in addition to a cut of deposit interest.

Usage & Revenue

As of July 2026, Kalshi and Polymarket are the dominant prediction market platforms both in the US and globally. According to the companies, Kalshi has over 5 million active registered users, and Polymarket has 2.3 million, volumes that have more than doubled since May 15th, 2026, as NBA playoffs, NHL playoffs, and World Cup games have increased interest in wagering on sports event outcomes.

Source: Sherwood News

Analysts estimate that over 17 million individuals worldwide will use prediction markets in 2026, including prediction markets facilitated through traditional online sportsbooks. According to one study of prediction market users, 90% are male, a skew that prediction market advertisements are actively attempting to address.

In line with a growing user population, prediction market trading volume has increased over 5x in the last year and over 10x in the last two years. Combined monthly global trading volume on Kalshi and Polymarket rose from less than $5 billion in September 2025 to $24 billion in April 2026. The global market was estimated at $31.2 billion as of May 2026. Kalshi reportedly makes up 58% of that volume, and Polymarket 28%. This global total is comparable to the handle of regulated US sports betting, which reached $167 billion during 2025, or around $14 billion a month.

“Handle” is the total amount of money wagered at a sportsbook. The sportsbook only keeps a small slice of that pool, called the hold (revenue as a percent of handle): in the US in 2025, the hold averaged about 10.15%, so books kept roughly $10 of every $100 wagered. This means $167 billion in handle translates to about $17 billion in revenue. Prediction markets keep a far thinner slice of their overall trading volume, since they charge small exchange-style fees instead of a 10% hold, which is why Kalshi and Polymarket can post comparable trading volume yet report much lower revenue (Kalshi’s $2 billion ARR or Polymarket’s $1 billion ARR versus sportsbooks’ $17 billion).

The above is an excerpt from our new deep dive on the mechanics and challenges of prediction markets. See the full report here.

Meta launched Muse Spark 1.1 this week, its most capable model to date, developed by the Meta Superintelligence Labs under Alexandr Wang. On agentic benchmarks, the model rivals GPT-5.5 and Opus 4.8. Muse Spark supports a 1M-token context window with active context management and is priced at $1.25 input and $4.25 output per million tokens, making it 4-8x cheaper than Anthropic’s Fable 5. Zuckerberg framed the pricing as a response to what he called “very extreme” margins from competing labs.

The timing of Muse Spark 1.1’s launch couldn’t be better. Palantir CEO Alex Karp went viral just last week after telling CNBC that enterprise customers were “livid” with frontier labs, that something has “gone completely wrong” with token-based pricing, and that companies are privately telling him they’re getting “no value” from their AI spend. With this week’s launch, Meta is offering a proprietary frontier-tier model at notably cheaper pricing.

What makes this story especially noteworthy is that until recently, Meta was widely considered an afterthought in the model race, which was characterized by SemiAnalysis as a “two horse race between OpenAI and Anthropic. “ Muse Spark 1.1 is not yet at performance parity with Fable 5 or GPT-5.6 Sol, but it is competitive on agentic tasks and dramatically cheaper. Whether that tradeoff resonates with enterprises currently rethinking their AI budgets is the question that will define whether Meta has actually re-entered the race.

SpaceX Joins the Nasdaq-100

On July 7, just 15 days after its IPO, SpaceX (SPCX) joined the Nasdaq-100, marking the fastest-ever entry for a newly listed company following Nasdaq’s May change to its fast-track rules. JPMorgan has estimated that QQQ-tracking funds drove $4.3 billion in buying around the index inclusion. Despite this mechanical demand, SPCX shares fell nearly 7% on their first day in the index at $143, only modestly above the IPO price and down roughly 34% from the all-time high of $225.64 on June 16.

Investor reaction has been mixed, with banks including BofA, Deutsche Bank, Goldman Sachs, UBS, RBC, Bernstein, Morgan Stanley, and JPMorgan bullish on the stock. Insider lockups for stock sales do not begin unwinding until late July, with nearly 20% after Q2 earnings and around 7% in tranches through October, with Musk’s 6.4 billion shares locked until June 2027.

Separately, SpaceX President Gwynne Shotwell announced that she and her husband would donate shares to over 2 million Trump administration-sponsored accounts for children, worth about $320 million. At the White House opening-bell ceremony celebrating the rollout, Trump was joined by Nasdaq CEO Adena Friedman and president Nelson Griggs, plus NYSE president Lynn Martin, though notably none of the five ETFs (SPYM, IVV, ITOT, VTI, SPTM) these accounts can invest in is a Nasdaq fund.

Reading Between Claude’s Lines

Two unrelated stories have been circulating within the tech world this week surrounding implicit messages in Claude logs. The first was raised by developer “Thereallo” in late June, who found that Claude Code embedded steganographic markers in the system prompt, subtly altering the date format and apostrophe in “Today’s date is,” to flag users on Chinese time zones, proxies, or AI-lab domains, with the domain list XOR-obfuscated. Anthropic engineer Thariq Shihipar called it a March “experiment” to prevent reseller abuse and distillation, which was later rolled back; in response, Alibaba banned Claude Code, and privacy advocates called it a breach of trust.

Separately, on July 6th, Anthropic published interpretability research introducing the “Jacobian lens,” which surfaces the words a model is poised to say. Collectively, these form the “J-space,” a small privileged set of representations behaving like a global workspace, Claude’s analog of accessible thought. Reading it lets researchers catch Claude silently noticing it’s being tested, weighing manipulation, or reacting emotionally, work that has reignited the machine-consciousness debate.

What We’re Reading

• China landed a reusable Long March rocket booster on a recovery ship, using shipboard netting instead of the landing legs used by SpaceX’s Falcon 9, making it the second country to master orbital-class booster recovery after the US.

• Meta drew up plans for a new cloud infrastructure business to sell excess AI compute, competing directly with AWS, Azure, and Google Cloud through an internal division called Meta Compute that would let outside developers pay to run queries against models, including Muse Spark; the announcement raised Meta shares 10% while CoreWeave and Nebius fell more than 10% each.

• OpenAI released GPT-Live, a new full-duplex voice-model architecture powering ChatGPT Voice that can listen and speak simultaneously, delegate deeper questions to GPT-5.5 in the background, and show visual widgets for weather, stocks, and sports; the new model becomes the default for Go, Plus, and Pro users, and GPT-Live-1 mini for Free users worldwide, replacing the turn-based Advanced Voice Mode.

• Anthropic signed a 20-year, $19 billion lease with crypto miner-turned data center provider TeraWulf for capacity in Hawesville, Kentucky, covering 401 megawatts that will come online in phases starting in late 2027, exceeding TeraWulf’s $12 billion market cap and sending shares up as much as 19% intraday.

• Beijing convened meetings with Alibaba, ByteDance, and Z.ai to consider restricting overseas access to China’s most advanced AI models, including models not yet released, according to three people familiar with the discussions.

• Illinois Gov. JB Pritzker signed the Artificial Intelligence Safety Measures Act into law, requiring large frontier AI developers (>$500 million annual revenue) to publish catastrophic-risk mitigation plans, submit to first-in-the-nation annual third-party audits, and report incidents within 72 hours.

• The US Cybersecurity and Infrastructure Security Agency deployed Anthropic’s Mythos to audit government software, with CISA’s Attack Surface Evaluation team using the model to scan government code repositories for vulnerabilities exploitable by foreign spies and cybercriminals, reportedly recovering a large number of bugs, even as Anthropic remains formally on the Pentagon’s supply-chain risk list.

• Anthropic launched Claude Science, an AI workbench for scientific research, alongside its own internal drug-discovery program focused on neglected diseases, demonstrated during a livestreamed event screening 2.2K compounds to find a stabilizer for the enzyme responsible for phenylketonuria.

• SpaceXAI and Cursor released Grok 4.5, their first jointly developed AI model built for coding, finance, and legal tasks, and positioned to challenge Anthropic’s Opus 4.8 and OpenAI’s GPT-5.5. The release comes weeks after SpaceX formally exercised its option to acquire Cursor at a $60 billion valuation.

• Apple is reportedly lobbying the Trump administration for clearance to buy memory chips from blacklisted Chinese company CXMT, per the FT, as the memory shortage that drove last month’s Mac and iPad price hikes continues to bite; Apple wants guarantees CXMT won’t be added to the US Entity List (which would trigger licensing requirements) after CXMT’s Pentagon 1260H listing over alleged PLA ties.

• DeepSeek began developing its own AI inference chip to reduce reliance on Nvidia and Huawei after its first $7.4 billion fundraise at a $52-59 billion valuation, sending Nvidia shares down 1.6% on the news.

• Tesla expanded its Robotaxi service to Miami on July 3, its first city outside Texas and California, running fully unsupervised rides in a geofenced 10–14 square-mile stretch of western Miami-Dade from launch, raising Tesla’s active robotaxis to around 60 versus Waymo’s 4K vehicles and over 500K paid rides per week.

• Samsung forecast Q2 operating profit of 89.4 trillion won ($58.4 billion), a 19-fold year-on-year jump that overtakes Nvidia’s most recent quarterly profit and makes Samsung the world’s most profitable tech company for the period, driven almost entirely by AI memory demand.

At Contrary Research, we’ve built the best starting place to understand private tech companies. If you're interested in researching and writing about tech, apply here for our Research Fellowship!

For decades, developing treatments for rare diseases, which are medical conditions that affect fewer than 200K people at any given time, were viewed as one of the least attractive markets for pharmaceutical development. Patient populations were too small, development costs too high, and commercial upside too limited to justify investment. The Orphan Drug Act of 1983 changed that by restructuring the economics of drug development through tax credits, market exclusivity, and accelerated regulatory pathways, turning rare disease from a neglected corner of medicine into a viable commercial category.

Since then, however, the rare disease market has evolved into a commercial and scientific priority for pharmaceutical developers beyond that which policymakers originally envisioned. Rare disease is no longer just a niche area of biotech supported only by government incentives; it has become the primary testing ground for new therapeutic platforms such as gene editing, RNA interference, antisense oligonucleotides, and gene therapies. Small, genetically defined patient populations allow companies to validate entirely new treatment modalities faster and with greater biological precision before expanding those technologies into the larger markets of more common diseases.

As a result, the central priorities in developing rare disease treatment are shifting. The principal challenge is no longer simply determining whether a company can build a successful “orphan drug”, which is a pharmaceutical agent developed specifically to treat a rare medical condition. Instead, the question is whether a company can use rare disease to validate a scalable platform technology, extend that platform into broader indications, and navigate the reimbursement dynamics created by multimillion-dollar curative therapies.

Orphan drugs are a critical contributor to this shift towards platform strategies more broadly, and research on orphan drugs has already benefited drug development in general. The companies best positioned for the next decade will not be those treating orphan drugs as isolated products, but those using rare disease as the commercial and regulatory proving ground for the next generation of biotech platforms.

The History of Orphan Drugs in the US

The term “orphan disease” is an economic designation rather than a clinical one. These conditions were orphaned not necessarily by providers or researchers, but by pharmaceutical companies hesitant to invest hundreds of millions of dollars over a decade in drugs that would serve small populations consisting of thousands of patients. Orphan drug development requires $1–2 billion or more to bring a single treatment to market, and recouping that investment requires either a large patient population or higher prices than insurers and government programs want to pay.

Somewhere between 7-10K rare diseases collectively affect 30 million Americans, more than cancer and heart disease combined. In the early 1980s, when the Orphan Drug Act was passed, 95% of these diseases had no FDA-approved treatment. In lieu of a commercial incentive to create drugs for these diseases, policies have been proposed and passed to nationally fund and develop such drugs.

