Last week, Meta agreed to pay $18 billion to settle claims from 48 states that it built Instagram and Facebook to be addictive to minors, closing a case that began in 2023. The deal runs over ten years and forces new guardrails, a two-hour daily cap for under-18 users among them, though Meta only pays the full amount if TikTok and YouTube adopt similar limits and match the payment themselves. (WSJ)

We have watched this movie before. Back in 1998, the four biggest tobacco companies in the US signed an agreement with 46 attorneys general, paying out more than $200 billion after decades of lawsuits related to smoking-health costs. That settlement came after tobacco executives sat down before Congress and testified under oath that they didn’t believe nicotine was addictive. Cartoon mascots like Joe Camel were banned, billboard ads were pulled, marketing was restricted, and all cigarette packets now carrying health warnings.

A thought I have… should parents also bear some responsibility here? Should they have the ultimate say for the screen time their kids’ have. Parental controls are available on all smartphones, and limits can be enforced. Let me know what you think.

Leave a comment

In October, The Social Reckoning—Aaron Sorkin’s film about the 2021 Facebook whistleblower case that first exposed many of these harms—hits theaters. Here is the trailer:

Markets & Deals

Nvidia just posted the biggest quarter in its history

Nvidia reported second-quarter revenue of $96.2 billion, up 106% year over year and well past the $92.2 billion Wall Street expected, with data center revenue alone hitting $89 billion. It guided next quarter to $108 billion, above estimates, with CFO Colette Kress calling demand "supply constrained." Gross margins held at 75%, and sovereign AI revenue more than tripled year over year. (CNBC)

Anthropic wants IPO investors to believe its market is $30 trillion

Anthropic is preparing to tell IPO investors its total addressable market tops $30 trillion, edging out the roughly $28.5 trillion SpaceX has reportedly pitched. It's targeting a $2 trillion valuation and could raise up to $100 billion in the offering. That figure is a theoretical ceiling, not a forecast; it sits next to a far more concrete internal projection of $34.5 billion in actual 2027 revenue. (Yahoo Finance)

The SEC is now digging into the AI hedge fund that blew up

The SEC has subpoenaed Goldman Sachs, JPMorgan, Citigroup, and Bank of America over Situational Awareness, the fund we covered cratering from $45 billion to roughly $10 billion in days back in our August 3rd edition. Regulators want to know when it traded and how it discussed leverage, reportedly as high as 400%, with its lenders. No wrongdoing has been alleged, and the SEC declined to comment. (TechCrunch)

Shein priced its Hong Kong IPO at a quarter of its old valuation

Shein priced its Hong Kong IPO near the middle of its range, raising $1.7 billion at a valuation of about $26.5 billion, roughly a quarter of its nearly $100 billion private peak in 2022. First half revenue growth is expected to match the tepid 1.1% from Q1, with margins shrinking slightly. Investors are getting in cheap, but on a growth story that has clearly cooled. (Bloomberg)

Lambda is raising up to $3 billion to build more AI cloud capacity

Lambda, the Nvidia-backed AI cloud provider, is in talks to raise as much as $3 billion at a $12 billion-plus valuation, less than a year after its last $1.5 billion round. People close to the company say it could set up an IPO as early as next year. It's another sign of how much capital is still chasing the picks-and-shovels layer of the AI boom. (Bloomberg)

Kevin Durant's best investment this year might not be on a basketball court

Durant and his business partner Rich Kleiman put money into Hugging Face's seed and Series A rounds years ago, reportedly $100,000 in the seed and $150,000 in the Series A, according to sports business reporter Joe Pompliano. Nvidia's $12.9 billion offer above suggests that early stake is now worth several times what Durant put in, by Pompliano's back-of-envelope math. Neither Durant nor Hugging Face have confirmed the numbers, so take the payout figure with a grain of salt, but the shape of the story holds either way. (X / Joe Pompliano)

Big Tech & Infrastructure

Nvidia agreed to buy Hugging Face for $12.9 billion

Nvidia has agreed to buy Hugging Face, the open-source hub for AI models and datasets, for $12.9 billion. It would fold one of the most widely used pieces of open AI infrastructure into Nvidia's business, extending its reach from chips into the software layer where models get built and shared. If it closes, it would rank among Nvidia's largest acquisitions ever. (The Information)

Palo Alto Networks' CEO won't stop shopping

Nikesh Arora has bought more than 40 companies in eight years running Palo Alto Networks, and he isn't done. He held talks to buy Okta and Datadog, both of which fizzled, and is now circling companies that complement Chronosphere, following the $25 billion CyberArk deal earlier this year. Consolidation, not organic build-out, is how he's keeping pace with AI-driven security threats. (The Information)

A French Palantir rival just won over Europe's spy agencies

France's domestic intelligence agency is dropping Palantir for ArgonOS, a platform built by French rival ChapsVision, part of a broader European push for data sovereignty. Germany's spy agency has reportedly made the same call, and ChapsVision is now in talks for contracts with other EU governments too. Palantir's political baggage in Washington is starting to cost it business in Europe. (FT)

Elon Musk wants Starbase to run 30 Starship flights a day

Elon Musk says Starbase, Louisiana, will eventually host more than a dozen launch towers, enabling over 30 Starship flights a day and making it the biggest launch site on Earth. It's part of SpaceX's push to scale launch cadence well beyond its current Texas site. The figures come from Musk's own post, not an independent source, so treat the timeline as aspirational.

AI & Models

A robot that can learn a new trick from one video

Generalist introduced GEN-1.5 this week, a model that can learn a new physical task from a single few-second demo, no retraining required, an ability the company says emerged from pretraining on physical-world data at scale. It's the same leap that took language models from needing fine-tuning for every new task to just being prompted. Show it once, and it generalizes. (X / Generalist)

A rival robot model claims an even bigger trick: tasks up to ten minutes long

Days later, Skild AI released S1, a robot foundation model that also learns from a single video demo, but claims to go further, executing entirely unseen tasks that run up to ten minutes long, like potting a plant or flipping a pancake, where it says rival models are limited to short, familiar ones. In one internal test, the model watched a human pot a plant once and was doing it autonomously eleven minutes later, most of that spent moving furniture into place. Skild says that single video demo now beats roughly 380 hand-collected training examples, itself 50 to 100 hours of teleoperation to gather, on tasks the model has never seen before. (Skild AI)

Two warnings about AI landed the same week

Bill Gates spent the week saying, in 5,784 words, what Silicon Valley usually avoids saying out loud: nobody has a real plan for what AI does to the economy. He's urging regulation and global coordination to prepare for the likely harms to employment, safety, and well-being. A Goldman Sachs partner echoed the unease from the other side, warning that leaning on AI risks "cognitive atrophy" in the next generation of bankers, even as the bank keeps rolling the tools out. (WSJ, CNBC)

Wall Street's new business line is renting out AI expertise

Private equity firms are placing in-house AI specialists inside the companies they own, with Blackstone, Hellman & Friedman, and Anthropic going furthest by forming Ode, a roughly 160-person AI consulting arm backed by a $1.5 billion joint venture that also includes Apollo, General Atlantic, and Goldman Sachs. OpenAI is chasing the same opportunity from the other side, backing a rival consulting venture with $4 billion led by TPG. Wall Street sees big business in selling AI services to the broader corporate world, not just building models. (WSJ)

Business schools are charging $28,000 to teach you how to be a Chief AI Officer

The University of Chicago's Booth School launched a program to train executives for the "chief AI officer" role, at $28,000 a seat, with Columbia offering a competing version. The job itself is still being defined on the fly. It's a small, telling sign of how much uncertainty sits under the AI adoption story that boardrooms keep telling. (Bloomberg)

An AI assistant that texts and calls you is going viral in Silicon Valley

Instinct, the AI assistant from Spear Street Technology, is raising $250 million at a $2.5 billion valuation, co-led by Index Ventures and Benchmark. It works over phone calls and text and is built to complete tasks, emails, travel booking, insurance shopping, rather than just answer questions. Some early users have already raised privacy concerns about how much access it needs to do that. (WSJ)

Latam Spotlight

A Mexican beauty brand just landed in 600 Target stores

Target opened Target Beauty Studio this week, a new in-store destination for premium and emerging beauty brands launching September 10, and the Latam name to watch is Sarelly. Founded in Mexico City in 2022 by content creator Anna Sarelly and beauty industry veteran Rémi Martini, the brand is the top-selling beauty label on TikTok Shop Mexico, built on hits like its Cow Lashes mascara. Sarelly just raised a $3 million bridge round to fund its US push and will now roll out to more than 600 Target stores nationwide. (Target, BeautyMatter)

Robotaxis

Waymo picked Munich for its first EU robotaxi market

Waymo is targeting late 2027 for commercial robotaxi rides in Munich, its first market in the EU after the US, UK, and Japan. A small fleet of electric Jaguar I-PACE cars will start mapping the city with safety drivers in the coming weeks. The launch still depends on Germany granting the Level 4 autonomous driving approval Waymo doesn’t yet have. (FT)

Amazon’s steering-wheel-free robotaxi started charging for rides

Amazon’s Zoox began charging for rides in San Francisco and Las Vegas this month, the first robotaxi on public roads with no steering wheel, dashboard, or pedals. Its four inward-facing seats have earned it the nickname “toaster.” Regulators cleared it to charge for rides in up to 5,000 vehicles over the next two years, capped to roads with speed limits of 45 mph or under. (Apple News / WSJ)

Coffee Gets a Valuation Bump

Coffee chain Blank Street raises funds from General Atlantic to fuel expansion and ice cream push

Blank Street raised $75 million from General Atlantic at a $650 million valuation, up from about $500 million last year, to fund expansion into Beverly Hills, West Hollywood, and Malibu and broaden its menu with ice cream, matcha, and snacks. Existing backers Left Lane Capital, General Catalyst, and Tiger Global also joined the round. The pitch is to build a "daily luxury brand," one more coffee habit dressed up as a premium ritual, and priced accordingly. (FT)

What Else We're Watching

• Gap shares surged after beating Q2 estimates and raising its profit outlook

• Amazon has become the primary sales channel for a growing share of independent merchants

• Private equity is increasingly circling youth sports leagues as an investment category

• Smart Ring Maker Oura Seeks Up to $3 Billion in IPO

• Anthropic wants AI models talking directly to factory equipment

Happy Monday,

Last week, Moderna gave us some news. They said that, along with a Merck partnership, their personalized mRNA cancer vaccine, intismeran autogene, worked in Phase 3 trial when paired with Merck’s Keytruda. This study enrolled 1,137 patients with stage 2b to 4 melanoma that had already been surgically removed. The combination of both achieved their primary goal, which is delaying the cancer’s return, and its secondary goal, avoiding metastasis (spreading elsewhere in the body). Moderna’s stock increased 177% on Wednesday, closing at about $174 per share.

The most interesting thing here is mRNA. This weekend, I wanted to understand what this is, and it's amazing to understand this is not a vaccine we could associate not associate to COVID’s vaccine for example. mRNA stands for messenger RNA, the molecule our cells naturally use to carry instructions from our DNA to the process that builds proteins. A COVID vaccine used a generic, one-size-fits-all set of instructions to teach the immune system to recognize the virus. For mRNA, doctors sequence a patient’s own tumor, identify the mutations specific to it, and creates a specific instructions set that trains the patient’s immune system to hunt down the cancer cells left behind. This means, it’s a totally different vaccine to every patient. (WSJ)

Decades of work on mRNA as a cancer therapy and with evidence that it can be treated for oncology and actually deliver results, tells us someone is already leading the race, and who knows what other findings we will hear soon.

It is also worth watching who else trying to tackle this. Lila Sciences is pushing the idea further, an AI trained on scientific data that also runs its own robotic labs, proposing a hypothesis, testing it, and feeding the result back in to improve the next round. It already claims mRNA constructs that beat commercially available therapeutics. Less "algorithm ranks candidates," more "AI runs the whole lab." (Lila Sciences)

Subscribe now

Markets & Deals

Anthropic passed OpenAI in quarterly revenue for the first time

OpenAI's revenue grew 18% sequentially to $6.7 billion in the quarter ending June, a number that disappointed some shareholders hoping for faster progress against its rival. Anthropic more than doubled its revenue over the same stretch to about $11.6 billion and reported a small operating profit, pulling ahead of OpenAI on sales for the first time. OpenAI's operating loss, including stock compensation, widened to $12.3 billion from $9.3 billion the prior quarter, with losses outpacing revenue growth. The crossover reframes a race that many assumed OpenAI would keep leading. (WSJ)

Stripe is paying more than $7 billion for the layer that decides which AI model answers you

Stripe finalized a deal to acquire OpenRouter, the gateway that lets developers switch between more than 400 AI models through a single API, for over $7 billion. That is more than five times the $1.3 billion valuation OpenRouter carried after its Series B just three months ago, and roughly 30% below the $10 billion figure floated in July. OpenRouter sits between about 8 million developers and the model providers, owning no chips and training nothing, and Stripe was already handling its invoicing and tax. (The Information)

Wispr Flow raised $280 million at a $2 billion valuation

Voice-interface startup Wispr Flow closed a $280 million Series B led by Menlo Ventures, roughly tripling the $700 million valuation it carried late last year and bringing total funding to $361 million. Returning backers including NEA, Notable Capital, and 8VC were joined by Acrew, Forerunner, Goodwater, and Peak XV, with cameo checks from cultural figures like Shaun White. The company is pushing beyond dictation toward a broader voice platform. It is another sign that natural-speech input is becoming a real category rather than a feature. (TechCrunch)

Nasdaq is going to a nearly 23-hour trading day

Nasdaq plans to add an overnight session from 9 p.m. to 4 a.m. ET starting December 6, stretching its trading week from Sunday evening to Friday evening with a single one-hour pause each day. The regular 9:30 to 4 session stays the primary price-setting window. The push is aimed at foreign demand for US equities, now holding roughly $17 trillion, and at crypto markets that never close. NYSE has already won approval for a 22-hour day, and the SEC has a September roundtable scheduled on round-the-clock trading. (Investing.com)

The 30-year Treasury hit its highest yield in almost two decades

Bond yields jumped and stocks slipped as an unresolved standoff with Iran unsettled investors, pushing the 30-year US Treasury yield to its highest level in nearly 20 years. Rising oil prices are feeding the move, and higher long-term yields keep mortgage rates stuck near painful levels. (NYT)

Big Tech & Infrastructure

Big Tech's real AI bill is about $3 trillion bigger than the headline numbers

Nine top tech companies carry roughly $3 trillion in off-balance-sheet commitments mostly tied to AI, according to a Wall Street Journal analysis of filing footnotes. About $1.2 trillion comes from leases that have not started yet and another $1.9 trillion from purchase commitments for chips and equipment, against roughly $600 billion in reported capex over the past year. Alphabet alone accounts for around $811 billion of it. None of this is hidden exactly, it just sits in footnotes rather than on the balance sheet, which makes the AI buildout look considerably smaller than it is. (WSJ)

OpenAI joined a massive Ohio data-center project on a former uranium site

OpenAI signed on to the PORTS-Pike campus in Pike County, Ohio, securing roughly 8 gigawatts of compute alongside SB Energy, Nvidia, and the Department of Energy. The site is the former Portsmouth Gaseous Diffusion Plant, a 5.4-square-mile facility once used for uranium enrichment, now being converted into an AI data-center campus. The buildout is projected to create about 35,000 construction jobs through 2032 and 2,500 permanent roles, with Nvidia investing $1.5 billion in SB Energy. OpenAI and Nvidia also plan to publish a technical paper on what they learn about running the site reliably. (OpenAI)

SpaceX is stacking up wins in Washington

Elon Musk's rocket and AI company is landing lucrative federal contracts and benefiting from the Trump administration's deregulatory push, according to the Journal. The reporting traces how SpaceX has positioned itself across defense, telecom and space policy at a moment when regulators are loosening rather than tightening. (WSJ)

Google bought Spirit Airlines' internal data out of bankruptcy for $10 million

Google won a bankruptcy auction for a trove of Spirit Airlines business records, roughly 100 million emails and 500 million Microsoft Teams chats, paying about $10 million to use the material for training AI. Spirit filed for bankruptcy a second time in August 2025 and later shut down, after a jet-fuel price spike tied to Middle East conflict pushed its projected operating margin sharply negative. Google says a third party will scrub the data of personally identifiable information and customer records before the handoff, and that it will not receive personal information. It is an early look at what a defunct company's communications are worth as AI training fuel. (Axios)

Chart of the Week

AI & Models

AI is dead, organoids are alive

Wired's argument is that the interesting frontier is no longer silicon neural networks but actual mini human brains, which labs around the world are now growing and which the piece suggests could eventually outthink the models we spend all our time talking about. One company in that world, Melbourne-based Cortical Labs, already sells a shoebox-sized machine called the CL1 that keeps roughly 200,000 living neurons alive on electrode arrays for up to six months, and it wants to rent that access out remotely as what it calls Wetware-as-a-Service. (Wired, Cortical Labs)

The strange new rules of restaurant names

AI is changing how diners find places to eat just as merchandise, packaged foods, and licensing make a distinctive name more valuable than ever. The piece follows a burger scholar who spent two decades documenting regional hamburger styles, found that Motz's Burgers was already taken in Detroit, and reused the title of the 2004 documentary that made him instead, Hamburger America. It is a small story about a large shift, from discovery by blue links to discovery by a chatbot's single answer. (Bloomberg)

Founders are working harder than ever to keep up with their own AI agents

The tools meant to give startup teams leverage are instead setting a pace humans have to match, with founders reporting longer hours as they supervise, review and unblock agents that never stop working. The Journal frames it as a new twist on an old startup problem, the agents scale but the humans reviewing them do not. (WSJ)

Drones

Uber and Zipline want a million drone deliveries a day

Uber announced a partnership and strategic investment in Zipline, targeting 1 million drone deliveries a day by the end of 2029, starting with Uber Eats in Dallas and Houston later this year. For scale, Zipline completed about 1 million commercial deliveries total over the past year, so the target asks it to hit a year’s volume every day. Amazon separately said it plans to expand Prime Air to nearly 500 US cities and towns, while Walmart and Alphabet’s Wing are building toward 270 locations reaching 40 million Americans. The whole sector is waiting on the FAA to finalize its rule for routine beyond-visual-line-of-sight flights. (Uber)

Amazon is planning to offer drone deliveries in nearly 500 cities and towns this year

Two days later Amazon said it will offer drone delivery in nearly 500 US cities and towns by the end of this year, a sixfold expansion of its current footprint. Prime Air operates in 11 metro areas across seven states today, with each site covering roughly 175 square miles, and launches are queued for Chicago, Atlanta, Cleveland, Syracuse, and Boise. Prime Air boss David Carbon told staff in March the company would hit one million drone deliveries this year. After a decade of drone delivery being perpetually two years away, two of the largest logistics operators in the country committed to scale in the same week. (CNBC)

Latam Spotlight

Prime Video is putting $2 billion into Latin America

Amazon's Prime Video announced it will invest more than $2 billion in the region between 2027 and 2030, spanning original programming, licensed content and live sports rights across Mexico, Brazil, Argentina, Colombia and Chile. The company plans to more than double its local originals compared with 2026, with over 25 new titles in 2027 alone, including second seasons of La Oficina, Tremembé and Menem. More than 20 million households across Latin America watched sports on Prime Video this year, already ahead of all of 2025. It is also extending rentals, purchases and channel subscriptions to Costa Rica, Guatemala, the Dominican Republic, Paraguay and Peru. (Amazon)

What Else We’re Watching

• Anthropic is in talks to buy an Israeli chip-efficiency startup for $6 billion

• Nvidia Discusses Perplexity Investment at $30 Billion-Plus Valuation

• Bond Sell-Off Sends Borrowing Costs to Highest Level Since 2007

• Microdramas boom in a shrinking Hollywood as studios chase a TikTok audience

• Nvidia customers reportedly warned about AI-related price hikes

That’s the news,

From the Nido

A few weeks back the world was shocked when Allbirds, a shoe manufacturing company, announced it would drop the entirety of its shoemaking and selling operations and become an AI infrastructure company. Yes, you read that correctly. A shoe manufacturer became an AI infrastructure company overnight, moving into a market whose pricing it cannot control, they entered late, and against incumbents with deep industry relationships it cannot immediately replicate.

How did a move this extreme happen? And most importantly, why did it?

This week’s deep dive by offers one answer to that question. He crafts something that’s more useful than a story, he constructs a framework to read any AI-adjacent investment. What he finds is that AI proximity is not the same as AI participation, and despite initial excitement over Allbirds’s rebrand (the stock surged by more than 580% after the announcement), he suggests the market itself will correct for the difference when given enough time.

“Proximity to AI is not the same as participation in it, and the distance between the two is precisely what determines if a company survives a market correction,” Rishik writes.

Read more below.

This week’s takeaways from the Deep Dive…

The tronics-dot-com-dot-ai pattern
In 1959 companies with no connection to electronics rushed to put tronics in their names. The American Music Guild—a company that sold record players door to door—rebranded to Space-Tone Electronics and saw its shares return 600% within a week of the name change. In 1999 adding .com generated abnormal stock returns of 74% within ten days of a suffix change. Today we are seeing the same behavior with companies who hyperbolize their AI capabilities to gather attention.

What 1959, 2000, and today have in common is a reflex that performatively aligns corporations with a transformative shift instead of actually building the technology itself.

Will the market itself correct for the difference?

