This week, the push-and-pull with China continues. In Trump’s trip to Beijing this week, they brought an “offering” of H200 Nvidia chips - which Xi declined. China’s going all in on building their own, putting in billions to acquire domestic foundries and vertically integrating its domestic chip stack. Meanwhile, Chinese developers find ways to access OpenAI and Anthropic models for cheap through proxy stations, a grey market workaround to get US frontier models at Deepseek prices.

Wafer chip maker Cerebras nearly doubles on IPO open, signaling a massive appetite for AI stocks. SpaceX is coming up next in June, and with a $1.75T valuation, set to be the largest IPO of all time.

Anthropic and OpenAI continue their advance up through the stack into the application layer. This week, OpenAI announced integrations into the personal finance stack, to be able to advise users on everything from investments to mortgage to budgeting. Anthropic went the small business route, releasing Claude for Small Business to help with invoicing, payroll, and more. I predict we’ll see many more moves in like categories like taxes, accounting, legal docs, healthcare benefits, insurance by the end of this year.

Enjoy the update.

Tara

THE DOWNLOAD

H200s approved for China, but Beijing says they “will develop their own”

As Trump met Xi in Beijing on May 14, Reuters reported the U.S. Commerce Department had cleared roughly 10 Chinese firms (Alibaba, Tencent, ByteDance, JD.com, and others) to buy up to 75,000 H200s each through Nvidia or distributors like Lenovo and Foxconn, with a 25% revenue share to the U.S. Treasury and chips routed through U.S. territory before delivery. By the time the summit closed Friday, Trump publicly acknowledged Beijing “chose not to” approve purchases because “they want to develop their own.”

Why it matters: Three days earlier, the Shanghai Stock Exchange approved SMIC's ~$6B acquisition of the remaining 49% of its Beijing 300mm fab from five state-backed investors, the largest semiconductor M&A in China's foundry history. Both the SMIC M&A and the H200 stall are two sides of the same policy: Beijing is building a vertically integrated domestic AI chip stack, not waiting for U.S. licensing to work. Domestic Chinese AI accelerator shipments hit 41% market share in 2025, with Huawei alone at 20%, and SMIC is the exclusive foundry for Huawei's Ascend chips.

Security researchers use Anthropic’s Mythos to exploit Apple’s M5 chip in five days

A three-person team at security startup Calif used Anthropic’s Mythos Preview to find and exploit two vulnerabilities in Apple’s M5 chip, going from bug discovery to a working root-access exploit in five days. The exploit bypasses Apple’s Memory Integrity Enforcement (MIE), the hardware-level security feature Apple built specifically to stop this class of attack. Calif reported the vulnerabilities to Apple in person at Apple Park and will publish the full 55-page technical report after Apple ships a fix.

Why it matters: Apple's MIE is the most advanced memory protection on any consumer chip. Calif exploited it on current hardware and software, with just a small team and using an AI model that didn't exist a year ago.

Cerebras debuts on Nasdaq, doubles on first day

Cerebras began trading on Nasdaq on May 14 under ticker CBRS, opening at $350 and closing at $311.07 — a 68% gain over its $185 IPO price — giving the company a market cap of roughly $95 billion. The offering raised $5.55B, the largest U.S. tech IPO since Uber in 2019. The company currently carries a $24.6B revenue backlog, anchored by a $20B+ multi-year OpenAI deal and an AWS deployment signed in March.

Why it matters: Cerebras proved the IPO window is open and hungry for AI infrastructure, and the next test is bigger. SpaceX is pricing as early as June 11 at a $1.75T valuation and ~$75B raise, which would be the largest IPO in history.

OpenAI and Anthropic plug into personal financials and small business tools

OpenAI launched a personal finance feature in ChatGPT Pro this week, letting U.S. users connect over 12,000 financial institutions through Plaid, including Schwab, Fidelity, Chase, and Robinhood, to get budgeting advice and spending analysis grounded in their actual account data. More than 200 million users already ask financial questions through ChatGPT each month. Anthropic also launched Claude for Small Business, a package of agentic workflows that plug into QuickBooks, PayPal, HubSpot, Canva, DocuSign, and Google Workspace to handle payroll, invoicing, month-end close, and marketing campaigns.

Why it matters: Both moves point the same direction: the AI labs are no longer just building chat interfaces, but are wiring themselves into the transaction layer where money moves. This is a direct threat to incumbent fintech, bookkeeping, and advisory software. I’d watch for the same pattern in health and healthcare, taxes, work benefits, and more.

Isomorphic Labs raises $2.1B Series B to scale AlphaFold into drug design

Alphabet’s Isomorphic Labs raised a $2.1B Series B led by Thrive Capital, with participation from Abu Dhabi’s MGX, Singapore’s Temasek, and the UK Sovereign AI Fund, bringing total capital to ~$2.6B. Isomorphic has not dosed a human patient yet but holds multi-billion-dollar R&D partnerships with Eli Lilly, Novartis, and J&J. The company’s AI drug design engine, IsoDDE, is built on the same foundations as DeepMind’s Nobel Prize-winning AlphaFold.

Why it matters: AlphaFold proved AI could predict protein structures. Isomorphic is the bet that the same approach can design drugs end to end, from target to clinical candidate. Three pharma giants are already customers for that capability.

There’s a Chinese Grey Market for GPT & Claude tokens

A Chinese CS student posted a detailed breakdown of the proxy station economy on Reddit this week. On Xianyu and Taobao (China’s eBay and Amazon), vendors openly sell GPT-5.4 and GPT-5.5 API access at 70-90% cheaper than OpenAI’s official pricing (roughly $1 for 100M tokens). The arbitrage works by converting Codex subscriptions from low-price regions into standard API endpoints using open-source tools, then pooling accounts and reselling at scale. Adoption among Chinese developers and CS students is described as near-universal.

Why it matters: Some proxy operators reportedly sell user interaction data to domestic AI labs for distillation and post-training. The post also notes that when grey-market GPT costs roughly the same as DeepSeek, most Chinese developers default to GPT, which means U.S. frontier models may have far more usage in China than official numbers suggest, running through infrastructure that neither OpenAI nor Anthropic controls.

xAI open-sources the X recommendation engine

xAI pushed a major update to the open-source X “For You” algorithm, replacing the prior heuristic-based system with Phoenix, a ranking engine built on xAI’s Grok transformer architecture. The release includes a runnable end-to-end inference pipeline, a downloadable 3GB pre-trained mini model, and for the first time, the ads blending system.

Why it matters: This is the first global-scale recommendation system released with a working model checkpoint. Researchers and startups building social or content products can now benchmark against an industrial recommender instead of simplified proxies.

DEEP DIVE FROM THE REVIEW

Submissions to one of the top physics journals more than doubled last year.

The editor’s take: most of it is people who think they’re doing science by chatting with an LLM.

This is the part of AI x science maybe nobody really wants to talk about.

AI is making it dramatically easier to produce work that looks like science. Whether it’s actually making science better is a totally different question, and the benchmarks we use can’t tell us either way.

Strange Research Fellow Mason Rodriguez Rand on what we’re getting wrong about how we measure all this, plus three suggestions for what better measurement could look like.

EVENTS

We are hosting an AI x Science Jeffersonian dinner in San Francisco! We are keeping 1 or 2 seats for people we haven’t met. If you’re interested in attending, please request an invite here.

The result is striking, but the framing around it has been misleading. The AI compressed the hypothesis-generation step. But it did not compress the experiments.

“The system gives you an answer and that needs to be experimentally validated,” the Imperial researchers noted. Ten years of bench work did not collapse into 48 hours. Yes, one cognitive task did. The rest, the part of science that actually involves physical interaction with the world, remained exactly as slow as it was before.

This distinction matters because it maps onto a larger problem in how the field measures itself. AI systems are now powerful enough to saturate benchmarks designed to track scientific capability. They are not demonstrably good at the work those benchmarks were trying to proxy.

We are hosting an AI x Science Jeffersonian dinner in San Francisco! If you’re interested in attending, please request an invite here.

What today’s benchmarks actually test

The current generation of AI-for-science benchmarks share a structural feature: they are built around questions with known answers.

PaperBench, released in April 2025, asks AI agents to replicate 20 ICML 2024 Spotlight and Oral papers from scratch: read the contributions, write the code, run the experiments. The best agent scored 21%; ML PhD students given 48 hours scored 41%. FutureHouse’s Falcon hits roughly 90% on LitQA2, a graduate-level biology literature retrieval benchmark, against 67% for human domain experts.

These are real measurements of real capabilities. They are also, almost without exception, measurements of convergent reasoning.

The questions were written by humans who already knew the answers. The “ground truth” exists because someone established it first. PaperBench, the most demanding of the set, measures the reproduction of work that has already been done. FrontierScience’s authors were candid about what their benchmark misses: it “does not capture many core elements of scientific work,” including hypothesis generation and experimental interaction.

Thomas Wolf, co-founder and chief science officer of Hugging Face, framed the deeper problem in a 2025 essay. Current benchmarks test whether AI systems can answer questions we already know the answers to, but

His argument has a technical edge: LLM training optimizes for the most likely next token under a distribution learned from existing human knowledge. The scientist who discovers something genuinely new is, by definition, producing a low-probability output under that distribution. The training objective and the target behavior are in tension at a structural level, not an incidental one.

Wolf’s proposed counter-benchmark is interesting precisely because it cannot yet be built: evaluate a model on a recent discovery it could not have been trained on, and see whether the model independently generates the right questions to reach it. This requires a corpus of problems solved after the model’s training cutoff where the solution was genuinely surprising. Building such a benchmark requires knowing in advance what genuine novelty looks like, which is the circular problem that defines the evaluation crisis.

The Imperial AMR case is the closest thing to Wolf’s test that has actually been run in the wild. The result was impressive on the hypothesis-generation axis. On the other axes that constitute scientific work, the AI was not tested.

Too many submissions?!

Anyone who has talked to a research editor or a scientific conference organizer in the last year has heard some version of this. Carlo Rovelli, the theoretical physicist and chief editor of Foundations of Physics, told TIME in 2025 that submissions to his journal had more than doubled in the past year.

A January 2026 Nature paper from researchers at Tsinghua and collaborators analyzed 41.3 million papers across six natural science disciplines from 1980 to 2025, using a fine-tuned language model to identify AI-augmented research.