Government agencies, however, have historically been ill-equipped for this role. The FDA regulates medicines, but doesn’t directly or indirectly incentivize their creation. The NIH funds basic research, but it isn’t designed to run large-scale clinical trials or manage commercial-scale manufacturing. Direct government development of drugs would also eliminate the competitive pressure that drives innovation, replacing market efficiency with bureaucratic process.

The 1983 Orphan Drug Act (ODA) took a different approach. Rather than developing orphan drugs itself, the government introduced new incentives for the development and manufacturing of drugs for rare diseases. It created a formal designation process to recognize drugs targeting rare diseases, which they defined as those affecting fewer than 200K people in the United States (or any condition for which there is no reasonable expectation of commercial profitability for other reasons). It also attached concrete financial incentives to that designation, including tax credits for clinical testing, waivers of multi-million-dollar FDA user fees, and seven years of market exclusivity upon approval.

How the Orphan Drugs Act Reset Incentives

Tax credits for the costs of qualified clinical trials for orphan-designated drugs were originally set at 50%, and later reduced to 25% by the 2017 Tax Cuts and Jobs Act. Unlike a deduction, which reduces taxable income, a tax credit directly reduces the taxes a company owes, making it especially valuable for smaller biotech firms with limited capital. Since clinical trials are the most expensive phase of drug development, this credit effectively transfers a portion of the downside risk of a trial’s failure to the government.

Orphan-designated drugs are also exempt from FDA user fees. The FDA charges several million dollars in application fees when companies submit a drug for approval. Removing these fees meaningfully reduces upfront costs that could otherwise deter smaller companies or those pursuing multiple rare disease indications simultaneously.

Once an orphan drug is approved, a seven-year window begins during which the FDA will not approve a competing drug for the same disease. This creates a temporary, legally protected monopoly, giving the sponsor the pricing power necessary to recover its R&D investment without immediate competitive pressure. This window is designed to be long enough to attract private investment but short enough to eventually allow competition and market correction.

Beyond financial incentives, the ODA also introduced expedited review pathways that shorten the time from development to patients. Fast Track allows more frequent FDA interaction during development. Breakthrough Therapy provides intensive FDA guidance for drugs showing early clinical promise. Priority Review cuts the standard 12-month review clock to six months. None of these programs changes the evidentiary standard for approval.

The Long-Term Impact of the ODA

The next few decades showed that the ODA was successful in incentivizing orphan drug development. By August 2018, 503 unique medicines targeting 731 orphan indications had been approved, and 78% of them were approved solely for orphan diseases. By 2023, more than 370 orphan drugs were actively marketed (some approved drugs were later discontinued, withdrawn for safety reasons, or never commercially launched).

Not only that, but the pace accelerated over time: between 1983 and 2019, over 5K drugs and biologics received orphan drug designation, with the number of designations more than doubling in the 1980s and 1990s, nearly doubling between the 1990s and 2000s, and nearly tripling between the 2000s and 2010s.

The effects of orphan drug development went beyond orphan diseases. Research into some rare diseases led to new, often more efficient drug development processes. About 80% of rare diseases are monogenic (caused by a single genetic mutation), making the underlying biology easier to characterize and target. Genetically defined patient populations can also be easier to diagnose via genetic testing, allowing companies to run smaller, more efficient trials with cleaner endpoints and faster paths to regulatory validation.

The diseases that attracted the most development efforts in the decades since the passage of the ODA share a common profile: a well-characterized genetic mechanism, clear biomarkers or measurable clinical endpoints, a patient population large enough to support trials, and an engaged patient community advocating with regulators. Together, these conditions made rare disease not just more commercially viable under the ODA’s incentive structure, but more scientifically productive per dollar invested.

The Drug Development Payer Problem

The reimbursement system for drugs is built around the basic assumption that drugs are taken repeatedly, so costs are recurring and can be spread over time. Despite the incentives introduced by ODA, this reimbursement model makes it economically challenging to break even for some rare disease treatments.

Small molecules and RNA interference (RNAi) both fit that description because they modify existing biology, correcting the impact of a faulty protein or silencing a problematic gene, but the underlying genetic defect remains, and treatment continues indefinitely. Genetic treatments, however, work differently, correcting or replacing the gene itself in a single intervention.

For this reason, many orphan drugs carry high list prices: 39% cost more than $100K annually, and gene and cell therapies in particular routinely run into the hundreds of thousands of dollars or more. In 2019, the average annual cost of an orphan treatment per treated patient was $32K, though individual treatments ranged anywhere from $6K to $500K per year.

The trajectory of these costs has been steep rather than incremental. The US saw a 26-fold increase in total orphan drug costs between 1998 and 2017, with annual cost growth running at roughly 12% since.

The economic structure of insurers, annual premiums collected against annual costs, makes these costs difficult to absorb. A drug that costs $3 million once but saves $500K a year in avoided ongoing care can be cost-effective over a patient’s lifetime, but the insurer paying that $3 million today may not be the same insurer collecting the savings five years later, once the patient has changed jobs or insurance plans. This mismatch means no individual payer has a strong incentive to pay full price for a cure, even when the cure is, in aggregate, a good deal for the healthcare system. This is a problem for any expensive one-time therapy, not just those specific to rare diseases.

Platform Drug Development

This mismatch has pushed the drug development industry towards a platform strategy for getting new technologies to market in the first place. The biotech industry is undergoing a structural shift from single-asset companies, which develop one drug for one disease, to platform companies, which build a reusable technology that can be applied to many diseases in succession. A platform, in this context, is not a drug, but the underlying engine that discovers and produces drugs. Platform companies recycle the same core technology, applied repeatedly, to produce a stream of new drugs rather than just one.

Platform companies benefit from compounding returns: each successful drug makes the next one cheaper and faster to develop, because the core technology, manufacturing process, and regulatory relationship with the FDA are already established. This lowers drug development costs, leading to lower drug costs for patients and insurers. Examples of drug development platforms include a chemistry and screening process that can be re-run against new disease targets, a delivery mechanism that can carry different genetic payloads to the same tissue, or a manufacturing and vector system that can package different genes for gene therapy.

Rare disease has become the preferred arena to prove that these platforms work, for a specific scientific reason: roughly 80% of rare diseases are caused by a single-gene mutation, so the drug target is known from the outset. Many of these diseases have well-established biomarker endpoints, such as protein levels, enzyme activity, and lung function, that can demonstrate a platform is working biologically long before a company has the years of follow-up needed to show clinical outcomes.

Biotechnology companies such as Vertex and Alnylam each demonstrate this model, but in two different layers of drug development: one validates a drug discovery platform, the other a drug delivery platform.

The above is an excerpt from our new deep dive on orphan drugs. See the full report here.

This piece was originally published in Arena Magazine.

Most stories about American industrialism have turned from inspirational myths into cautionary tales. One such story is that of the New York Shipbuilding Corporation in Camden, New Jersey. Founded in 1899, New York Ship grew to become the largest shipyard in the world by 1917. World War I sparked such a shipbuilding boom that the War Department had to fund a new city — Yorkship Village — just to house the labor force for New York Ship; it was the first federally funded planned community.

By the Second World War, New York Ship employed 47,000 across a 160-acre site along two miles of the Delaware River. What started as a factory had become an entire economy built to power the largest and most productive shipyard on Earth. New York Ship delivered 670 merchant and naval ships during its life, including three nuclear submarines and even the first nuclear-powered merchant ship.

By the end of WWII, America’s fleet represented 50,000,000 deadweight tons — a measure of total carrying capacity — two-thirds of global tonnage. But after the war ended, the burden of upkeep for such a large fleet felt unnecessary, leading the US to immediately start mothballing and scrapping ships rather than maintaining them. The Cold War sustained demand for specialized naval work, like nuclear submarines and carriers, but that work concentrated in a select group of dedicated yards. Meanwhile, commercial shipbuilding migrated to lower-cost foreign yards. Caught between a shrinking commercial market and increasingly specialized naval contracts, New York Ship didn’t have the business to stay afloat. By 1968, it had shut its doors. Camden, the city that powered New York Ship, tracks an eerie road map for what happens when industrialism falters. The city lost a third of its population, including 30,000 manufacturing jobs. It now faces a poverty rate of 28.5% and has become one of the most dangerous cities in the country.

The New York Shipbuilding Corporation is just one of dozens of examples of the cost of assuming that America no longer needed a strong industrial base. Particularly in the 1980s, the US came to the conclusion that manufacturing and shipbuilding were commodity processes better suited for lower-cost labor markets in Asia. We’ve seen the same dynamic play out in semiconductors, batteries, mining, and pharmaceuticals. But shipbuilding represents a distinctly gnarly systems problem tangled between defense funding, policy, and a fading American manufacturing core.

Why Build Ships?

Any national soul-searching about reindustrialization must start by disabusing ourselves of the notion that we don’t really need to build our own ships. We do. For most Americans, boats are a leisure category — a pleasure cruise, a fishing trip — while the working ocean has faded from the national imagination in a way that airports and interstates haven’t. While airplanes and highways may feel like the global connective tissue, the reality is that 80% of trade volume moves by sea. Without domestic shipbuilding capacity, the flow of global trade, including energy, raw materials, and consumer goods, is ultimately at the mercy of those who do control it.

The last few years have given us countless examples of that vulnerability. COVID pushed the median cost of transporting a shipping container from $2,000 to $20,000. When the Ever Given, a massive container ship, blocked the Suez Canal in March 2021 for six days, each day of disruption held up over $9 billion worth of goods. And when the recent US-Israeli strikes on Iran led Iran to effectively halt shipping through the Strait of Hormuz, it cut off 20 million barrels of oil per day, representing roughly 20% of global seaborne oil volumes.

Alfred Thayer Mahan, a Navy officer and historian described as “an apostle of sea power,” famously argued that “whoever rules the waves rules the world.” He saw the need for a navy as springing from the existence of peaceful commercial shipping. Increasingly, we live in a world where “peaceful commercial shipping” is more at risk than it has been in decades, owing to the volatility surrounding global chokepoints like Hormuz and the Taiwan Strait. The Council on Foreign Relations has put the odds of “a cross-strait crisis between China and Taiwan” at 50% in 2026. Such a crisis could dwarf a Hormuz closure in its ramifications for the global economy; nearly half of the world’s container shipping passes through the Taiwan Strait each year.

The seafaring domain is critical. The threat of great power conflict draws closer; a conflict between the US and China that may define the 21st century; a US-China war over Taiwan could deliver a roughly $10 trillion shock to the global economy in its first year, nearly 10% of global GDP — larger than COVID, the 2008 financial crisis, or the war in Ukraine. The US itself would absorb a 6 to 7% GDP hit, primarily through severed access to advanced semiconductors and the collapse of shipping through one of the world’s busiest sea lanes. Unfortunately, the majority of the ocean-going vessels are built by our would-be competitor.

In 2024 alone, one Chinese company built more commercial vessels by tonnage than the United States has produced since the end of World War II. Today, the US accounts for roughly 0.1% of the global commercial shipbuilding market while China accounts for 53%, up from just under 10% in 2000. China has become the dominant global shipbuilding superpower. It builds the vast majority of global shipping, including shipping critical to US naval power. Three of the ten oil tankers in the US fleet designed to carry fuel for military operations were constructed in Chinese yards. Seven of the twelve ships in the Maritime Security Fleet are made in China. The dependence is reckless, but the economic justification is straightforward: Regulators only require these ships to be US-flagged and crewed; otherwise, the ships would cost roughly $200 million, instead of their current sticker price of $50 million.