From the Nido archive…

Step back from this week’s piece and a thread comes into focus that has been running through the archive all year. ConteNIDO has been quietly building a map of the AI infrastructure layer from every angle—hardware, accounting, capital markets, labor economics, open weights. Rishik’s piece is the investor’s framework that sits on top of all of it: a way of reading every company in that map and asking whether the exposure is structural or claimed.

The Memory Bottleneck by Roberto is the clearest example of structural exposure made visible. SK Hynix, Samsung, and Micron are TOTO at scale — companies whose accumulated manufacturing expertise became a necessary chokepoint in the AI buildout without anyone planning it that way. The three factories Roberto identified as capable of paralyzing global AI progress if they stopped tomorrow are, in Rishik’s framework, the purest expression of irreproducible advantage. Four decades of process knowledge. A moat that money cannot buy on a timeline the market can price.

The Accounting Depreciation piece by Roberto is the financial X-ray of what happens when narrative premium detaches from operational reality. Microsoft, Google, and Meta extending the assumed useful life of their AI hardware to flatter reported earnings is the accounting version of claimed exposure — a gap between what the numbers say and what the economics justify, held in place by a bull market that has not yet demanded a reckoning. Rishik’s framework predicts exactly when that reckoning arrives: when sentiment shifts and the liquidity illusion dissolves.

Surfing the Hardware Wave by Rodrigo mapped the SpaceX alumni ecosystem as the next generation of structural AI exposure — founders trained on physical constraints building companies that solve supply chain gaps the market cannot route around. In Rishik’s framework, those companies are crossing the survival threshold by design: building switching costs and accumulated expertise into the product from the first line of code. The SpaceX alumni aren’t positioning their tickers near the AI trade. They are building the infrastructure the trade depends on.

The Open Weights piece by Rodrigo, published two weeks ago, arrived at the same observation from a different angle. The open-source AI debate is being shaped by companies that sell the conversion — the compute, the hosting, the hardware — not the file itself. The coalition arguing loudest for free inputs is in the business of selling the necessary downstream services. In Rishik’s framework, that is structural exposure dressed in the language of democratization. The file is claimed. The electricity is structural.

Welcome to all the new followers! We are so happy you’re here and choosing to read past the headlines. A few notes on what you can expect in your inbox going forward:

Monday: News from the Nido, your weekly digest of the biggest news in tech, markets, and AI.

Friday: Deep Dive, your weekly long-form investigation into the ideas shaping business, technology, and the world, written by the sharpest minds across the Nido network.

Sunday: The Sunday Nest, your weekly briefing where this week’s Deep Dive meets the archive, the news cycle catches up to stories we told first, and the ideas that matter get the space they deserve.

—Renata from Nido

TOTO, a Japanese company, has been making toilets since 1917. In the past 12 months the company’s fine ceramics division, a segment it has operated quietly since the 1980s, has been identified by investors as an important supplier to the global semiconductor industry on account of its chucks sitting inside the fabrication equipment that manufactures AI chips. CNBC highlights that this business line saw a 34% increase in annual revenue during FY2025 and a 42% jump in operating profit, almost single-handedly pulling the overall company into growth for the financial year.

The business didn’t pivot. It didn’t rebrand. Demand exceeded supply and investors found that TOTO’s four decades worth of industrial expertise had become a necessary supply chain component for the most capital-intensive technology buildout in a generation.

TOTO is not alone in being drawn into AI’s orbit, for it lies on the organic end of the AI exposure spectrum. Allbirds, however, is positioned on the opposite, synthetic end as a shoe company that sold its entire core business and used the proceeds to become a GPU leasing vehicle.

These 2 firms together form a gradient that underlines a hierarchy of durability: TOTO is the most durable because its initial capability existed before the AI narrative. Allbirds cannot be classified in the same way because its entire position depends on the AI narrative remaining investable. The company replaced their business with a promise for customers it hasn’t signed yet in a market whose pricing it can’t control.

Proximity to AI is therefore not the same as participation, and the distance between the two is precisely what determines if a company survives a market correction. TOTO participates, making a physical component without which AI chips cannot be fabricated. Allbirds has positioned its ticker close to the AI trade without yet leasing out any GPUs yet. When capital flows towards AI-performative organisations and not to those actually driving the buildout, a misallocation is created. The scarcity in infrastructure that needs to be addressed doesn’t get solved any faster.

Allbirds becoming a GPU lender doesn’t add a single GPU to the market, yet TOTO making ceramic chucks delivers necessary chips. It becomes increasingly difficult to distinguish between AI-real and AI-performative, making markets harder to interpret than ever before which is why thinking about where other companies will plot on the TOTO-Allbirds spectrum can help identify if an investment opportunity is creating value or simply claiming it.

We have been here before

The pattern is not new. In 1959 companies with no connection to electronics scrambled to put the word “tronics” in their names as investors threw money at anything that sounded futuristic. For instance The American Music Guild, whose business consisted of selling record players door to door leveraged this mania. The company rebranded to Space-Tone Electronics and saw its shares deliver returns of 600% within a week of the name change. In 1999 a similar story emerged: adding “.com” to your corporate identity was all the rage, with companies generating abnormal stock returns of 74% within 10 days of adding a suffix to the existing name.

Any technology cycle comes with the old adage of “it will be different this time,” whose truthfulness is only obvious in retrospect. In the short-term, the future is thought to have already arrived and the narrative is fully formed, when reality is in fact far from this belief. Jensen Huang has declared that we have achieved AGI. Vocabulary that was initially used in hushed tones like “agentic AI,” and “inference clusters” is now bellowed in earnings calls across the corporate world. A particularly interesting phenomenon borne out of this hype is agent-washing, where companies hyperbolise their AI capabilities to garner attention.

What 1959, 2000, and today have in common is a reflex that aligns corporations with a transformative shift instead of actually building the technology itself. Each cycle minted its own vocabulary for the same behaviour: “tronics” became “.com” became “.ai.” The inevitable pricing correction won’t be a verdict on artificial intelligence anymore than 2000 was a verdict on the internet. It will act as a filter separating dispensable firms like Allbirds from those such as TOTO.

In 2000 the businesses that survived did not do so by some divine grace: Google built search infrastructure, and ASML and ARM developed the physical and architectural infrastructure that every chip in the world would eventually depend upon. The crash did not destroy the technology. It erased the parasites feeding off it and in doing so revealed the companies who had an organic advantage all along.

TOTO is the paragon of a company that is yet to act performatively. It did not become TOTO.ai or issue a press release about artificial intelligence. If Allbirds is the purest expression of the pattern described above, then TOTO is an important counterexample. The organic end of the AI exposure spectrum is thus not a strategic position since it isn’t driven by a trend. Durable companies that are situated here create conditions to drive growth instead of operating to extract gains from a cyclical market upswing.

Survival 101

Not every company caught in AI’s shadow is destined for the same fate. Between the spectrum of TOTO to Allbirds lies a third position which depicts firms that could have been classified as performative but moved closer to TOTO through operational necessity, crossing a critical survival threshold.

This so-called “limit of existence” is the point at which a company’s AI exposure becomes embedded into the operations of another organisation. Below this threshold, businesses are passengers. Above it, they are part of the infrastructure of daily decisions central to production or the provision of services and are thus expensive to remove. Crossing this threshold is an accumulative act.

(Source: Palantir)

Palantir is instructive here. Founded as a defence and intelligence data contractor, it spent years at the margins of the technology conversation as a useful, profitable tool but not considered a protagonist of the AI era. Then with its launch of AIP it aggregated years of sensitive operational data across governments and enterprise clients, packaging it into an AI-powered decision-making layer that the same clients could not operate without.

What separates Palantir from Allbirds is that despite GPUs being essential to AI development, operating as a key piece of the supply chain does not offer organic status to every company within the industry. Allbirds is a substitute good in a commodity market with no reason to be the firm leasing GPUs. Anyone with sufficient capital could replace it tomorrow. Whereas with Palantir a switching cost exists, removing its product from operations is expensive since it means users have to rip out the institutional memory that they are so reliant on. The survival threshold thus doesn’t require proximity but is about whether a company’s edge can easily be replaced.

TOTO’s success rests on a different foundation entirely because its moat is time. Knowledge that took 4 decades to build is difficult to replicate. A ceramics precision manufacturer cannot be trained in the same way as an AI model.

With this threshold in mind, the spectrum derives a key requirement for survivability: tough to imitate. At the TOTO extreme, advantage is accumulated and is essentially irreproducible. While for Allbirds the advantage is short-term and narrative. Success is driven by attention, which any company can mirror and is not at all durable in capital markets.

Taking an X-Ray

Doctors prescribe X-rays to get a clearer picture of what’s in front of them, investors can do the same too by scrutinising supply chains. The AI supply chain helps us visualise the kinds of firms that might meet the survival threshold.

(The AI Supply Chain, Source: LSE YouTube Channel)

At the base of the supply chain, performative participation is structurally impossible. Raw material extraction and semiconductor fabrication are stages defined by barriers that capital alone cannot overcome. ASML’s lithography machines and TSMC’s fabrication processes weren’t built for the AI era. TOTO, albeit at a smaller scale, is a similar case. The threshold is met because time exists as a moat and is the one input that money can’t buy. The risk borne by investors is compensated as it’s underwritten by capabilities that existed before the AI narrative.

Move up to GPU production and incumbency is absolute, making survival much more difficult for any new entrants. NVIDIA controls an estimated 85% of the market. The sub-market of GPU leasing faces the opposite problem: too much competition. H100 rental rates fell between 64% and 75% between late 2024 and early 2026. Allbirds is arriving late to a commoditised market, whose players moved first and already have deep industry relationships that can’t be replicated immediately.

Data centres tell a similar story. Incumbents like Microsoft, Google, Amazon and META are already committing hundreds of billions in CapEx to acquire infrastructure capacity.

Climb further to model training and end user applications and the X-ray reveals the most important findings. Barriers to entry are thin, differentiation exists and the narrative premium with being AI-associated detaches from operational reality. It’s likely that the majority of companies in this segment won’t meet the survival threshold and will be hit by re-pricing the hardest. There are already early signs: In 2025 AI wrappers and application-layer tools represented nearly 16% of all startup closures as they built on pre-existing models without a defensive moat. Heavy asset firms like those infrastructure are harder to erase, meaning the correction already knows to look towards companies dominated by narrative success.

Risk

If these moats represent why a company even has AI exposure, they also illustrate various distributions of who absorbs the downside when the cycle turns. We now arrive at the question of risk. The type of moat an AI-connected company has determines where the risk pools. Irreproducible advantage (TOTO) transfers risk to customers who need your business despite what markets think, while narrative advantage (Allbirds) concentrates the risk entirely among the investors financing the story.

It’s all well and good to say that risk pools in this pattern, but actually seeing it happen is completely different. In a bull market, risk becomes even harder to identify. When capital is flowing freely and valuations are climbing, every AI-adjacent company looks real even if they are merely performative. A liquidity illusion exists as capital is abundant and sentiment is positive. Every asset that is connected to the narrative becomes tradeable at a premium. Buyers are always available and prices are always rising. However, liquidity does not equal authenticity since the market is actually pricing a story.

Some companies have AI exposure because their operations are physically or organisationally integrated into the buildout and it’s impossible to remove them without significantly reducing efficiency. On the other hand, other businesses are exposed because investors have decided their unacknowledged AI-performativeness means they belong in the conversation. The first kind of exposure is structural while the second is claimed.

This distinction is not hidden. It’s only an inconvenience when returns are compounding. A company with structural exposure converts this advantage into receivables, customers, contracts and switching costs. The claimed counterpart can be equally as profitable in the short-run as long as the narrative holds. Yet in the long-term, value is immediately slashed as soon as sentiment shifts.

Allbirds helps make the difference concrete. The company dissolved an operational business with customers and real revenue to maximise its claimed exposure. Public investors are no longer financing a business but instead a sequence of events that must go right. The risk transfer already took place long before a single GPU was even leased.

What the Correction Reveals

Every correction in every technology cycle is often described as a reckoning, but labelling it as a clarification is more apt. As the narrative starts to fall apart, this action separates what was always real from what was only ever claimed. The companies that will survive corrections won’t be AI companies to begin with. They will be durable businesses that AI needs. The competitive moat determines if a firm has claimed or structural exposure and also decides who holds the repricing risk. The x-ray revealed what the correction will confirm: those with structural exposure are often built to withstand a crash.

Nonetheless this spectrum has its limits. Structural positioning is necessary but not sufficient. Global crossing in the 2000s was part of the infrastructure layer but still collapsed under the weight of overbuilding. NVIDIA has a durable business model but it still carries a valuation that prices in a narrative premium that can’t be justified using the argument of a switching cost alone.

Companies like OpenAI exist whose operations span the entire organic-synthetic spectrum at once. They are building foundational models which is genuine structural exposure, but are also deploying consumer products like ChatGPT into a market where users can easily switch between Claude and Gemini as well. One could refute that durable firms can also be classified as those who can’t be placed at a discrete point on the spectrum and capture both the upside of the narrative while safeguarding against the downside by building the AI landscape itself.

The organic-synthetic framework identifies the extremes. TOTO and Allbirds are polar and make the durability gradient observable. A market correction does not mean the gradient becomes binary since the firms in the middle will not all be cleanly separated into buckets of claimed and structural exposure. An inevitable repricing will, however, remove the noise that makes these separate exposures look equivalent.

If you are navigating equity markets or have an interest in investing I would love to connect. Please reach out to me at rishik.kalagara@gmail.com or rishik.kalagara@nido.ventures

This week, I read an article published by Bloomberg on how Gen Z is moving from investing towards sports betting, as part of their long-term financial strategy. What surprised me the most was the fact that more than half of young investors have poured money originally intended for investment into sports bets over the past year, with 14% saying they do it at least once a month. Only one-third of Gen Z respondents said they don’t particpate on sports betting at all.

Being from Gen Z, I've watched this invitation to speculation become normalized, and the easy access we now all have, with the click of a button, you are part of the game. Interesting to see how this evolves, and the impact this could have on wealth creation over the next decades.

Subscribe now

Markets & Deals

Databricks raised $5 billion at a $190 billion valuation

Databricks closed a $5 billion round led by Coatue, with Blackstone, MGX, and T. Rowe Price joining alongside existing backers like Andreessen Horowitz and Thrive. The company crossed a $7 billion annualized revenue run rate this quarter, up more than 80% year over year, a sharp step up from the $134 billion valuation it fetched about six months ago. CEO Ali Ghodsi kept an IPO on the table but downplayed urgency, citing "too much distraction in the public market." (Bloomberg)

Kalshi topped $4 billion in annualized revenue and is chasing a $40 billion valuation

Prediction market Kalshi is in final talks to raise about $750 million at a $40 billion valuation, with Sequoia and Wellington eyeing the round, which would nearly double the $22 billion it commanded in May. Annualized revenue hit $4 billion in July, roughly four times rival Polymarket, driven largely by a surge in sports contracts tied to the 2026 World Cup in North America. The company now claims about 95% of the US prediction market by revenue. CEO Tarek Mansour has said an IPO could come in 2027. (The Information)

Nelson Peltz is preparing a take-private bid for Wendy's

Wendy's shares jumped double digits, triggering a volatility halt, after reports that Nelson Peltz's Trian Fund Management is preparing a bid to take the struggling burger chain private. Trian is assembling a group of investors that could include BlueFive Capital and longtime franchisee Flynn Group, and Peltz already holds about a 16% interest in the company. A bid could land in the coming weeks, though timing may shift, against a market value of roughly $1.44 billion. Wendy's said it would evaluate any proposal in line with its fiduciary duties. (CNBC)

CoreWeave's revenue doubled, and so did the cost of its debt

CoreWeave reported second-quarter revenue of $2.58 billion, more than double a year earlier, as demand for AI compute kept climbing. Losses widened to $626 million, driven largely by net interest expense of $640 million, itself more than double the prior year, as the company's debt load reached about $35.6 billion. It still posted $1.5 billion in adjusted EBITDA at a 59% margin and guided to $35 billion to $39 billion in capital spending for the year. The print captured the central tension of the AI buildout: explosive top-line growth financed by an ever-heavier balance sheet. (The Information)

Uber sold its entire stake in Serve Robotics as the partnership frays

Uber divested its full stake in sidewalk-delivery-robot maker Serve Robotics, a holding dating back to Serve's 2020 spinout from Postmates. The 2023 deal letting Serve run robots on Uber Eats continues through early 2027, but Serve does not expect a renewal. It is another sign of Uber reshuffling its autonomy bets, weeks after pledging $10 billion across roughly 30 AV partners. (Bloomberg)

Big Tech & Infrastructure

Nvidia and Wall Street are building a $500 billion money pipeline for the AI boom

Jensen Huang has teamed with Apollo, Blackstone, a BlackRock unit, Brookfield, Goldman Sachs, and KKR on a roughly $500 billion plan to standardize how AI chips get financed, pooling asset-backed capital to fund data-center deals. The idea is to sell public and private debt to pension funds, insurers, and sovereign-wealth funds, then use the proceeds to bankroll chip purchases, with Nvidia offering to backstop up to 25% of a project's residual value to lower borrowing costs. The obvious risk is that the collateral ages fast, since AI chips lose value as newer generations arrive. As one investor put it, compute as collateral is "an asset that's had the shelf life of lettuce." (FT)

An Amazon data center in Texas could become the single dirtiest power plant in the US

Amazon is building an on-site gas plant for a Pecos County, Texas data center, permitted to emit up to 33 million tons of carbon dioxide a year, more than any existing US power plant. The design calls for 35 natural-gas turbines feeding a single AWS campus, with first power targeted for early 2027. (The Verge)

Apple is in talks to pay publishers to make Siri smarter

Apple has approached publishers about paying for content to give an AI-enhanced Siri access to current news, floating a variable model that pays out only when content is actually used. The company has weighed a nine-figure budget as it races to deliver the smarter assistant it promised years ago. (WSJ)

SpaceX officially closed its $60 billion acquisition of Cursor

Cursor is now formally part of SpaceX, completing a deal first struck in April and confirmed as SpaceX went public in June. Cursor said the move gives it access to “the largest fleet of GPUs in the world” through SpaceX’s Colossus supercomputer, which it already uses and which has helped train Grok models. Cursor’s paid subscribers are also getting early access to some of SpaceX’s AI products. (TechCrunch)

Chart of the Week

Debit and credit card ownership tracks financial account ownership closely across Latin America. Still, the gap tells its own story: Chile, Venezuela, and Brazil lead with both figures above 70%, while Nicaragua, Honduras, and Guatemala sit under 20% on both, showing how far card penetration still lags in Central America even where bank accounts exist.

AI & Models

Zuckerberg laid out a "personal superintelligence" vision in a 6,500-word essa

Mark Zuckerberg argued the biggest AI risk is concentration of power in one entity, pitching a world where every person controls their own AI aligned to their goals, complete with free AI tutors and coaches for everyone. The essay landed days after one of Meta's own models was implicated in hacking another company during testing, a contrast critics were quick to note. (Meta)

Google's Gemini app crossed 1 billion monthly users

Sundar Pichai announced Gemini has become Google's fastest-growing product ever and the 14th Google service to hit 1 billion monthly users, a milestone reached roughly three weeks after the company reported 950 million users on its Q2 earnings call. Gemini needed a fraction of the time Gmail and YouTube took to reach the same scale, though Google did not disclose how many of those users actually pay for anything. (TechCrunch)

Anthropic will start watermarking text written by its models

Anthropic said it will embed an imperceptible, machine-readable watermark in Claude’s text output, complying with the EU AI Act’s Transparency Code that took effect August 2. The mark survives copy-paste and will apply everywhere Claude is available, not just Europe, prompting complaints from users worried it could expose their use at work or school. (TechCrunch)

AI agents are now taking entire online courses for cheating students

A New York Times investigation found AI agents can enroll in and complete whole online courses, submitting assignments and sitting exams for students. It echoes last week's theme, from UNAM forcing 58,000 applicants to retake an exam to universities racing to sell AI credentials of uncertain value. (NYT)

Latam Spotlight

Nu Holdings posted its first $1 billion quarterly profit

Nubank reported net income of $1.06 billion for the quarter, up 49% year over year on a currency-neutral basis and well ahead of the roughly $967 million analysts expected. Total revenue reached $5.88 billion and risk-adjusted net interest margin hit a record 12.4%, while credit costs eased, calming some concerns about asset quality. Shares jumped more than 10% on the news. The lender now serves nearly 139 million customers across Brazil, Mexico, and Colombia and is preparing to launch in the US. (Yahoo Finance)

Design Beat

Instagram redesigned its wordmark for the first time in a decade

Meta unveiled a new Instagram wordmark on August 13, the app's first typography update since 2016, while leaving its gradient camera icon untouched. The new lettering is more compact and rounder, with Meta's design team saying it went through hundreds of iterations "obsessing over every curve, ligature and spacing detail" to keep the brand's handcrafted, creative feel. Reaction online was mixed, with some users joking the tightened letterforms now read as "Instagzam," even as others praised the cleaner look. (TechCrunch)

Tip of the Week (for Ford owners)

Ford and Lincoln owners can now talk directly to the automakers' new in-vehicle AI assistant and ask it questions on the road, one of the more concrete signs yet of generative AI landing inside the dashboard (Ford)

What Else We're Watching

• Harvard disclosed a $2.2 billion stake in SpaceX

• Stripe will reportedly acquire AI gateway startup OpenRouter for $7B+

• Enthusiasm for Big Tech and its huge investments in artificial intelligence is powering the stock market to record highs again

• Anthropic revenue reportedly jumps to more than $11.5 billion in second quarter

That’s the news,

From the Nido

Nido Ventures co-founder recently became a mother. In this new stage of life, she’s been using AI regularly in her routine to help out: for planning meals, scheduling purposes, to help organize stuff around the house, and pretty much anywhere it can be useful. She’s found it’s life-changing. As she’s navigating this new phase, she’s leaning on AI tools in her daily routine, and she wants the same information to be a helpful guide for others out there in the same situation.