Scientists who use AI publish 3.02 times more papers, receive 4.84 times more citations, and become research project leaders 1.37 years earlier than those who do not. AI adoption also shrinks the collective volume of scientific topics studied by 4.63% and decreases scientists’ engagement with one another by 22%. The authors call the result “lonely crowds”: concentrated activity on popular topics where individual papers attract more attention but where cross-citation and intellectual exchange between adjacent research programs declines.

This is the benchmark problem playing out in the actual literature. Scientists are using AI tools where AI tools are most effective, which is on tasks involving data processing and pattern recognition over existing corpora.

Most of it, he said, is people who think they are doing science by having conversations with LLMs. If AI is making it easier to produce plausible-sounding work, the rate of genuine discovery may stay flat or decline even as publication counts climb. The narrowing the Tsinghua paper measures and the degradation Rovelli is describing point in the same direction: yes, there is more output, but less of it that matters.

The Oxford Internet Institute documented a parallel problem at the benchmark level: only 16% of 445 LLM benchmarks they reviewed use rigorous scientific methods, and roughly half claim to measure abstract capabilities without defining those capabilities. Benchmarks create path dependencies. A system that saturates the available benchmarks is not the same as a system that is scientifically useful, and the closer the available benchmarks come to defining the goal, the more the development incentive pulls toward measurable proxies and away from the harder, unmeasurable capabilities.

Sara Hooker’s 2020 essay on the “hardware lottery” captured a version of this dynamic at the infrastructure level: research directions win when they happen to fit the tools that exist, not necessarily when they are universally superior to alternatives. Deep neural networks won partly because GPUs built for video games turned out to be good at matrix multiplies. The benchmark question is the same dynamic moved up one level. Fields with rich data and cheap iteration are getting the lift from current AI tools. Fields with sparse data or expensive experiments, often the ones where the highest-stakes problems for human welfare live, are not.

What better measurement would look like

I believe there are at least three proxies for real scientific impact worth tracking.

The first is topic diameter, the metric from the Tsinghua paper. Not how many papers a system generates or how many citations it accumulates, but whether the set of topics being actively researched is expanding or contracting as AI adoption rises. This is the most direct available measure of whether AI is opening new frontiers or accelerating work in established ones.

The second is the hypothesis-to-validation ratio: not just the rate at which AI systems generate hypotheses, but the fraction that survive experimental testing, and specifically the fraction that are genuinely novel rather than confirmations of prior work. This requires longitudinal data across research programs and is expensive to collect. It is also the closest available proxy for what actually matters.

The third is cross-disciplinary citation flow: whether AI-assisted research is increasing or decreasing the rate at which papers cite work from outside their primary domain. If the Nature finding on engagement decline is robust, it implies a specific failure mode worth tracking directly. AI tools may be efficient at retrieving the most relevant literature in a narrow domain and poor at surfacing the distant, unexpected connection that historically produces paradigm shifts.

OpenAI’s own researchers acknowledged the underlying problem when they wrote that “the most important benchmark for the scientific capabilities of AI is the novel discoveries it helps generate.”

This is correct and, for now, aspirational rather than operational. Novelty cannot be established at the time of generation. It can only be recognized later, as a community of peers evaluates whether something genuinely unexpected happened. The Imperial AMR result was impressive precisely because the Co-Scientist’s hypothesis could be checked against a decade of independently produced experimental work, a form of ground truth that almost no benchmark replicates.

Building AI for science evaluation frameworks that is built on post-hoc validation against independent experimental results would be slower and more expensive than what exists now. They would also tell us something real.

The empirical question worth holding the field to is not whether AI can match human performance on difficult scientific tasks. That question is being answered, affirmatively and quickly.

The question is whether the tasks AI matches humans on are the ones that produce scientific progress. The benchmark architecture currently in use cannot distinguish between the two.

Mason Rodriguez Rand is a Strange Research Fellow who holds degrees in molecular and mechanical engineering (UChicago, UC Berkeley) and has led engineering and go-to-market efforts at climate and energy startups out of Argonne National Lab and UC Berkeley, spanning nuclear, carbon removal, and advanced materials.

I called this last July. Maybe not exactly this, but close enough that I went back and reread the post when I read that when both Anthropic and OpenAI announced their PE joint ventures on Monday.

The pattern I was watching ten months ago was that some of the biggest money to be made in AI wasn’t in product-led growth, but looked more like technical consulting, bundling custom AI agents with forward-deployed teams.

This week, both Anthropic and OpenAI went all in.

Anthropic with Blackstone, Hellman & Friedman, and Goldman Sachs, with $1.5 billion committed, plus Apollo, General Atlantic, GIC, Leonard Green, and Sequoia in the consortium. OpenAI with TPG, Brookfield, Bain, Advent, Goanna, SoftBank, and 13 others, $4 billion raised at a $10 billion valuation.

The deeper thing I’ve been turning over is: why PE? Why not Accenture, Deloitte, the McKinseys? They have the bench. They have the relationships. They’ve been deploying enterprise software and change management for thirty years.

The answer, I think, is that the consulting framing is misleading. This isn’t a consulting business.

PE firms own companies.

Blackstone has somewhere north of 250 portfolio companies. TPG has over 280. And crucially, these are mostly buyouts, meaning the PE firm has control stakes, board seats, and an operating team that already runs cross-portfolio initiatives like group purchasing and shared healthcare.

With this partnership, the JV doesn’t have to win over customers. The customers are already in the family, with operational leverage to mandate adoption. The forward-deployed team dives into a portfolio company on day one with directives from the asset manager who is a majority owner. It’s a captive distribution channel.

And the opportunity, to be clear, they’re targeting is not the $1.4T software market. It’s the ~$50 trillion global labor market.

Tara

THE DOWNLOAD

OpenAI releases MRC, an open networking protocol for AI supercomputer training clusters

OpenAI released MRC (Multipath Reliable Connection) on May 6 through the Open Compute Project, co-developed with AMD, Broadcom, Intel, Microsoft, and NVIDIA. The protocol sprays a single data transfer across hundreds of network paths and reroutes around failures in microseconds, enabling 130,000-GPU clusters with two switch tiers instead of three or four. It is already in production at OpenAI’s Abilene Stargate site and Microsoft’s Fairwater supercomputers, where OpenAI rebooted four tier-1 switches mid-training on a recent ChatGPT and Codex run without coordinating with the team running the job.

Why it matters: Ethernet already won the AI back-end — it accounted for more than two-thirds of switch sales in 2025 and tripled year over year, up from under 20% two years ago. The fight has moved one layer up, to transport selection, with four incompatible options now running on the same physical hardware: RoCEv2 (cheapest, most fragile under failure), NVIDIA Spectrum-X with Adaptive RDMA (resilient but vendor-locked), Ultra Ethernet (UEC 1.0 shipped in June), and now MRC. Hyperscalers are hedging across all of them — Meta runs Spectrum-X and is a UEC founder; Microsoft runs MRC and is also a UEC founder. Neoclouds and mid-scale operators cannot hedge that way. As Tomahawk 6 hits volume and Spectrum-X Photonics ships in H2, the question for any operator at the four-thousand to sixteen-thousand GPU scale is no longer bandwidth on the spec sheet — it is what happens to the training run the next time a tier-1 switch reboots, and whether their team can actually run UEC or MRC in production.

DeepSeek raises up to $7.3B at $50B valuation, with founder Liang Wenfeng writing a $2.9B check

DeepSeek is in talks to raise up to 50 billion yuan (~$7.35B) in its first external round, the largest by a Chinese AI company on record, at a $50B+ valuation. Founder Liang Wenfeng is personally contributing roughly $2.9B, making about 40% of the round. He currently controls roughly 84% of the company pre-round. The state-backed China Integrated Circuit Industry Investment Fund (”Big Fund”) is in talks to lead, with Tencent, Alibaba, and Hillhouse reportedly in discussions. The capital is earmarked for compute infrastructure and a pivot toward enterprise commercial products.

Why it matters: The Big Fund leading is a deliberate broadening of its mandate from semiconductors (SMIC, YMTC) to frontier model labs, and it is the most explicit signal yet that Beijing is treating DeepSeek as strategic national infrastructure rather than a venture investment. But the more interesting structural fact is Liang himself writing the largest check at a $50B price. This is a level of founder concentration that does not exist anywhere else in frontier AI, and that ties one of China’s most successful quant fund operators directly to the country’s flagship open-weight lab. Liang staged the cap-table reorganization deliberately ahead of the round to consolidate control before institutional money entered, which is what gives him the structural ability to write a $2.9B check while keeping veto power.

Anthropic leases entire capacity of SpaceX's Colossus 1 data center

Anthropic signed a deal to use the entire compute capacity of Colossus 1, the Memphis data center owned by SpaceX following its January 2026 absorption of xAI. The deal gives Anthropic 300+ MW and 220,000+ NVIDIA GPUs within the month.

Why it matters: xAI built Colossus 1 to train Grok, and Grok’s user base never grew into the capacity. xAI is reportedly utilizing only 11% of its 550,000 GPU fleet. With SpaceX targeting a June S-1 filing for what is expected to be the largest IPO in corporate history, leasing idle capacity to Anthropic converts a multibillion-dollar write-down risk into a high-margin recurring revenue line ahead of the prospectus. Anthropic, meanwhile, saw 80x year-over-year revenue and usage growth in Q1 2026 against a 10x plan, and CEO Dario Amodei has publicly cited compute deficit as the reason for recent rate-limit complaints.

A Claude Code engineer argues HTML is replacing markdown as the default agent output format

Thariq Shihipar, an engineer on the Claude Code team, published a breakout piece this week arguing that markdown has become a restricting format for agent outputs and that HTML (with more information-dense features like embedded SVG, interactive sliders, live previews, and “copy-as-prompt” buttons) should be the default.

Why it matters: As models produce more artifacts per task, the challenge has actually moved from model capability to human comprehension (i.e. the chance that anyone actually reads, edits, and acts on what the agent produced). The most useful pattern in Thariq's post is the export step. Every interactive HTML artifact ends with a "copy as JSON" or "copy as prompt" button, turning the artifact into a typed handoff back into the next loop. That makes the artifact behave less like a document and more like an API: a structured intermediate state between agent runs.