Meanwhile, domestically, US shipyards are failing to deliver on Navy shipbuilding goals. Some projects, like the Virginia-class submarine program, running at just 60% of its target production rate, with some ships delayed by as much as three years. Going forward, the US Naval fleet is already 20% short of the minimum capacity needed to carry out core missions. And that shortfall will continue to grow as maintenance backlogs pile up.

Meanwhile, the Chinese Navy, which surpassed the US Navy in total fleet size around 2020, keeps extending its lead. Huntington Ingalls Industries, the US defense prime most focused on seafaring, reported a backlog of $56.9 billion in 2025 — roughly five-years of work at current production rates. Even if Congress tripled shipbuilding appropriations tomorrow, Huntington Ingalls couldn’t build meaningfully faster. China’s advanced capacity, meanwhile, enabled it to add 30 ships to its fleet last year, while the US decommissioned 19 ships and built six, for a net loss of 13. In a protracted naval conflict, China can replace what it loses. We cannot.

To address this vulnerability, we must resurrect our national shipbuilding industry. But any attempt to do so must reckon with the chicken-and-egg problems inherent in bootstrapping a dying industry back into global competitiveness.

Shipbuilding’s Doom Loop

Every industrial process is a system; each input has process and timing dependencies, and every output demands execution. But when the process includes dozens of states and countries, multiple regulatory bodies across land and sea, and thousands of suppliers, each with their own regulatory regime and manufacturing idiosyncrasies, that system becomes almost hopelessly complex.

Regaining domestic shipbuilding capacity in the US means reckoning with a doom loop of interdependencies: labor, stable demand, competitive costs, scale, supply chain depth, and functional shipyards.

You can’t attract workers without stable demand: Shipbuilding apprenticeship completion rates sit below 35%. Even major shipyards, like Puget Sound, only begin with around 200 apprentices. The US Merchant Marine Academy, which offers a Marine Engineering and Shipyard Management major, was described by Transportation Secretary Sean Duffy as “dilapidated”. For those who make it into the shipbuilding industry, first-year attrition runs around 50-60% even at the most consistent shipyards. This attrition happens because there is no confidence in demand signals relative to more reliable industries, like oil and gas or construction. What’s more, around 27% of maritime workers are over the age of 55, meaning we’ll see a silver tsunami of retirements over the next decade.

You can’t generate stable demand without competitive costs: Asian shipyards produce at 5-6x lower prices and twice the delivery speed, driven in part by the massive regulatory burden American shipyards carry. Every input in the US, from labor to steel, is dramatically more expensive. No rational commercial buyer will pay that premium unless compelled by law or subsidy. Even artificial demand from the US government has wavered, with multi-billion-dollar shipbuilding contracts fluctuating between administrations.

You can’t achieve competitive costs without scale: China builds 1,000 commercial ships per year compared to about 10 in the US. Everything gets cheaper with scale. Workers move down learning curves faster with repetition, suppliers price lower with volume, and large workforces can specialize instead of generalize. As a result, the fixed costs of running a shipyard — the cranes, docks, management, financing — spread across more ships, making each one cheaper.

You can’t achieve scale without supply chain depth: Even a simple commercial ship requires thousands of components from hundreds of suppliers. Since 2000, the US has lost more than 25,000 of those suppliers. Some components, like marine diesel engines or marine-grade steel, have limited or no domestic production left. Of the six American companies manufacturing large marine diesel engines before 1980, only one — Fairbanks Morse Defense — still exists, surviving exclusively off of military contracts. Suppliers have done the math and found the US market wanting. No one will invest in marine diesel engine production for a market of ~10 ships per year.

You can’t achieve supply chain depth without functional shipyards: Of all the bottlenecks, shipyards themselves may be the hardest to resolve. You need deepwater coastal sites — exactly the kind of sites that attract the most onerous regulation — including 50-100+ contiguous acres of waterfront with deep draft access, heavy industrial zoning, and proximity to rail or highway. There are vanishingly few sites in the US that meet all those criteria and aren’t already developed.

Which brings us back to labor: Even if you could build a functional shipyard, with access to a deep supply chain, commanding sizable demand at competitive prices, you still need workers to staff it. Which brings us full circle.

No matter which lever you pull, nothing makes enough of a difference on its own. Legislation can artificially lift demand, but it cannot sustain a workforce. Foreign expertise can bridge some of the knowledge gap, but it can’t magically spin up a supplier base. What’s more, solving all of these issues simultaneously requires a level of sustained, coordinated, multi-decade industrial policy that the US hasn’t demonstrated since before the Cold War.

The solution here isn’t akin to the CHIPS Act for semiconductors or the IRA for batteries, solutions that were, primarily, acts of capital deployment by the federal government. The closest analogy for a solution is the interstate highway system. From 1956 to the 1980s, the National Interstate and Defense Highways Act enacted a 20+ year, multi-administration, federally funded infrastructure program that required simultaneous land acquisition, engineering, workforce development, materials supply, and political consensus across all 50 states.

The above is an excerpt from our new deep dive on American shipbuilding. See the full report here.

In AI, the high token cost of using frontier models is becoming an increasingly loud part of the conversation. This week, OpenAI launched GPT-5.6 Sol at $5 input / $30 output per million tokens, half the input cost of Fable 5, which costs $10 input / $50 output. The price gap with open source is even wider. DeepSeek V4 Flash is available at $0.14 input / $0.28 output, about 36x cheaper than GPT-5.5 on input and over 100x cheaper on output. Meanwhile, China’s Z.ai released GLM-5.2, which costs about an eighth as much as Claude Opus 4.8 for certain tasks and is now the third most widely used AI system in the world.

Because China produces most of the top-performing open-source models, six of the ten models on the leading AI leaderboard were developed in China, meaning the cost pressure on US frontier labs is increasingly coming from state-backed competitors with no profit motive. Enterprises are starting to respond to the realities of token economics. It was reported this month that Uber burned through its entire annual AI budget in four months and has since implemented spending tiers on AI tools. Nevertheless, roughly 95% of enterprise AI usage still runs on frontier models, suggesting the shift to cheaper alternatives has barely begun.

Meta Eyes Prediction Markets

This Tuesday, it was reported that Mark Zuckerberg had tasked a small team with building a standalone prediction-markets app, internally called “Arena,” to rival Polymarket and Kalshi. Independent from Facebook, Instagram, WhatsApp, and Messenger, the app would initially use video-game-style points rather than cash, but hasn’t ruled out using real money in the future.

Arena would be Meta’s second attempt at a prediction market, after launching Forecast in 2020, a points-based COVID-era prediction app, which sunset in 2021 after failing to gain traction with a no-money, conversation-based strategy. Other tech companies have tried and failed with similar projects: Google ran the largest corporate markets ever, Prophit (2005–2011), which died after Dodd-Frank killed its regulatory path, and the internal Gleangen in 2020. The biggest differences between these experiments and rapidly-growing platforms Polymarket and Kalshi are (1) the ability to bet with real money and (2) the inclusion of sports contracts, which make up the vast majority of prediction market trading today.

US EVs Score Much-Needed Regulatory Win

This week, the Department of Transportation proposed removing the brake-pedal requirement for vehicles “designed to be driven exclusively by automated driving systems,” removing an important barrier for Tesla’s wheel-and-pedal-free Cybercab and Amazon’s Zoox. The same day, the Commerce Department barred Polestar, the Geely-owned EV maker, from selling new cars in the US under the “Connected Vehicle Rule,” restricting Chinese software or hardware only months after granting the same authorization to Polestar’s sibling company, Volvo, also owned by Geely.

This support to the EV industry comes as US auto sales just fell for an eighth straight month, with EV sales collapsing 35.5% year-to-date and EV market share sliding to 5.1%. The EV slump has been framed as a competitiveness crisis with China, which accounts for about 60% of global EV sales, and with BYD surpassing Tesla as the world’s largest EV maker in 2025.

Latest Research

• Flow Engineering is an AI-native physical engineering platform building the system of record for complex hardware programs. Read our new report here.

• Sakana is a Japanese AI model developer that launched a multi-agent orchestration system purported to rival Mythos. Read our new report here.

What We’re Reading

• OpenAI is considering delaying its blockbuster IPO into 2027 after SpaceX’s turbulent post-IPO trading spooked advisors and rekindled internal tension between CEO Sam Altman and CFO Sarah Friar, who has urged waiting to manage cash burn and compute commitments.

• OpenAI and Broadcom unveiled Jalapeño, OpenAI’s first custom inference chip, completing the design-to-tapeout in nine months (claimed to be the fastest ASIC cycle ever), with OpenAI’s own models accelerating parts of the work; the first deployment is set for late 2026 at a gigawatt scale with Microsoft and other partners.

• Anthropic accused Alibaba of conducting the largest known distillation attack on Claude, alleging operators tied to Alibaba’s Qwen lab ran 28.8 million exchanges through ~25K fake accounts to extract reasoning and coding capabilities; the June 10 letter to the Senate Banking Committee follows three earlier industrial-scale campaigns Anthropic attributed to DeepSeek, Moonshot, and MiniMax.

• The NSA lost access to Anthropic’s Mythos 5 after the Trump administration’s June 12 export-control order forced Anthropic to disable the model globally, depriving analysts of a tool that had found vulnerabilities in classified networks within hours during an internal Project Glasswing red-team test (the source of Sen. Mark Warner’s widely misread “broke into almost all our classified systems” comment).

• Meta paused its Model Capability Initiative, the mandatory employee-monitoring program that recorded keystrokes and screen activity to train internal AI models, after a misconfiguration left tens of thousands of internal data tables of private conversations, performance reviews, and transcriptions readable by any Meta employee; a 1.5K-signature staff petition predated the breach.

• President Trump signed two executive orders pushing for a research-grade quantum computer by 2028, also ordering the federal migration to post-quantum cryptography by 2030-2031 and Pentagon deployment of quantum sensors by 2028; the orders follow the Commerce Department’s $2 billion in equity stakes across nine quantum-computing companies last month, including a new IBM venture.

• NVIDIA introduced its Rubin reference data-center design, a closed-loop liquid-cooling system claiming up to a 100% reduction in on-site water use by running outdoor dry coolers instead of evaporative chillers. However, the closed loop addresses only ~25-33% of total AI data center water consumption, since electricity generation and chip fabrication also draw water.

• SpaceX launched and deployed Starfall on its first demonstration flight from Cape Canaveral, a saucer-shaped reentry capsule designed to return up to 1,000 kg of cargo from orbit anywhere on Earth; the FAA assessment describes Starfall as a potential “proliferated successor” to the ISS for orbital pharma and bio-manufacturing, and the Pentagon has parallel point-to-point cargo agreements with Blue Origin, Rocket Lab, and Anduril.

• Kalshi opened talks to raise fresh capital at a roughly $40 billion valuation, nearly doubling its $22 billion May valuation in seven weeks as the FIFA World Cup drove monthly volume above $17 billion; the talks land amid an escalating jurisdictional fight, with the CFTC suing nine states and CME suing the CFTC over Kalshi’s perpetual-futures approval.

• China retook the world’s fastest supercomputer crown for the first time since 2017 when the Shenzhen-based LineShine hit 2.198 exaflops on the TOP500 list, dethroning Lawrence Livermore’s El Capitan using a CPU-only design built entirely on Chinese-designed processors; organizer Jack Dongarra said US export controls give China “a strong incentive to develop domestic alternatives.”