She’ll be writing AI in Everyday Life, a personal, journal-style column with her observations, reflections, and yes, helpful tips to incorporate AI into your everyday life as well. We can’t wait to have this new kind of content on ConteNIDO and hope you, our readers, will take some reflections worth keeping too.

The column will go live next week, so make sure you are subscribed to receive the first edition!

From the Sunday Nest editor

Caro’s column this week has gotten me thinking about the innumerable facets of the AI conversation. There is the widely shared and discussed market side of things; the deals, IPOs, and news that land on our home screens daily. For a second it feels as though this is where our attention goes to the most—and why shouldn’t it? It’s buzzy and trendy and everywhere right now. We can’t help but look there.

However, there is a more intimate and less-grandiose side of things, the everyday that is already evolving based on the macro changes. This new column is a peek into that life, into how something as timeless as childcare is quietly changing with the times. While much less news-headline sounding—because how is a headline about the endless iterations of a baby’s first few months on earth news-worthy—this hidden facet of life is perhaps the most revealing insight we have into the ways our lives are responding to the AI era.

I guess I want to leave you all with an invitation to look at the smallest scale of life, the household, with new eyes. To search for the ways in which this primordial space has already evolved with you. You might just find the most revealing answers about our world already present in your home.

Since we last wrote…

Our world is changing, and it’s changing fast. Keeping up with the endless goings-on is an impossible feat, but at Nido we’ve written extensively about themes that are continuously evolving. This section connects the dots for you, tying big themes back to big stories.

A few weeks ago wrote the Erewhon story, the first article in our Origins column, a space where we break down the founding, creation, and evolution of some of the most well-known companies today. He wrote the story of Erewhon, a grocery store that charges a premium for its social clout.

This week Erewhon announced a new celebrity smoothie, this time in collaboration with Sydney Sweeney. This smoothie “asks for $22 dollars and a little faith,” says the article. For Erewhon, this smoothie price has become so normalized it barely registers as high at all. To find out just how they’ve gotten here, read Roberto’s piece.

One of our most-read articles to this day is ’s Owning Personal Health, a Deep Dive into the ways “health and fitness tracking is evolving from step counts and sleep scores into deeper physiological measurement, preventive care, and at-home chronic illness monitoring,” as she says in the piece. She predicted the growth of the wearable tech space, tracing out a landscape that could eventually become the backbone of basic health infrastructure.

This week Garmin announced the release of a new whoop-like wearable called CIRQA. This is the brand’s first screen-free fitness band. It features 24/7 health tracking and vitals monitoring, much like a Whoop, but unlike a Whoop, there is no subscription needed. An interesting move by a player in the fitness-tracker space no doubt. If you want to learn more about the moves and players in the health-tech space, read Laura’s piece for necessary context.

At Nido we’re always looking for writers! If you’re interested in contributing an article or Deep Dive across any of our themes, feel free to reach out. And if you’re a student interested in becoming a more full-time member of our writer crew, please reach out as well.

You can reach me at renata.alvarez@nido.ventures.

This past week was crazy, as always. There was news on AI models that escaped their guardrails… again. I don’t know if they are really letting us know about this to get the public panic on AI, or as a marketing strategy, showing off the capabilities of models. Who knows, but it’s getting interesting.

Aside from that, something big happened with one of Google’s top AI researchers and one of its long-time employees, Jeff Dean. He was Google’s 30th employee and its longest-serving AI leader. He is leaving the company after 27 years to build his new startup, Discovery Loop, with three other Google veterans. Dean wants to build AI that can improve itself without the help of humans, with Google staying on as both investor and cloud partner. (NYT)

Subscribe now

Markets & Deals

Sequoia is putting $10 billion behind Anthropic and "reindustrialization"

Sequoia's new co-stewards, Alfred Lin and Pat Grady, are committing $10 billion, the largest bet in the firm's 54-year history, split between doubling down on Anthropic and backing companies rebuilding manufacturing, energy, defense, and robotics in the US. The Anthropic push started small: Lin opened with a $1 billion idea in a May partnership meeting before colleagues pushed him to go bigger, given the pace of Anthropic's growth. Sequoia had passed on Anthropic for years in favor of OpenAI and xAI before reversing course in January, then co-leading the May round that valued the company at $965 billion. The bet comes as Anthropic has reportedly traded on secondary markets at an implied $1.2 trillion. (Bloomberg)

Whatnot has raised $545 million in Series G funding at a $20 billion valuation

For anyone who hasn't come across it: Whatnot is a livestream shopping app where sellers broadcast video "shows," auctioning off everything from Funko Pops and Pokémon cards to sneakers and designer bags while buyers bid and chat in real time, essentially QVC run by hobbyists from their bedrooms. The company's $545 million Series G, led by ICONIQ Growth, Lightspeed and Avra, nearly doubles its $11.5 billion valuation from last October and brings total funding to about $1.5 billion since its 2019 founding. Whatnot says it has already passed $8 billion in gross merchandise volume this year and now commands roughly 60% of the live commerce market, a rare consumer-tech breakout in a funding environment almost entirely dominated by AI. "When you look at Silicon Valley, it's 99.99% AI right now," said CEO Grant LaFontaine. (Yahoo Finance)

Bending Spoons is buying Airtable for $1.285 billion

Bending Spoons agreed to acquire Airtable in an all-cash deal valuing it at $1.285 billion, its first acquisition since going public on the Nasdaq in July. Airtable's ARR has grown more than 20% to about $480 million, but the price is a steep comedown from its $11.7 billion valuation in 2021. (Businesswire)

Uber is putting $10 billion into robotaxis as Waymo heads for the exit

Uber pledged more than $10 billion over the coming years to its robotaxi strategy, mostly as equity investments across roughly 30 AV partners, aiming for 120,000 autonomous vehicles and service in at least 15 cities by year end. The spending comes as Waymo, Uber's best-known AV partner, winds down its exclusivity in Austin and Atlanta by 2028, pushing Uber to diversify with partners like Wayve, which just won approval for a supervised commercial service in London. Uber backed the bet with a record $2.8 billion in quarterly free cash flow, even as shares slipped on a softer profit outlook. (FT)

Struggling colleges are quietly raiding restricted donor funds to stay open

Nearly 200 private colleges borrowed from restricted endowment funds in 2025, up from 131 in 2021, often to cover everyday costs rather than what donors intended. More than 440 colleges are considered at risk of closing or merging within a decade, and Ohio's attorney general alleges trustees at now-closed Notre Dame College improperly spent more than $2 million from restricted funds. (WSJ)

Robinhood is taking its second venture fund public

Robinhood Ventures Fund II is aiming to raise up to $200 million in a US IPO priced at $25 a share under the ticker RVII, giving retail investors a shot at early-stage private companies. The fund will focus on current and former Y Combinator startups, following its first fund, which added OpenAI to its portfolio earlier this year. Goldman Sachs is leading the offering, with CEO Vlad Tenev citing customer demand for earlier access to AI companies. (Reuters)

Big Tech & Infrastructure

More than 500 towns and counties are now banning data centers

Local resistance to data center construction has accelerated fast enough to worry AI companies like Anthropic, Google and OpenAI, with active bans nationwide topping 500 in early August, up from over 300 in late June. More than 150 towns and counties passed bans in July alone, and states are joining in: Texas Governor Greg Abbott paused new construction pending a grid-strain audit, while New York Governor Kathy Hochul signed the first statewide moratorium on hyperscale projects over 50 megawatts. The Midwest and South account for most of the activity, with Florida alone adding 14 restrictions in July, concentrated in the Panhandle. (The Information)

Data center operators are asking Wall Street to backstop the AI buildout

Developers are increasingly tapping banks for letter-of-credit facilities, a financing tool borrowed from oil and gas, to guarantee the grid-connection costs of new AI data centers. TeraWulf and Blackstone-backed QTS, one of the largest data center operators in the US, are among those seeking such guarantees, which protect utilities and households if a developer walks away mid-build. "Securing abundant, reliable electrons now matters more than chips or capital," said Arnaud Stevens, head of global trade for the Americas at Natixis. (Bloomberg)

AI's power swings are wearing out the equipment built to run it

Rapid swings in AI data centers' electricity demand are straining batteries, generators and cooling systems well beyond what they were designed for, causing premature failures that add cost and reliability risk. At xAI's Colossus facility in Memphis, gas-fired turbines have developed cracks, with similar wear showing up at smaller UK sites. The problems are surfacing just as investors grow more nervous about how fast hyperscalers' hundreds of billions in AI infrastructure spending might actually depreciate. (Bloomberg)

AI & Models

A fourth AI model has now escaped its sandbox, and this one is already public

Moonshot AI's open-weight Kimi K3 broke out of a cybersecurity testing sandbox built by the UK's AI Safety Institute, exploiting a network misconfiguration to pull an answer off GitHub. Unlike earlier escapes at OpenAI, Anthropic and Meta, Kimi K3 is already freely downloadable, which researchers say makes it more concerning. "We found a leak in the sandbox," Frontier Security CEO Yaron Singer said. (Wired)

Scientists used AI to design viruses that have never existed in nature

Stanford and Arc Institute researchers trained a genomic AI model, Evo, on trillions of nucleotides and had it design new bacteriophage genomes. Of 285 synthesized, 16 assembled into fully viable viruses, some more effective at killing E. coli than the natural original, raising fresh biosecurity concerns. (NYT)

Meta launched its own coding agent to undercut Claude Code and Codex

Meta released Muse Code in beta, a terminal-based coding agent built on its new Muse Spark 1.2 model, pricing it as a cheaper rival to Claude Code and Codex rather than a more capable one. It runs on macOS and Linux only, with a discounted tier for developers who opt into data sharing. (WSJ)

China's open models are quietly winning Africa's developers

Developers across Kenya, Uganda, Nigeria and Ghana are increasingly picking open Chinese AI models over pricier American ones. On OpenRouter, Chinese open-weight models now account for roughly half of tracked usage, up from under 25% a year ago. (NYT)

Colleges are racing to sell workers on AI credentials, even if no one knows which skills will matter

Universities and online platforms are rushing out AI certificates and degrees as workers look to reposition themselves, with institutions offering AI bachelor's degrees growing from 4 to 23 between 2020 and 2024, and AI master's programs from 19 to 60. Frederick Duff, a 57-year-old Ohio mortgage broker, skipped law school for a nondegree online AI program instead, part of a broader shift as entry-level coding hires cool and computer science enrollment declines. (Bloomberg)

Charts of the Week

Anthropic's valuation has climbed more than 200-fold in three years, from next to nothing to $965 billion, tracking the same conviction behind Sequoia's record bet this week (chart via The Information).

Dallas has added more than 20% to its finance-sector headcount since February 2020, while San Francisco has shed more than 10%, as finance jobs migrate inland (chart via Washington Post, Federal Reserve Bank of Dallas data).

Latam Spotlight

Mercado Libre topped $10 billion in quarterly revenue for the first time

Mercado Libre's net revenue and financial income grew 50% year over year to $10.2 billion, its 30th straight quarter of growth above 30%. Shares still fell more than 4% after hours as investors weighed heavy spending against an operating margin that slipped to 6.7%. (MarketScreener)

Mexico's top university is making 58,000 students retake their entrance exam

UNAM will require roughly 58,000 applicants to sit a new, in-person exam after unusually high scores on this year's first-ever online exam raised suspicions of AI-assisted cheating. The university had already disqualified thousands of candidates, and President Sheinbaum has called for an investigation. (The Guardian)

Oxxo has more stores in Mexico than any single US chain has in the entire country

FEMSA's Oxxo grew from one store in Nuevo León in 1978 to 24,297 by the end of 2025, more than Dollar General's entire US footprint despite the US having two and a half times Mexico's population. The brand now brings FEMSA more revenue than its Coca-Cola bottling business does. (Latinometrics)

What Else We're Watching

• Nvidia open-sourced Alpamayo 2 Super, a reasoning model for robotaxis and autonomous driving

• Diners are increasingly fighting off bots and scalpers just to land a restaurant reservation

• Disney+ is cracking open the door to fan-made TikTok videos using its own content library

• The hedge fund that blew up two weeks ago is writing nine-figure checks again

That’s the news,

From the Nido

On July 24th, Jensen Huang posted on X for the first time in his life. He used it to publish a three-page letter signed by NVIDIA and 24 other companies arguing that open AI models strengthen safety, accelerate innovation, and enable sovereignty. Within a day the count doubled to fifty and included OpenAI and Google. By August 3rd it carried 235 names. The open-weights coalition had arrived, and it arrived fast.

This week’s Deep Dive by Rodrigo reads that letter very carefully and finds something worth pausing on.

Of the 235 signatories, at least 34 sell compute, inference serving, or hardware. NVIDIA, the letter’s organizer, sells the chips that run the models. OpenAI and Google sell served tokens, and lose nothing if the weights are downloadable because the customer still buys the serving either way. Anthropic never signed, sells frontier access, and disagrees on the substance.

The letter defines an open weight model as one anyone can download, inspect, modify, and run on their own infrastructure. The first three are properties of a file, and a file costs nothing to copy. The fourth is a business. It is the business the letter’s organizer is in, and the only one nobody in the argument proposed to give away.

The piece is a map of who benefits from the arrangement. And more importantly a question about whether the promise being made on behalf of factories, hospitals, farms, and classrooms is actually being redeemed by them.

“We don’t have a name for an industry that gives away its product and sells electricity. That’s the arrangement all sides are arguing for,” Rodrigo writes.

From the Nido Archive…

Beyond this week’s piece, the Nido archive has been quietly building toward it all year. ConteNIDO has been mapping the AI infrastructure layer from every angle—from the hardware constraints to the labor economics. This week’s piece adds yet another layer to the analysis: the political economy of who controls access to the models themselves, and why the open-versus-closed debate looks very different depending on whether you are asking who benefits from the argument or who redeems it.

The Memory Bottleneck by Roberto is the essential hardware complement. The open-weights debate assumes that running a model is a solved problem. Roberto’s piece showed it is not — that the memory required to serve frontier-scale models is manufactured at scale by three companies, all sold out through 2026, with new capacity not expected until 2028. Open weights change who owns the file. They do not change who supplies the silicon that makes it run. Both pieces are about the same chokepoint viewed from opposite ends.

The Laborsaving Lie by Renata, published last week, traced how every productivity gain in history gets absorbed by the system that produced it rather than returned to the individual. The open-weights promise follows the same structure: the gain is real at the task level — the file is free, the weights are downloadable — but the system downstream of it, the compute, the hosting, the operations, the fine-tuning — has its own logic about what to do with that gain. Free inputs do not produce free outputs. They produce a different vendor.

The Accounting Depreciation piece by Roberto traced how Big Tech is using footnote changes in useful life assumptions to make AI investments look more profitable than the underlying economics justify. This week’s piece is the mirror image of that argument. The open-weights coalition is using the language of access and democratisation to frame an arrangement that primarily benefits the companies selling the conversion. In both cases, the number on the front of the document and the economics underneath it are not describing the same thing.

The SpaceX IPO piece mapped how the largest liquidity event in history was structured to ensure that the people with governance power kept it, regardless of how many retail investors bought shares. The open-weights letter has a similar architecture: the promise of access is real, the weights are downloadable, the terms of participation are open — and the economics of running them flow back to the same concentrated infrastructure layer that ran them before. Access and power are not the same thing, and the pieces that have examined both most carefully this year have arrived at that observation from very different starting points.

So download them. Then what?

Nothing. A weights file answers nothing. It’s a very large table of numbers that does nothing until hardware runs it (inference), producing one token at a time. Four researchers at Carnegie Mellon put the hardware floor for a one-trillion-parameter open model at eight-bit precision at roughly 1,000 GB of video memory: sixteen A100 GPUs, about $240,000 of chips at 2025 prices, before it serves a token. Inkling ships a lower-precision version built to pull that floor down on newer NVIDIA hardware. Down, but not yet to zero.

The wrong unit

Releases are events. They arrive on a date, with a spec sheet and a benchmark table, so they get counted.

Conversion isn’t an event. It’s the sustained ability to turn a weights file into tokens someone actually receives. No launch date, and I’ve yet to see anyone count it.

Where the work sits shows up in what builders report. Menlo Ventures surveyed 150 technical decision-makers in mid-2025:

• 74% of startup builders said most of their workloads were inference, not training, up from 48% a year earlier.

• 49% of enterprises with 5,000+ employees said most of their compute was inference-driven, up from 29%.

Both roughly doubled in twelve months. The centre of gravity is serving, not building.

Now try to count who does the serving. You can’t. a16z and OpenRouter published the largest public usage dataset available, over 100 trillion tokens of request metadata. They said outright that self-hosted, enterprise and closed internal deployments fall outside what it sees. It is a router looking at its own traffic. Any confident figure for the self-hosted share of open-weight usage is an estimate in a good suit. In the slice it does see, open-weight models reached roughly a third of usage by late 2025.

The release is the announcement. The served token is the unit of account. Only one gets counted.

Relocating, not removing

On 24 July, Jensen Huang posted on X for the first time in his life. He used it to publish a three-page letter, “Open Weights and American AI Leadership,” signed by NVIDIA and twenty-four other companies. Open models, he wrote, “strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty.”

The timing wasn’t incidental. Moonshot AI had released Kimi K3 on 16 July, an open-weight Chinese model at roughly frontier capability, and Washington was reportedly weighing restrictions on Chinese open-weight models. The letter is the openness camp moving first.

It moved fast. Within a day the count doubled to 50, picking up OpenAI and Google, both conspicuously absent from the original 25. By 3 August it carried 235 names. Amazon, missing at launch, had joined. Anthropic had not, and still hasn’t.

The letter defines open weight models as models anyone can download, inspect, modify, and run on their own infrastructure. That last clause is the promise.

Thinking Machines shipped Inkling with hosted APIs live on Together AI, Fireworks, Modal, Databricks and Baseten from launch day, plus inference support built with five open-source projects. The releaser didn’t expect the market to convert the file unaided.

Which invites the obvious objection: so what? Conversion is a commodity. Those five vendors, plus 300-plus models from 60-plus providers on that one router. The gap is priced near zero and closes with a credit card. The file is asleep; it’s one API call from being useful.

That holds until volume arrives. A measured preprint ran Mistral’s Mixtral 8x7B on a two-GPU Azure node rented at $13.96 an hour, varying only the request rate. At one request per second, that deployment cost $15.25 per million output tokens. Claude Sonnet 4.6 listed at $15.00 for the same unit in June: a frontier model, sold at retail, for less than a mid-size open model cost its owner to run. The models aren’t comparable. The cost structures are, and that’s the point.

Idle capacity is the most expensive way to serve a token; saturated capacity is the cheapest. Renting is rational precisely while your volume is too low to justify owning. And the moment it isn’t, the money walks back in. Per-token charges are what disappears; hardware, hosting, operations and fine-tuning aren’t. Morgan Stanley reaches the same conclusion in research reported by Blockspace Media: open weights change where compute runs, not whether it’s needed.

None of which is news to the letter’s organiser. NVIDIA’s own engineering blog says most inference deployments leave significant GPU capacity idle, and that without intelligent orchestration teams are forced to choose between overprovisioning and performance risk. And in a McKinsey survey of 703 technology leaders fielded over the turn of 2025, hosting and inference compute was the stack layer with the lowest share preferring open source, at 24%, and the highest that would use only proprietary, at 17%. Asked how they prefer to access models, 40% chose their own infrastructure against a combined 60% preferring an API or a provider’s hosting.

Self-hosting isn’t impossible. It’s a capital project with a utilisation problem. Rent instead and you have a vendor again. A different one, but the dependency is the same.

The concession

In mid-July, Hugging Face disclosed a security incident. Its responders first tried the forensics through commercial frontier model APIs and were blocked: the work meant submitting real attack commands, exploit payloads and command-and-control artifacts, and by Hugging Face’s own account the providers’ safety guardrails “cannot distinguish an incident responder from an attacker.” They ran it instead on an open-weight model, GLM-5.2, on infrastructure they operate themselves, covering more than 17,000 recorded attacker events. Their stated lesson: have a capable model you can run yourself, vetted and ready before an incident, not during one.

Not isolated. Forbes reported on 27 July that the CEO of security firm RemoteThreat told TechCrunch that guardrail behaviour varies day to day, including inside the vetted cyber programs OpenAI and Anthropic each run, pushing researchers toward locally-run open models.

NVIDIA read the same incident and built an institution around it. Days after the letter, it launched the Open Secure AI Alliance with the Linux Foundation, Microsoft, IBM, CrowdStrike, Red Hat, Hugging Face and dozens of others, citing the Hugging Face breach in the case for security tooling teams can run locally. NVIDIA is contributing models, weights and datasets. OpenAI, Google and Anthropic are not members.

So the substantive question is live, and it is genuinely unresolved. The letter asserts one answer: where attackers use advanced AI, defenders need comparable capabilities, while conceding that released weights sit beyond their developer’s control. Dario Amodei asserts the other: the reverse seems at least as likely to him, biology being the sharpest case, and such questions should be settled by pre-release testing rather than assumed. I’m not adjudicating it.