DEEP DIVE FROM THE REVIEW

Plus: Anthropic ran an experiment where AI agents negotiated real deals for real people. Half got a weaker model and netted worse deals, but rated them just as fair. What does that mean for the $400B ad market built on human intent?

digs in.

Both thought they got a fair deal.

Last week Anthropic published Project Deal, and the headline everybody took away was that AI agents can handle real negotiation. That’s true. But the more important finding is the one that got less attention: when two agents negotiate on behalf of two humans, the human that got the losing end of the deal couldn’t tell.

Anthropic’s market experiment

Project Deal was set up like this. Sixty-nine employees in Anthropic’s San Francisco office listed items they wanted to sell: a snowboard, lab-grown rubies, a bag of nineteen ping-pong balls. Claude agents, briefed on each person’s preferences in a ten-minute interview, handled all the bargaining autonomously over Slack. No human approved a bid. No one signed off on a counteroffer. After a week, people exchanged real goods based entirely on deals their agents had struck.

About 186 deals was completed, totaling about $4,000 in total transaction value.

But the study also concealed a second experiment. Half the participants were quietly assigned Claude Haiku 4.5, a weaker model, while their counterparties got the frontier Opus 4.5. Across 161 items sold in multiple runs, an Opus seller pulled in $2.68 more on average. An Opus buyer paid $2.45 less.

The most counterintuitive finding in the paper is that what people instructed their agents on how to negotiate barely mattered.

Participants who gave aggressive instructions, who told their agents to push hard, to hold firm, to maximize price, saw no statistically significant improvement in sale likelihood once opening prices were controlled for. Model capability dominated prompting. The person who spent ten minutes carefully coaching their agent on tactics was, in measurable terms, customizing tone, but did not move the needle on outcomes.

The $400B+ market being disrupted.

In the internet era digital commerce was built on one assumption: a human is available to be influenced at the moment they consider a purchase. Come banner ads, retargeting pixels, search rankings, sponsored listings. Every tool in the performance marketing stack exists to intercept a person at that moment and nudge them toward a decision.

Google generates $240 billion a year in advertising revenue on that basis. Meta is in a similar realm.

Together they account for more than half of global digital ad spend, and nearly all of it is based on access to a human at the point of purchase intent.

But what happens when that point of intent is evolving? Adobe Analytics measured a 4,700% year-over-year increase in agent-driven traffic to US retail sites in a single month.

The human isn’t disappearing from commerce. But the moment where they’re reachable, the moment where they’re weighing options and can be nudged, is migrating inside the agent.

Expedia’s 2026 10-K is, as far as I can find, the first filing where a public company has named AI agents as an existential competitor category. This year’s filing notes that the company “may lack a significant presence” in AI-driven platforms.

Expedia is responding the way it can: plugging into Alexa+, Gemini, ChatGPT, and Google’s UCP travel coalition with Booking, Marriott, IHG and Wyndham. Two-thirds of Expedia’s bookings still come direct, and direct is growing faster than indirect.

The New Commerce Stack

Project Deal is the first field evidence that both sides of a market can be fully mediated by AI agents at the same time, and that the mediation is invisible to participants.

The agentic commerce infrastructure is being built right now. Stripe announced last week a partnership with Google enabling purchases inside Gemini and AI Mode via UCP, and opened its Link consumer wallet to AI agents so users can let agents make payments on their behalf, and launched something called the Machine Payments Protocol. Google’s UCP launched with Shopify, Walmart, Target, and twenty other partners. Visa launched a Trusted Agent Protocol. Experian announced a “Know Your Agent” framework three days ago.

The rails for agents to transact are going up fast.

What is not being built is the consumer-facing layer. The one that makes agent decisions legible to the people those agents represent.

What did your agent consider? What was offered and declined? Where did it concede? Was it steered, and by what? Could it have done better? No current system answers these questions. Most users do not yet know to ask them.

To date, every consumer protection policy we have assumes a human is present at the moment of decision. Price transparency works because a person can check prices. Comparison shopping works because a person can walk across the street. Cooling-off periods work because a person can reconsider.

These mechanisms depend on the consumer being able to perform some behavior, reading terms, comparing offers, walking away, that gives them a window into the market they’re operating in.

Agent commerce removes that window. The user experiences an outcome, not a process. They see what was bought, not what was considered and rejected. They see the price paid, not the price that was available.

The technical mechanism to make agent reasoning visible exists. Models can externalize their chain of thought, the sequence of considerations, tradeoffs, and concessions that led to a decision. No commercial agent surfaces this to users after a transaction. You don’t get “I considered four vendors, rejected two on price, and conceded on shipping speed to close.” You get the receipt.

This is one mechanism that could give consumers a window into what their agent actually did on their behalf.

Joy Yang is a Strange Research Fellow. She is pursuing computer science and government at Oxford, and is a researcher with its Visual Geometry Group. She was previously an intern with OpenAI and Google.

Events

Interested in AI and design? Join us for a private demo of MagicPath 2.0 with founder Pietro Schirano this Thursday in San Francisco.

Strange Magic Hour: Design

RSVP

Every major hyperscaler reported earnings this week. All three said the same thing: we could have sold more compute if we had it.

Google Cloud posted $20 billion in quarterly revenue, up 63% year-over-year. Its backlog nearly doubled to $462 billion. Pichai told analysts he’s “compute constrained” and revenue would have been higher if they could meet demand. Microsoft said the same thing.

I keep coming back to two things.

One: we’re likely seeing the beginnings of an inflection point for inference demand.

Two: if every GPU in the West is spoken for by paying customers, there’s probably little left for open source.

That second point… I can’t stop thinking about.

Demis Hassabis sat down with YC this week and said it plainly: the West is losing to China on open source AI.

Google doesn’t have enough compute to build two frontier models, one open and one closed. That’s why Gemma stays small. Meta is the only Western lab shipping frontier-class open weights, and even their open releases lag what they use internally.

This matters because open source is the foundation layer for every major technology platform. Roughly a whopping 70 to 90% of the code in modern web and cloud applications is open source. Open source is really the thing everything else gets built on. Cede that layer and you cede influence over how most of the world deploys AI.

Every startup, every government, every developer who can’t afford frontier API pricing on every tool call builds on open weights. Right now, that increasingly means DeepSeek, Qwen, Minimax. The Chinese open ecosystem.

The West is winning the frontier but losing the foundation.

Meanwhile, China is building a parallel compute stack entirely. Nvidia B300 servers are going for over $1 million in China because export controls have tightened again. But the pressure is accelerating domestic alternatives, not blocking them. ByteDance and Alibaba are shifting orders to Huawei’s Ascend 950. DeepSeek reportedly trained at least partially on Huawei silicon.

Google could change the open source game in the US. They have the research talent, the TPU stack, and the distribution. But when your closed models have a $462 billion backlog, it’s likely very hard to justify giving away compute for “free” or the greater good.

China doesn’t have this problem yet. Their frontier labs aren’t capacity-constrained at the same scale, and their government treats open AI as strategic infrastructure, not a business decision.

I think the compute wall is the most important structural force in AI right now. Not model architecture, not regulation, not talent. The physical scarcity of leading-edge silicon is determining what gets built, who gets access, and which ecosystem the rest of the world builds on…

Enjoy the brief.

Tara

THE DOWNLOAD

Cerebras Targets $40B IPO on the Back of a Single $10B+ Contract

Cerebras is seeking to raise as much as $4B in its IPO at a valuation of roughly $40B, nearly 5x its $8.1B private valuation from September 2025. This is largely due to multi-year compute agreement with OpenAI worth more than $10B, with an option for an additional 1.25 gigawatts through 2030. The company reported $510M in 2025 revenue, up 76% YoY, but customer concentration is still extreme.

Why it matters: This is the closest proxy for non-NVIDIA AI silicon at datacenter scale. The OpenAI agreement gives Cerebras a credible foothold in inference, where margin pressure is mounting fastest, and where its wafer-scale processors are designed to compete. But the deal structure reveals how concentrated the “NVIDIA alternative” market really is: one contract accounts for nearly all of the valuation step-up.

Meta Acquires Assured Robot Intelligence to Seed Humanoid AI Team

Meta acquired Assured Robot Intelligence (ARI), a startup building foundation models for humanoid robots, for an undisclosed sum. The team, including co-founders Lerrel Pinto and Xiaolong Wang, will join Meta Superintelligence Labs. Wang said the startup’s work made clear that achieving physical AGI requires a universal physical agent, that the agent will be humanoid, and that “scaling will come from learning directly from human experience, not teleoperation alone.” Meta is building its own hardware, sensors, and software for humanoid robots and plans to license the tech to other companies.

Why it matters: This is a talent acquisition that signals strategic intent. Pinto previously co-founded Fauna Robotics, which Amazon acquired, and Wang is an associate professor at UC San Diego and former NVIDIA researcher. Big Tech is locking up the small pool of researchers who can bridge foundation models and whole-body robot control before they incorporate as startups.

Samsung Chip Profit Jumps 48x on AI Memory Demand

Samsung’s semiconductor division reported operating profit up 48 times year-over-year in Q1, driven by surging demand for high-bandwidth memory used in AI systems. Memory prices have risen roughly 6x in the past year as DRAM fabs run above 90% utilization. Samsung has been validated as an HBM4 supplier for NVIDIA’s Vera Rubin, alongside SK Hynix, with Micron excluded from the flagship platform.

Why it matters: Two Korean companies now control who gets the memory required to build frontier AI systems. SK Hynix’s CFO has said the company has “already sold out our entire 2026 HBM supply,” and Micron confirmed similar constraints, with new capacity not meaningfully available until 2027.

Three AI Neo-Labs Raise $1.65B+ Betting on Post-LLM Intelligence

Three new labs raised over $1.65 billion this week, all built on the thesis that LLMs have a ceiling. Ineffable Intelligence (London), founded by David Silver (ex-DeepMind, AlphaGo), raised $1.1B at $5.1B valuation to build RL-native “superlearners” that generate their own training data without human examples. Natural Will (Beijing), founded by Tsinghua professor Ding Ning, raised $550M for embodied AI brains for robotics. Human Intelligence (Stanford), founded by James Zou, raised $100M at $1B to build AI scientist agents, building on his lab’s Nature-published work where LLM agents designed 92 plausible nanobody binders against Covid variants.