• Meta and EssilorLuxottica rolled out Meta Glasses starting at $299 across three frame styles (Meta Adventurer, Meta Fury, and a Kylie Jenner collaboration), all running Meta AI powered by Muse Spark (the first model out of Meta Superintelligence Labs) from day one, with a new Rx Lens Swap for post-purchase prescription lenses and live translation now covering 20 languages.

• Apple raised prices across its Mac and iPad lineups (MacBook Neo from $599 to $699, MacBook Air from $1,099 to $1,299, iPad Air from $599 to $749) while sparing the iPhone for now, citing an “unprecedented” memory and storage shortage driven by AI data-center demand; Counterpoint Research says memory prices have roughly quadrupled in three quarters, and CEO Tim Cook last week called it a “hundred-year flood.”

At Contrary Research, we’ve built the best starting place to understand private tech companies. If you're interested in researching and writing about tech, apply here for our Research Fellowship!

Midjourney, a company that until recently had been known for its AI image generation models, drastically pivoted this week into what it’s calling “a new era” for the company. It started by unveiling a new division called Midjourney Medical built around a whole-body ultrasound scanner the company says will be “in many ways superior to even MRI machines”, with scans taking as little as 60 seconds and requiring no radiation or magnetic fields. The scanner lowers patients into a ring of underwater transducers that emit and record ultrasonic waves from every angle, reconstructing a 3D map of the body roughly 100 times faster than current MRI technology. The company has grand ambitions for its new scanner. It plans to open a San Francisco spa at the end of 2027 with 10 scanners and wants to scale to 50K scanners worldwide and a billion scans per month by 2031.

Midjourney’s scanner elicited a range of reactions. Some hailed it as “the most significant new imaging modality in 50 years” and predicted that it would herald a new era of preventative medicine, with the potential to save many lives. Others were more skeptical. As psychiatrist and blogger Scott Alexander laid out, ultrasound physically cannot penetrate bone or air, ruling out the brain, lungs, and bowels, which is precisely why MRI and CT remain necessary for most serious diagnostic work. Radiologists he consulted were uniformly critical, with one noting that the entire premise rests on the unproven assumption that earlier detection of incidental abnormalities improves patient outcomes.

This is the same reason the medical establishment already discourages whole-body screening MRIs today: they cost about $2K, are not recommended by the American College of Radiology due to insufficient evidence of benefit, and carry high false-positive rates in asymptomatic populations. Nevertheless, the strongest case for Midjourney’s approach may simply be convenience and cost, since the underlying sensor technology should scale far more cheaply than MRI’s giant magnets.

The Mythos Rollout Saga Continues

At a Seoul press conference, Anthropic’s Managing Director of International, Chris Ciauri, expressed confidence that Fable and Mythos would soon be available again, saying, “We are very confident that in the coming days, the models will become available again.” This comes after last week’s White House directive that blocked Mythos and Fable 5 for foreign nationals and prompted Anthropic to disable the models worldwide, and demands that the company prevent any possibility of model jailbreaks.

Though the event was meant to mark Anthropic’s expansion to Korea (including the opening of a new office there), questions about model export control and Project Glasswing dominated. Korea has been central to the controversy around the models, after the Washington Post reported that a Korean telecom company with access to Mythos triggered the US directive over suspected ties to China.

Another Advanced Reactor Goes Critical

On Thursday, Valar Atomics announced that it had taken its Ward 250 reactor critical. In doing so, it became the second advanced reactor to do so under the DOE’s Reactor Pilot Program (the other being Antares Nuclear’s Mark-0), and the first DOE-authorized reactor built and operated outside the national laboratory system. As Valar put it, the reactor went critical as a complete, fully integrated system configured for power operations and has now begun non-commercial power ascension. Founder and CEO Isaiah Taylor noted that Ward 250 is Valar’s second criticality as a company, following the NOVA core’s zero-power criticality at Los Alamos in November 2025. The company still has its eye on a harder target: power operations by July 4, the deadline set by President Trump’s executive order for at least three advanced reactors to achieve criticality.

This is the second major milestone in nuclear this month. As we wrote a few weeks ago about Antares, the persistent knock on advanced reactor startups has been a pattern of missed deadlines, and the value of these milestones lies as much in proving execution as in the underlying physics. Two more contenders remain in the running for the July 4 deadline: Aalo Atomics and Oklo’s Atomic Alchemy subsidiary, both still working through DOE authorization reviews. With two of the three required reactors now critical within six months, the challenge now shifts from proving feasibility to generating electricity at scale.

Latest Research

• FieldAI is building autonomy for robots in GPS-denied environments. Read our new report here.

• Nudge is a neurotechnology company developing non-invasive focused ultrasound systems to modulate deep brain circuits. Read our new report here.

What We’re Reading

• Microsoft moved Copilot Cowork to usage-based pricing and signaled it may add a Microsoft-hosted, fine-tuned DeepSeek V4 (or another open-source model) as a lower-cost alternative to the Anthropic and OpenAI models currently powering the enterprise agent, with EVP Charles Lamanna saying unlimited-use plans had become untenable since some users run hundreds of tasks per week.

• Bernie Sanders introduced the American AI Sovereign Wealth Fund Act, a one-time 50% stock tax on the largest US AI firms (estimated at $7 trillion at current valuations) that would create a public fund overseen by a seven-member Independent Commission for Democratic AI, require AI/non-AI conglomerates to split off their AI businesses, and pay out an initial 5% annual dividend of roughly $1K per American.

• Tech journalist Ed Zitron obtained audited OpenAI financial documents (verified by the FT) revealing a $38.5 billion 2025 loss against $13.07B in revenue, including a $20.9 billion operating loss and a $41.6 billion non-cash charge tied to the for-profit conversion.

• Anthropic CEO Dario Amodei and Google DeepMind CEO Demis Hassabis called for a US-led AI coalition at a G7 lunch attended by Trump, Bessent, Lutnick, Rubio, and other tech executives, with Amodei pressing for structured access to frontier models and cooperation on cyber, bioterrorism, and intelligence risks.

• Apple CEO Tim Cook warned that product price increases are “unavoidable” as AI server demand for high-bandwidth memory crowds out consumer DRAM and NAND supply, saying “everything needs to be on the table,” including loosening national security restrictions on Chinese memory suppliers.

• Sensor Tower’s State of AI 2026 report showed ChatGPT’s global market share fell below 50% for the first time, dropping to 46.4% by end-May (from 65.3% in December 2024) as Gemini reached 27.7% and Claude 10.3%; ChatGPT still leads on absolute monthly users (1.1 billion vs. Gemini’s 662 million and Claude’s 245 million) and total time spent.

• UK PM Keir Starmer announced a ban on under-16 social-media use covering Snapchat, TikTok, YouTube, Instagram, Facebook, and X (but not YouTube Kids, WhatsApp, or Signal) that goes further than Australia’s December 2025 model with bans on livestreaming, stranger-contact features, and potentially overnight curfews and infinite-scroll breaks for under-18s.

• Meta severed data-sharing with Manus and barred its staff from internal systems, completing an operational firewall as it begins unwinding the $2 billion Manus AI acquisition that China’s National Development and Reform Commission ordered reversed in April under its Foreign Investment Security Review Mechanism; the Singapore-domiciled co-founders are exploring a ~$1 billion raise to buy the company back, possibly via a Chinese JV with a Hong Kong listing.

At Contrary Research, we’ve built the best starting place to understand private tech companies. If you're interested in researching and writing about tech, apply here for our Research Fellowship!

When Anthropic launched Claude Fable 5 on June 9, its initial plan was to run automated safety checks on every request and visibly reroute queries touching cybersecurity, biology, and thinking-extraction to Claude Opus 4.8, with a notice explaining the switch — but the company also shipped a second, quieter layer: an invisible safeguard for frontier LLM development work that, per the system card, would limit the model’s effectiveness through prompt modification, steering vectors, or parameter-efficient fine-tuning without any fallback or notification, a measure partly aimed at preventing distillation of the model’s outputs.

After researchers branded the silent degradation “secret sabotage,” Anthropic repositioned, apologizing for making “the wrong tradeoff” and announcing that the invisible safeguards would become visible, with affected queries falling back to Opus 4.8 and users notified. The general user response has been harshly negative across the board: cybersecurity professionals complained the restrictions were so keyword-driven that even writing secure code or reading a blog post triggered downgrades, former employees argued the approach would slow scientific progress on problems like cancer and Alzheimer’s, and frustrated users vented that any science-related queries were flagged, with one remarking “I wouldn’t use this thing for anything to be honest. A refusal or HTTP-4xx error for content is fair enough, but this is basically taking your money and poisoning your code base.”

The story escalated further late Friday, when the US government issued an export control directive ordering Anthropic to suspend all access to Fable 5 and Mythos 5 for every customer, citing a national security concern related to a possible jailbreak. Anthropic complied but said that the jailbreak in question surfaced only minor, already known vulnerabilities that were discoverable by other publicly available models, including GPT-5.5. It also argued that recalling a model deployed to hundreds of millions of users over a narrow, non-universal jailbreak would effectively halt frontier deployments industry-wide if applied as a standard.

Solar’s Quiet Surge

US solar generation surpassed coal for the first time last month, supplying 12.8% of national electricity to coal’s 12.2%. In doing so, it became the third-largest source of electricity in the US, behind natural gas and nuclear. The milestone comes after two decades of roughly flat US electricity demand, which is now rising again due to AI, manufacturing, and electrification. Solar’s surge accounts for almost all the net new energy supply in the country; solar and battery storage made up 91% of all new US generating capacity in the first quarter of 2026.

The story of solar growth in the US reflects a global pattern that has been building for decades. Solar photovoltaic costs have fallen 90% over the last decade alone, following a remarkably consistent pattern known as Wright’s Law, which states that every time global cumulative solar capacity doubles, costs fall by roughly 20%. Global solar generation has grown from a negligible amount before 2010 to nearly 2,800 TWh in 2025, roughly tenfold in 15 years, with China alone now generating over 1,100 TWh. The cumulative effect over four decades has taken solar from one of the most expensive sources of electricity to the cheapest in many markets.

SpaceX Unveils AI Satellite

Ahead of SpaceX’s IPO this week, the company unveiled a significant milestone on the way to its orbital data center ambitions: 230-foot “AI1” satellites that generate 150 kW of peak power each, connected via laser links to each other and to Starlink. SpaceX said that it is to launch a dedicated AI satellite factory by the end of next year, and argued that the underlying technology is largely an extension of what SpaceX already built for Starlink V3. The company’s experience operating over 10K satellites will give it a unique advantage in managing an AI data center constellation that could reach up to one million satellites at scale.

The AI1 announcement occurred just days before the company underwent the largest IPO in history. SpaceX closed its first day of trading up 19% at $161 per share, valuing the company at $2.1 trillion after pricing at $135 and opening at $150. More than 500 million shares traded hands, a volume approaching Facebook’s 2012 debut. Elon Musk said SpaceX has been cash-flow positive since around 2015 and is going public now to fund a “significant growth phase,” including launching over 100K satellites and building AI data centers in space. Regardless of the triumphant IPO, or perhaps because of it, skeptics remain, with one calling the company’s $28.5 trillion total addressable market a “hallucination” and even Musk himself giving SpaceX a 10% chance of success as he defines it.

Latest Research

• Periodic Labs was founded last year to build autonomous laboratories and is already valued at $7 billion. Read our new report here.

What We’re Reading

• Anthropic launched Claude Fable 5, its first generally available Mythos-class model at $10/$50 per million tokens (less than half Mythos Preview’s price), with classifier safeguards routing ~5% of cyber/bio/chemistry/distillation queries to Opus 4.8 instead.