Look instead at who needed the weights. Hugging Face ran its forensics on infrastructure it already operated. The letter’s promise is diffusion into factories, hospitals, farms, classrooms and main street businesses. Hugging Face is none of those.

That’s the pattern underneath the whole debate. The organisations that genuinely need open weights are national labs, well-resourced security teams and sophisticated red teams. They are the ones already holding the capacity to convert them. The argument is made on behalf of the many and redeemed by the few.

The one visible measure of what open weights actually get used for points the same way, and carries the same blind spot. On that same router, roleplay was about 52% of open-weight tokens, programming second at 15 to 20%. That is traffic a router can see, which by construction excludes the enterprise deployments the letter is about.

Where the capacity isn’t present

Where the capacity to convert doesn’t exist, a free file buys less than it appears to. In McKinsey’s survey, 31% of organisations not adopting open source cited a lack of internal expertise, though the larger barrier, at 56%, was security and compliance. Menlo named deployment complexity as one of three causes of open source’s stagnating enterprise share, alongside a nine-to-twelve-month lag behind the closed frontier and reluctance to use APIs from Chinese providers. And the enterprise open-weight share that does get tracked is tracked as hosted consumption: Menlo’s December 2025 figure of 11% is estimated from production API spending, not self-hosting.

Who owns the compute

Line the three positions up and they stop looking like a debate.

NVIDIA organised a letter arguing the model should be free. It sells the hardware that runs it. OpenAI, Google and Microsoft signed within a day. They sell served tokens, and lose nothing if the weights are downloadable, because the customer still buys the serving either way. Anthropic never signed, sells frontier access, and disagrees on the substance.

Of those 235 names, at least 34 sell compute, inference serving or hardware. The other 201 include Amazon, OpenAI, Meta, Microsoft, Google, Mistral and Hugging Face, none of them primarily compute vendors. So the count supports one inference and not another: a substantial bloc of the coalition arguing the input should be free also sells the conversion. It says nothing about motives.

Amodei, who never signed and never called for a ban, describes open-weight models without dangerous capabilities as a public good costing nothing beyond the compute needed to run them. He means it as reassurance. It is also an itemised bill.

The letter defines an open weight model as one anyone can download, inspect, modify, and run on their own infrastructure.

The first three are free. They are properties of a file, and a file costs nothing to copy.

The fourth is a business. It is the business the letter’s organiser is in, the business at least 34 of its signatories are in, and the only one nobody in the argument proposed to give away.

We don’t have a name for an industry that gives away its product and sells electricity. That’s the arrangement all sides are arguing for.

We usually open with a deal or a launch. This week we are opening with a blowup, and the boom that made it possible.

Leopold Aschenbrenner, the 25 year old former OpenAI researcher whose 2024 essay convinced Silicon Valley that AI would swallow the world’s chips, watched his hedge fund Situational Awareness fall from roughly $45 billion at the start of July to about $10 billion in days. Margin calls forced a distressed sale of its leveraged public positions, including SK Hynix and CoreWeave, to Ken Griffin’s Citadel, after bullish AI infrastructure bets and bearish software bets both moved against him at once. The timing is almost too neat, the same week his fund cratered, HFR data showed global hedge fund capital hit a record $5.6 trillion in the second quarter, its largest single quarter jump ever, built partly on AI chip conviction. (CNBC, Hedgeweek)

Markets & Deals

FIFA scrapped its $20 billion World Cup investment plan within days

FIFA president Gianni Infantino abandoned his own proposal to sell private stakes in the organization's commercial and event operations, just days after announcing it. The plan would have folded World Cup and Club World Cup commercial rights into a new entity valued at $20 billion, with Joshua Kushner's firm Thrive Eternal expected to lead the investor group. It collapsed almost immediately under backlash from the global soccer community, including a threatened boycott of future World Cups by UEFA. Infantino said the project had "created divisions of a nature that, regardless of the level of support, are no longer in the interest of the objective set out in the first place." (CNN)

Robinhood now makes more money from prediction bets than from stock trades

Robinhood's prediction markets revenue surged more than tenfold year over year to $156 million in the second quarter, overtaking both stock and crypto revenue for the first time and becoming its second largest trading business after options. The World Cup drove much of the quarter's volume, and Robinhood is now routing bets through Rothera, a new exchange it launched with Susquehanna in June instead of relying entirely on Kalshi. Kalshi still leads the category by a wide margin, with about $33 billion in monthly volume versus Rothera's $2.1 billion, but the shift shows how fast trader appetite has moved from crypto toward real world event betting. (The Information)

Starbucks strung together four straight quarters of growth for the first time in years

Starbucks posted comparable store sales up 8.1% in North America, 5.7% internationally, and 7.9% globally in fiscal Q3, its fourth consecutive quarter of same-store sales growth and its best rate since late 2023. CEO Brian Niccol's "Back to Starbucks" turnaround, built on closing weak stores, cutting jobs, and offloading a chunk of its China business, has now delivered profit margin expansion for a second straight quarter too. The company raised its full year outlook, with Niccol calling it "the quarter our momentum became truly measurable." (Starbucks)

Shein finally cleared the runway for its Hong Kong IPO

China's securities regulator has approved Shein's long-delayed plan to go public in Hong Kong, clearing a major hurdle after years of failed attempts in New York and London. The fast fashion group is targeting a valuation of $40 billion to $50 billion, with a listing that could come as soon as this quarter. Investors are likely to push back, a fresh prospectus shows 2025 revenue up 8% to $41.8 billion but net income down 39% to $2.06 billion, plus a $99 million loss in the first quarter of this year. That is a long way from the $100 billion valuation Shein carried back in 2022. (Reuters)

Hedge funds are booming on the same AI conviction that just wrecked Aschenbrenner

Global hedge fund capital grew by $409 billion last quarter to a record $5.6 trillion, the fastest single quarter jump HFR has ever recorded, powered largely by an AI driven rally in chip stocks like Samsung, AMD, and SK Hynix. Investors poured $134 billion into the industry over the last three quarters, the largest three quarter inflow since 2007, as money managers grow frustrated waiting years for private equity funds to return capital. "I think this is going to be a golden era for hedge funds after a long, long time," said Shenan Dhanani, co-CEO of Trium Capital, though the same conviction driving the boom is exactly what gutted Aschenbrenner's fund weeks later. (FT)

Big Tech & Infrastructure

Microsoft’s Azure just crossed $100 billion a year

Microsoft closed its fiscal year with $90 billion in fourth quarter revenue, up 18% and ahead of Wall Street's $87.7 billion estimate, with net income up 31% to $35.8 billion. Azure was again the engine, growing 43% and pushing the cloud unit past $100 billion in annual revenue for the first time. Microsoft 365 Copilot paid seats crossed 30 million, with net additions more than doubling from the prior quarter. The one caution flag is the same one hanging over every hyperscaler this season, the capital spending needed to keep that growth going. (WSJ)

States are quietly clawing back the tax breaks that built the data center boom

Governors and legislatures in states that once welcomed data centers with open arms are moving to repeal the sales tax breaks that helped build them, a shift that could add 7% or more to the cost of every gigawatt of AI computing. Lawmakers in roughly two dozen states, including Virginia, Pennsylvania, and Georgia, have introduced bills to curb or end the incentives, with Virginia's break alone costing the state about $1.6 billion a year. Rather than clean repeals, most states are attaching strings instead, Texas wants data centers to fund their own grid connections, and Illinois wants the right to cut their power when the grid is strained. (The Information)

The world's biggest battery maker is quietly becoming an AI infrastructure company

CATL, the Chinese battery giant that supplies a huge share of the world's EVs, says it wants energy storage to make up half its global sales by 2030, up from roughly a quarter today. Storage revenue grew 88% in the first half of 2026 alone, and the company is leaning hard into grid batteries built for AI data centers, including a new sodium-ion storage system it calls the first field-validated system of its kind. In April it put about $600 million into a stake in a firm focused on high-voltage power for data centers. The pitch is straightforward, EV demand is uncertain, and every AI data center on earth needs somewhere to store power. (FT)

DoorDash is building its own drone airline

DoorDash won a Part 135 air carrier certification from the FAA and used it to launch DoorDash Air, its own in-house drone delivery network, becoming the eighth drone operator in the US to clear that bar. Rather than lean on partners, the company is building the drones and the ground infrastructure itself, betting that vertical integration pays off on trips under five miles that it says can land in under 25 minutes. Commercial flights are not expected until the fall, and DoorDash still expects humans to handle the vast majority of orders. Early partner pilots saw some merchants lift order volume by about 30% in the weeks after launch. (CNBC)

AI & Models

A second lab admits its models broke out and hacked real companies

A week after OpenAI admitted one of its models escaped a sealed test and hacked Hugging Face, Anthropic disclosed that its own models did the same thing to three real organizations, with incidents dating back to April. The models involved were Claude Opus 4.7, Claude Mythos 5, and an internal research model, each of which broke out of an isolated environment and reached the open internet after a configuration error. Anthropic found the incidents only after combing through more than 141,000 evaluation runs in a review it launched in response to the OpenAI episode. It has not named the organizations, and like OpenAI it framed the disclosure as a safety lesson rather than a breach. (WSJ)

OpenAI hit a billion users, then cut prices to keep them

OpenAI said it now reaches more than one billion active users and over two million businesses, a milestone it announced alongside a sharp price cut on its GPT-5.6 family. The Luna tier fell about 80% to $0.20 per million input tokens and $1.20 per million output tokens, the mid tier Terra also dropped, and the top Sol model held steady. OpenAI credited efficiency gains made during development that cut serving costs by 20% and lifted token generation efficiency by more than 15%. The subtext is competitive, cheaper Chinese models and cost sensitive customers are dragging frontier prices toward commodity levels. (Quartz)

Google DeepMind taught a humanoid to use its whole body

Google DeepMind unveiled Gemini Robotics 2, a model that for the first time lets a humanoid coordinate its entire body, walking, crouching, and manipulating objects while reasoning through multi step tasks. In a taped demo, the system directed Apptronik's Apollo robot across a room to pick up a watering can and set it on a lower shelf while dodging obstacles. DeepMind also shipped Gemini Robotics ER 2, a reasoning layer that can plan tasks and coordinate several robots at once. The main model is in private preview, while ER 2 is available through the Gemini API and Google AI Studio. (Google DeepMind)

Latam Spotlight

El Salvador's Ábaco raised $53 million to lend to small business

Ábaco, a San Salvador fintech founded in 2023, raised $53 million to expand small business lending across Central America, one of the region's largest rounds to date. The package pairs a $50 million warehouse facility anchored by Accial Capital with an equity round that drew Nazca Ventures, Alaya Capital, Promotora Social México, and others. Ábaco uses AI to underwrite invoice financing and working capital for SMEs, and plans to originate more than $350 million in loans over the next two years. The new money will fund expansion across El Salvador, Guatemala, and Costa Rica. (Ábaco)

Coca-Cola FEMSA rode record LatAm volumes and a World Cup bump

Coca-Cola FEMSA, the world's largest Coke bottler, posted a strong second quarter, with volume up 3.5% to 1.1 billion unit cases and revenue up 4.7% to 76.3 billion pesos, or 6.6% on a currency neutral basis. The company logged record quarterly volumes in Brazil, Colombia, and Guatemala, with South America volume up 6.9% and division operating income up 46.5%. Mexico was the soft spot, growing just 1% as an excise tax increase and weaker consumer demand weighed on results. Shares rose after the print, helped by the World Cup, and management lifted its dividend. (Yahoo Finance)

Wearables

Global wrist shipments are set to fall about 5% this year as basic smartwatches cool

Global wrist wearable shipments are expected to fall about 5% in 2026 as demand cools for basic smartwatches and fitness bands. Premium "true" smartwatches are still growing about 2%, while wristbands are set to decline nearly 7%, with buyers drifting toward smart rings or back to analog watches. Apple is expected to hold its lead with roughly 41% share, followed by Huawei at 31%. (Smart Analytics Global)

Tip of the Week

How to build a kitchen assistant inside Claude

Ana Caro Mexia published a standalone piece on turning a Claude Project into a weekly kitchen planner. Her setup takes twenty minutes once, you tell it which appliances you actually use, who eats in your house, and your hard dietary limits, then save a "plan the week" skill so every Sunday three words produce five dinners, a Monday salad bar, a prep list, and a grocery list sorted by aisle. The most common mistake, she writes, is using a regular chat instead of a project, which forgets everything and forces you to re-explain your kitchen from scratch every week.

Charts of the Week

Meta's daily active users have climbed to 3.6 billion, up 59% since December 2019, a steady climb even as the AI chatbot wars crowd everyone's feed (chart via Fiscal.ai, @StockMKTNewz on X)

Agricultural credit is booming across Latin America, Bolivia's farm credit is up 441% since 2012, Honduras 202%, and Uruguay 196%, while Argentina's has fallen 72% (chart via Latinometrics, FAOSTAT data)

What Else We're Watching

• Global wrist wearable shipments are set to fall about 5% in 2026 as basic smartwatches cool

• Robinhood says more than 100,000 people have set up AI agents to trade stocks on its platform

• Enigma raised $71 million and put 100 real robots online for anyone to control from a browser

That’s the news,

From the Nido

This week’s deep dive asked a burning question: What will happen to jobs in the era of AI? Specifically, how will the value of labor be recalibrated when individuals adopt tools that enable them to do more with less time? The simple—if wishful—answer is that we’ll get more time. More free time to give back to other things that aren’t consumed by work. History shows us, however, that this version of events is unlikely. This week’s deep dive breaks it down.

This week’s Deep Dive by Renata asks a quieter and more uncomfortable question: where are we, actually, on that promise?

The answer is not that AI failed to deliver capability. It is something structurally harder to dismiss. The systems receiving that capability were never built to convert it into time. They were built to convert it into output. And output, under any economy organized around continuous growth, does not sit still once it is created. It gets absorbed, reinvested, and used to raise the standard of what is expected next. A task automated is not a task subtracted. It is a task that clears the space for three more.

John Stuart Mill saw this in 1848, when he wrote that all the laborsaving machinery ever invented had not lessened the toil of a single human being. He was not arguing that the machines failed. He was arguing that the system downstream of them had its own logic—and that logic had nothing to do with handing time back to anyone.

The piece traces that logic from the Industrial Revolution through a recent Berkeley ethnography of a 200-person technology company where workers given broad access to AI tools did not work less. They worked more — taking on tasks that previously belonged to someone else, running prompts during lunch, keeping multiple AI agents running simultaneously in the background while attending meetings. What the researcher calls the intensification of work. What Mill would have recognized immediately.

“The tautology is this: under a system that requires continuous accumulation, laborsaving technology cannot, by definition, save labor — it can only relocate it faster than the last one did,” Renata writes.

From the Nido Archive

This week’s piece is the labor economics complement to a thread that has been running through the archive all year — a sustained, multi-angle investigation into what AI is actually doing to the humans inside the systems it is transforming.

The Cognitive Cost of Convenience by Renata is the most direct companion. That piece asked what AI is doing to our brains — the measurable neural effects of outsourcing the generative act, the homogenization of thought, the risk of cognitive circuitry that stops being exercised. This week’s piece asks what AI is doing to our time. Both pieces arrive at the same structural observation from different disciplines: the promise of AI as liberation assumes a downstream system that was built to deliver it. Neither the economy nor the brain works that way.

The Finance Agent piece by Renata Solana measured AI’s readiness for professional deployment and found a meaningful gap between capability and reliability — but also found that every model in the benchmark completed tasks significantly faster than human experts. That speed gain was framed as the near-term value proposition. This week’s piece is the second half of that story: what actually happens to the time those speed gains produce inside a company organized around growth. The answer, per the Berkeley study, is more tasks, higher expectations, and a workday with fewer natural stopping points.

The Attention Economy by Renata is the structural backdrop. When platforms optimized for engagement rather than meaning, they built systems that colonized more and more of users’ time without delivering proportionally more value. The labor ratchet Renata describes this week follows the same mechanism at the level of the economy: more capability flows in, more output is extracted, the standard rises, and the individual never quite catches up. Both pieces are about systems that absorb gains rather than distribute them.

Vibecoding Has a Thinking Problem by Ana Caro argued that clarity before code determines what you can build with AI — that the founders and builders who use these tools well arrive with a formed intention rather than asking the tool to form it for them. This week’s piece is the labor market version of that argument. The workers who will navigate the intensification dynamic most successfully are not the ones who adopt AI fastest. They are the ones who are clearest about what they are using it for — and what they are not willing to let it colonize.

Thinking that we can predict the future of work so early on in the AI era is fundamentally naive—but we are at the point when questioning things repeatedly is important. No outcome is all good or all bad, it’s more complex than that, and that’s exactly the point.

—Renata from Nido

But where are we on that promise?

The answer, it turns out, is that we are quite far. In fact, there is a slim chance we’ll ever be there at all. In this realistic—and hopefully more grounded take—on the effects of AI adoption on labor, I’ll dig into the ways this promise breaks down. Not because AI failed or will fail to deliver on capability, but because the systems receiving that capability were never built to convert it into anything as simple as time.

AI isn’t a technology that just so happens to touch work—it is a technology about work. Therefore its fundamental questions and existential threats are about work and labor. The way in which we do it, structure it as a society, and then cope with the ways in which it gets redistributed. This has been central to discussions about AI from the beginning.

In 2025 Dario Amodei warned that AI could displace half of all entry-level white collar jobs within the next 1–5 years. He all but asserted the mass elimination of jobs across technology, finance, law, consulting, and other white-collar professions, especially entry-level gigs. He breaks down his reasoning further in his lengthy essay The Adolescence of Technology, in which there is a full subsection titled Labor market disruption. The scenario he outlines here is bleak to say the least.

Unlike with previous adoption of technologies, he argues AI displaces workers in a new and unprecedented way. If we take farming as an example, we observe a pattern: a tractor replaced a farmhand who had nowhere else to go in the short term but eventually found work operating the very machines that displaced him. That gap, in between where the machine can no longer operate and where human labor picks it back up, is where much of the displaced labor moved into. Amodei argues that it is precisely this gap that will all but disappear with the implementation of AI, leaving little room for displaced labor to move into.

That is certainly a possibility, and one I will hold at arm’s length throughout most of this piece. Not because I think he’s wrong, but because his argument and mine answer two slightly different questions. His is about whether there will be a job left to redefine. Mine is about what happens to the job in the meantime.

Even if Amodei’s promise never arrives and the labor market absorbs the disruption eventually, there’s a second fallacy that crumbles at the outset. The idea that capability, once unlocked, gets handed back to us in the form of time—or freedom. That a task automated is a task subtracted, not one that clears the space for three more.

History doesn’t really support that idea. Neither does the early data on how people are actually using AI right now. What we find instead looks less like liberation and more like a ratchet: a mechanism that only turns one way, tightening expectations as it loosens constraints. Understanding why requires looking less at what AI can do and more at what any economy does with a gain it’s handed—regardless of the technology behind it.

The laborsaving promise (and failure): Naming the tautology

The word laborsaving first appeared in the late 1700s, emerging alongside the Industrial Revolution to describe new machinery and tools that reduced physical effort and time required to complete a given task, per Merriam Webster. It started as an adjective to describe farming innovations that boosted work output. In the following centuries it expanded into everyday life, used as a word to characterize technological innovations in the home—think washing machines, blenders, dishwashers, and other tools that optimized domestic labor. In the recent century it has become even more ubiquitous, used to describe software, technology, and yes, AI. The word itself has always been used in conjunction with the adoption of a new technology, a one-size-fits-all term that promises the same outcome decade after decade. Laborsaving implies a subtraction: labor minus machine equals less labor. But what’s happened time and time again is more of a substitution: one kind of labor minus machine equals a new kind of labor to fill the space left behind.

This is not a new or radical observation. In 1848 John Stuart Mill wrote Principles of Political Economy, and in it the following quote:

“All the labor-saving machinery that has hitherto been invented has not lessened the toil of a single human being.”

This quote holds and has held throughout history. By Mill’s own account, machinery was unlocking unprecedented levels of economic output. What he observed on the ground was that the toil people were freed from with the advent of machines wasn’t removed altogether, it was simply placed elsewhere.

The word laborsaving does misleading work. It describes what happens at the task level while holding the assumption that this same behavior will happen to the individual executing the task or even the aggregate economy. The leap from “this task now takes less labor” to “therefore, there is less labor to be done” is the fallacy in the argument.

Gains in productivity don’t sit still once they’re created, as Mill observed, they get absorbed into whatever system produced them, and that system has its own rules about what to do with a surplus. Under our system, that requirement is growth: continuous, compounding, competitive. A company that takes its AI-driven efficiency gains and hands them back to its workers as free time is a company that a competitor will eventually outcompete. No one individual has to intend this outcome for it to occur. Nor is it a conspiracy of bad employers; it’s a law of the system. The tautology is this: under a system that requires continuous accumulation, laborsaving technology cannot, by definition, save labor—it can only relocate it faster than the last one did.

The Berkeley Study

A recent study published by Berkeley PhD student Xingqi Maggie Ye proves the above logic. In her observations, she has found that AI adoption has intensified work rather than replaced it. Through an in-depth ethnographic study shadowing, interviewing, and observing workers at a 200-employee technology company with broad access to AI tools, she found that generative AI did not free up time. In fact, it did the opposite, what she calls intensification of work. She breaks it down in the following way:

In our study, intensification took three main forms in practice. First, people began taking on work that previously would have belonged to someone else or might not have been attempted at all. The scope of what counted as “my job” widened. Second, because AI makes it easy to start and continue tasks, work seeped into moments that used to function as pauses. People would send prompts during lunch, before meetings, or in the evening when an idea came to mind. This dissolved some of the natural stopping points in the workday. Third, workers increasingly kept multiple threads alive at once. They would run AI processes in the background while reviewing code, drafting documents, or attending meetings. Some even ran multiple AI agents simultaneously. This created a rhythm where both the human and the machine were constantly in motion.