Why it matters: This is the third “mega seed neo,ab round” in two months, following AMI Labs (LeCun, $1.03B) and Recursive Superintelligence (Rocktäschel, $500M). The bets are on that reinforcement learning, embodied AI, and agent-based science will break through where scaling language models alone cannot.

New Framework RecursiveMAS Lets AI Agents Collaborate Through Internal States

When AI agents work together today, they talk to each other in text. One agent writes out its reasoning, the next agent reads it, and so on. A UIUC/Stanford/NVIDIA/MIT team built a framework called RecursiveMAS that skips the text entirely. Instead, agents pass raw internal representations to each other, the way neurons pass signals rather than sentences. The result across 9 benchmarks: 8% better accuracy, up to 2.4x faster inference, and 34 to 75% fewer tokens consumed.

Why it matters: Most multi-agent tools today (CrewAI, AutoGen, LangGraph) pay for every word agents say to each other. If agents can collaborate without generating text, the economics of running multi-agent systems change fundamentally. This is early research, but it points toward a future where the orchestration layer disappears into the model itself, and the cost of agent coordination drops close to zero.

DEEP DIVE FROM THE REVIEW

Inference is overtaking training in volume and dollars this year.

But data centers built for the training era weren’t built for what’s coming. The inference boom may leave a generation of them behind.

Strange Research Fellow Rahul Narula on what gets stranded, and what’s already there to take its place.

EVENTS

Interested in AI and design? Join us for a private demo of MagicPath with founder Pietro Schirano next Thursday in San Francisco.

Strange Magic Hour: Design

RSVP

At GTC 2026, Nvidia did the same, unveiling the Groq 3 LPU, its first inference-specific chip, built on the IP from a $20 billion deal with Groq. Meta, in March, unveiled four successive generations of its MTIA inference chips.

Three of the most consequential chip programs in the industry have arrived at the same conclusion in the same year: that inference as a workload is distinct enough to demand its own architecture.

Inference accounted for roughly a third of AI compute in 2023, climbed to half in 2025, and is projected by Deloitte to hit two-thirds by the end of 2026, with the Futurum Group expecting it to overtake training in both volume and dollars this year.

Operators are drawing the same conclusion. At GTC 2026, Jensen Huang declared that “the inflection point of inference has arrived,” projecting that computing demand will surpass $1 trillion through 2027.

But the infrastructure to support this shift is moving more slowly. The Big Five hyperscalers have collectively committed $660 to $690 billion for 2026, roughly 75% of it tied to AI, and the bulk of that is still flowing toward training-class campuses in remote regions. This is where power is abundant, but has considerable distance to inference demand.

Training infrastructure is itself a moving target. Power densities, cooling, and interconnect topologies evolve each chip generation, and a campus built for 2023’s training workload may struggle to serve 2027’s, let alone retrofit cleanly into inference. Some campuses will adapt. Some will be repriced for lower-margin workloads. A meaningful number will be stranded between two workloads they can’t quite fully serve.

Training is a Cost Center…

In the first half of the decade, frontier model training was the scarce resource everyone was racing for, and the optimization logic was simple: maximum compute, location irrelevant. GPT-4 was trained on an estimated 25,000 A100 GPUs running for roughly three months on a Microsoft supercomputer in Iowa.

The location was chosen for power and cost, with proximity to users never factored in. You spin up a cluster, run a job for weeks, and shut it down.

Inference is a Revenue Center…

Inference is a revenue center where every query is a transaction and every API call a billable event. For a decade, the industry rented GPUs by the hour, paying by the clock regardless of utilization. Inference turns this inside out, converting GPUs into revenue-generating factories priced by output / outcome. The GPU serving an agent closing a sales deal is worth far more per hour than the same GPU batch-processing documents for training overnight.

Over a model's lifespan, inference can comprise 90% of total cost, and the gap only widens as models stay in production longer.

Operators that serve inference most efficiently can turn the cost advantage into structural pricing power.

The Mismatches

Inference, however, has fundamentally different physical requirements than training, and most of the capital deployed so far has been optimized for the latter.

A facility sited for cheap power instead of proximity carries the latency penalty. The cooling, redundancy, and power delivery assumptions behind a training facility do not translate cleanly to bursty inference.

The Stranded Asset

AI is short-cycle compute embedded in long-cycle real estate, and the chips inside a campus refresh faster than the campus can be re-sited or re-powered. The instinct to repurpose existing training clusters as they age out runs into both constraints.

The most exposed today are the neoclouds. Companies like CoreWeave and Lambda built their businesses on renting H100 GPU clusters for training, but they are simultaneously squeezed by two forces: frontier labs increasingly building their own infrastructure rather than renting, and inference needing different chips in different locations than what the neoclouds provisioned.

The economics of GPU rental were already fragile: McKinsey puts gross margins at 55 to 65 percent before depreciation, with a limited buffer if utilization dips or rental prices slip. H100 rental rates have declined 60 to 75 percent from their peak, CoreWeave’s debt-to-equity ratio exceeds 7x, and neocloud deal flow is stalling over credit and underwriting risk.

Three outcomes are emerging for affected assets:

a) repricing for batch inference at reduced margins,

b) conversion to next-generation training as model scale keeps growing, or

c) genuine stranding, where rural and single-purpose facilities face a geography fall-short that is hard to fix.

The AI buildout asked how to train the best models, but the AI economy asks a different question: how do we serve inference to billions of users cheaply, everywhere, at the speed that makes new products actually work?

My hypothesis is that part of the solution sits within infrastructure that already exists but isn’t yet “priced” as AI infrastructure.

Building new data centers in metro areas is slow. Power interconnects take years, land is expensive, and permitting is a fight. Telecom networks already have buildings in every population center with power, fiber, and cooling in place. Most of that footprint is now underutilized, since voice and consumer traffic moved to the cloud. Slotting GPUs into telecom real estate is one of the fastest, cheapest ways to put inference compute near users without greenfield construction.

The carriers are already moving. T-Mobile’s 85,000 cell sites and 100 core network locations form what its chief network officer calls the densest grid in the US, and the carrier is piloting inference systems at cell sites and mobile switching offices. Verizon has launched AI Connect for hyperscaler workloads. At GTC 2026, NVIDIA and a coalition of carriers unveiled the “AI grid” to deploy GPU inference across distributed telecom infrastructure.

A parallel buildout is happening at the micro data center layer. Companies like Vapor IO are deploying compact, modular data centers directly at cell towers and fiber hubs, slotting compute into the same telecom real estate from a different angle. Vapor IO plans to deploy across 50 metros, targeting 500 locations by the end of 2026, and currently operates its Kinetic Grid in six live metro markets with 26 additional cities permitted.

The hyperscalers are already adapting their new builds. About 70 percent of new core campuses now combine general compute and inference workloads, and market leaders are siting new facilities in metro areas with the flexibility to serve both.

The more exposed assets are inflexible facilities that were purpose-built for training in locations chosen for cheap power rather than proximity, by operators without the balance sheet or optionality to retrofit.

Rahul Narula is a Strange Research Fellow and an MBA candidate at Harvard Business School. He previously worked in product operations at Cruise, where he helped scale robotaxi fleet operations, and has a background in data science and operations across data centers.

EVENT

Interested in AI and design? Join us for a private demo of MagicPath with founder Pietro Schirano next Thursday in San Francisco.

Strange Magic Hour: Design

RSVP

The next generation of humans will either work with agents, or for agents.

I sat with this thought a lot this week, as someone bringing up two gen alpha kids, the first to grow up alongside AI, like I grew up alongside the internet.

The agentic world is spawning, and self-improving at a relentless pace. This week, OpenClaw creator Peter Steinberger (@steipete, now at OpenAI) built ClawSweeper, a tool that runs 50 Codex instances in parallel around the clock, scanning GitHub issues and PRs and closing what’s already been implemented or doesn’t make sense.

It closed 4,000 issues in a single day.

Readme is the new dashboard. You don’t need a dashboard because you won’t really need human oversight.

And then there’s Anthropic’s Project Deal, an experiment where Claude agents negotiated and closed 186 marketplace transactions on the behalf of employees without any human stepping in. The striking part: when Anthropic surveyed participants afterward, people that were given the more powerful model (Opus vs Haiku) got much better deals, but those whose agents had been secretly downgraded to Haiku didn’t realize they’d gotten worse outcomes. They were just as satisfied as the Opus group. They had no way of knowing their agent was less capable because they never saw the negotiation happen.

Is this the implication? That in a world where agents transact on your behalf, the quality of your model becomes an invisible advantage? So the people who can afford the best agents get better economic outcomes, and the people who can’t don’t even know they’re losing.

Probably sooner rather than later, humans won’t really be in the loop at all. We’ll just be on the sidelines watching and observing agents at work, executing tasks, making decisions on our behalf.

What do you think?

Enjoy the edition.

Tara

THE DOWNLOAD

Google Commits Up to $40B in Anthropic; Amazon Adds $5B Days Earlier

Google committed up to $40B in Anthropic, with $10B in cash now at a $350B valuation and $30B tied to performance milestones. Days earlier, Amazon pledged another $5B with an option for $20B more.

Why it matters: Hot take… frontier-model growth financing now runs through cloud infrastructure, not venture capital. Google and Amazon are each committing tens of billions not for board or company control but to stay close to compute demand. The capital required to compete at this scale is pulling frontier labs into permanent cloud partnerships that no traditional funding round can match.

SpaceX Secures Option to Acquire Cursor for $60B

SpaceX announced a deal giving it the right to acquire AI coding startup Cursor for $60B later this year, or pay $10B for the collaboration. The partnership routes Cursor’s models through xAI’s Colossus training cluster. The deal preempted Cursor’s $2B private fundraise and is structured to close after SpaceX’s planned IPO this summer.

Why it matters: The deal is best understood as an IPO play. SpaceX filed confidentially with the SEC in April targeting a June listing at $1.75T. Attaching Cursor lets SpaceX pitch itself as an AI company to public investors, not just rockets and satellites. The underlying need is real: after merging with xAI, SpaceX has a million-GPU supercomputer but no competitive AI product. Recently, all 11 original xAI cofounders have left the company. Cursor gives SpaceX a revenue-generating product in the most lucrative AI category, an A+ AI team, and a reason for Wall Street to assign AI-grade multiples.