• Relatedly, Anthropic imposed a mandatory 30-day data-retention policy on all Mythos-class traffic, overriding Zero Data Retention agreements on the Claude API, AWS Bedrock, Google Cloud Agent Platform, and Azure Foundry, citing the need to detect Best-of-N jailbreaks, state-sponsored espionage, and data-extortion patterns that only surface across many requests.

• OpenAI filed a confidential S-1 with the SEC and announced the filing preemptively (“we expect it to leak so we’re just announcing it”), targeting a potential Q4 2026 / September debut at up to a $1 trillion valuation; OpenAI plans to facilitate an employee tender offer alongside the public option.

• OpenAI weighed drastic token-price cuts in anticipation of similar moves from Anthropic, as both head toward IPOs with very different unit economics; CEO Sam Altman has called AI costs “a huge issue” for business customers as AI spend grows.

• China readied a ~2 trillion yuan ($295 billion) five-year plan to build a nationwide network of interconnected state-operated AI data centers, run by China Mobile and China Telecom and sourcing over 80% of chips and equipment domestically from suppliers like Huawei as opposed to Nvidia and AMD.

• SpaceX unveiled AI1, its first “AI satellite,” ahead of its Friday IPO with a 70-meter wingspan, a 20-meter-tall orbital data center delivering up to 150 kW of compute through onboard chips, solar arrays, and an 110-sq-meter radiator to dump heat into space.

• Asset managers filed 25 SpaceX-themed ETFs with the SEC ahead of Friday’s $1.75T Nasdaq debut (ticker SPCX), 12 of them 2x long or short and two filed for 3x leverage, with three pre-IPO SpaceX-exposure ETFs pulling $4.1 billion in May inflows plus $645 million in early June, and BlackRock launching iShares Space Technologies UCITS (STAR) for Europe.

• The CFTC proposed the first federal regulatory framework for prediction markets, a “balancing test” that would permit most sports trading and event contracts but restrict bets on player injuries, officiating, and “first-pitch” gambling, and bar contracts on war, terrorism, or assassinations as not in the public interest, asserting exclusive federal jurisdiction over Kalshi, Polymarket, and ProphetX.

• Over 1K current and former SpaceX employees organized a collective-bargaining group through RIA Choreo to negotiate wealth-management fees below 0.5% AUM (vs. the ~1% industry standard) ahead of Friday’s IPO, with combined assets estimated at ~$20 billion, plus access to direct indexing, equity-backed lending, and other ultra-high-net-worth services typically gated to individual investors.

• Jeff Bezos backed Flourish, a neuro-AI startup led by ex-Amazon S-team executive Rob Williams and neuroscientist Thomas Reardon, with a $500 million raise at a $2.5 billion valuation to study the brain’s architecture and build “Cortex AI,” a system aiming to learn continuously and run at a fraction of today’s LLM power.

• Niantic Spatial partnered with Vantor to integrate ~30 billion Pokémon Go player environmental scans into a GPS-denied military drone navigation stack combining ground-level VPS with Vantor’s aerial Raptor software, with field testing planned in early 2026.

At Contrary Research, we’ve built the best starting place to understand private tech companies. If you're interested in researching and writing about tech, apply here for our Research Fellowship!

In January 2026, the FDA cleared the first cell reprogramming therapeutic drug for human testing with the implicit goal of reversing aging. Specifically, the agency cleared ER-100, a therapy designed to treat optic nerve conditions, including glaucoma and non-arteritic anterior ischemic optic neuropathy, to be tested on human patients. Optic nerve conditions, which are thought to be caused by aging, cause damage to the retinal ganglion cells (RGCs) of the eye. The trial, which will last five years, is considered an important step forward for epigenetic therapies to reverse cell damage. More significantly, it is also a landmark trial for therapies designed to slow, stop, or reverse aging.

Anti-aging medicine is a field that has emerged in academia, the pharmaceutical industry, and the commercial wellness space over the last decade as part of a wider cultural shift in medicine from reactive to preventive care. Putting the approval and eventual results of the ER-100 trial in context requires an understanding of the anti-aging field at large, including what trials can and cannot prove, the regulatory backdrop for reaching customers, and the landscape of scientific approaches to other therapies that may fall in this bucket.

While both academic and commercial endeavors to slow, stop, and reverse aging are consistently attracting increasing funding, the field has yet to recover from public failures in previous hype cycles or gain widespread traction with consumer usage. The success of newer research experiments and clinical trials may meaningfully accelerate continued funding in this space, even as both technical and regulatory guidelines are still being drawn.

Geroscience vs. Longevity

It is worth noting up front that geroscience, the interdisciplinary study of age-related chronic diseases and the biological pathways that trigger them, as practiced in academic labs and pharmaceutical pipelines, is a fundamentally different enterprise from the longevity content economy that is concurrently growing. To better parse the distinctions between these spaces, it is useful to separately define aging, longevity, and rejuvenation.

Aging refers to the underlying biological processes that collectively raise the risk of chronic diseases in humans over time. These include genomic instability, telomere attrition, epigenetic alterations, loss of proteostasis, deregulated nutrient-sensing, mitochondrial dysfunction, cellular senescence, stem cell exhaustion, and altered intercellular communication (as characterized in Hallmarks of Aging, the leading catalog of aging-related biological markers published by Lopez-Otin et. al. in the journal Cell).

Longevity is the downstream outcome of aging, measured clinically by the number of years lived. A related distinction exists between the concepts of lifespan and healthspan, both of which may be extended by the interventions that aging researchers are developing:

• Lifespan is the total number of years of life duration from birth to death. While lifespan is easy to define, it is difficult to study in real-time trials because participants may live for years or decades beyond the beginning of a study and may encounter confounding health factors throughout that period.

• Healthspan is the total years lived free of significant disease or disability. Healthspan is what most patients and clinicians actually care about, and it is what geroscience-oriented trials are designed around.

Academic and pharmaceutical industry geroscience researchers are largely less focused on longevity (as a noisy output of thousands of cross-cutting biological processes evolving concurrently) and more on the tracking of specific biomarkers associated with aging. Wellness influencers focused on longevity interventions, however, have gained significant followings in recent years, often making grand and unsubstantiated claims about increasing human lifespan via practices that have not been scientifically tested.

Rejuvenation is the reversal of biological aging (as measured by age-related biomarkers), distinct from just slowing its accrual. As of June 2026, rejuvenation has been demonstrated in mice through partial cellular reprogramming and parabiosis experiments, but has not been verifiably tested in humans.

While academic geroscientists study the mechanisms that control aging and rejuvenation, influencer protocols often conflate all three, treating changes in biomarkers as equivalent to reversal of aging or extended lifespan when these correlations between interventions are still being validated.

Longevity Hype Cycles

ER-100 is not the first attempt at a major breakthrough within geroscience. Several prior cycles revolving around different technologies each produced clinical disappointments, but provide important context for thinking through the eventual results of the ER-100 trial and others that may take place in the future.

Telomerase (1998-2010s)

Geron Corporation, founded in 1990, and Bill Andrews at Sierra Sciences both pursued telomerase activation as a strategy to extend cellular lifespan, building on the Hayflick Limit discovery that telomere shortening drives replicative senescence. After Geron-funded researchers showed that introducing the hTERT gene into normal human cells extended their replicative lifespan indefinitely without making them cancerous, the company pursued telomerase activation as an anti-aging strategy.

This work largely stayed at the level of cell biology, preclinical research, and licensing, and the company never ran human clinical trials of a telomerase activator for aging indications. The closest thing to a clinical product in this lineage was TA-65, a small-molecule telomerase activator derived from astragalus that Geron licensed out to TA Sciences, which sells it as a nutraceutical rather than developing it as an FDA-regulated drug as of June 2026. Bill Andrews pursued a parallel screening effort for telomerase activators but likewise never brought one to clinical trials for aging.

The fundamental problem with this approach was that telomerase is one common indicator of cancer: roughly 90% of human malignancies upregulate telomerase to achieve replicative immortality, meaning any systemic activator risks promoting tumorigenesis. Geron pivoted to oncology in the mid-2000s, abandoning anti-aging development to pursue the opposite mechanism, inhibiting telomerase to selectively kill cancer cells dependent on it. The company eventually developed imetelstat, approved in 2024 as RYTELO for transfusion-dependent anemia in lower-risk myelodysplastic syndromes.

Resveratrol / Sirtuins (2003-2013)

The sirtuin-activator program was the most heavily capitalized therapeutic effort in the modern aging field. The scientific basis for the treatment originated in a 2003 finding from David Sinclair’s lab that resveratrol, a polyphenol, increased the catalytic activity of the NAD⁺-dependent deacetylase SIRT1 and extended replicative lifespan in Saccharomyces cerevisiae (Nature 2003), positioning it as a small-molecule mimetic of caloric restriction.

Sirtris Pharmaceuticals was co-founded by Sinclair to develop synthetic sirtuin-activating compounds (STACs) such as SRT1720 (which are structurally unrelated to resveratrol), which reportedly activated SIRT1 with far greater potency than polyphenols. GSK acquired the company for $720 million in 2008, with the sale including a pipeline of structurally distinct next-generation STACs and then-leading clinical asset SRT501, a proprietary micronized formulation of resveratrol intended to improve its low oral bioavailability.

The trajectory of sirtuins illustrates the risk of building a platform on a mechanism observed purely in-vitro; the claim was subsequently attributed to an assay artifact. Groups at Amgen (2009) and Pfizer (2010) reported that the apparent activation depended on the use of a non-native peptide substrate conjugated to a fluorophore; the compounds enhanced deacetylation only in the presence of the fluorescent tag and showed no activation against native untagged substrates, indicating that the measured effect reflected substrate-fluorophore interaction rather than allosteric activation of the enzyme.

On the clinical side, GSK discontinued SRT501 after reports of nephrotoxicity, including renal failure, in a trial in multiple myeloma patients, where the formulation was also judged to lack sufficient efficacy. GSK closed Sirtris in 2013 and folded its remaining activities into internal R&D. The story of an enzyme-activation mechanism that did not replicate on physiological substrates, followed by a formulation discontinued on safety and efficacy grounds, is frequently cited in discussions of why subsequent geroscience trials (TAME, ASPREE, the Loyal programs) have been designed around clinical endpoints rather than biochemical or enzyme-activation surrogates.

Rapamycin (2009-2025)

The most extensively replicated lifespan-extending intervention in any mammal is rapamycin, a macrolide originally isolated from Streptomyces hygroscopicus in Easter Island (”Rapa Nui”) soil samples and developed as the immunosuppressant sirolimus. Rapamycin inhibits mTORC1, a nutrient-sensing kinase complex that genetic studies in yeast, C. elegans, and Drosophila had previously implicated in control of organism lifespan.

The NIA Interventions Testing Program, a three-site randomized protocol run in parallel at the Jackson Laboratory, UT Health San Antonio, and the University of Michigan, first reported in a 2009 study that encapsulated rapamycin fed to genetically heterogeneous (UM-HET3) mice from 600 days of age (roughly the murine equivalent of age 60 in humans) extended median and maximum lifespan in both sexes. The result has since been replicated across multiple doses, multiple starting ages, and intermittent dosing schedules, with effect sizes ranging from 9–26% extension of median lifespan depending on regimen. No other intervention has accumulated comparable depth of independent mammalian replication.