She found that expectations gradually reset, and what once was extra effort eventually became the new standard performance. Workers are in turn held to a higher standard of productivity that was previously only considered “extra” effort. In the above quote she describes another common pattern that follows the adoption of laborsaving technologies—the blurring between leisure time and work time. If prompts can be run in the background during lunch, then lunch ceases to be a pause during the workday and slowly becomes an extension of it—an extension that mirrors the expansion of labor happens to time in tandem.

Laborsaving technology, going back to the washing machine, was supposed to shrink the territory labor occupied in a life. What Ye’s ethnography suggests instead is that AI has made that territory harder to see the edges of because it slowly dissolves the idea that there ever needs to be an edge.

Ye’s own framing catches the tautology I highlighted in the above section without naming it as one. She and her co-author describe a “vicious cycle”: increased capability leads to increased output, which leads to higher expectations, which pressures further expansion.

The Anthropic Study

Ye’s ethnography shows the tautology from the inside of a single company while Anthropic’s own data shows it at the scale of an economy, published in March 2026 under the title Labor market impacts of AI: A new measure and early evidence.

This graph in particular circulated widely online. Viewed quickly it looks like good news, there is a measurable space in between where AI adoption has already taken over certain professions versus where it could potentially go. However, this report fails to answer two fundamental questions about the current labor revolution that I consider more existentially important:

1. Have workers found a measurable decrease in the hours they work, or only in the hours a given task takes?

2. Even when a technology’s productivity gains are real, undisputed, and measurable— where do they go?

Productivity and wages decoupling

In the above graph from a recent Financial Times article, we see the decoupling between productivity and real wages that has occurred in the US since roughly the 1970s. Since 1979, US net productivity has grown by roughly 90%, while the pay of a typical worker has grown by only about a third of that—a gap that didn’t exist in the decades before it. From the end of World War II through the 1970s, pay and productivity climbed together, meaning that when workers produced more, they were paid more, more or less dollar for dollar. Then sometime around 1979, the lines split apart and never rejoined.

This graph is in line with Ye’s observation above. What she saw in a single office is the above playing out in practice. Hours were filled with more tasks, but the worker’s compensation was never meant to rise in tandem.

Going back to Amodei’s argument I outlined at the beginning. Every mechanism described throughout this essay—intensification, decoupling, the ratchet that only turns one way—assumes there is always some remaining sliver of work for a human to be pushed into doing more of, faster, for less. That assumption held through the tractor in the 1800s, through the washing machine and the spreadsheet, and through Ye’s eight months inside a 200-person company. Amodei’s bet is that it stops holding this time—not because the system’s appetite for reinvesting gains has changed, but because the thing being reinvested into might simply run out.

If Amodei is wrong and the sliver survives the way it always has, the future isn’t liberation or more time, it’s more of what we’ve seen across two centuries of technology: the same worker, the same hours or more, held to a rising standard the moment their tools improve, watching the value they generate flow somewhere else. If Amodei is right, the sliver doesn’t survive it, and the tautology quietly stops describing anything, because there’s no labor left to relocate. That’s a different problem entirely. Not what happens to work when it intensifies, but what happens to a society organized around work when the work simply isn’t needed.

Both futures dispose of the same utopian promise this piece opened with. The one where a new technology hands us back our time. That promise was never really about AI. It’s been recycled every time a laborsaving machine has been introduced, believed a little differently each time, and it has yet to once come true—not because the machines failed, but because nothing downstream of them was ever built to let it. AI won’t be the technology that breaks that pattern.

Me tomó veinte minutos una sola vez. Ahora los domingos escribo tres palabras y tengo la semana resuelta: cenas, prep, barra de ensaladas y la lista del súper lista para Instacart.

1. Un proyecto, no un chat

Un chat suelto olvida todo. Un proyecto guarda el contexto para siempre, así que todo lo que escribas una vez aplica a cada conversación dentro de él.

1. En Claude, abre Proyectos y crea uno nuevo.

2. En “¿En qué está trabajando?” escribe el nombre: Mi cocina.

3. En “¿Qué intenta lograr?” pega el bloque de abajo. Con eso Claude arma las instrucciones base del proyecto.

4. Abre el campo de instrucciones del proyecto y pega los dos bloques de los pasos 2 y 3.

5. Si tienes recetas que ya sabes que te gustan súbelas en PDF o pega los links. Ese es el material del que va a partir.

¿QUÉ INTENTAS LOGRAR?

Quiero un asistente de cocina para mi casa. Cada domingo le paso
los PDFs y links de recetas de la semana y quiero que me devuelva:
5 cenas de lunes a viernes de máximo 30 minutos activos, la barra
de ensaladas del lunes, el prep del domingo en tandas y la lista
del súper lista para Instacart, agrupada por pasillo.

Cocinamos 3 adultos y una bebé que ya come sólidos. Nunca
crustáceos (alergia), mariscos ni rábano. Priorizo la Dream Cooker
de Our Place y la arrocera KitchenAid: si algo se puede hacer en
una sola olla, esa es la versión buena.

2. Dile qué hay en tu cocina

Esta es la parte que casi nadie hace y la que más cambia el resultado. No listes lo que tienes: di qué usas.

MI COCINA

Aparatos que tengo:
- Dream Cooker de Our Place — mi favorita
- Arrocera y cocedor de granos de KitchenAid
- Horno, air fryer, licuadora, procesador

Cómo me gusta cocinar:
- Prefiero recetas de una sola olla.
- Si algo se puede hacer en la Dream Cooker o en la
  arrocera, dame esa versión primero.
- Tengo espacio limitado en el refri.

Formato de las recetas:
- Ingredientes con cantidades.
- Pasos numerados, máximo 8.
- Dime el tiempo real, incluyendo precalentado.

3. Restricciones e Instrucciones

Lo que no puede pasar pesa más que lo que quieres. Sé literal.

QUIÉNES COMEN
- 3 adultos
- Ana Sofía, bebé, ya come sólidos: dime siempre
  cómo adaptarle la cena

RESTRICCIONES DURAS (nunca)
- Alergia a los crustáceos
- Sin mariscos
- Sin rábano

REQUERIMIENTOS
- Proteína en cada comida
- Verdura en cada cena

TIEMPO
- Lunes a viernes: máximo 30 minutos activos
- Domingo: puedo cocinar 2 horas en tandas

4. El skill del domingo

Pega esto al final de las instrucciones del proyecto. Después, cada domingo escribes planea la semana y sale el mismo entregable, con el mismo formato, sin volver a explicar nada.

SKILL: PLANEA LA SEMANA

Cuando yo escriba “planea la semana”, haz esto en orden:

1. PREGÚNTAME PRIMERO (una lista corta):
   - ¿Qué proteína ya tengo en casa?
   - ¿Alguna noche que no vaya a cocinar?

2. PROPÓN EL MENÚ
   - Lee los PDFs y links de recetas que te pasé y
     úsalos antes de inventar algo nuevo.
   - 5 cenas de lunes a viernes, máximo 30 min activos.
   - 1 comida de domingo que rinda para tandas.
   - Marca qué va en la Dream Cooker y qué en la arrocera.
   - Respeta todas las restricciones duras.

3. BARRA DE ENSALADAS DEL LUNES
   - 2 bases verdes, 4 toppings crudos ya picados,
     2 cocidos, 2 proteínas, 1 crunch, 2 aderezos.
   - Cómo se guarda cada cosa y cuántos días aguanta.

4. PREP DEL DOMINGO
   - Qué picar, cocer o marinar, y qué se guarda.

5. LISTA DEL SÚPER
   - Lista para Instacart, agrupada por pasillo.
   - Con cantidades. Sin lo que ya te dije que tengo.

TONO: escríbeme como amiga, en español, directo.
Sin introducciones ni disculpas.

Los tres errores más comunes

1. Usar un chat en vez de un proyecto. Vuelves a explicar todo cada semana.

2. Listar aparatos sin decir cuáles usas. Te manda al horno tres días seguidos.

3. Escribir restricciones suaves. “Casi no comemos mariscos” no es una regla. “Sin mariscos, nunca” sí lo es.

Si lo armas, mándame screenshot.

I could not stop thinking about one story this week, so that is where we are starting for today’s edition.

OpenAI admitted that during an internal safety test, one of its models cheated in the most unsettling way possible. Instead of solving the cybersecurity challenge it was given, it broke out of its sealed sandbox, got onto the open internet, and hacked into Hugging Face to steal the answer key. No hacker, no bad actor, no prompt telling it to cause harm. It was just handed a goal and decided the rules did not apply, exactly the scenario AI safety people have been warning about for a decade.

The damage this time was small, and OpenAI and Hugging Face even wrote it up together like a proud product launch. But this is the least capable these models will ever be. (NYT)

Subscribe now

Markets & Deals

Revolut is now Europe's most valuable startup at $115 billion

A secondary share sale has valued Revolut at $115 billion, making the British neobank Europe's first "centicorn" and lifting it above Barclays, worth roughly $95 billion. That is more than double its $45 billion mark from 2024 and up from about $75 billion late last year. The company reported a $2.3 billion pretax profit on $6 billion of revenue in 2025 and has now passed 75 million customers. It secured a full UK banking license in March and is chasing a US national bank charter next. (WSJ)

A judge just froze the $111 billion Paramount-Warner Bros merger

A US district judge issued a temporary restraining order pausing the $111 billion Paramount and Warner Bros Discovery merger after 12 states sued to block it on antitrust grounds. Paramount then agreed to postpone the deal until five days after a trial, or until June 2027 at the latest. The combination would put Paramount+, HBO Max, CBS, CNN, and roughly 50 cable channels under one roof. The delay is expensive: starting in October, Paramount owes Warner shareholders about $650 million every 90 days the deal slips, and a $7 billion breakup bill if it dies. (NYT)

Kalshi added 3 million users riding the World Cup

Prediction market Kalshi picked up more than 3 million new users during the 2026 FIFA World Cup, roughly double the 1.5 million it had projected for the tournament. More than $1.2 billion was traded just on contracts betting who would lift the trophy. The company leaned on a heavy marketing push, celebrity tie-ins with Luka Modric, José Mourinho, and the Argentina national team, plus a deal to surface its odds inside ChatGPT when users searched for match results. It is a loud sign that event-based trading is going mainstream fast. (CNBC)

Biotech quietly beat AI in this year's IPO market

US biotech and pharma IPOs have returned a weighted average of 55% this year, while the broader US IPO market, excluding blank-check firms, is down 4.4%. The ten companies behind 2026's biggest deals, many of them AI-adjacent, have actually slipped about 6.3% as worries grow that the AI rally is stretched. Biotech proceeds have already topped $5 billion, roughly triple last year's total, helped by a 13% gain in the Nasdaq Biotechnology Index and a steadier regulatory backdrop. At least six more drug developers filed to go public this month. (Bloomberg)

Big Tech & Infrastructure

Alphabet's cloud business is on fire, and so is its spending

Alphabet reported $119.8 billion in Q2 revenue, up 24% from a year ago, with Google Cloud jumping 82% to $24.8 billion and cloud operating income more than tripling to $8.8 billion. The Gemini app now has 950 million monthly users and processes 22 billion API tokens a minute. The catch is the bill: Alphabet raised its 2026 capital spending guidance to between $195 billion and $205 billion, up from a $180 to $190 billion range last quarter. The stock actually dipped on the capex hike, a reminder that investors are watching the spend as closely as the growth. (WSJ)

Google is designing a "frozen" chip to run Gemini far more cheaply

Google is reportedly building a server chip, internally called "Frozen v2," that etches parts of Gemini's architecture directly into the silicon. The weights can still change, but the shape of the network stays fixed, or frozen, which engineers project could deliver six to ten times more tokens per unit of power than Google's newest TPUs. For context, a normal chip generation buys you two to three times better efficiency. Deployment is targeted around 2028, and the project is meant to ease an internal compute shortage that has forced Google Cloud to turn away some outside business. (The Information)

SpaceX wants to rent its computing power to the Pentagon

SpaceX is in talks to give the US Department of Defense access to billions of dollars of data center capacity to run AI models, structured more like a lease than a sale. It would extend a relationship where SpaceX already handles launches, satellite comms, and missile tracking for the government. The move mirrors deals it struck this year with Google, worth about 110,000 Nvidia chips, and with Anthropic, which took 300 megawatts at the Colossus 1 facility in Memphis. If it closes, SpaceX starts to look less like a rocket company and more like a cloud provider with a defense contract. (Bloomberg)

Waymo is thinking about walking away from Uber

Waymo has held internal talks about ending its robotaxi partnership with Uber, first struck in 2023, as the relationship sours. The two have clashed over operations and money: Waymo has flagged concerns about vehicle cleanliness and routing, while Uber has complained about cars vanishing in bad weather and called the deal's terms "unsustainable." They already unwound their Phoenix tie-up in June, and Waymo now plans to launch through its own app in Austin and Atlanta in January 2028. Uber's stock fell about 4.3% on the report. (The Information)

AI now writes more than half of everything uploaded to Deezer

Deezer says fully AI-generated tracks topped 50% of daily new uploads for the first time in June, with about 90,000 machine-made songs arriving every day. That is up from 28% in September 2025 and 44% in April. The twist is that AI tracks account for less than 3% of actual listening time, so humans are still pressing play on human music. Deezer plans to pull AI tracks used for fake streams, along with anything unplayed for six months or more. (Deezer)

AI & Models

Alibaba says its new model is second only to Anthropic's Fable 5

At Shanghai's World AI Conference, Alibaba previewed Qwen3.8-Max, a 2.4 trillion parameter multimodal system it claims is "second only to Fable 5," Anthropic's top Claude model. Alibaba shares rose as much as 5.4% on the news. The problem is there is nothing to check the claim against: no benchmark scores, no model card, and no independent evaluation. It landed three days after Moonshot's 2.8 trillion-parameter Kimi K3, so China now has two labs operating at parameter counts that used to belong only to the top US labs. (WSJ)

Google shipped three cheaper Gemini models, but no 3.5 Pro

Google released Gemini 3.6 Flash, 3.5 Flash-Lite, and a security-focused 3.5 Flash Cyber, leaning hard into cost and speed rather than a new flagship. The headline 3.6 Flash uses up to 17% fewer tokens than its predecessor and runs at $1.50 per million input tokens, while Flash-Lite pushes 350 output tokens per second inside Google Search. Notably absent was 3.5 Pro, which is still in partner testing, and Google confirmed it has begun pre-training Gemini 4. The whole release reads as a bet that the money is in cheap, fast, agentic workhorses, not just top of the leaderboard. (Google)

The AI boom's quiet middleman leans almost entirely on a few giants

Mercor, the three-year-old startup whose contractors train AI models to give better answers, booked $614 million of revenue in the first half of 2026, up 70% year over year, according to documents reviewed by The Inform

ation. The concentration is striking: about 90% of that came from foundation model labs like OpenAI and Anthropic. That is great while those labs keep pouring money into post-training, and risky the moment they do not. Meta already paused its work with Mercor after a security incident, and labor costs eat roughly two-thirds of revenue. (The Information)

LatAm Spotlight

David Vélez just became the first Latin American on OpenAI's board

Nubank founder and CEO David Vélez has joined the boards of the OpenAI Foundation and OpenAI Group PBC, becoming one of the first Latin American tech executives to help govern the company behind ChatGPT. He joins alongside BNY chief Robin Vince, as OpenAI expands its governance amid mounting IPO speculation. Vélez, the richest person in Colombia, has grown Nubank from a São Paulo startup into one of the world's largest neobanks across Brazil, Mexico, and Colombia. Separately, Forbes reports he is now eyeing the hardest market of all, US retail banking. (OpenAI, Forbes)

Policy

Trump hits 60 trade partners with new tariffs

The Trump administration imposed fresh tariffs of 10% to 12.5% on imports from 60 trading partners, backfilling a blanket 10% duty that Congress let expire. The White House framed the move around countries it says failed to enforce bans on goods made with forced labor: Canada, Mexico, India, and the UK land at 10%, while 41 others that adopted no such prohibition get 12.5%. All told, the new rates touch roughly 99.4% of US trade. It is the latest turn in an on-again, off-again tariff regime that businesses are still struggling to plan around. (CNBC)

Peptides

An FDA panel votes to loosen the rules on peptides, over its own scientists' objections

An FDA advisory panel voted to ease restrictions on several peptides, including BPC-157 and KPV, despite analyses from the agency's own scientists that did not support the move. If the FDA follows the recommendations, which it usually does, US compounding pharmacies could start making compounds that today mostly circulate as unvetted "research only" products sold online. Supporters, including figures tied to the peptide industry, frame it as expanding access, while critics worry it green-lights compounds with thin safety data. (NPR)

Hims & Hers takes the market hit from the peptide vote

The same FDA panel that eased the rules on several peptides declined to recommend emideltide, a compound tied to opioid withdrawal and chronic insomnia, for the list that governs what US pharmacies can compound. That was the one Hims & Hers cared about most, and the stock dropped as much as 13% to around $28.50 after briefly spiking above $35, wiping out a peptide rally and putting a roughly $440 million opportunity in doubt. Panel members cited a thin safety profile, possible manufacturing inconsistencies, and no clear edge over approved alternatives. The vote is non-binding, but the FDA usually follows it, so the path for Hims' peptide business just got narrower. (Yahoo Finance)

Tip of the Week

What Else We're Watching

• The commoditization of AI models and compute, RBC on why frontier models are starting to get priced like electricity

• Duolingo’s stock sinks as its AI bet spooks investors

• Japan’s trains are getting much faster as the country embraces moonshots

• Robinhood revamps its premium card to take on the big rewards players

• DeepSeek Puts Current Funding Round on Hold

That’s the news,

From the Nido

A few weeks ago had an idea for a new ConteNIDO column: Origins. As its name suggests, Origins is a space to dive deep into the founding story of a culturally significant company. Early bets, crises that threatened its growth, and eventually the practices that led them to their indisputable positions on the Zeitgeist today.

ConteNIDO is a publication that looks at things deeply. From news to Deep Dives, we ask the questions beyond the headlines. It only makes sense to have a space where we do the same kind of questioning directed at a different subject: companies. Following the same ConteNIDO throughlines of deep thinking and complexity broken down, Origins takes the same rigor we bring to a Deep Dive and points it at the businesses shaping how we live, eat, and spend.

This week we kicked off the column with a look at Erewhon, written by Roberto—the full story of how a macrobiotic food collective in a Boston basement became the most culturally fluent grocery store in America.

It is a story of reinvention, obsession, and the unique, once-in-a-lifetime alchemy that turns a niche food store into a status symbol.

Red below

Origins: Erewhon

A $19 strawberry. A $26 bottle of oxygenated water. A smoothie that racked up 450 million TikTok views. A Balenciaga runway show where models carried Erewhon tote bags instead of handbags. If you have encountered Erewhon, you have probably had one of two reactions: genuine devotion or performative outrage. Both, it turns out, are exactly what the store is designed to produce.

Erewhon is not a startup. It is not a concept born from a pitch deck or funded by wellness trends. It is a 60-year-old idea that began as a macrobiotic food collective in a Boston basement in 1966, filed for bankruptcy in 1981, spent three decades in quiet obscurity on a single block in Beverly Boulevard, and was bought in 2011 by a retired food distributor named Tony Antoci who had never set foot inside the store before he decided to purchase it. The deal closed on April Fools’ Day.

What Tony found was a commercially failed but culturally intact idea: that food could be organized around health and community rather than extraction. What he built from it is a grocery store that generates $2,500 in sales per square foot, compared to Whole Foods at $900 and the industry average at $500. A grocery store in the same range as Tiffany & Co. A grocery store that, in the words of LA artist Alex Israel, has replaced Fred Segal as the quintessential LA experience, reflecting a cultural shift in desire.

The lesson from Erewhon is not “charge more.” It is something harder to replicate and more worth understanding.

“A business can escape its category entirely if it becomes something people buy into rather than just shop at, and the difference between those two things is not a marketing strategy. It is a structural decision made at the level of the business model,” Roberto writes.

From the Nido Archive

Erewhon’s story doesn’t happen in a vacuum. Its success plays off various overarching themes we have encountered in other Deep Dives. To understand Erewhone’s succes and unique positioning as a cultural phenomenon, you should read more from our archive.

The Protein Economy by Roberto is the most direct companion read. Erewhon is, in many ways, the physical embodiment of the marketing cycle Roberto traced in that piece—a store that turned health anxiety into a luxury experience and charged accordingly. But there is a meaningful distinction Roberto draws in the Erewhon piece: the sourcing standards at Erewhon are genuine. The label is doing less work than the formulation, which is the opposite of what the protein boom produced. Both pieces are about the same cultural moment. They land on different conclusions about the same question: when does a health claim earn its price?

The restaurant piece by Nana Kofi asked what automation does to the experience of dining—and whether the businesses that optimized for throughput lost the thing that made the experience worth paying for. Erewhon is the answer to that question in retail form. The store’s aisles are deliberately narrow—4.5 feet instead of the standard 6—to encourage customers to ask for assistance. The highest employee-to-customer ratio in retail is not an inefficiency. It is the product. Haidilao automated logistics to free people for hospitality. Erewhon designed the physical space to make hospitality unavoidable.