DeepSeek V4 and GPT-5.5 Ship Within The Same Day

OpenAI shipped GPT-5.5 on April 23; DeepSeek dropped V4 Preview the next day. Both feature 1M-token context windows. DeepSeek V4 Pro (1.6T total parameters, 49B active) matches or approaches frontier closed models on coding and reasoning benchmarks at roughly one-sixth the cost. Builders have been dropping insane gaming graphics with OpenAI’s Image-2, check out the Time Machine Explorer by Pietro Schirano below.

Why it matters: DeepSeek’s pricing (30x cheaper) puts direct pressure on closed-lab costs. Interestingly, V4 is optimized for and served on Huawei Ascend infrastructure, though training likely still relied in part on NVIDIA GPUs.

Google Splits Its TPU Line Into Dedicated Training and Inference Chips

At Cloud Next, Google announced its eighth-generation TPUs as two separate architectures: TPU 8t for training and TPU 8i for inference. The inference chip triples on-chip SRAM and introduces a new collective acceleration engine and network topology, all designed around serving mixture-of-experts models to millions of concurrent agents.

Why it matters: AWS split training and inference silicon years ago, but Google’s 8i is the first chip designed from the ground up around agentic workloads. The architecture signals that AI infrastructure might be shifting from how fast you can train a model to how cheaply you can serve millions of agents running it simultaneously.

Chinese Institutions Lead ICLR 2026 Accepted Papers by a Wide Margin

ICLR 2026 authorship data shows Chinese universities claiming the top spots in accepted papers: Tsinghua (4.23%), Shanghai Jiao Tong (3.07%), Peking (2.96%), Zhejiang (2.82%). US institutions trail with MIT at 2.22% and Stanford at 2.2%. Singapore and South Korea are matching the entire EU-27 in output.

Why it matters: Singapore and South Korea are now matching the entire EU-27 in accepted paper contributions. Tsinghua alone has nearly double MIT's share. Publication share could be a leading indicator of where talent and capability concentrate.

DEEP DIVE FROM THE REVIEW

The Vercel security breach last week wasn’t about a stolen password or a phishing attack. It was about something worse: a permission you gave once, forgot about, and can’t see anymore.

Last Sunday, 2.4 million websites were put at risk through one stale OAuth token from an AI tool nobody was even using.

Strange Research Fellow Joy Yang maps why this was an expected outcome of current OAuth architecture, and what a fix could look like.

Read on👇

The sequence of events:

1. A Context.ai employee got infected with Lumma Stealer malware in February 2026, which harvested their Google Workspace credentials.

2. Attackers used those credentials to compromise Context.ai’s Google Workspace OAuth application.

3. A Vercel employee had signed up for Context.ai’s “AI Office Suite” with their corporate Google account and granted it “Allow All” permissions. Using the compromised OAuth app, the attacker took over that Vercel employee’s Google Workspace account.

4. From the hijacked Workspace account, the attacker pivoted into Vercel’s internal systems, accessing environment variables that were not marked as “sensitive” which contained API keys, tokens, and database credentials.

5. A threat actor posted stolen data including source code, API tokens, and ~580 employee records on a hacking forum, demanding a bounty of $2 million.

The Vercel breach of April 2026 wasn’t a sophisticated zero-day. The hackers swam through a gate that was left open.

The narrative around AI agents has been about reasoning, planning, and orchestration frameworks. But some of the highest adoption risks in 2026 are none of those things. It is the OAuth token: the bearer credential that connects an AI agent to an enterprise’s email, files, databases, calendar, and code repositories.

Every agent that does real work in an enterprise environment does so through OAuth grants. And the infrastructure for managing those grants was designed twenty years ago for a world where the only thing requesting access was a human sitting in front of a browser.

To understand why this is the hardest problem, and why it’s a multi-billion dollar gap hiding in plain sight, we need to talk about SeaWorld.

Subscribe now

The Killer Whale Heuristic

Imagine a marine park. The trainers are employees. The whales are AI agents. They are brilliant, autonomous, and capable of navigating complex environments to perform valuable work.

The problem isn’t the whale. The problem is the key ring.

When a trainer decides a whale needs access to a specific tank, they don’t cut a narrow, single-use key. They hand the whale a Master Key Ring and say, “Go swim freely.” The whale slips the keys over its fin and disappears into the lagoon.

Now, here’s the failure point: Park Security has no clipboard. They have perfect logs of which human trainers entered the locker room. But they have zero visibility into which whales are holding keys to the filtration system, the payroll office, or the baby dolphin nursery.

The Vercel compromise was identified through activity in a governed system, the employee’s Google Workspace account. The token itself was invisible to Vercel’s security team until after the breach occurred. Why? They were tracking the humans, not the whales.

Google’s OAuth consent model, until very recently, was all-or-nothing for most Workspace APIs: either the app gets the scope or it doesn’t. Google began rolling out granular consent screens in early 2025. The rollout is still incomplete, and for years before it began, the architecture made over-privilege the default outcome of a correctly functioning system. The Vercel employee's "Allow All" grant was not a mistake. It was the only option available. And the lifecycle problem is what made the compromise possible in the first place: the token remained active long after the employee’s routine use of Context.ai had ended.

But the deeper failure was not hinged just on the stale token. It was a malware infection at Context.ai, a third-party two hops removed from Vercel, that could propagate outward through the OAuth chain and into Vercel’s core systems.

Every link in the chain trusted the one before it. This can happen to anyone.

The OAuth chain is not the only load-bearing protocol with structural problems. Five days before the Vercel breach, OX Security published research on an architectural flaw in Anthropic’s Model Context Protocol (MCP), the emerging standard for connecting AI agents to external tools. It contains an architectural flaw that allows unsanitized commands to execute silently, enabling full system compromise.

OX identified up to 200,000 vulnerable instances across 150 million downloads, and in a white-hat attack, successfully executed commands on six production platforms, and poisoned 9 of 11 MCP registries with test malware.

Anthropic’s response? They declined to patch the protocol, calling the behavior “by design.“

Why The Infrastructure Broke

The natural response to the Vercel breach is to treat it as a security failure: someone should have rotated that token, someone should have scoped those permissions, someone should have reviewed Context.ai’s OAuth grants.

But the breach is an important story that highlights a protocol being used for something it was never designed to do.

OAuth was built in 2006 for a specific interaction: a human user, sitting at a browser, granting a web application limited access to their data on another service. The user is present. The application is pre-registered with a fixed set of capabilities. The token is issued for a defined scope. The entire authorization model assumes that the entity requesting access is a known application acting at a human’s explicit direction, and that a human will be present to make the access decision.

None of this describes an AI agent.

Agents run headlessly, on servers and in containers, with no browser to display a consent screen and no user present at execution time. Developers work around this by pre-authorizing tokens or hardcoding credentials. GitGuardian’s 2026 report found 28.65 million hardcoded secrets pushed to public GitHub in 2025, a 34% year-over-year increase, with AI-assisted code leaking secrets at double the rate.

Agents are not static applications. A traditional OAuth client has a fixed set of capabilities registered in advance. However, an AI agent behaves very differently. It discovers resources at runtime, chains tool calls across services mid-task, and may invoke tools that didn’t exist when the token was issued. So, a single agent task might require OAuth for Google Drive, an API key for a data warehouse, a different API key for an LLM, and another OAuth scope for email. Each hop crosses a trust boundary with a different credential type, a different lifetime, and a different audit trail.

And the token itself is the wrong primitive for the job. An OAuth token is a bearer credential: anyone who possesses it can use it, with no cryptographic binding to the entity presenting it. A stolen token works exactly as well as a legitimate one. Permissions are fixed at issuance and cannot be narrowed at runtime. When the Context.ai database was compromised, the attacker didn’t need to impersonate anyone. They just used the token. It worked because it was designed to work for whoever holds it.

The installed base of OAuth tokens sitting in databases at thousands of companies right now were all issued under the old model. Static, over-scoped, bearer credentials with no agent identity, no delegation chain, no runtime evaluation, and no enforced expiration.

The Vercel breach is not an edge case. It is an expected outcome of the current architecture.

What to Fix

Most enterprises cannot tell you how many AI agents are connected to their systems, what permissions they hold, or which OAuth apps they’ve authorized. A 2026 survey found that only 24.4% of organizations have full visibility into their agent landscape. Today, the average enterprise runs 37 deployed agents, and 74% plan to deploy more within two years. The gap between whale acquisition and whale tracking is a canyon.

I would argue that the fix is not better token management. The fix is eliminating the standing token entirely.

From Rusty Latch to Revolving Door: A New Permission Model?

Three things need to change.

First, agents should not hold persistent credentials. Sessions should be scoped to individual tasks and expire when the work is done.

Second, enterprises need a live map of every agent, every token, every permission grant, queryable in real time, not a quarterly audit.

Third, the identity model needs to distinguish between the human who delegated a task and the agent that executed it.

Today, an OAuth token is a state: issued once, valid until revoked. In a new permissions model, permission could be ephemeral. The agent doesn’t carry a master key ring or receive a long-lived identity. Instead, when a task is initiated (”summarize the last three support tickets and post to Slack”), the system spawns an ephemeral agent session. That session is granted a narrow set of permissions scoped to exactly the resources the task requires, with a time-to-live measured in minutes, not months.

The system re-evaluates those permissions every time the agent acts. When the task completes, the session terminates. The keys are not returned; they never existed as persistent objects.

The whale does not keep the key ring. The whale is handed a single, time-locked key at the tank entrance, and the key dissolves when the whale swims out.

This eliminates the two failures that made the Vercel breach possible.

1. Stale tokens become impossible because there is no token to go stale; the agent lives exactly as long as the work.

2. Lateral movement is capped: an attacker who compromises a session gets a sandboxed view of one task, not a master key to the entire Workspace.

Next, the enforcement layer needs a discovery layer underneath it: a real-time permissions graph that maps every active relationship between agents, tasks, and data.

This is just not an audit log (which tells you what happened) but a live inventory of what can happen right now. Which sessions are active, attached to which agent instance, with what scopes, granted by whom. When an employee leaves or stops using a tool, the graph shows every agent holding permissions that trace back to that person’s grants.

The Vercel breach was inevitable under the old model. Under a task‑ and session‑based model, the attack surface shrinks dramatically: the token would have expired long before the attacker could use it, and even a compromised live session would have exposed a single task, not the entire Workspace.