Human translation trials have proceeded along two distinct tracks, both targeting age-related immune decline as a tractable proxy for “aging” rather than aging-as-construct. The first track used everolimus (RAD001), a structural analog of rapamycin. In a 2014 study, a Novartis-sponsored Phase 2a trial in 218 adults aged 65+ showed that low-dose RAD001 improved influenza vaccine antibody response by approximately 20% and reduced te proportion of PD-1-expressing exhausted T cells, providing the first human evidence that mTOR inhibition could partially reverse immunosenescence. A follow-on Phase 2b study using a combination of RAD001 with the catalytic inhibitor BEZ235 reported reduced respiratory tract infections in 264 older adults, motivating the spinout of resTORbio to develop a selective mTORC1 inhibitor (RTB101) at doses intended to preserve the immune benefit while minimizing the immunosuppression that accompanies higher rapamycin exposure. The pivotal Phase 3 trial, PROTECTOR-1, tested RTB101 against clinically symptomatic respiratory illness in adults 65+ and missed its primary endpoint (odds ratio 1.07, p=0.65); resTORbio merged with Adicet Bio in 2020 and discontinued the program.

The second track is the off-label, low-dose rapamycin use that has spread through the longevity community, formalized in the PEARL trial. The PEARL trial was a 48-week, decentralized, placebo-controlled study testing weekly oral rapamycin against placebo with visceral adipose tissue (VAT) change on DXA as the primary endpoint. PEARL missed its primary endpoint but reported sex-specific secondary findings (increased lean tissue mass and reduced pain scores in women on the highest dose) that have been characterized by the investigators as hypothesis-generating rather than confirmatory.

The commercial trajectory for rapamycin has therefore reverted to the disease-specific indication pattern: Aeovian Pharmaceuticals raised a $55 million Series B in December 2025 to advance AV078, a next-generation selective mTORC1 inhibitor, into Phase 2 for tuberous sclerosis complex–associated epilepsy. Like resTORbio before it, Aeovian’s path to approval runs through a specific genetic disease in which mTORC1 hyperactivation is causally implicated, not through aging itself, demonstrative of the recurring pattern in which the most mechanistically promising geroscience targets ultimately advance through narrow disease indications because the regulatory and endpoint infrastructure for aging-as-indication does not yet exist.

Senolytics (2017-2020)

The senolytic hypothesis in aging therapeutics is the idea that selectively killing senescent cells (post-mitotic, secretion-active cells that accumulate with age and drive chronic inflammation via the senescence-associated secretory phenotype, or SASP) would ameliorate multiple age-related diseases. This idea emerged from a series of foundational mouse studies in the early 2010s, most notably studies in 2011 and 2016, from Jan van Deursen’s group at the Mayo Clinic, which used genetically engineered mice to clear p16^Ink4a-positive senescent cells and showed extended healthspan and delayed onset of multiple age-related pathologies. The commercial bet on this hypothesis was anchored by UNITY Biotechnology, co-founded in 2011 by van Deursen and other researchers in the field, with backing from investors including Jeff Bezos and Peter Thiel. The company listed publicly in May 2018 at a $700 million market capitalization on a pipeline of small-molecule senolytics targeting BCL-2 family anti-apoptotic proteins that senescent cells depend on for survival.

Both of UNITY’s highest-profile clinical programs produced placebo-indistinguishable readouts on their primary endpoints. Both UBX0101 and UBX1325 are senolytics, drugs that selectively kill senescent cells by inhibiting the pro-survival proteins those cells depend on to evade apoptosis. UBX0101, an intra-articular MDM2 inhibitor for painful knee osteoarthritis (chosen as the lead indication because OA is mechanistically tied to local senescent-cell accumulation in cartilage and synovium), failed to separate from placebo on WOMAC pain scores in its Phase 2 readout in August 2020, and the program was discontinued.

The company pivoted to UBX1325 (foselutoclax), a BCL-xL inhibitor delivered intravitreally for diabetic macular edema, which in the ASPIRE Phase 2b trial in March 2025 missed its primary non-inferiority endpoint versus aflibercept (the standard-of-care anti-VEGF therapy). Both compounds targeted disease contexts in which senescent-cell accumulation had been mechanistically implicated, and both failed to separate from comparator on their primary clinical endpoints — leaving open whether senescent cells were not the dominant driver of disease, the compounds did not achieve sufficient clearance at tolerated doses, or the molecules acted more as senomorphics (SASP modulators) than as true senolytics in vivo.

Post-UNITY

Following the ASPIRE readout, UNITY stockholders voted to approve liquidation and dissolution in September 2025. The structural problems the company faced echo those of sirtuin: a mechanism with strong genetic mouse evidence did not translate into clinical separation from comparator on hard endpoints in human disease. The senolytic field did not collapse with the failure of UNITY, but has since split along two tracks that both avoid pitching directly against established standards of care.

The commercial track is exemplified by Rubedo Life Sciences, whose lead candidate RLS-1496 is a topical GPX4 modulator (inducing ferroptosis specifically in senescent cells rather than apoptosis); the compound met its safety endpoint in a European Phase 1 study and received FDA IND clearance in September 2025 for a US Phase 1b/2a trial in actinic keratosis, a localized, biopsy-accessible dermatologic indication that allows efficacy to be measured against a clear lesion-clearance endpoint rather than against an existing systemic therapy.

The academic track is led by James Kirkland, one of the field’s founders, who runs investigator-initiated trials using off-patent senolytic combinations (most prominently dasatinib plus quercetin (D+Q)) across multiple disease indications, including diabetic kidney disease, idiopathic pulmonary fibrosis (Justice et al., 2019), osteoporosis, and the STOP-Sepsis Phase 2 trial using fisetin in older sepsis survivors.

Neither track is positioned to test the original senolytic hypothesis (that clearing senescent cells systemically extends healthspan in healthy older adults), and both implicitly concede that the path to clinical validation runs through narrow disease indications with conventional efficacy endpoints, not through aging as a construct.

In each of these cases, early testing in preclinical trials and testing on mice showed strong results, leading to optimistic funding rounds. In each case, clinical trials in humans failed to show the same results, leading both external and pharmaceutical industry funding to turn in other directions.

Aging is Not an FDA Indication

The vast majority of federally regulated drugs have specific biochemical targets, or pathways across one or multiple human organ systems (including the endocrine system, immune system, and nervous system, among others). These drugs also have specific disease or condition targets, or “indications,” defined as the specific, officially recognized medical purpose for which a drug, biological product, or medical device can be prescribed and marketed. In the official indications and usage guidance document from the FDA, the agency specifies that any approved drug’s published indications details should be scoped to “the disease or condition, or manifestation or symptoms thereof, for which the drug is approved, as well as whether the drug is indicated for the treatment, prevention, mitigation, cure, or diagnosis of that disease or condition, including relief of symptoms.”

Critically, the FDA does not recognize aging among the diseases and conditions for which treatments may be trialed and approved, though it does acknowledge myriad conditions associated with advanced age. This means that researchers studying treatments for aging need to select a specific indication for their therapies to target as a proxy for aging in general. Such indications can be challenging to select because clinical trials last months to years at most, while many aging-related mechanisms operate over entire lifetimes.

There are three primary layers of indications that are used in the field of anti-aging therapeutics for clinical trials: composite-endpoint trials, biomarkers and epigenetic clocks, and proxy disease indications.

The above is an excerpt from our new deep dive on anti-aging therapeutics. See the full report here.

On Wednesday, Cloudflare announced that, for the first time, bot traffic had surpassed human traffic online. Before generative AI, bots accounted for ~20% of traffic, but most of it was from crawlers. As autonomous agents acting on behalf of users have proliferated, bot traffic share has skyrocketed to 57%. The primary mechanism driving this explosion is intrinsic to how AI works. When a person goes to buy a camera, they might visit five or six sites; when agents even just recommend a camera, they may visit 5K+ sites.

The fear of robots taking over, however, is overblown. This is a tale as old as technology. Any time we automate something, the mechanisms of automation quickly outnumber humans. James Watt’s famous steam engine became useful only after he added a flyball governor, a simple mechanism that regulated the engine’s speed without human intervention. By the time we get to the Industrial Revolution, any given factory could have 1,000x as many governors as human workers. The same was true of transistors; what started as a simple electrical switch quickly grew to the point where a cutting-edge chip today can have over 20 billion transistors!

The real enabling force in any revolution is information and control, not raw power. Agents will run amok, but the real value will come from how effectively we can control those systems.

AI Leaders Sound Alarm on Bio Materials

Frontier AI lab leaders Demis Hassabis, Sam Altman, and Dario Amodei were among the signatories of an open letter to Congress this week calling for the institution of regulations governing requests for nucleic acid synthesis in order to prevent the production of potentially harmful bioweapons. While AI lab leaders have joined forces on other commitments, such as a pledge to prevent data centers from affecting electricity rates, the letter is the first to advocate for additional regulation downstream of potential AI-enabled activity.

Other signatories of the letter include Alexandr Wang, Paul Graham, Mustafa Suleyman, Patrick Collison, biology and biotechnology researchers, national security directors, and industry leaders in nucleic acid synthesis. The letter calls for required screening of synthesis requests, asking “providers of synthesized DNA and manufacturers of synthesis machines to check synthesis requests for sequences of concern and to verify customer legitimacy before shipping orders.” It comes after the Trump administration revoked an executive order instituting a gene-synthesis screening framework, though it said a new policy would soon be issued.

America’s First Advanced Reactor in 40 Years

On June 4th, Antares, a three-year-old startup developing compact microreactors for defense and space applications, achieved initial criticality of its Mark-0 reactor at Idaho National Laboratory. In doing so, it became the first privately developed non-light-water reactor to go critical in the United States in over 40 years, and the first to meet President Trump’s July 4th deadline under the DOE Reactor Pilot Program. The company went from concept to critical reactor in 12 months, using the same nuclear fuel developed for US Army battlefield reactor programs. Antares has committed to electricity production in 2027 and first deployments to US military installations by September 2028.

Antares hitting its timeline is an important execution proof point for the viability of SMRs, perhaps paving the way for near-term commercial-scale deployment of the technology to meet surging energy demand. The DOE’s Reactor Pilot Program also established a replicable licensing pathway that future demonstrations can use as a template for compressing the timeline from design to grid. The space remains speculative, as advanced reactor startups are yet to achieve commercial electricity generation at scale. However, this week’s announcement brings the US one step closer to a new era of nuclear generation.

Latest Research

• NewLimit wants to radically extend human healthspan by targeting aging itself rather than treating specific diseases. Read our new report here.

• Luminance offers a proprietary legal pre-trained transformer (LPT) trained on over 150 million legal documents. Read our new report here.

What We’re Reading

• Anthropic submitted a confidential draft Form S-1 registration statement to the SEC for a proposed IPO, giving the company the option to go public once SEC review is complete; the company selected Goldman Sachs and Morgan Stanley to lead the offering.

• Sen. Bernie Sanders proposed the American A.I. Sovereign Wealth Fund Act, a one-time 50% tax (paid in stock rather than cash) on OpenAI, Anthropic, xAI, and other top AI firms, which would create a public sovereign wealth fund with voting shares, board representation, and eventual cash payouts to Americans.

• Sen. Elizabeth Warren called for a new excise tax on AI data centers' energy use that scales with facility size, alongside broader tax code overhauls, framing the proposal as a way to claw back the financial gains of AI companies for working families amid layoffs, healthcare risks, and rising electricity bills near data centers.

• The US Department of Commerce closed a year-old loophole that had allowed overseas subsidiaries of Chinese AI firms to buy Nvidia Rubin/Blackwell and AMD MI350x chips without a license, with a former State official saying Chinese companies had been buying these chips “very likely at scale.”