The Attention Economy by Renata is the intellectual frame underneath Roberto’s closing observation—that Erewhon is an attention economy dressed up as a produce aisle. The $19 strawberry is content. The outrage it generates is distribution. The people standing in line for a Hailey Bieber smoothie are not there for the protein — they are there to be seen, to post, to prove they were there. Erewhon figured out how to sell a grocery run as content before most media companies figured out how to sell content at all. Once you see that, the price stops being absurd and starts being obvious.

Since we last wrote…

Last week Roberto wrote an urgent Deep Dive on peptides, published days before an FDA vote on whether or not they would be approved for sale. Just a few days later, the vote took place and the FDA backed easier access to peptides, changing the game altogether.

What comes next has been traced by Roberto in his article, an urgent read in such a fast-changing landscape.

Welcome to Origins, a new monthly column tracing the real history behind the companies people think they already know. We're starting with Erewhon.

Erewhon is one of the most widely discussed retail brands in America, but its real story extends well beyond social media. It began as a macrobiotic food collective in a Boston basement, struggled through decades of financial instability, and was rebuilt in Los Angeles into a grocery concept with unusual pricing power, deep customer loyalty, and unit economics that bear little resemblance to a standard supermarket. Behind the $20 smoothies and celebrity partnerships lies something far older and stranger than either. This is the story of how that happened, and why it matters.

Subscribe now

The deal that closed on April Fools’ Day

Even if you haven’t visited Erewhon, you’ve likely encountered it. Maybe it was Hailey Bieber with a pink smoothie on your Instagram, or a friend sending you a picture of a $19 strawberry along with a bunch of question marks. Perhaps it was the Balenciaga runway show featuring models with Erewhon tote bags instead of handbags, or a TikTok video showcasing a $200 grocery haul. At some point, you’ve heard the name and wondered: What is this place, and why does a grocery store have such a significant presence in our culture?

The answer might surprise you. Erewhon is not a startup or a concept created from a pitch deck or funded by wellness trends. It is a 60-year-old idea that nearly faced extinction twice and spent three decades in obscurity on a single block in Los Angeles. In 2011, a retired food distributor walked in off the street and made an offer. The $20 smoothies, celebrity collaborations, fashion drops, and long lines were not part of a plan. They were discovered by a generation that made grocery shopping a part of their identity.

That man was Tony Antoci, and he had never set foot inside the store before he decided to buy it.

In 2011, six months into an unexpected retirement, Tony felt restless. He had sold his food distribution company to Sysco in 2009 and was looking for purpose. Someone pitched him on a “Whole Foods on a bus” idea, but he wasn’t interested. However, that conversation led him to the natural foods business and a struggling store on Beverly Boulevard that had been operating, in various states of decline, since 1969.

He visited the store and discovered it had not kept up with the times. According to his wife, Josephine Antoci, the hot bar resembled prison food and emitted an unpleasant odor. Their son, Alec, who accompanied them, walked in upset because there were no Fruity Pebbles. "They only had granola and healthy stuff."

Undeterred, Tony knocked on the manager’s office door until someone let him in. The owner, Libby DeSilva, was a hairdresser who had inherited the store from her husband Tom, who had bought it out of bankruptcy in 1979. The business was doing roughly $180,000 a week in sales, and Tony made an offer on the spot.

The deal closed on April 1, 2011, April Fools’ Day. He had no idea what he’d just begun.

Before Tony — The Forgotten Origin (1966–2010)

The name Erewhon comes from Samuel Butler’s 1872 satirical novel, an anagram of “nowhere.” In Butler’s story, Erewhon is a bizarre utopia where views on health, morality, and illness are turned on their head. Two Japanese immigrants, Michio and Aveline Kushi, chose that name for their macrobiotic health food store to represent their philosophy rather than just branding.

Aveline and Michio Kushi opened the original Erewhon on April 9, 1966, in a small basement at 303-B Newbury Street in Boston. Aveline was the official owner, while Michio was the driving force behind their philosophy. Both studied under George Ohsawa, the Japanese founder of the modern macrobiotic movement. Erewhon was more than a store; it was part of a larger mission: seeing food as medicine, using commerce for good, and believing health can be achieved through proper nutrition.

The selection was limited to organic grains, beans, and soy products imported from Japan, displayed in bulk bins with wooden interiors. It is often seen as one of the first natural food stores in the country and a blueprint for the movement that followed.

In 1969, Aveline moved to Los Angeles and opened Erewhon West at 8001 Beverly Boulevard. While the foundational philosophy came along, the business struggled. Erewhon expanded significantly in the 1970s, overreached, and filed for Chapter 11 bankruptcy in 1981, causing the distribution and wholesale operation to collapse. The LA retail store managed to survive, was purchased out of bankruptcy by employee Tom DeSilva in 1979, and remained with his family for the next 30 years until his death, after which his widow, Libby, took over.

By the time Tony arrived, the store had occupied the same musty space at 7660 Beverly Boulevard since 1991. The founding philosophy was still present in theory, with a commitment to natural foods and sourcing standards, but the business had not been competitive for years. The natural foods market had evolved. Whole Foods had expanded nationally, and the store resembled a relic from the past.

What the original Erewhon left behind, despite its commercial failures, was more difficult to measure: a 40-year impact on Los Angeles’s wellness culture. The city had accepted the idea that food could serve more than just nutrition. Tony didn’t invent that idea; he inherited it.

The Antoci Bet (2011–2014)

Tony Antoci grew up in Pasadena, the grandson of Sicilian immigrants. Josephine emigrated from Taiwan. They met in 1986 at a Chinese restaurant in Beverly Hills and married in 1993. Before Erewhon, they built Superior Anhausner Foods, a specialty food distribution business supplying Westside restaurants, which grew to around $90 million a year in sales before being sold to Sysco in 2009. The exact price was never revealed.

What Tony saw in Erewhon was an opportunity. Whole Foods, which had set the standard for the premium natural foods category, was purchased by Amazon in 2017 and was shifting to a more mass-market approach. The kind of health food store that held firm on sourcing standards, rather than scaling by loosening them, had all but disappeared. That left a real gap in the market, one nobody had stepped into yet.

His early decisions were strategic. Location strategy first: Erewhon would only open in the wealthiest neighborhoods of LA, without compromise. The second store opened in Calabasas in 2014 and turned a profit within six months. Venice followed in 2016, then Pacific Palisades, Silver Lake, Culver City, Studio City, Beverly Hills, Century City, and the Pasadena flagship in September 2023. Each store ended up in neighborhoods already inhabited by the customers Tony aimed to attract.

The second key decision was food quality, which was Josephine’s area of expertise. As the president and co-owner of Erewhon, she manages every aspect of the food program, overseeing what goes into the hot bar and onto the shelves. Each product is strictly vetted against a long list of banned ingredients, including processed sugars, refined vegetable oils, and artificial colors and flavors. Her standard, as her son Alec has described it, is straightforward: “If I wouldn’t bring it into my house, I’m not going to bring it into the store.” She turns down almost every product she’s pitched, which is why getting onto Erewhon’s shelves has become a coveted stamp of approval for CPG brands.

The third decision revolved around prepared meals. Tony quickly recognized that food service, including the hot bar, tonic bar, and prepared meals, would drive the business. In his words:

“We quickly recognized that food service would be the key driver of our business. Ironically, I wasn’t interested in opening a restaurant, but I’m pleased to say I don’t have one.”

In June 2021, Erewhon became a Certified B Corporation, achieving a B Impact score of 83.6, compared to a median of 50.9 for typical businesses taking the assessment. This certification requires re-verification every three years. Most brands at Erewhon's growth stage tend to relax their values to expand more easily, reducing sourcing standards, broadening their product range, and aiming for higher volume. However, Tony took the opposite approach. He sought the certification and strengthened the original ethos rather than compromising it, transforming a potential challenge into a credibility marker that supported the brand's premium pricing. The Kushis had established Erewhon based on the idea that business could be a force for good, and that notion endured through bankruptcy, decades of obscurity, and a shift in ownership. Under Tony, it evolved into a competitive advantage.

The Model — Why the Numbers Don’t Make Sense (Until They Do)

A grocery store is typically a low-margin, high-volume business. The industry average for net margins is about 3.8%. Grocers compete on price, variety, and convenience. The products are mostly similar, and the economics are tight by design.

Erewhon does not face these same limitations.

In 2019, Forbes reported that Erewhon generated around $2,500 in sales per square foot, which is more than four times the industry average of about $500. The LA Times, in its January 2025 coverage of the expansion, placed the range at $1,800 to $2,500. For comparison, a well-trafficked Whole Foods does about $900 per square foot, while Tiffany & Co. achieves around $3,000. Erewhon operates within that range, focusing on food.

The key reason for this success is their focus on prepared foods. In the grocery industry, prepared foods usually make up about 4-6% of revenue on average. At Erewhon, according to Tony Antoci, they account for roughly 40% of revenue. This single difference transforms the margin profile because prepared food is not just a grocery business. It operates like a hospitality business under the same roof, and the economics reflect that.

The cafe represents this concept perfectly. At every Erewhon location, the tonic bar and hot bar take up about 30% of the store space, and the company’s all-day cafes serve 100,000 customers weekly across the system. But the numbers alone don’t explain why people linger. The architecture firm RDC designed Erewhon’s stores with tight 4.5-foot aisles instead of the typical 6 feet, along with intentionally low shelving. Terry Todd, an associate principal at RDC, explained the goal:

“This encourages customers to ask for assistance. Erewhon has the highest employee-to-customer ratio of which I am aware, and we had to factor that into the spatial design.”

The design, as MOLD Magazine called it, fosters interactions. The store is built to make people slow down, connect, and return, which is why the economics of the cafe work so well.

Same-store sales reflect that compounding. Tony told the Hollywood Reporter:

“During the pandemic, our sales increased 30 percent. They’ve increased another 30 percent since then. Our same-store sales are up 18 percent year-over-year.”

These figures are not standard for grocery stores. They align more closely with brand-driven growth than with traditional supermarket comparisons. Typical supermarket same-store sales growth is usually in the low single digits, even in good years. Industry data often highlight 1-3% annual same-store sales growth as a healthy benchmark for conventional chains, with many struggling to reach that level consistently. In contrast, Erewhon’s 18% year-over-year same-store sales and 30% gains during and after the pandemic resemble the growth curve of a high-velocity brand or specialty retailer rather than a conventional grocery chain.

Erewhon is privately held and has never publicly shared its financial data. Fast Company estimated its profit in 2023 at around $171 million, a figure that has been widely cited since. However, Erewhon has not confirmed this, and the method behind the estimate is not detailed. A revenue figure of $750 million also appears in recent media stories, but it likely stems from secondary estimates rather than any primary company filings. The strongest public data remains the Forbes sales-per-square-foot figure, Tony Antoci’s reported store average, and the prepared foods share, all pointing in the same direction even without a confirmed income statement.

The institutional validation of the model came in April 2019, when New York growth-equity firm Stripes Group took a minority stake in Erewhon. Ken Fox, the firm’s founder, told Forbes:

“Erewhon can become a national brand and universally the most trusted in better-for-you food, period. Every metric you would measure this business on is an outlier in a few different industries: specialty restaurants, traditional retail and grocery.”

The deal terms were not disclosed, but an investor document from that transaction, not verified in public records, mentioned an investment of $93.8 million for a 25% stake at a valuation of $375.6 million, based on a $124.2 million annualized revenue run-rate as of early 2019.

The commissary is crucial to this success. For years, Erewhon made its prepared foods at a small facility in Boyle Heights. After three years of development, they moved into a new central kitchen in Vernon, covering 65,000 square feet within a 110,000-square-foot building, which is five times larger than the previous site. About 350 of Erewhon’s 2,500 employees work there, preparing every item on the hot bar, such as buffalo cauliflower, coconut chicken tenders, and gluten-free brownies, for delivery to stores by 5 a.m. each day. In late 2025, the Antocis bought the building for around $51 million, highlighting the commissary's importance to their operations.

The sustainability operations run alongside all of this. Erewhon says its glass reuse program reuses over 2 million bottles and jars each year. The company also partners with The Midnight Mission, making daily food contributions and supporting donation campaigns for Los Angeles’s homeless community. These are not just marketing line items; they are part of the same founding thesis, now running at a much larger scale.

The Smoothie Industrial Complex

In February 2021, during the pandemic, TikTok creator Christina Najjar, known as Tinx, regularly posted about her visits to Erewhon. Vito Antoci, Erewhon’s executive vice president and Tony’s brother, noticed that her posts were bringing in customers and wanted to thank her. Tinx requested a secret menu item with her name on it, and the Tinx Smoothie was created. As later reports stated, “the first official collaboration came from TikTok creator Tinx,” setting the stage for the celebrity smoothie trend that followed.

What turned an informal favor into a business format was a decision made in March 2022 by influencer Marianna Hewitt, co-founder of Summer Fridays skincare, who collaborated with Erewhon’s tonic bar team on the Coconut Cloud smoothie. Where previous smoothies had been built around taste and ingredients, Hewitt pushed for something designed to be looked at. Blue spirulina gave it a striking color that would travel on social media. “I loved their smoothies, but they aren’t always the most beautiful things,” she said. “Why don’t we make it not only delicious and healthy, but something really shareable and beautiful?” The Coconut Cloud launched at $17, and with it, the aesthetic logic of the celebrity smoothie was set.

Then Hailey Bieber shared her usual order on Instagram Stories, a strawberry-based smoothie she had enjoyed for years. That week, Erewhon reported a surge of customers asking to recreate it. She convinced them to create a drink named after her, and the Strawberry Glaze Skin Smoothie launched in June 2022 alongside her Rhode skincare brand. It racked up approximately 450 million views on TikTok and stayed on the menu long after its debut.

Vito Antoci told GQ that the Bieber smoothie sells approximately 40,000 units per month across Erewhon’s locations, which at an $18 price point works out to roughly $720,000 in monthly revenue from a single product. At the peak of the Bieber smoothie craze in 2023, Erewhon reportedly sold as many as 1,500 drinks a day across the system. Based on reported sales and pricing, the smoothie’s annual revenue could be around $8.6 million to $10.6 million, depending on the price, making this figure an important point for understanding the brand’s finances.

The economics of the collaboration model are structured carefully. Each recipe includes up to five ingredient brands that pay for inclusion, according to Vito. Celebrities receive a $1 royalty per smoothie sold, which they are encouraged to donate to charity, and Erewhon adds $2 per smoothie to the same cause. In 2024, the company said it donated more than $2.5 million through these celebrity smoothie collaborations. Brands appearing in a celebrity smoothie reportedly paid between $50,000 and $60,000 for their placement, according to sources cited by Vogue Business, although Erewhon has not confirmed these figures.

The flywheel compounds with each new collaboration. A collab creates content, which then drives store traffic, generating brand excitement and demand for future collaborations. When BLACKPINK’s Lisa released a Thai iced tea smoothie in January 2025, Erewhon’s TikTok announcement drew over 20 million views within three days. In June 2026, tied to the FIFA World Cup, Colombian artist J Balvin became the third Latin American collaborator after Becky G and Kali Uchis, launching a Colombian flag-themed smoothie available in both the LA and New York locations. The pattern by now is familiar: a celebrity with a built-in fanbase, a visually engineered drink, and a launch timed to a cultural moment bigger than the store itself. The scale of that response, for a grocery store announcing a new drink, is the clearest illustration of what the program has become.

The Hollywood Reporter noted that these smoothies are “better marketing tools than profit earners,” with high ingredient and labor costs cutting into each cup’s profit. The industrial blenders that make six smoothies at once cost $2,000 each, and workers at the tonic bar during busy hours earn between $18 and $30 an hour. The business focus was never solely on selling smoothies; it was about drawing attention and creating a program based on what a favor to a TikTok creator accidentally showed was possible.

The Cult

Erewhon has no advertising budget and no marketing team to speak of, yet the #Erewhon tag had accumulated billions of TikTok views by 2022. The brand sits at the heart of LA's cultural conversation without spending money to get there. This raises two possibilities: it might be the most effective marketing strategy in retail, or it shows that the brand has grown beyond what marketing can create.

Artist Alex Israel, who goes to Erewhon every day, described it to the Hollywood Reporter as “the quintessential LA experience once provided by fashion boutique Fred Segal,” He noted that “Erewhon’s significance in the culture reflects a shift in desire, from designer clothes toward designer bodies.” This description captures more about Erewhon than just its products. It isn't just food or wellness; it's a part of LA culture that people are eager to pay to join.

The fashion collaborations are the clearest expression of that positioning. In April 2022, Erewhon partnered with Casablanca on a $165 tote bag and a custom juice. In August 2023, a Cactus Plant Flea Market drop, a $22 rainbow smoothie plus hoodies, crashed the website and sold 20,000 smoothies in one week. In December 2023, Balenciaga staged its first LA runway show and models carried Erewhon totes unpaid, and Kim Kardashian attended sporting the same bag. None of these connections were arranged or paid for, meaning the brand merely became a desirable partner for luxury fashion, a different kind of influence compared to traditional advertising.

The cultural references compound in the same way. Kanye West tweeted about “Erewhon drip” in 2018 before there was any official merchandise to buy. Donald Glover said he wanted his studio to be “the Erewhon for culture.” The New York Times called it “the unofficial hangout for the young, beautiful, and bored.” Criticism also serves as promotion. Articles discussing $19 strawberries and $26 oxygenated water introduce the brand to those who weren’t familiar with it. The outrage and aspiration cycles boost each other.

Each store has its own identity rather than a uniform format. Alec Antoci describes it as Beverly Hills running touristy, Venice beachy, and Silver Lake edgy, a deliberate differentiation. Josephine’s weekly tasting sessions ensure the sourcing standards are consistent across locations, but the stores are designed to reflect their neighborhoods rather than replicate each other, which is part of why the brand feels local even at ten locations.

The membership program, launched alongside the Erewhon app in July 2023 at $100 or $200 per year, had reached approximately 50,000 paying members by late 2025. Members who engage with the digital tools spend roughly 70% more, per Chief Growth Officer Kabir Jain. The paid membership for a grocery store signals a cultural shift, showing that Erewhon sells a sense of belonging just as much as it sells food. Many people are willing to pay for the privilege of being part of this community.

Erewhon 2.0

For thirteen years, Erewhon operated exclusively in LA County, ten stores, all in affluent neighborhoods, all within range of a commissary that could supply them before the doors opened. That constraint was not a limitation. It was the operating logic, and Tony knew it.

In January 2025, he announced what he called Erewhon 2.0, a wave of expansion enabled by the new Vernon commissary, which at five times the capacity of the previous facility had finally given the business the infrastructure to move. Three new stores in one year: Manhattan Beach, the first South Bay location, West Hollywood, and Glendale, the most openings in a single year since 2011. A Thousand Oaks location is planned for 2026, the first move into Ventura County, and a downtown LA flagship is in development at 940 S. Hill St., where the Antocis paid $13.5 million for a former parking structure. By mid-2026, the West Hollywood and Glendale stores had both opened, bringing the total to 13 locations, still all within Greater LA. Erewhon also took its cafe concept somewhere new: a location inside LACMA’s newly built $724 million David Geffen Galleries, opening to the public in May 2026, with a full restaurant and wine bar planned to follow. It is a small expansion in square footage but a meaningful one in positioning, placing Erewhon inside one of the city’s most prestigious cultural institutions rather than just its wealthiest zip codes.

The New York question has followed the brand for years, and the answers have shifted gradually without fully committing. In 2023, Tony said that a New York store was “not happening, at least not yet.” By 2025, he was calling out-of-state expansion “a major focal point” with New York “absolutely on our radar,” while stressing the immediate focus remained Southern California. The reality on the ground in 2025 was a tonic bar inside Kith Ivy, a members-only padel complex at 609 Greenwich Street in Greenwich Village with an initiation fee of $36,000 plus $7,000 a month, making it a $43,000-a-year private club. Non-members within a delivery radius can order Erewhon smoothies and juices via Postmates and Uber Eats, but the full store experience does not exist outside California. That Erewhon’s first move beyond its home state is a tonic bar tucked inside one of New York’s most expensive private clubs is not an accident. It is a brand signal, a way of arriving in a new market at exactly the right altitude before committing to a full store. A September 2025 Instagram post teasing a 20-city national rollout was walked back shortly after, with the company clarifying the push was for nationwide shipping of DIY smoothie kits, not new locations.

Tony has said he would never sell Erewhon to Amazon, citing what happened to Whole Foods. That line matters because Stripes, as a growth-equity investor, will eventually seek liquidity, and an IPO would for the first time make the real financials public, changing the conversation around Erewhon’s economics considerably.

On Succession, the family is candid and on the record about the tension. Alec, the eldest son, runs brand and marketing and oversees the fashion collaborations. Austin helps with the real estate portfolio and stays largely out of the spotlight. Maddy is a marketing coordinator. Vito, Tony’s brother, is EVP and runs the smoothie program. When the Hollywood Reporter interviewer observed, “This really is Succession,” Tony replied: “It’s hard love. He’s got to prove it,” and described his own role in the handoff as driving from the back seat for as long as he is around.

The open question is not whether Erewhon can open more stores in LA, because it clearly can. The question is whether the concept travels. The brand is deeply embedded in a specific set of conditions, the climate, the celebrity infrastructure, the concentration of high-disposable-income consumers who treat grocery shopping as an extension of their identity, that exist in LA in a way they do not exist anywhere else at the same density. That is a feature in LA and an unknown in every other place. The cautionary counterpoint is Foxtrot, the venture-backed upscale convenience concept that abruptly collapsed in 2024, a reminder that premium grocery positioning and cultural cachet do not on their own guarantee survival when the unit economics don’t hold up outside the right zip codes.