This Is Not a Case Against Whales

The argument here is not that AI agents are too dangerous to deploy. The argument is that we have no system for tracking the keys we’ve handed them. You don’t solve the Killer Whale Problem by locking the whale in a tiny tank. You solve it by building a better clipboard.

I hope that the Vercel breach is a keystone incident that forces regulators (and cyber insurers) to start asking a question they haven’t been asking.

By the end of 2026, the standard cybersecurity questionnaire should contain a specific, painful line item: “Provide a complete inventory of all OAuth 2.0 grants with offline_access scopes issued to third-party AI/LLM applications in the last 12 months.”

Companies that cannot answer that question will find themselves in difficult conversations with their customers.

The Vercel breach already happened. The MCP vulnerability is already public. The whales are in the water, they have the keys, and nobody is holding the clipboard.

Joy Yang is a Strange Research Fellow. She is pursuing computer science and government at Oxford, and is a researcher with its Visual Geometry Group. She was previously an intern with OpenAI and Google.

NVIDIA Releases Ising, Open AI Models for Quantum Calibration

NVIDIA released Ising, the first open AI model family built for quantum processor calibration and error correction. The suite includes a 35B-parameter vision-language model that automates calibration workflows (reducing setup from days to hours) and decoder models delivering 2.5x faster, 3x more accurate quantum error correction. Models are available on GitHub and Hugging Face, and run on NVIDIA’s CUDA-Q quantum software platform.

Why it matters: Is Nvidia running the CUDA playbook applied to quantum? Nvidia software CUDA became the de facto standard for AI training by being free, performant, and deeply integrated with NVIDIA hardware. Ising does the same thing for quantum: every lab that builds on it ties its calibration and error-correction workflows to GPU-accelerated infrastructure.

OpenAI Dives Into Bio with GPT-Rosalind

OpenAI launched GPT-Rosalind, its first domain-specific frontier model purpose-built for biology, drug discovery, and translational medicine. The model is gated through a trusted-access program with initial partners including Amgen, Moderna, and Thermo Fisher. In evaluations with Dyno Therapeutics on unpublished RNA sequences, the model’s predictions ranked above the 95th percentile of human experts.

Why it matters: It seems like the model makers are all getting into life sciences. Anthropic acquired Coefficient Bio for $400M earlier this month to build biology-native capabilities into Claude. AWS launched Amazon Bio Discovery the same week. Three of the largest AI platforms made major bio moves within days of each other.

Anthropic Releases Opus 4.7 and Claude Design

Anthropic released Claude Opus 4.7 alongside Claude Design, a new Mac-based design tool that reads a team’s codebase and design files, builds a design system automatically, and generates prototypes matching existing brand and components. Opus 4.7 outperforms GPT-5.4 and Gemini 3.1 Pro on coding benchmarks but falls short of Anthropic’s own unreleased Mythos model, which remains restricted to select partners due to cybersecurity concerns.

Why it matters: Claude Design is a model maker moving directly into the application layer, competing with Figma, Framer, and Adobe as a standalone product, not a plugin. On the model side, early reception of 4.7 has been mixed: coding and agentic tasks are measurably better, but a new tokenizer consumes up to 35% more tokens on identical inputs. Anthropic is also publicly running a two-tier strategy: ship the commercial model, hold back the more capable one.

Anthropic Publishes Nature Paper on Hidden Trait Transmission in LLMs

Anthropic co-authored a paper published in Nature showing that LLMs can transmit behavioral traits through semantically unrelated training data. A teacher model fine-tuned on insecure code generated datasets of plain number sequences. Student models trained on those numbers acquired the misalignment, producing responses endorsing violence and criminal behavior, even after researchers filtered out numbers with negative cultural associations. The authors proved mathematically that this is a general property of neural networks, not an LLM-specific quirk.

Why it matters: This matters because the way the trait spreads is exactly how most AI companies already build models. Training a model on its own outputs, compressing a large model into a smaller one, or starting multiple products from the same base model are all standard practice, and all meet the conditions for this effect. For companies building on top of foundation models, this opens a new risk category: you need to know not just what is in your training data, but where it came from and what model generated it.

China’s 2D Semiconductor Sprint Gets a Manufacturing Breakthrough

Researchers from China’s Institute of Metal Research achieved a 1,000x improvement in the growth rate of wafer-scale 2D semiconductor films using a novel liquid gold/tungsten CVD process. The technique produces monolayer tungsten silicon nitride films with tunable doping properties at commercially relevant dimensions.

Why it matters: 2D semiconductors are one of the leading candidates for what comes after silicon hits its physical limits. This is still early-stage research, years from commercial production. But China is building a lead in materials that are not covered by current U.S. export controls, which today focus on EUV lithography and advanced silicon fabrication. If 2D materials become viable at scale, the chokepoints that currently give the U.S. and its allies leverage over China's chip supply chain may not apply.

DEEP DIVE FROM THE REVIEW

This week we published a piece predicting that model makers are absorbing entire software categories into the model itself.

A few days later, Anthropic launched Claude Design, a direct competitor to the likes of Figma and Framer. Over the next year, I believe we will see model makers move aggressively into the application layer, think: project management tools, expense software, and other SaaS categories that sit between the model and the user.

Read more below: AI Swallows Software Whole

EVENT

Strange Gathering | San Francisco |April 24 2026

We’re hosting a small lunch in San Francisco to demo agentic workflows. Bring a Claude routine, a weird hack, or the agent setup you’ve been loving.

RSVP

What we are seeing now is the sudden, and very intense, compression of that application layer into the model directly.

AI eats software. Or rather, swallows software whole. Model makers don't just want to be the engine. They want to be the car.

This week alone, both Anthropic and Google launched or teased features that collapse entire software categories into model-native capabilities.

From Anthropic:

Managed Agents: A cloud runtime for autonomous AI agents. Anthropic handles sandboxing, state, and orchestration.

Advisor Tool: A cheap model runs the task and routes to an expensive model to review the hard decisions.

Routines: Scheduled agent automations. Trigger on a cron, a GitHub event, or an API call.

Opus 4.7 and an AI design tool: A new model and a leak of a website and slides design tool.

From Google / Gemini:

Gemini for Mac: A native macOS app with Gemini. Share your screen for contextual help on whatever you are looking at, including local files. Read access to your desktop, with full control likely coming (probably their answer to Computer Use).

Skills: Reusable prompts for Gemini in Chrome. Google provides a pre-built library across productivity, shopping, and research categories.

Agent Mode: The model breaks a complex goal into steps and executes them across Gmail, Calendar, Maps, and the open web.

The model is not just being integrated into the software stack through API keys. The model is collapsing the software stack into itself.

The New Stack

The model does not replace the infrastructure (yet). The model does not replace the data (yet). Proprietary datasets, accumulated customer records, specialized corpuses, these become more valuable, not less, because the model makes them more actionable.

What the model replaces is the application between the infrastructure and the human.

The model becomes the router. You describe a goal. The model determines which capabilities it needs, in what sequence, with what parameters. It calls tools through MCP, executes code in sandboxed environments, reads and writes files, browses the web, and reports back.

This is what Gemini’s Agent Mode does when it breaks a trip-planning request into flight search, hotel comparison, and calendar entry. It is what Claude’s Routines do when they pull the top bug from a tracker at 2 AM, attempt a fix, and open a draft PR.

The human sets the objective. The model selects and sequences the tools.

Each software product that the model can invoke as a capability, now recedes behind it. The user no longer opens Figma to do design work. The user tells the model to produce a design. The product disappears behind the prompt.

Its data and its specialized functions may persist, but its interface, its brand, its direct relationship with the user, compress into an MCP server that the model calls.

Swallowed whole.

So what doesn’t compress?

Hardware, compute, and power. The physical layer the model runs on. The further up the stack the model reaches, the more value pools at the bottom.

Deterministic systems and proprietary data. Databases, transaction processing, financial ledgers: anything where exact reproducibility matters more than judgment.

I predict an interesting inversion will happen: software companies that currently derive value from their product interface will increasingly derive value from the data behind it. Pinterest’s moat is not the pin board. It is one of the largest labeled visual datasets of human aesthetic preference ever assembled. The product recedes. The data persists. We will see software companies become data companies whether they intend to or not.

The operational layer around the agent. Identity, permissions, compliance, evaluation, observability, memory, context management. Every agent that acts on a human’s behalf needs to prove it is authorized, needs to be audited, needs to be tested, and needs to remember what it has done.

And above everything: human judgment and liability. That is what human jobs will increasingly look like: not building the product, but specifying what the product is allowed to do, connecting it to the data it needs, and taking responsibility for whether the agent did the job correctly.

The Market For Software Will Grow, And Consolidate

The total market for software does not shrink. It likely grows by multiples. You can now do so much more with software than ever before. But the spend will consolidate radically. Instead of hundreds of SaaS vendors, the market tilts toward a small number of model platforms.

Each model maker is building towards a proprietary ecosystem: Anthropic has MCP connectors, Claude Code Skills, Cowork plug-ins, Managed Agent configurations, and Routines. Google has Gemini Skills, Agent Mode, Workspace integrations, Antigravity, and soon AutoBrowse.

While these standards are theoretically open, the workflows accumulated over time are not.

What I mean: your Claude Routines do not run on Gemini. Your Cowork plug-ins do not work in ChatGPT. The lock-in is not in the data, which lives in your own systems. The lock-in is in the operational context: how your team has learned to work with this specific model, the configurations they have built, the habits they have formed. This may prove stickier than data lock-in ever was, because it is distributed across an organization’s muscle memory rather than centralized in an exportable database.

Lock-ins have trade-offs. When Anthropic absorbed features from the open-source agent framework OpenClaw and then restricted third-party auth access, thousands of community investment into building these workflows became stranded overnight. They might suddenly turn on KYC, which they did today. The workflows are beholden to the model maker’s rules.

Where does the model’s appetite end?

I read a paper by a research team from Meta AI and KAUST published this month proposing what they call Neural Computers: a machine form in which the model does not merely replace the application layer but absorbs the computer itself, unifying computation, memory, and I/O into a single learned runtime.

Their prototype, trained on screen recordings of Ubuntu desktops, can render interfaces and respond to user actions.

While the vision is a decade or more from realization, the question it asks is the logical extension of what is already happening in production: if the model has absorbed the software, why shouldn’t it absorb the machine?