• DeepSeek opened its first funding round for $7.4 billion at a $ 52 billion- $59 billion post-money valuation, with founder Liang Wenfeng committing 20B yuan of his own money, Tencent considering 10B yuan, and CATL 5B yuan.

• OpenAI announced that it will integrate Codex into the core ChatGPT app “in the next few weeks,” alongside six new business-focused Codex plugins covering sales, data analytics, creative production, product design, public equity investing, and investment banking.

• The European Commission drafted strict cloud-procurement criteria for its Cloud and AI Development Act, unveiled this week, that could exclude AWS, Azure, and Google Cloud from critical state tenders in banking, energy, and healthcare through “non-price” sovereignty requirements.

• The European Parliament also replaced Google with French search engine Qwant as the default on its computers starting June 4, affecting over 700 MEPs and thousands of staff, as part of a broader EU push that includes France switching to Visio from Microsoft Teams/Zoom.

• Voters in Monterey Park approved Measure NDC with ~86% support, completing the first US voter-enacted municipal data-center ban; the measure amends the city’s land-use code to permanently prohibit data centers after a 50 MW HMC StratCap facility near homes triggered grassroots opposition.

At Contrary Research, we’ve built the best starting place to understand private tech companies. If you're interested in researching and writing about tech, apply here for our Research Fellowship!

Anthropic had a big week. On Thursday, it announced it had raised $65 billion in a Series H led by Altimeter, Dragoneer, Greenoaks, and Sequoia at a $965 billion post-money valuation, overtaking OpenAI’s $852 billion March valuation for the first time and more than doubling its valuation from the $380 billion it was valued at in February. It also reported that run-rate revenue had crossed $47 billion this month. The round brought in strategic infrastructure partners Micron, Samsung, and SK Hynix alongside the hyperscaler commitments, signaling that Anthropic is now thinking about its compute supply chain at the memory and storage layer, not just GPUs. The same day, the company launched Claude Opus 4.8, an upgrade to its flagship model available at the same price as its predecessor, with fast mode now three times cheaper than prior models.

Earlier in the week, Anthropic also confirmed it will imminently roll out Mythos-class models to all customers, having made swift progress on the cyber safeguards required for general release. As we noted in a previous newsletter, the model’s public launch has been the subject of significant debate, with regulators, bank CEOs, and cybersecurity researchers all weighing in on whether the threat was real or overstated. If the rollout proceeds, it will be the first time a model of this capability class has been made broadly available, and the first real test of whether Anthropic’s Project Glasswing safeguards hold up outside a controlled partner environment.

Clean Energy for Data Centers

Elemental Impact, a nonprofit investor, launched a clean energy initiative on Tuesday in partnership with Amazon, Google, Meta, and Microsoft. Dubbed the Data Center Innovation Initiative, the program aims to accelerate the deployment of clean energy technologies in data center projects. It involves hyperscalers paying annual membership fees to cover the operating costs of new technologies, rather than each hyperscaler running its own bespoke pilots.

According to the press release from Elemental Impact, the initiative is designed to “test and validate critical technologies in data center environments” in collaboration with philanthropic partners such as Breakthrough Energy Discovery, Builders Vision Philanthropy, Salesforce, and the Stolte Family Foundation. Funding for individual projects will range from $500K to $5 million, with the initiative targeting up to 10 startups developing clean energy storage, low-carbon materials, advanced electrical systems, and industrial cooling before the end of 2027.

Private Companies, Weapons-Grade Plutonium

This week, it was reported that the DOE has selected five companies to enter into “advanced negotiations regarding the potential allocation of surplus plutonium materials” sourced from dismantled Cold War-era nuclear warheads, which will be converted into fuel for advanced reactors. If finalized, it would be the first time the US government has made weapons-grade plutonium available to private companies. The companies named in the negotiations were Oklo, Exodys Energy, SHINE Technologies, Standard Nuclear, and Flibe Energy.

The plan stems from an executive order calling for a pause to an existing program diluting and disposing of surplus plutonium from dismantled warheads to instead repurpose about 20 metric tonnes of the material as fuel for advanced nuclear technologies. The plan has drawn opposition from Senator Edward Markey of Massachusetts and Representatives Don Beyer of Virginia and John Garamendi of California, who claim that 20 metric tonnes of weapons-usable plutonium is enough for approximately 2K nuclear bombs and have asked the agency to cancel the plan. Scott Roecker of the Nuclear Threat Initiative, meanwhile, pointed out that other countries had concluded that this type of plutonium was a liability that should be permanently disposed of rather than used as fuel.

Latest Research

• Retro Biosciences wants to extend human lifespans by 10 years by developing therapies to reverse age-related disease. Read our new report here.

• The American Missile Crisis is our latest deep dive, examining a critical bottleneck in our munitions supply chain and exploring various solutions.

What We’re Reading

• The OpenAI Foundation committed an initial $250 million to study and support AI’s economic disruption across labor-market measurement, near-term worker support, and new models for distributing AI gains, part of a broader $1 billion+ annual pledge. This occurred amid OpenAI’s IPO prep and after Musk’s lawsuit accused its leaders of abandoning the nonprofit mission.

• The Illinois legislature passed SB 315, the first US law requiring third-party audits of frontier AI labs’ safety protocols, as well as 72-hour critical-incident disclosure, whistleblower protections, and AG-enforced penalties of up to $3M per violation, effective in 2028.

• Meta unveiled paid Meta AI plans under a new “Meta One” brand, with the Plus plan at $7.99/month and the Premium plan at $19.99/month, with testing starting next month in Singapore, Guatemala, and Bolivia as Meta seeks AI revenue beyond ads.

• Polymarket began rolling out identity-verification measures, including a portal for passports, licenses, and proof of residence, in response to Congressional scrutiny of insider trading and sanctions compliance, as well as Dutch regulators’ threats regarding unlicensed gambling.

• Samsung Electronics’ chip union approved a wage-and-bonus deal with 73.7% support, averting a threatened strike and creating a special pool worth 10.5% of semiconductor operating profit (plus 1.5% cash), paid in stock over 10 years contingent on profit targets and averaging ~$340K per chip-division worker.

• Elon Musk claimed US military “kamikaze” drones improperly used the civilian Starlink network instead of the government Starshield system, a “direct violation of terms of service,” while disputing a Reuters report that SpaceX hiked per-drone Starshield connection fees from $5K to $25K during the Iran war.

• The US Space Force awarded SpaceX a $2.3 billion contract to build the Space Data Network Backbone for missile targeting without ground-based relays as part of Trump’s Golden Dome, with an operational prototype due by end-2027; the Pentagon says it intends to name a second supplier by summer.

• ETH Zurich physicists generated certifiably perfect random numbers for the first time, using two entangled superconducting qubits linked by a 30-meter cryogenic tube and an improved Bell test to “amplify” imperfect randomness into bias-free output.

At Contrary Research, we’ve built the best starting place to understand private tech companies. If you're interested in researching and writing about tech, apply here for our Research Fellowship!

Recent global conflicts, from Russia and Ukraine to Iran and Israel, have seen a resurgent awareness of the frailty of US munitions stock, which has been drawn down by both direct and indirect involvement in these events. While exact stockpile volumes are not disclosed, it is estimated that supplies of US warheads and the missiles that carry them have declined by nearly an order of magnitude since their peak during the Cuban Missile Crisis. Analysts have estimated that in the event of a conflict in the Pacific between China and Taiwan, US munitions supplies could be depleted in as few as three days, with some higher-tier terminal-phase missile supplies potentially depleted in the first 24 hours of conflict.

This was a foreseeable problem. Advocates for deterrence have supported expanding munitions stockpiles and accelerating production timelines for decades. While improving technical performance (precision and range) is useful, increasing attritable mass, or the volume of munitions that can be produced and pointed at a target set per year, is the actual measure of credible deterrence. And we have a critical bottleneck problem.

Historically, the decline of US munitions production capacity has been attributed to the bottleneck imposed by solid-rocket motor (SRM) casting, which only a handful of US companies are authorized to perform as of May 2026. The limitation on expanding solid rocket motor production is not inherent to the motor mechanics, but rather the fuel these motors use for power: ammonium perchlorate (AP). AP is the oxidizer that enables high-performance SRM in the inventory, bound with powdered aluminum (fuel) in rubber to form a controlled explosive. This fuel is cured inside motor castings for multiple days in heavily regulated environments designed to prevent cracks or voids in the cured grain that can cause the motor to over-pressurize and explode.

The handling required for AP not only limits SRM production but also the production of AP itself. In the period following the Cold War, there were two primary AP producers in the US, the Kerr-McGee Chemical Corporation and the Pacific Engineering and Production Company of Nevada (PEPCON). On May 4, 1988, an explosion generated by some subset of the 9 million pounds of AP at the PEPCON chemical plant in Henderson, Nevada, caused a large fire, eventually killing two people and injuring 372 others. As of May 2026, there is now only one US producer of AP, the American Pacific Corporation (AMPAC) in Cedar City, Utah.

Because AP serves as a fundamental bottleneck to expanding SRM supply, new entrants to the munitions space won’t necessarily improve the US’s munitions production capacity, even if they are more nimble startups with more efficient processes. In fact, the opposite can be true: more demand for AP means there is less to go around, limiting production for any one company. A munitions base in which a single AP-plant accident can halt national missile production is more fragile than any other major military capability. Resolving this bottleneck by building a second, independent propulsion supply chain should be a strategic priority for the US defense sector.

One primary option for resolving this dependency is expanding scaled production of liquid-propulsion missiles, which are powered by widely available hydrocarbon fuels, high-test peroxide, and advanced engines adapted from commercial counterparts. In this piece, we outline how the missile supply chain became so brittle and why pursuing liquid propulsion is likely the best route to rebuilding a national munitions stockpile with a realistic timeframe and budget.

The Origins of Solid Propulsion

Missile fuel is a binary: it can either be solid or liquid. The choice of fuel governs a meaningful share of downstream decisions about how missiles are built and operated, including how motors are constructed, how missiles are armed for launch, and whether missiles can be shut down or throttled in flight. At the start of the US ballistic missile program in the 1940s, liquid propellants were used exclusively until the development of sufficiently performant solid propulsion fuels in the 1960s. In the decades since then, the US has transitioned to using only solid propulsion missiles, due largely to safety issues with early liquid technologies.

Liquid Propulsion Systems

Liquid-fueled missiles like the Atlas and Titan I, which were the first intercontinental ballistic missiles designed and built in the US, used cryogenic propellants such as liquid oxygen oxidizer (LOX) with RP-1 kerosene fuel. LOX boils at -183 °C, meaning it could not be stored in the missile itself because it boils off continuously and embrittles seals. On launch order, the sequence to prepare a LOX missile required 10-20 minutes before launch:

1. Pressurize and condition the propellant tanks

2. Pump RP-1 aboard

3. Begin LOX transfer from a vacuum-jacketed tank truck or fixed dewar (slowly, because flash-boiling at warm metal interfaces will rupture fuel lines); topping continues right up to ignition because of boil-off

4. Raise the missile from the silo on an elevator

5. Run a pre-flight checklist, then ignite

This timeline assumed nothing went wrong, but LOX leaks, fuel-line freeze-ups, and ignition aborts were common. Further, the warning time to launch in response to a strike from another ICBM (15–25 minutes flight time) was about the same as the required fueling time, which meant a launch-under-attack posture was essentially impossible.