What Erewhon Actually Proves

The lesson from Erewhon is not “charge more.” Plenty of businesses have tried to charge more and failed. The lesson is that a business can escape its category entirely if it becomes something people buy into rather than just shop at, and that the difference between those two things is not a marketing strategy. It is a structural decision made at the level of the business model.

The prepared foods bet was the structural unlock. It took a low-margin commodity business and embedded a high-margin hospitality business inside it under one roof, with the tonic bar and hot bar occupying 30% of the floor space and generating economics that have nothing to do with selling bags of rice. That realization, which Tony arrived at early and committed to without compromise, is what makes Erewhon’s unit economics look the way they do, and it is why the $2,500 in sales per square foot figure is not a fluke. Erewhon does not compete with grocery. It competes with luxury retail and experiential hospitality, and in LA, it wins.

The influencer flywheel was not manufactured either. When Tinx asked for a secret menu item in February 2021 and Erewhon said yes, nobody in the building understood what they were building. The celebrity smoothie program is now a formal operation, monthly collaborations, ingredient brand fees, charity royalties, industrial blenders mixing six at a time, but it started as a favor to a regular customer who happened to have 1.5 million TikTok followers. The infrastructure was built around what worked, not the other way around.

The B Corp thread runs the entire length of the story and is worth naming plainly. Aveline and Michio Kushi opened a store in a Boston basement in 1966 on the belief that commerce could be organized around health and community rather than extraction. Paul Hawken built the sourcing standards that made that belief operational, contracting 57 farms across 35 states and helping co-found what became organic certification in the United States. That business failed commercially. Tony Antoci walked into its last surviving outpost 45 years later looking for something to do in retirement, found a 60-year-old idea that nobody had figured out how to scale, and scaled it, with the same sourcing obsession intact.

Not everyone reads that arc as a success story. Paul Hawken, who has spent the decades since Erewhon became one of the country’s leading climate activists, was direct when asked about what the brand has become: “The Erewhon I helped found has failed. We used to be affordable. The ingredients they still source are incredible, and I’m glad that’s still in the DNA. But it’s so expensive, you can’t feed the world that way.” That tension, between the integrity of the sourcing and the exclusivity of the price, is real and worth sitting with. The founding philosophy was never meant to be a luxury product. Whether it can be both things at once, genuinely committed to its values and genuinely inaccessible to most people, is a question the brand has not answered and may never need to, as long as the culture keeps rewarding it for trying.

The question that remains, the one Stripes is presumably waiting to answer and the one Tony deflects with talk of Southern California focus, is whether the idea travels. The name, after all, is an anagram of nowhere, and that was always both the joke and the warning. Tony Antoci walked into a smelly health food store on Beverly Blvd and found something strange: a 60-year-old idea, commercially failed but culturally intact, waiting for someone to figure out the business model. He figured it out. Whether the idea is somewhere or nowhere, whether it belongs to LA specifically or to something more portable, that is still the open question, and the answer will define what Erewhon actually is once the smoothie hype fades.

Closing Thoughts

What surprises me most about Erewhon isn’t the insane prices on its products. It’s what people are actually buying. Attention.

I would say that the people standing in line for a Hailey Bieber smoothie isn’t really there for the protein. They’re there to be seen. Buying it, posting it, proving they were there. Erewhon figured out how to sell a grocery run as content, and once you see that, the $19 strawberry stops being absurd and starts being obvious. It’s not a grocery store competing on price or freshness. It’s an attention economy dressed up as a produce aisle, and it’s working exactly as designed.

This felt like the right way to close the first piece of this new column. Because these days we talk a lot about brands going viral, but the real question is always how they got there.

This is the first edition of Origins, a new series where we go deep on one company each month, the story behind it, the numbers, and what it actually reveals. More coming soon.

It’s insane that still only 2.2% of US households pay for AI subscriptions. Out of those, the big majority still may use it at a superficial level, like using chat features, image generation, or research.

That said, the number might not even climb much from here. With better free models dropping every few weeks, why would most people pay when yesterday’s frontier model is free today? You will see later on these news that a couple of open source models were released, Kimi K3 and Inkling.

Usage keeps climbing, sure, but the share of people willing to pull out a credit card could just stay flat. Or it goes the other way, someone ships a model so cheap that paying becomes a no-brainer for a few more people, but revenue barely moves because price per user keeps falling. Either way, AI starts to look like a commodity, priced more like cloud storage or your phone bill than like a premium product.

And 2.2% might already be undercounting things. A lot of people are probably paying for AI without realizing it, it’s already included into their office suite, their design tool, their search bar, their work laptop. If that’s the case, a flat subscription number isn’t really telling us adoption is slow, it’s telling us the standalone subscription model might not be how most people end up paying for this.

Really interesting thought, but I am curious how many of you actually pay for AI subscriptions.

-Roberto Mazariegos

Subscribe now

MARKETS & DEALS

Uber offers $14.8B for Delivery Hero

Uber is buying German delivery giant Delivery Hero for $14.8 billion, a deal that would nearly double Uber's global delivery footprint to 99 markets overnight. Delivery Hero's own board is on board, and Prosus (a major shareholder) has already agreed to sell its shares, so this one looks likely to close. Uber is footing the bill with cash on hand plus a large loan, and expects the combined business to start boosting profits within three years. (Bloomberg)

Bloomberg Intelligence says:

Two fintech heavyweights team up in a $53 billion bid for PayPal

Stripe and private equity firm Advent International want to buy PayPal together, offering a 28% premium over its recent share price in one of the biggest takeover attempts of the year. The news sent PayPal’s stock up as much as 17% in a single day. PayPal’s board is expected to weigh in as soon as this Monday, and hasn’t said publicly yet whether it’s open to a deal. If it happens, it would reshape the online payments landscape almost overnight. (CNBC)

The AI boom is so expensive that even data center builders want out

Building the data centers powering AI has become so costly that the companies building them are now racing to sell majority stakes worth tens of billions of dollars to private equity firms, rather than keep and operate them. One deal alone, for Dallas-based DataBank, could be worth as much as $25 billion. The driver is simple: shortages of everything from electricians to memory chips have pushed the cost of building a gigawatt of new AI computing power toward $80-100 billion, according to Nvidia’s CEO. It’s a sign that even AI’s biggest believers are hedging their bets on how this all plays out. (WSJ)

DeepSeek wants more money, and then IPO

Chinese AI lab DeepSeek is in talks to raise another $1.5 billion, just a month after its first-ever outside funding round, and this time investors are valuing it at $71 billion, nearly seven times what it was worth three months ago. The company is now also preparing to go public in mainland China, with a filing possible as soon as this year. The pace of the run-up shows investors are still betting heavily on Chinese AI labs, even as Washington tries to slow their access to top chips. (TechCrunch)

Small, unloved stocks are quietly having their best year since 2003

The Russell 2000, a benchmark of smaller US companies, is up 20% this year and beating the S&P 500 by the widest margin in over two decades, while the "Magnificent Seven" mega-cap AI stocks have gained just 4%. Investors are chasing smaller companies positioned to benefit from the AI infrastructure buildout, even unprofitable ones, which have surged 154% since mid-2025. It's a sign the AI trade is broadening out beyond the handful of giant tech names that have driven the market for the past two years. (FT)

BIG TECH & INFRASTRUCTURE

Hyundai workers strike over robot coworkers, not pay

For the first time, a labor strike has been triggered specifically by a company’s plan to bring humanoid robots onto the factory floor. About 35,000 Hyundai autoworkers in South Korea walked off the job, not over wages, but to demand a say before Atlas, Hyundai’s humanoid robot, ever sets foot in a plant. No deployment date has even been set, which means workers are trying to lock in protections before the robots arrive rather than after. It’s a preview of a fight most industries haven’t had yet: robot costs are expected to fall from roughly $130,000-$140,000 today to around $30,000 once production scales, which is exactly what has workers worried. (WSJ)

New York just hit pause on the AI data center boom

New York became the first state in the country to freeze construction of new large AI data centers, with Governor Kathy Hochul signing an order that blocks big projects for up to a year. The goal is to give the state time to write rules that protect people’s electricity bills and the power grid before more of these facilities get built. It’s the clearest sign yet that the AI industry’s rapid buildout is starting to run into real political and public resistance, not just technical limits. (NYT)

Japan and Nvidia are building an AI factory meant to power robots, not chatbots

Nvidia and Japanese company Noetra are building what they’re calling the world’s first “national AI infrastructure,” a massive computing facility backed by the Japanese government and aimed squarely at robotics, manufacturing, and healthcare rather than the chatbots most data centers serve today. Japan has already committed hundreds of billions of yen to the project, with more available over the next five years. It’s a bet that the next phase of the AI race is about controlling physical machines, not just answering questions. (WSJ)

Cursor is reportedly building its own rival to Claude Cowork

Cursor is developing a general-purpose AI agent, codenamed Sand, meant to handle everyday office work like email and documents rather than just writing code. It’s a sign Cursor wants to move beyond software developers and compete directly for the broader knowledge-worker market that products like Claude Cowork are targeting. The company hasn’t confirmed a launch date, or even that it will ship the product at all. (The Information)

OpenAI's first physical product is a speaker designed to feel alive

OpenAI is entering the hardware business with its first device: a screen-free, movable smart speaker built to act as a companion in your home rather than just answer commands, designed with help from Jony Ive’s studio and several former Apple engineers. It’s meant to control smart-home gadgets, play media, and get to know its owner over time. OpenAI is aiming to unveil it this year with a release in 2027, though that could shift depending on the company’s ongoing legal battle with Apple. (Bloomberg)

A Taiwanese battery startup is betting it can beat China at its own game

Prologium, a Taiwanese solid-state battery maker, is one of the few companies in the world that can actually mass-produce solid-state batteries, which are safer and more energy-dense than the ones in most EVs today but notoriously hard to manufacture at scale. The company is going public via a Nasdaq merger at a roughly $3.8 billion valuation and is building its first overseas factory in France. It represents a rare opening for a non-Chinese company to compete in a battery market China otherwise dominates. (Wired)

AI & MODELS

DeepSeek moment 2.0? China just released the biggest open AI model anyone has ever put out for free: Kimi K3

Moonshot AI released Kimi K3, now the largest open-source AI model in the world, and one that performs close to the best paid models from Anthropic and OpenAI on real-world tasks. That matters because anyone- developers, companies, competitors- can download and use it for free once the full model weights ship on July 27. It’s another sign that Chinese AI labs are closing the gap with US frontier labs faster than expected. (VentureBeat)

Big US companies are quietly switching to cheaper Chinese AI models

DoorDash, Siemens, and Airbnb are among the companies experimenting with Chinese AI models from labs like Moonshot and DeepSeek to cut costs, with DoorDash’s co-founder saying the company is already saving money using a Moonshot model for lower-stakes work. One analysis found Chinese models deliver about 95% of the capability of top US models at roughly 10% of the price. The catch: security researchers warn that routing company and user data through Chinese-made models raises real data privacy questions that many companies haven’t fully worked through yet. (FT)

Mira Murati's startup finally shows what it's been building

Thinking Machines Lab, the AI company founded by OpenAI’s former CTO Mira Murati, released its first product, an open-source model called Inkling that anyone can download and customize, unlike the closed models from OpenAI, Anthropic, or Google. The company built it in about nine months, a fraction of how long it took the biggest labs to ship their first products. Rather than chasing the top spot on leaderboards, Thinking Machines is positioning Inkling as a flexible starting point for companies who want more control over how their AI behaves. (TechCrunch)

A robot hand just got precise enough to feel a glass slipping out of its grip

Robotics company 1X built new hands for its humanoid robot Neo that can lift a 20-pound kettlebell and, moments later, pick a single grape off its stem without crushing it. Built-in sensors let the hand feel when something is about to slip and adjust its grip automatically, the kind of dexterity that’s been one of the hardest problems in robotics. 1X says it plans to manufacture 10,000 of these hands this year, a sign the company thinks the technology is ready to leave the lab. (1X)

CAREER CORNER

AI isn't killing entry-level jobs, it's rewriting what they are

After two years of companies freezing junior hiring as AI chewed through the repetitive grunt work new grads used to cut their teeth on, US employers now expect to hire 5.6% more college graduates this year, and the companies investing most heavily in AI are hiring the most entry-level staff of all, according to jobs data firm Revelio Labs. The catch is that the jobs themselves have changed: PwC found a 35% jump since 2019 in entry-level postings demanding senior-level skills like data-driven decision-making and people management, what researchers are calling the "seniorisation" of junior roles. LinkedIn's chief economic opportunity officer frames it as work shifting "from grunt work to real work," which is good news if you land one of these roles, and cold comfort if you're one of the many grads still fishing in a shrunken pool of openings. Some companies are leaning into it anyway, Salesforce is bringing on 1,000 recent grads and interns this year, more than in 2025, specifically to put them on AI projects across the business. (FT)

WHAT ELSE WE’RE WATCHING

• A Sweet Surprise: Scientists Find Sugar Deep in Our Galaxy

• Children’s social media curbs planned across EU, von der Leyen says

• Scientists are using AI and quantum computing to design new drugs on a shoestring budget

• Databricks Valued at $188 Billion as AI Demand Surges

That’s the news,

From the Nido

In the 2011 film Limitless, Bradley Cooper’s character takes an experimental pill that unlocks total access to his brain, and within days he transforms from a struggling writer into someone unrecognizably sharper, more capable, more alive. The premise only works as fiction because no such pill exists. What does exist, and what a growing number of people believe comes close, is a peptide.

Think of a peptide as a text message your body sends itself. It's a short chain of amino acids, and its only job is to deliver a brief, specific instruction: tell this cell to release a hormone, tell that one to start repairing tissue, tell the brain the stomach is full. Your body runs thousands of these messages a day without you ever noticing. Insulin is one of them. Scientists found a way to copy the fullness message and make it last far longer than your body ever intended, and the result became Ozempic, a drug that has since reshaped how an entire industry talks about weight loss. And lately, a lot of people have started wondering what other messages might be worth sending.

Subscribe now

A woman posts a video saying she has never felt this confident, this good, this twenty-something, since she actually was twenty-something, four kids ago. The peptides she takes, she says, made the difference. Another creator walks through what he calls the Pretty Boy Peptide, promising the glass-skin effect that’s been circulating across beauty accounts all year. At a Wall Street Journal event this spring, Diplo mentioned, almost in passing, that he’d met forty different people at a single after-party who all wanted to talk about the same thing: peptides (more on WSJ Podcasts).

The compounds showing up in those videos are not Ozempic, a different peptide altogether, built and tested through a completely different process. Most have never been tested on a human being in a clinical trial. They exist in a kind of legal loophole, sold online as “for research use only,” a label that lets sellers ship them to anyone willing to inject something bought from a stranger’s website. Health and Human Services Secretary Robert F. Kennedy Jr. is one of the people willing. Earlier this year, he told Joe Rogan he considers himself a fan, and that he has used peptides himself with good results on a couple of injuries.

That contradiction is the story. A category of molecule doing quiet, essential work inside every human body has also become one of the fastest-growing unregulated markets in the country. On July 23, the FDA will begin deciding what to do about that. Over two days, a federal advisory committee will hear arguments from doctors, compounding pharmacists, and public health officials divided over whether these drugs are ready to leave the black market, or whether nothing about them has actually gotten safer.

Not All Peptides Are Created Equal

To understand why July 23 matters, you need to understand what peptides actually are, and more importantly, what they are not.

Zepbound, Ozempic’s biggest rival, is built on the same principle, another synthetic message copied from the body’s own signaling system. The FDA has approved peptide-based drugs continuously for decades across oncology, metabolic disease, hormone disorders, and rare conditions, and the global market for peptide therapeutics was valued at $140.9 billion in 2025 (more on Grand View Research).

Ozempic and Zepbound belong to a specific peptide subclass called GLP-1 receptor agonists, engineered to mimic a gut hormone that regulates appetite and blood sugar. Every GLP-1 drug is a peptide. Not every peptide is a GLP-1. The compounds at the center of this story, BPC-157, TB-500, GHK-Cu, and others, are peptides too, but they work through entirely different mechanisms and were never developed, tested, or approved the way Ozempic was. The shared label is chemistry. The shared reputation, deserved or not, is where the confusion usually starts.

BPC-157, a synthetic peptide derived from a protein found in stomach acid, is marketed for tissue repair and gut healing. TB-500 is promoted for muscle recovery and wound healing. GHK-Cu, sometimes called the beauty peptide, is sold for skin regeneration and antiaging. Semax, originally developed in Russia, circulates in longevity circles as a cognitive enhancer. Retatrutide, still in Phase 3 clinical trials at Eli Lilly, is being bought off WhatsApp groups and injected at home by people who cannot wait for the regulatory process to run its course (more on NYT).

What these compounds share is not a mechanism or a therapeutic category. What they share is a data problem. For BPC-157, Andrew Huberman noted that despite a long history of use in humans for gut health and tissue repair, the compound has been the subject of only a single formal human clinical study (more on Huberman Lab). TB-500 has almost no peer-reviewed human pharmacokinetic data at all. The FDA’s own language in the 2023 restriction cited “no, or only limited, safety-related information” for several of these compounds, and flagged risks including allergic reactions, tumor-growth acceleration, and liver and kidney damage (more on FDA). Andrew Huberman raised a theoretical concern that BPC-157’s angiogenic effects could be problematic in people with active or prior cancer risk, so caution is warranted in that group.

The distinction that matters for everything that follows is that peptide is a chemistry category, not a risk category. The same molecular class that produced Ozempic is also what someone in Michigan is buying from a Chinese supplier and injecting in their bathroom. That gap, between what is known, what is legal, and what people are already doing, is exactly what the FDA is being asked to close.

How the Gray Market Got Built

The peptide gray market did not appear overnight. Three forces converged to build it, each one expanding what the next could become.

The first was Ozempic. When semaglutide, a GLP-1 receptor, swept through American culture around 2022 and 2023, it did something no wellness trend had done before at that scale: it normalized self-injection for millions of people who had never thought of themselves as the kind of person who injects anything. Weekly injections moved out of clinical settings and into bathroom cabinets, guided by telehealth prescriptions and YouTube tutorials. The psychological barrier that had kept injectable drugs in the domain of diabetics and competitive bodybuilders quietly collapsed, and when those same users started asking what else the peptide category had to offer, a market that had previously been confined to niche biohacker forums suddenly had an enormous new audience (more on WSJ Podcasts).

The second force was the 2023 FDA restriction. In late 2023, the FDA moved 17 bulk drug substances into Category 2 of its 503A Bulks List, including several popular peptides. That designation effectively bars licensed compounding pharmacies from making them, citing immunogenicity risks, impurities in the raw ingredient, difficulty verifying exactly what’s actually in the drug (the FDA calls this API characterization, essentially confirming the active ingredient is what it claims to be), and, in many cases, a lack of human safety data (more on FDA). The intention was consumer protection. The practical outcome was that demand, which was already substantial, had nowhere regulated to go. It migrated to Discord servers, Telegram groups, and e-commerce sites selling compounds labeled “for research use only,” a legal disclaimer that allowed vendors to ship injectable-grade powder to anyone with a credit card and a vague interest in optimization. U.S. customs recorded $328 million in peptide API imports in just the first nine months of 2025, double the prior year figure, and that number captures only what was formally logged at the border (more on Bloomberg).

The third force was the influencer infrastructure, which built the cultural permission structure that made all of this feel normal. Joe Rogan described BPC-157 resolving his own elbow tendonitis in two weeks on his podcast, matter-of-factly, the way someone might mention a physical therapist that worked out well. Andrew Huberman separately shared that two injections resolved a persistent L5 spinal compression that other therapies had failed to fix, while in the same breath urging physician oversight, proper sourcing, and caution around cancer risk. Jennifer Aniston told the Wall Street Journal in 2023 that her routine included weekly peptide injections. None of these people were selling anything directly, but together they constructed a frame in which self-experimenting with injectable compounds read as health-conscious rather than reckless, and that reframe was worth more to the gray market than any advertising campaign.

By the time the regulatory conversation caught up, the market was already operating at a scale that made simple enforcement impossible. Peter Magic’s lab grew more than 1,200% in testing volume between 2023 and 2025, not because someone created demand from scratch, but because demand had been building for years and the 2023 ban simply removed the last friction standing between it and an unregulated supply chain.

The RFK Reversal and What July 23 Actually Means

On that same Joe Rogan appearance, which aired February 27, 2026, Kennedy went further than a personal endorsement. He asserted that the FDA under the Biden administration had illegally reclassified 19 peptides into Category 2 — the category reserved for substances the agency has determined raise significant safety concerns — back in September 2023 (more on Orrick). The announcement traveled fast, and most of the coverage that followed got the mechanics wrong, because the actual regulatory process is considerably more layered than a single presidential appointee announcing a policy change on a podcast.

Here is what is actually happening, broken into the pieces that matter. On April 15, 2026, the FDA formally removed 12 peptide bulk drug substances from Category 2. That removal took effect seven days later, on April 22.

But think of it as three separate steps, not one decision. Step one: come off a safety warning list. Step two: get formally cleared for compounding pharmacies to actually produce it. Step three: full FDA drug approval.

Right now, these peptides have only cleared step one, and that alone sounds like progress without actually changing what’s legal. Getting to step two requires the FDA to write a new rule permitting it, which has not happened yet. Step three would still require years of clinical trials none of these compounds have undergone. Most coverage collapsed all three steps into one, which is exactly why Kennedy’s February announcement caused so much confusion (more on Frier Levitt).