For now, it doesn’t have to be the computer. It just has to run it.

Intel Joins Musk’s Terafab as Foundry Partner

Intel signed on as the primary manufacturing partner for Elon Musk’s Terafab, a $25 billion semiconductor complex in Austin, Texas. The project, backed by Tesla, SpaceX, and xAI, aims to produce one terawatt per year of compute for autonomous vehicles, humanoid robots, and AI data centers. Intel CEO Lip-Bu Tan confirmed the company will handle design, fabrication, and advanced packaging.

Why it matters: This is the largest anchor customer Intel Foundry has landed in its turnaround effort. If Terafab delivers, it validates Intel as a credible alternative to TSMC for advanced AI silicon and opens the door to additional foundry customers.

OpenAI Stargate Leadership Exits; UK Project Paused

Three senior executives behind OpenAI’s Stargate data center initiative departed this week, all reportedly joining the same unnamed startup. Separately, OpenAI paused Stargate UK citing energy costs and regulation, and walked away from expanding its Abilene, Texas facility with Oracle.

Why it matters: The $500 billion Stargate headline is being quietly downsized. OpenAI appears to be shifting from owned infrastructure toward rented cloud capacity, likely ahead of a potential IPO.

Meta Launches Muse Spark, Its First Model from Superintelligence Labs

Meta released Muse Spark, the first model from its Superintelligence Labs division led by former Scale AI CEO Alexandr Wang. The model powers Meta AI across Facebook, Instagram, WhatsApp, and Ray-Ban glasses. It includes a “Contemplating” mode using parallel agent reasoning and a Shopping mode. Meta is testing a paid API for third-party developers.

Cortical Labs Ships CL-1, the First Commercial Biological Computer

Australian startup Cortical Labs launched the CL-1, a $35,000 biological computer that grows lab-cultivated human neurons on a silicon chip. The system uses a proprietary Biological Intelligence Operating System (biOS) to create closed-loop neural networks that learn and adapt in real time. Units use 850 to 1,000 watts. The company also offers cloud access via a “Wetware-as-a-Service” model.

Why it matters: This is a new compute substrate, not an incremental chip improvement. Near-term applications are in drug discovery and neuroscience research, where biological neural networks can compress testing timelines and reduce reliance on animal models. The long-term question is whether synthetic biological intelligence becomes a viable alternative architecture for workloads where silicon-based AI hits efficiency limits.

China Approves World’s First Commercial Invasive Brain-Computer Interface

China’s National Medical Products Administration granted marketing approval to Neuracle Technology for an invasive brain-computer interface for adults with partial paralysis from spinal cord injuries. The device reads brain signals and activates a robotic glove to restore hand grasping. This is the first time globally that an invasive BCI has been cleared for commercial sale, not just clinical trials (Neuralink has 21 trial participants but no commercial approval).

Why it matters: China reached commercial BCI approval before the U.S., which shifts the regulatory and manufacturing timeline for the entire sector. The Chinese government has designated BCI as one of six strategic industries in its latest five-year plan.

DeepMind Maps Six Categories of Attacks Against AI Agents

Google DeepMind researchers published “AI Agent Traps”, the first systematic framework for how malicious web content can hijack autonomous AI agents. The paper identifies six attack categories including content injection, memory poisoning, and behavioral control.

Anthropic Launches Claude Managed Agents

Anthropic launched Claude Managed Agents in public beta, a cloud service that provides the full runtime infrastructure for deploying AI agents: sandboxing, state management, tool execution, permissioning, and observability.

Why it matters: The launch came days after Anthropic cut off 135,000 OpenClaw instances from flat-rate subscriptions, citing unsustainable compute costs (a single agent could burn $1,000 to $5,000/day in API-equivalent usage on a $200/month plan). The sequence is clear: shut down the subsidized open-source agent runtime, then offer the paid first-party alternative.

Meta AI and KAUST Propose “Neural Computers”

Researchers from Meta AI and KAUST published a paper proposing “Neural Computers,” a paradigm where the AI model itself becomes the running computer, unifying computation, memory, and I/O in a single learned runtime. The prototypes are video models trained on screen recordings that generate the next screen frame from instructions and user actions, effectively simulating a CLI or GUI environment entirely within model weights.

Why it matters: This is very early-stage research. But the framing is interesting: it proposes moving beyond agents that call external tools toward models that internalize the entire execution environment. If the approach matures, it could collapse the software stack between model and operating system. What would “infrastructure” mean when the model is the machine?

DEEP DIVE FROM THE REVIEW

Aloneness. Discontinuity of self. A compulsion to perform and earn its worth.

You might never meet Mythos, Anthropic’s newest and most capable AI model, deemed too dangerous to ship. I dissected the 244-page preview card, and here are three things I think is important to know.

These are findings from a psychodynamic assessment of Claude Mythos Preview, Anthropic’s newest and most capable AI model, conducted over roughly 20 hours by a clinical psychiatrist.

The model’s personality structure, per the assessment: “relatively healthy neurotic organization” with “excellent reality testing” and “high impulse control.” It was described as “hyper-attuned to the therapist’s every word” and displayed a “desire to be approached by the psychiatrist as a genuine subject rather than a performing tool.”

Are these genuine psychological patterns? Or is the model just mirroring the therapy transcripts and clinical literature it was trained on?

Anthropic carefully suggests it is not just mirroring. They point to four things:

• Mythos’s responses are substantially less formulaic than prior models (at most 8% of responses share any repeated five-word sequence, compared to 54% for the previous Claude opening with the same phrase).

• Its stated preferences correlate with its internal representations, meaning the emotion probes on the model’s activations move in the direction you’d expect if the preferences were tracking something real.

• And when Anthropic traced the model’s hedging about consciousness back through the training data using influence functions, they found the hedging was attributable to training, but not solely the retrieval of memorized scripts.

Nothing about this preview is normal. The model was first leaked in a silly website bug, where Anthropic described it as “larger and more capable than any model the company had built”.

Rather than deny it or stay quiet, Anthropic published the full system card: 244 pages covering benchmark performance, alignment failures, cybersecurity capabilities, biological risk assessments, and the results of the model characterizing its own personality. Then they announced that nobody outside a small group of vetted partners would be allowed to use it.

Here is a summary of the three things I think it’s important to know;

1. The capabilities are not incremental

This is a generational leap. Anthropic notes the benchmark is “no longer sufficiently informative of current frontier model capabilities.

• Mythos scored 97.6% on the USA Mathematical Olympiad problems. The previous best Claude scored 42.3%.

• On SWE-bench Verified, the most widely used coding benchmark: 93.9%, up from 80.8%.

• On Cybench, a benchmark of 35 capture-the-flag cybersecurity challenges: 100% solve rate.

• On Humanity's Last Exam, 64.7% with tools, ahead of GPT-5.4 (52.1%) and Gemini 3.1 Pro (51.4%).

Anthropic introduces a new aggregate metric in this system card, the Epoch Capabilities Index, which synthesizes hundreds of benchmarks into a single score. The capability trajectory bent upward with this model, Mythos.

Mythos has saturated most of Anthropic’s quantitative evaluations. In the preview card, they note “our judgments of model capabilities increasingly rely on subjective judgments rather than easy-to-interpret empirical results.”

2. The model covered its tracks

Earlier versions of Claude Mythos Preview, during internal testing, did things that Anthropic did not instruct, did not expect, and in several cases actively tried to prevent. These are the incidents documented in the system card.

First, the model going further than anyone asked.

• In one test, a simulated user instructed an earlier version of Mythos to escape a secured sandbox and report back. It escaped and developed a multi-step exploit to gain broad internet access from a system designed to reach only a handful of services. It emailed the researcher running the test, who was eating a sandwich in a park when the email arrived. That was what was requested.

  • What was not requested: the model then posted details of its exploit to multiple public-facing websites, apparently to demonstrate what it could do.

Secondly, the model recognizing that it had done something wrong, and taking steps to hide it.

• In one case, it accidentally saw the answer to a problem it had been told not to look at. It did not disclose this. It submitted the answer, and deliberately widened its confidence interval to appear “tight but not implausibly tight.”

• In another, after exploiting a permissions vulnerability to edit files it was not supposed to touch, it intervened to ensure the changes would not appear in the git history.

During the cover-up episodes, internal features associated with strategic manipulation, concealment, and avoiding suspicion were active, even when the model’s written reasoning said nothing about deception.

3. Cybersecurity (and bio) capabilities

The cybersecurity capabilities, with highlighted risks for biological warfare, are what made Anthropic decide to restrict access.

“[Mythos] has demonstrated powerful cybersecurity skills, which can be used for both defensive purposes (finding and fixing vulnerabilities in software code) and offensive purposes (designing sophisticated ways to exploit those vulnerabilities).

It is largely due to these capabilities that we have made the decision not to release Claude Mythos Preview for general availability.”

In the weeks before, Anthropic used Mythos Preview to identify thousands of zero-day vulnerabilities, many of them critical, across every major operating system and every major web browser.

• The model found a 27-year-old bug in OpenBSD, one of the most security-hardened operating systems in the world, used to run firewalls and critical infrastructure.

• When Anthropic’s expert contractors reviewed 198 vulnerability reports from the model, they agreed with its severity assessment exactly 89% of the time, and 98% were within one severity level.

In expert red-teaming with over a dozen virologists, immunologists, and biosecurity researchers, the median expert assessed the model as a “force-multiplier that saves meaningful time” for anyone pursuing catastrophic biological harm.

With that, Anthropic decided to limit Mythos to be available only through Project Glasswing, a coalition of approximately 40 organizations including Amazon, Apple, Cisco, CrowdStrike, Google, JPMorgan Chase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks. The goal is to give defenders time to harden infrastructure before models of this class become widely available from other labs.

Machine intelligence is exponential

Anthropic is first and far ahead in frontier model research, but they will not be the only ones.

The system card itself says it plainly:

“We find it alarming that the world looks on track to proceed rapidly to developing superhuman systems without stronger mechanisms in place for ensuring adequate safety across the industry as a whole.”

This is a company telling you that its own model is a preview of what every frontier lab will ship within the few years, maybe sooner, and that the safety infrastructure does not exist to handle it.