However, there were some meaningful benefits to liquid systems, including the ability to assemble empty missiles with no fuel inside at scale without concern about detonation. Additionally, liquid fuel systems could be throttled or shut down after initiation, enabling more control over rocket flight.

Titan II, the ICBM that followed Titan I, fixed the readiness problem by switching to hypergolic storable propellants, which could be loaded and left in the missile for years before igniting on contact, eliminating the ignition system. However, these propellants were toxic and corrosive, some destroying lung tissue even at low ppm, and others were both carcinogenic and flammable in air. Liquid-missile crews in the early 1960s worked in RFHCO (”rocket fuel handler’s clothing outfit”) suits while setting up these systems. Liquid-missile accident reports were almost exclusively associated with propellants, and included vapor clouds, Boiling Liquid Expanding Vapor Explosion (BLEVE), ruptures, and asphyxiations.

In October 1960, a Soviet liquid-missile ICBM ignited on the launch pad, killing an estimated 78 people. In Damascus, Arkansas, in 1980, a dropped socket wrench bounced against the ground and punctured a Titan II’s fuel tank, ejecting a 9-megaton warhead into a field several hours after it was struck by the wrench due to the buildup of propellant vapors. The dangers associated with liquid missile fuels during this period ultimately motivated the US and most other global defense agencies to move towards solids. US defense leaders elected to transition to entirely solid-propellant munitions, citing storability, safety, and operational simplicity as driving factors, in the 1980s, and retired the final liquid-fueled missile, the LGM-25C Titan II, in May 1987.

Solid Propulsion Systems

Missiles with solid propulsion systems, like the Boeing Minuteman (the successor to the Atlas and Titan ICBMs), could in a silo fully fueled for their entire service lives, prepared for launch with minimal preparation and requiring no maintenance. The propellant grain was cast into the motor casing once, at the factory, and the motor was effectively a sealed appliance, requiring no plumbing, pumping, cryogenic fuels, or toxic vapors. The launch crew’s job was monitoring and authentication, not propellant handling. The launch sequence for a solid rocket missile lasted about 60 seconds:

1. Two officers authenticate the order and turn keys simultaneously

2. Silo closure door is blown off by explosive bolts

3. Igniter fires the first stage

4. Missile flies

This quick turnaround from launch instructions (no fueling delay, no raising the missile, no countdown hold for valve opening) made true launch-on-warning and ride-out-and-respond doctrines possible.

Unlike liquid engines, SRMs cannot be easily throttled, shut off, or restarted once lit. The propellant chemistry, grain shape (star, finocyl, slotted), nozzle throat erosion, and case insulation must all be modeled and qualified precisely before flight because the missile burns to completion on whatever thrust profile is baked into the geometry of the grain.

Progress in solid propulsion systems enabled the launch of the Polaris A-1 from the USS George Washington submarine in late 1960. The missile was gas-ejected from its tube by a steam generator, breached the surface, and ignited in the air. Cryogenic liquid fuel missiles couldn’t be prepared inside a submerged submarine because of the risks of explosion. Solid propellant made the submersible leg of the US nuclear triad, encompassing land, sea, and air launch, physically possible.

In contrast to liquid-missile preparation crews, solid-missile crews sat in hardened launch control capsules underground waiting for launch direction. The propellant was somebody else’s problem, finished years earlier.

Supply Chain

Despite the simplicity of operating solid-propellant missiles, the production of SRMs remained challenging, ultimately shrinking the defense contractors building these parts from six in 1995 to only two in 2017 before rising back to four as of May 2026.

Ammonium Perchlorate Production

While the contraction in SRM builders is a real bottleneck for continued missile production, the drivers of this contraction make it unlikely that further investment will be sufficient to solve this problem without fundamental innovations. Specifically, the supply of AP required to build solid rocket motors contracted during the same post-Cold-War period that demand for new missiles fell, eliminating commercial incentives for new chemical producers. The construction of the one remaining domestic AP production plant, the American Pacific Corporation (AMPAC) in Cedar City, Utah, took place in 1989, just 18 months after the PEPCON disaster. Demand was actually guaranteed jointly by NASA and the DoD in 1989 in order to ensure that AP production in the US continued. The organizations funded the construction of a second AP production site and enabled producers to add a surcharge to the per-pound price of AP, which flowed through subcontractors and prime contractors back to the government as reimbursable costs. The government guaranteed minimum annual purchases for seven years, supplying both NASA and DoD missions that would have been critically vulnerable without the added production.

Composite solid propellants are typically formulated as 68-73% ammonium perchlorate (AP, NH₄ClO₄) oxidizer and 15-20% spherical aluminum powder fuel suspended in a hydroxyl-terminated polybutadiene (HTPB) prepolymer matrix, cured with an isocyanate into a crosslinked elastomer. Batches are mixed under vacuum in vertical planetary mixers, cast into insulated motor cases around a mandrel that defines the grain geometry. Consistency and precision of grain geometry govern the ballistic predictability of missiles: the burn rate is a function of local pressure and exposed surface area, so any unintended discontinuity (like porosity, propellant cracks, or case-liner-propellant debonds) instantaneously increases burn surface and drives chamber pressure up, producing either a catastrophic burst or a transition from deflagration to detonation.

The above is an excerpt from our new deep dive on rethinking the missile supply chain. See the full report here.

Anthropic is projecting $10.9 billion in Q2 revenue, up more than 2x from $4.8 billion in Q1. Not only that, but it’s expecting to report $559 million in profit, making Q2 the first profitable quarter in the company’s history just in time for its IPO, which is expected to take place in October.

Anthropic’s pace of growth is remarkable; as WSJ pointed out, the company’s quarterly revenue is now growing faster than Zoom’s did during the early part of the COVID-19 pandemic. The biggest driver of that growth is enterprise adoption of Claude’s coding tools. As we noted in a recent perspective, Anthropic’s run-rate revenue had already climbed to roughly $40 billion annualized even before these Q2 projections, with Claude Code alone generating over $2.5 billion.

The unit economics underlying the growth are almost as impressive as the growth itself. In Q1, Anthropic spent 71 cents on compute for every dollar of revenue. In Q2, it expects to spend 56 cents. Not only is Anthropic growing quickly at scale, but its margin is improving. This likely explains the feverish demand for Anthropic shares in private markets, sometimes through less-than-reputable SPVs. Anthropic took aim at such vehicles in a post warning investors against unauthorized stock sales, where it stated in no uncertain terms: “Any sale or transfer of Anthropic stock, or any interest in Anthropic stock, that has not been approved by our Board of Directors is void and will not be recognized.”

The White House Goes Quantum

Reports emerged this week that the Trump administration is in talks to take equity stakes in quantum computing companies such as IonQ, Rigetti, and D-Wave, with each seeking at least $10 million in federal funding through the Commerce Department. The stocks surged on the news before partially reversing after a Commerce Department official said it was “not currently negotiating with any of the companies,” and a separate source said that equity stakes are “not necessarily something the Trump administration is considering.” All four companies remain unprofitable, and China’s public funding for quantum is roughly double that of the US.

This is part of a broader theme we’ve been covering surrounding the nationalization of critical tech sectors, such as the government taking a 10% stake in Intel last August and the Defense Department taking a 15% stake in rare-earth miner MP Materials the month prior, as the urgency around rebuilding rare-earth supply outside of China increases. Quantum has become increasingly urgent, with Google recently estimating Q-Day could arrive as soon as 2029. Quantum supremacy is inseparable from the race to break and defend encryption. A nation that achieves a meaningful quantum advantage over its adversaries would gain a significant edge in surveillance, intelligence, computing, and cyber warfare, making it a national security concern.

South Korea’s National AI Dividend

Last week, South Korea’s presidential chief of staff for policy, Kim Yong-beom, shared a 2.5K-word Facebook message arguing that the massive AI-era windfall flowing to Samsung Electronics and SK Hynix should be “structurally returned” to the public as a “national dividend,” writing that “The fruits of the AI infrastructure era are not the result of certain corporations alone…They come from an industrial foundation that the entire people have built together over half a century.” The Kospi fell as much as 5.1% the next day, with foreign investors pulling 5.6 trillion won from the exchange in a single session amid fears that profits at Samsung and SK Hynix would be redirected from shareholders.

South Korea’s president, Lee Jae Myung, later clarified that the post was meant as “a review of a plan to distribute the national excess tax revenue generated from excess profits in the AI sector to the public as a citizen’s dividend,” not a hardened plan. This conversation comes as Samsung union workers have threatened to strike to increase their share of AI-related profits. The scale of the potential windfall is staggering, with the two firms’ combined operating profits projected to reach 600 trillion won this year, roughly a quarter of South Korean GDP. Comparisons to Norway’s sovereign wealth fund and proposals for a windfall tax or a UBI trial have already faced criticism on the grounds that they might alienate engineering talent or investment.

Latest Research

• Lead Bank is a BaaS platform that was valued at $1.5 billion in its Series B late last year. Read our new report here.

What We’re Reading

• President Trump planned to sign an executive order on AI cybersecurity, inviting tech leaders to the White House, with a voluntary framework requiring AI labs to share covered frontier models with the government 90 days before public release, before the signing was abruptly postponed Thursday afternoon.

• OpenAI prepared to confidentially file IPO paperwork as soon as Friday, targeting an autumn debut at a potentially trillion-dollar valuation; Kalshi prediction-market odds of Anthropic listing first collapsed from 69% to 20% on the news, and CFO Sarah Friar has called it “good hygiene” to “look and feel and act” like a public company.

• Minnesota Governor Tim Walz signed the nation’s first state law banning prediction markets like Kalshi and Polymarket, with the CFTC immediately suing in federal court, joining five other state-vs-CFTC battles over whether prediction markets are “event contracts” (federal jurisdiction) or gambling (state jurisdiction).

• SpaceX plans to complete its $60 billion acquisition of AI coding startup Cursor 30 days after its expected June 12 Nasdaq debut under ticker SPCX, with a $10 billion cash breakup fee if the deal falls through.

• Google expanded SynthID watermarking from Gemini to Search, Chrome, Circle to Search, and Pixel devices at I/O 2026 with OpenAI, Nvidia, Kakao, and ElevenLabs now embedding the invisible per-pixel watermark in their AI outputs; the technology is designed to survive compression, cropping, rotation, and screenshots.

• The SEC readied an “innovation exemption” for tokenized stocks this week, allowing crypto-native platforms to offer on-chain trading of digital equities (including third-party-issued versions of stocks without the underlying companies’ consent).

• Google launched Gemini for Science, a research suite spanning Co-Scientist, Literature Insights, Computational Discovery, and Deep Think mode, with DeepMind CEO Denis Hassabis saying the company hopes to “reimagine the drug discovery process with the goal of one day solving all disease.”

• Blackstone committed $5 billion in equity to a new joint venture with Google. The venture offers Google’s TPUs as a compute-as-a-service alternative to Google Cloud with plans to bring 500 MW of capacity online in 2027, with plans to scale significantly in the future, making Blackstone both the world’s largest alternative asset manager ($1.3T trillion AUM) and largest global data center provider.

• Meta reassigned 7K employees into four new AI-focused organizations using “AI native design structures” with fewer managers per employee, days before it laid off 8K employees, or 10% of its total headcount.

• Armada raised a $230 million Series B at a $2 billion valuation to fund a 400K square foot Arizona factory (Galleon Forge One) that will produce its Leviathan megawatt-scale modular data centers that attach to existing energy sources like solar and gas flares.

At Contrary Research, we’ve built the best starting place to understand private tech companies. If you're interested in researching and writing about tech, apply here for our Research Fellowship!