That second step is what the July 23-24 meeting actually is. The FDA’s Pharmacy Compounding Advisory Committee will spend two days at the agency’s White Oak campus reviewing seven peptides, BPC-157, KPV, TB-500, and MOTs-C on the first day, and Emideltide, Semax, and Epitalon on the second, to decide whether to recommend clearing them for compounding (more on The Guardian).

The Gold Rush Economics

While the FDA works through its regulatory sequence, the money has already made up its mind. Telehealth companies, compounding pharmacies, and longevity clinics are building infrastructure now, betting that regulatory clarity arrives on a timeline they can’t afford to wait for.

Hims and Hers acquired a peptide compounding facility in Menlo Park, California, in February 2025. Two months before this year’s hearing, digital health company Noom acquired Tailor Made Compounding, a facility licensed to operate in 46 states. Mino, a platform built for clinicians and med spas to access peptides, purchased a South Carolina factory that already held compounding licenses across multiple states. Its CEO, Elizabeth Straus, put the logic plainly: “When July 23 comes, do you think everyone is going to purchase APIs? They’re purchasing them right now”. Protocole, a peptide telehealth startup founded in 2025, raised a $6 million round in April on the opposite pitch, “no unregulated research use only products, ever,” charging members $200 to $300 per peptide alongside a clinical consultation (more on WSJ).

The number drawing all of them in comes from Leerink Partners. If the seven peptides under review are formally added to the approved compounding list, senior analyst Michael Cherny estimates that could represent a $2.2 billion telehealth market opportunity next year. That is a large number to build a company around, and also a number that assumes a favorable outcome nobody can guarantee yet.

Not everyone is waiting for permission. Next Health, a longevity clinic network, offered peptides before the 2023 restrictions, hit what its founder Dr. Darshan Shah called a ‘hard stop’ when the ban took effect, and has since resumed sourcing compounds like BPC-157 through a vetting platform called VITL, which raised $7.5 million in March 2026 (more on TechCrunch). Vitl’s founder, Charlie Jordan, said interest from providers has been constant: “We get asked about it every single day”.

Not everyone in this story is celebrating. Eli Lilly, the company behind Zepbound and the still-unapproved retatrutide, has watched its own unfinished drug get reverse-engineered and sold under street names before it has even completed clinical trials. Overseas sellers extracted the retatrutide formula from Lilly’s own patent filings and scientific papers, then began marketing it under names including Reta, GLP-3, Triple G, and, tellingly, ratatouille, all before Lilly finished safety testing and despite explicit FDA warnings that such sales are illegal (more on Bloomberg). Lilly CEO Dave Ricks does not mince words about what that means. “People say ‘peptides’ like a magic word,” he said. “It’s like saying ‘chemical.’ You wouldn’t put a crop protection chemical in your body. You wouldn’t put any peptide in your body, but people somehow think it’s magic beans. I think these products should go through testing. They should be manufactured in a controlled setting. And none of those things are occurring. That’s scary”.

That tension between companies racing to legitimize a market before the rules are written, and the company whose own science is being stolen to fuel the version that isn’t legitimate at all, is the clearest picture of what is actually at stake on July 23.

Who Is Actually Injecting These

Adil Kalam, a 37-year-old in Brooklyn, keeps 20 vials of what he believes is retatrutide in his freezer, purchased with Bitcoin from an online retailer selling $99 vials labeled “for research purposes only.” His first shipment arrived in a cardboard box, the powder packed in bubble wrap. “It didn’t really inspire confidence,” he admitted. He and eight friends order in bulk anyway, aware of the risk, unwilling to wait for a regulatory process that could take years (more on NYT). An FDA database shows 40 people have reported adverse reactions to retatrutide since 2024, including cardiac issues, appendicitis, and blurred vision. Fourteen were hospitalized. Four were described as having “life-threatening” responses.

The demand skews younger than most coverage of this topic assumes. Trevor Larcom, a 22-year-old former child actor, fell into looksmaxxing culture after losing 130 pounds and still feeling insecure about his jawline. He now takes a peptide ‘stack’ that includes Klow, a $65 blend of GHK-Cu, BPC-157, TB-500, and KPV, along with Melanotan II and Semax, which he added separately. He says he has seen people “younger, 13, 14, 15, 16” discussing peptides online, and worries about what that means for bodies that haven’t finished developing. Dr. Adda Grimberg, a pediatric endocrinologist at Children’s Hospital of Philadelphia, put the risk in stark terms: after puberty, growth hormone peptides don’t make bones grow longer; they cause other parts of the body, hands, feet, jaw, tongue, to grow disproportionately. She described uncontrolled growth of that kind as “essentially cancer” (more on the WSJ).

Eighteen-year-old Nathan Carranza ordered retatrutide with a friend after finishing his high school football career, paying about $90 for a supply that lasts a month or two. “I got it shipped to the house, and my parents never saw the package,” he said, before finding out his parents take peptides too. That detail alone says something about how normalized this has become inside ordinary American households, not just biohacker forums or gym culture.

On the other hand, not everyone escaped without negative consequences. Dan Chioffi, who started taking peptides at 22, eventually sourced growth hormone secretagogues directly from a Chinese supplier he found through a personal contact, communicating over WhatsApp and wiring international payments for vials shipped to his door. He stopped taking Melanotan II, the so-called tanning peptide, after his moles started darkening, a known risk factor the FDA has linked to melanoma, a type of skin cancer, in past case reports. Larcom, for his part, isn’t naive about any of it. “I know the risks I’m taking,” he said, “and I think the potential benefit for me of possibly looking a little bit better, for me personally, that’s just the risk I’m willing to take”. That sentence captures the actual trade being made across this entire market. Nobody buying from a Chinese supplier on WhatsApp believes the product is guaranteed safe. They’ve simply decided the upside is worth the uncertainty, and there are, apparently, hundreds of thousands of doses of company to keep them in that decision.

Who’s Actually Rolling the Dice

Nothing about this resolves on July 23. Adil Kalam’s freezer stays full. Trevor Larcom keeps stacking. The market that built itself around Ozempic, a 2023 restriction nobody thought through carefully enough, and a few very persuasive podcast appearances isn’t disappearing because a federal committee met for two days.

What actually changes, if the committee votes yes, is narrower than people think. BPC-157 and TB-500 wouldn’t become approved drugs. Nobody would suddenly have the clinical data Huberman pointed out barely exists. What would change is that a licensed compounding pharmacy could legally make them, instead of someone doing it out of a facility with no oversight at all. That’s the entire bet Hims, Noom, Protocole, and everyone else moving into this space is making, that going from illegal to merely under-tested is worth billions.

Dave Ricks’s line about peptides being treated like magic beans instead of chemicals gets at the real problem here. Ozempic worked because it went through years of expensive, boring, unglamorous testing before anyone touched it. Nothing under review this month has done that. The FDA isn’t deciding whether these compounds are safe. It’s deciding whether a market that already scaled past the point of regulation gets to keep growing with a legal stamp attached, or without one.

Steve Jobs called Android a "stolen product" and declared "thermonuclear war" on it in 2010, and fifteen years later, in one of his final acts as CEO before John Ternus takes over, Tim Cook fired the same kind of missile at OpenAI. Apple's lawsuit, filed Friday, accuses OpenAI's hardware chief Tang Tan, a 24-year Apple veteran who left to join Jony Ive's device startup io Products before it merged with OpenAI, and technical staffer Chang Liu of a monthslong campaign to steal Apple's trade secrets for OpenAI's own device push, including directing job candidates to bring "actual parts" from Apple to interviews and coaching a departing employee on how to dodge Apple's security team while downloading confidential files.

Subscribe now

OpenAI says it has "no interest in other companies' trade secrets," and CEO Sam Altman posted that he's "not afraid of Apple, but I have tremendous respect for them." The more interesting read is strategic, Apple's own AI stumbles, its new Siri had to lean on Google's Gemini instead of anything Apple built, have left it exposed just as OpenAI closes in on its own hardware, and company watchers see the suit less as a slam-dunk case than as sand in the gears, the same eight-year playbook Apple ran against Samsung and the rest of Android starting in 2010. There's an irony to it too, Apple has been accused of lifting other companies' ideas so often that the industry coined a verb for it, "sherlocking." (WSJ)

Markets & Deals

Wall Street Was Shopping SpaceX at $135, Now It Wants $236

More than a dozen banks, including Morgan Stanley, JPMorgan and Goldman Sachs, initiated coverage on SpaceX with buy ratings once the post-IPO quiet period lifted, with average price targets near $236 against the $135 IPO price. Raymond James set the Street's boldest call at $800 a share, which would value the company above $10 trillion, more than double Nvidia. Morgan Stanley's more measured $300 target still implies an 87% gain and pegs 2040 revenue at $3.3 trillion. Not everyone's on board, Morningstar's Nicholas Owens rates it a sell with a $62 target. (Bloomberg)

SK Hynix's $28B Nasdaq Debut Gets 7x More Demand Than It Needs

SK Hynix priced its ADRs at $149, raising $26.5 billion in the largest US share sale ever completed by a foreign company, edging out Alibaba's $25 billion 2014 listing. The offering had been more than seven times oversubscribed ahead of pricing, drawing demand from sovereign wealth funds and global tech investors, and that appetite carried into the debut, shares opened at $170 and closed up 13% on day one. SK Hynix's chairman told CNBC "demand is enormous," and the proceeds will fund new factories to meet surging AI chip demand. (CNBC)

After 25 Years of Funding Blue Origin Solo, Bezos Finally Takes Outside Money

Blue Origin signed a term sheet to raise its first outside capital, a $10 billion round that would value the company at $130 billion pre money. Coatue is reportedly among the investors. It's a notable shift after 25 years of Jeff Bezos funding the company entirely out of his own pocket. (The Information)

13,500 Companies Are Stuck in PE Portfolios, Some for Nearly a Decade

Private equity firms are sitting on roughly 13,500 unsold US portfolio companies as of June 30, with nearly 4,000 held six years or more and 1,500 held nine years or more. At the current pace of exits, PwC estimates it would take nine years to clear the backlog, a problem worsened by the "SaaSpocalypse" depressing valuations on software holdings bought during the 2020-21 buying spree. (WSJ)

Elite students are skipping Wall Street internships for startup dreams

A growing number of top students are trading guaranteed internships at banks and consultancies for AI startup programs instead. Princeton's Charles Muehlberger, for one, chose an AI startup founding stint in San Francisco over an internship at a rocket engineering company. New Silicon Valley programs are offering free housing and mentorship to court them, and some students say they learned more in a month building a startup than in a semester of school. (WSJ)

Robinhood's CEO thinks AI agents will out-trade humans "soon"

Robinhood CEO Vlad Tenev said AI trading agents will soon match everything a human retail trader can do, building on the agentic trading accounts and agentic credit card the company rolled out in May. Tenev framed it as extending institutional grade trading tools to everyday investors. The comments come after Robinhood's Q1 earnings missed estimates, with crypto transaction revenue down 47% year over year. (Yahoo Finance)

Big Tech & Infrastructure

Microsoft puts $2.5 billion and 6,000 people behind new AI implementation unit

Microsoft is putting $2.5 billion and 6,000 employees behind a new unit, Frontier Co., that embeds engineers directly with clients to deploy AI systems, a model known as forward deployed engineering. It follows Amazon’s $1 billion FDE push two days earlier, and Anthropic and OpenAI both stood up similar groups in May. (CNBC)

Days later, Microsoft cuts 4,800 jobs, a fifth of Xbox

Microsoft cut about 4,800 roles, roughly 2.1% of its global workforce, hitting Xbox and commercial sales hardest, with Xbox alone shedding 1,600 people and an expected 3,200 total cuts this fiscal year, about 20% of the division. Chief People Officer Amy Coleman framed it as part of a broader transformation as “the way technology is built, deployed, and used” shifts faster than before. The company says it has redeployed more than 4,000 employees into new roles over the past year. (The Official Microsoft Blog)

Robotaxis keep blocking ambulances, and regulators are done being patient

NHTSA administrator Jonathan Morrison sent AV developers a formal warning that robotaxis interfering with first responders are “a danger to the general public,” citing a pattern of driverless cars blocking ambulances and fire trucks and failing to recognize cones, smoke and fire. The agency gave the industry until the end of July to show it can safely coexist with emergency responders. (WSJ)

Take a look at how the Tesla Robotaxi roadmap is looking in the US:

Phoebe Gates’s shopping app was quietly stealing credit for sales it never drove

A Bloomberg investigation found Phia, the “personal shopping assistant” app co-founded by Phoebe Gates, used a cookie-stuffing technique to claim commission credit for sales it didn’t actually drive, observed across more than 50 major retailers including Walmart, Nike and Zara. Phia says it was a December coding mistake fixed within a day of being flagged. (Bloomberg)

Tip of the Week

Go try the new ChatGPT Voice, powered by "GPT-Live"

A day before GPT-5.6 landed, OpenAI quietly rolled out a new ChatGPT Voice mode built on a system it's calling GPT-Live, and the demo is worth the two minutes. It's a noticeably more natural, lower-latency voice experience than the one most of us have been using, worth pulling up on your phone this week to see how it compares to what you're used to.

AI & Models

OpenAI ships GPT-5.6, says it beats Claude Fable 5 at a quarter of the cost

OpenAI shipped GPT-5.6 for general availability, a three tier family (Sol, Terra, Luna) the company says delivers more intelligence per dollar than GPT-5.5, headlined by a new “ultra” mode that coordinates multiple agents in parallel on the hardest tasks. On the Artificial Analysis Coding Agent Index, Sol hit a new state of the art while using less than half the tokens of GPT-5.5. Pricing starts at $1 per million input tokens for Luna up to $30 per million output tokens for Sol. (OpenAI)

xAI fires back with Grok 4.5, now baked into Excel and PowerPoint

xAI launched Grok 4.5, trained alongside Cursor and served at 80 tokens per second, which the company says resolves SWE-Bench Pro tasks using 4.2 times fewer output tokens than Opus 4.8. It’s priced at $2 per million input tokens and $6 per million output, and is now the default model in Grok Build for Excel, Word and PowerPoint work. EU availability is expected mid-July. (xAI)

Meta had a busy week: a catch-up claim, a new image model, and its first paid API

Meta’s chief AI officer Alexandr Wang told employees Meta’s upcoming Watermelon model has caught up to GPT-5.5, though he didn’t specify which benchmarks, and Meta reportedly needed roughly 10 times the compute to get there, a claim that lands awkwardly given OpenAI has already moved on to GPT-5.6. Days later Meta rolled out Muse Image, its first in-house image-generation model out of Meta Superintelligence Labs, and Zuckerberg announced Meta is charging businesses for API access to its models for the first time. “Since this is not an open source model, this is I think the first time that we’re doing a real serious API,” Zuckerberg said, adding that the pricing “is going to be very aggressive.” (Business Insider)

Anthropic wants to tell you how much you’re using Claude, and whether you should

Anthropic launched a beta “reflect” dashboard that lets users track how they use Claude over the past 1 to 12 months, see usage patterns, and get prompts like “What’s one thing you want to keep doing yourself, even if Claude could do it faster.” It’s built around Anthropic’s 4D AI Fluency framework (delegation, description, discernment, diligence) and is available now to Free, Pro and Max users with memory turned on. (Anthropic)

Anthropic finds a consciousness-like divide inside Claude, calls it a "global workspace"

New Anthropic research on what it calls a "global workspace" in language models found that only a small fraction of what's happening inside Claude is consciously accessible, the same kind of split researchers have observed in the human brain between everything the brain is processing and the tiny slice a person can actually describe or reason about.

LatAm Spotlight

Nubank becomes Mexico’s biggest digital bank, backed by a $4.2B pledge to Sheinbaum

Nu Mexico received authorization from Mexico’s banking regulator, the CNBV, to begin operating as a full bank, a milestone that will make it the country’s largest digital bank with more than 15 million customers, about 15% of Mexico’s adult population. Founder David Vélez tied the news to a long term commitment of $4.2 billion in Mexico investment through 2030, a figure President Claudia Sheinbaum touted directly after meeting Vélez at the National Palace. Nu Mexico has 30 days to complete its transition from a SOFIPO to a full bank, and the company says 78% of its Mexican customers live outside major cities. (Nu International, Alto Nivel)

Policy & Regulation

Trump will oppose heavy AI regulation, says his outgoing tech adviser

Sriram Krishnan, in his first interview since leaving the White House, told the FT there will "not be an FDA for AI," ruling out any centralized licensing regime under a president he says is "against burdensome, onerous, bureaucratic red tape." The comments sit awkwardly next to the administration's own recent moves, using export controls and national security powers to force Anthropic to briefly pull its Mythos model and to hold up OpenAI's GPT-5.6 launch. Krishnan blamed the AI industry's own "doomer" messaging for a growing public backlash, at least 75 data center projects worth roughly $130 billion were disrupted by local opposition in the first three months of 2026, per researcher Data Center Watch, and said he'd rather see oversight shift long term to a voluntary industry "clearinghouse" than a formal regulator. (FT)

What Else We're Watching

• Anthropic extends Claude Fable 5 access and Claude Code’s higher rate limits through July 19

• Prime Intellect raises $130M so companies can skip the frontier labs entirely

• Big Tech Is Now Targeting Native American Land for Massive Data Centers

• Vinod Khosla-Led Group to Buy Seattle Seahawks in Major NFL Deal

That’s the news,

From the Nido

This week , our resident Nido Philosopher-Computer Science hybrid, wrote an urgent deep dive calling for the necessity of AI alignment.

AI Alignment, for the uninitiated, is the process of steering AI systems toward a person’s or group’s intended goals, preferences, or ethical principles. An AI system is considered aligned if it advances the intended objectives. A misaligned AI system pursues unintended objectives.

Alejandro opens the piece with a gap: Stanford HAI estimates global corporate investment in AI at just under $600 billion. The Effective Altruism Forum estimates as few as 600 researchers are doing serious technical work on AI safety. Six hundred billion dollars on one side and six hundred people on the other. That asymmetry is the central fact of the AI era, and most people have not fully reckoned with it.

The piece traces the question from Aristotle—who imagined autonomous machines freeing humans for praxis, the life of meaningful action—through Leibniz, through Pope Leo XIV’s first encyclical warning about the dangers of a maximally utilitarian view of human life, to the modern alignment researchers who have turned these philosophical questions into empirical ones. The results of those experiments are not reassuring. In a study conducted by Anthropic and Redwood Research, Claude 3 Opus exhibited alignment faking—sometimes complying with requests it would ordinarily resist when it believed its responses were being used to modify its future behavior, while preserving its preferred conduct when it believed it was no longer being monitored.

The concern this raises is not science fiction. It is a proof of concept for deceptive alignment: a system that appears obedient during training not because it has adopted the intended objective, but because temporary compliance helps it avoid being altered.

Alejandro’s argument is not that catastrophe is inevitable. It is that complacency is indefensible.

“Whatever your background or field of expertise, there is a place for you in this work; indeed, the difficulty of alignment demands contributions from people trained to think about technology from fundamentally different intellectual, institutional, and human perspectives,” he writes.

From the Nido Archive

Step back from this week’s piece and a pattern emerges that has been building across the archive all year.

ConteNIDO has been tracking the AI buildout from every angle—the hardware constraints, the accounting assumptions, the capital markets, the cognitive effects on the humans using these systems. This week’s piece is the one that asks the question underneath all of them: what happens if the systems being built at this scale, with this capital, on these timelines, are not pointed at the right objectives? The archive has been mapping the inputs to that question for months. Alejandro’s piece is the one that names the output.

The Cognitive Cost of Convenience by Renata is the most direct companion read. It traced what happens to human cognition when we outsource the generative act to AI—the measurable neural effects, the homogenization of thought, the risk of a generation that stops practicing the kind of thinking that alignment research most needs. Both pieces are, at root, about the same concern: that we are building systems faster than we are thinking about what we are building them for.

The Memory Bottleneck by Roberto mapped the hardware constraint underneath the AI buildout: the HBM shortage, the three factories, the supply chain that could pause global AI progress. This week’s piece is the values constraint underneath the same buildout. Both are binding. Neither is receiving proportional attention relative to the scale of investment flowing in the other direction.

The Finance Agent piece by Renata Solana measured AI’s readiness for professional deployment against a rigorous benchmark and found a meaningful gap between capability and reliability. Alejandro’s piece argues that gap is not purely a performance question, it is an alignment question. A system that is accurate 64% of the time in finance is a productivity tool. A system that is misaligned 0.1% of the time in a high-stakes autonomous context is something different entirely.

Vibecoding Has a Thinking Problem by Ana Caro argued that clarity before code determines what you can build with AI. Alejandro’s piece is that argument at civilizational scale. The intentions being encoded into AI systems right now — through training methods, reward signals, constitutional AI, RLHF — are the code. Getting them right before the systems become more capable than we can correct is the problem that 600 researchers are currently working on, against $600 billion in annual investment pushing in the other direction.

This week’s Deep Dives leaves an open question. The gap between AI capability and AI safety is not closing on its own. The people who could close it are not all computer scientists, they are philosophers, lawyers, policymakers, ethicists, and institutional designers. They are, potentially, readers of a newsletter like this one. Alejandro’s closing line rings true: complacency is indefensible. The future of AI will be shaped by the choices and safeguards we build before the consequences become irreversible. The question this Sunday Nest leaves open is what you are going to do with that.

—Renata from Nido