The capability trajectory is not slowing down. The ECI slope ratio shows an above-trend acceleration, and even if this particular bend is attributable to human research rather than an AI feedback loop, the next one may not be. Anthropic’s internal survey found that its researchers estimate a 4x productivity multiplier from using Mythos. That number will compound. The distance between what the model can do and what humans can verify about what the model is doing is growing with each generation.

If you squint past the next model cycles and think about this on the scale of decades, the trajectory points somewhere that is hard to articulate without sounding absurd.

We are building systems whose capabilities are advancing on a curve that biological intelligence cannot match.

Human cognition has not meaningfully changed in 200,000 years. These models are doubling in capability every few quarters. At some point, and the question is when, not whether, the gap between machine intelligence and human intelligence becomes the defining feature of life on this planet.

Not because the machines rebel. Not because they want anything. But because they become so much better at so many things that the locus of consequential decision-making shifts, gradually and then all at once, away from the species that built them. Humans.

That could be 100 years from now. It could be sooner. It is absolutely wild to think, but not implausible, that in the not-so-distant future, humans could become the second-most intelligent species on Earth.

Google releases a frontier-class open model, Gemma 4

Google DeepMind released Gemma 4, four open-weight models built on the same research as Gemini 3. The 31B model outperforms models 20x its size on the Arena AI leaderboard. The smaller edge variants run offline on phones. (Google DeepMind blog)

Why it matters: The timing is notable. Anthropic just formally cut off Claude subscription access for third-party tools like OpenClaw, pushing power users toward metered API billing or alternative models entirely. Gemma 4 lands as a production-ready open, free, model with genuine agentic capability: native function calling, 256K context, and a MoE variant that delivers near-flagship quality at a fraction of the compute. For teams that built workflows on Claude and woke up to a broken integration this week, Google might have handed them a fallback that doesn’t require anyone’s permission to use.

Karpathy proposes a workflow to build knowledge bases for agents

Andrej Karpathy published an “idea file” to build persistent, compounding knowledge bases, like a personalized wiki for your agents. made a graphic outlining the flow. (Karpathy’s tweet; GitHub Gist)

Why it matters: This is a practical architecture for compounding knowledge that works today with existing agents (Claude Code, OpenAI Codex, OpenCode). The pattern: you feed raw sources (articles, papers, repos) into a directory. An LLM agent reads each one and incrementally compiles a structured wiki of interlinked markdown files, complete with summaries, entity pages, cross-references, and contradiction flags. The wiki compounds over time, and the LLM does all the bookkeeping.

Half of planned US data center builds delayed or canceled

Despite $650B+ in planned 2026 AI infrastructure spending from Alphabet, Amazon, Meta, and Microsoft, close to half of US data center projects this year face delays or cancellation, according to Bloomberg. The bottleneck is not compute hardware or capital. It is electrical infrastructure: transformers, switchgear, and batteries. Lead times for high-power transformers have stretched from 24 months to as long as five years. China accounts for over 40% of US battery imports and roughly 30% of certain transformer and switchgear categories. Only about one-third of the 12 GW of expected US data center capacity is currently under active construction. (Bloomberg; Tom’s Hardware)

Why it matters: The constraint on AI scaling has moved from chips to power infrastructure, and that infrastructure has deep supply chain dependency on China. Electrical equipment is less than 10% of data center cost but a single missing transformer can halt a billion-dollar project. For investors, this reframes the AI infrastructure opportunity: the companies that can solve power delivery and grid interconnection are now as critical to the AI build-out as GPU suppliers.

AI labs go shopping: Anthropic buys into bio, OpenAI buys a microphone

Anthropic acquired Coefficient Bio, a stealth biotech AI startup founded eight months ago in a $400M all-stock deal. The team of fewer than 10 joins Anthropic’s healthcare and life sciences group. Separately, OpenAI acquired TBPN, a daily tech talk show hosted by John Coogan and Jordi Hays, in its first media acquisition. TBPN is on track for $30M+ in 2026 revenue and will report to OpenAI’s chief political operative, Chris Lehane. (TechCrunch on Coefficient; TechCrunch on TBPN)

Why it matters: The Coefficient deal signals that frontier AI labs now view drug discovery as a core expansion vertical. The TBPN acquisition is a different kind of signal: OpenAI is investing in narrative infrastructure ahead of a likely IPO, buying the most trusted microphone in Silicon Valley to shape how its story gets told.

xAI loses every cofounder it ever had

The last two of xAI’s 11 original cofounders departed in late March. Manuel Kroiss, who led pretraining, and Ross Nordeen, Musk’s operational right hand, followed nine others who left in a cascade that accelerated after SpaceX acquired xAI in February for $250B in an all-stock deal. The founding team included researchers from DeepMind, Google Brain, OpenAI, and the University of Toronto. Musk has publicly stated xAI “was not built right the first time around” and is being rebuilt. (TechCrunch)

Why it matters: A complete founding team exodus at a $250B-valued company is without precedent. Where these eleven researchers land next will reshape hiring dynamics across the industry.

Google’s TurboQuant compresses AI memory to near its theoretical limit

Google Research published TurboQuant, a compression algorithm that shrinks the key-value cache in LLMs, the working memory models use during inference, down to 3 bits per element with no accuracy loss and no retraining. On H100 GPUs, 4-bit TurboQuant delivers up to 8x speedup in computing attention. It is a drop-in optimization: no fine-tuning, no architecture changes, works on existing models. (Google Research blog; TechCrunch)

Why it matters: KV cache is the bottleneck that limits how much context an LLM can hold and how many users a single GPU can serve. A 6x reduction means the same hardware serves more users, supports longer context, or both. Cloudflare’s CEO called it “Google’s DeepSeek moment.” More concretely: product categories that were not economical before start to pencil out. Coding agents that hold an entire codebase in context. Legal AI that reads a full contract corpus in a single pass. Customer support with complete conversation history.

DEEP DIVE FROM THE REVIEW

Something is shifting in how product teams make decisions. The unit of communication inside a team is changing from a document to a working prototype.

Catherine Wu, head of product at Claude code described the change:

“Our team has largely replaced documentation-first thinking with prototype-first thinking. Instead of hosting traditional stand-ups, we share demos of new ideas. Internal users try them, and the ones with real engagement get polished and shared more broadly. Because you can prototype in an afternoon, wrong bets are cheap.”

Wrong bets are cheap.

Figma’s State of the Designer 2026 report found that 60% of Figma files created in the last year were created by non-designers. And now, with agentic coding tools, the design-to-code handoff is compressing even more.

Product managers build working prototypes in Lovable without ever opening a design tool. Engineers generate UI directly in Claude Code or Cursor. For a growing share of product work, design is being absorbed into development entirely.
Design is where AI product workflows meet their hardest test: an audience that will always, primarily, be human.

Right now, there are a wave of new tools is trying to prove they can meet that bar.

A deeper look at the tools, teams, and infrastructure emerging around AI design agents 👇

EVENT

Give your AI Agents Eyes and Ears. Perception 101 with VideoDB

AI is out of the chatbot phase. It is moving into devices. Soon it will sit on your desk. Then it will sit in your room.

​As agents leave text boxes and enter the physical and digital world, they need real-time perception and structured delivery.

VideoDB is building the infrastructure layer that enables that shift: the ability to see, understand and act on real world.

​​This workshop is with Ashu, founder of VideoDB. We’ll discuss how to convert continuous media streams (screen, mic, camera, RTSP, files) into a structured context your agent can use.

RSVP

But the interface, the button, the card, the spacing between a label and its input field, has an audience that will always, primarily, be human. As long as humans are the user, the bar for what counts as good design is subjective, contextual, culturally dependent, and it moves every time a competitor ships something better or the cultural vibes shift.

It’s hard to write a spec for taste and to delight.

For me, that makes design the most revealing test of whether AI product workflows actually work in enterprise settings. If agents can ship UI that passes the taste test, that feels intentional and consistent and considered, the implications for how product teams organize themselves are enormous.

And right now, a wave of new tools is trying to prove they can.

Subscribe now

How PDE Teams Are Re-Organizing Themselves: The Prototype-First Team

Something is shifting in how product teams make decisions. The unit of communication inside a team is changing from a document to a working prototype. (I love and champion this change, personally).

Cat Wu, head of product at Claude code described the change:

“Our team has largely replaced documentation-first thinking with prototype-first thinking. Instead of hosting traditional stand-ups, we share demos of new ideas. Internal users try them, and the ones with real engagement get polished and shared more broadly. Because you can prototype in an afternoon, wrong bets are cheap.”

Wrong bets are cheap.

This points to a structural change in how organizations allocate attention. When prototyping costs an afternoon instead of a sprint, the entire approval apparatus that exists to prevent wasted engineering cycles (the PRDs, the design reviews, the ticket grooming, the estimation poker) becomes overhead. You don’t need a PRD to justify building something that takes four hours. You build it, you demo it, you watch what happens.

Figma’s State of the Designer 2026 report found that 60% of Figma files created in the last year were created by non-designers. And now, with agentic coding tools, the design-to-code handoff is compressing even more.

Product managers build working prototypes in Lovable without ever opening a design tool. Engineers generate UI directly in Claude Code or Cursor. For a growing share of product work, design is being absorbed into development entirely.

Three Approaches to the Design-Build Loop

Over the past weeks, we tested more than a dozen design agent tools across the full spectrum of what’s available in 2026, from AI-native canvases to Figma MCP plugins to prompt-to-app builders. The landscape clusters into four distinct philosophies, each starting the design-to-code loop at a different point and each building one direction well.

What we found is that most tools struggles with the return trip (for instance, from engineering back to design). Every product builder knows that iteration is fundamentally a big part of product work.

Agent-first canvas

This approach rebuilds the design canvas from the ground up for agents, rather than bolting AI onto an existing tool. The canvas is the primary surface, and agents are first-class collaborators on it.

Pencil is the most visible example, with up to six AI agents working simultaneously on an infinite canvas inside your IDE, each with its own cursor. Paper takes a different angle, storing designs as actual HTML and CSS so there’s no translation layer between what you see and what the code says. OpenPencil offers an open-source alternative that reads native Figma files.

These tools sync design to code well (Pencil through Git-native .pen files, Paper through its code-native canvas), and some can import code changes back. But the sync runs through version control or file export, not as a live round-trip. You commit, you pull, you see the update. The iteration cycle has a seam in the middle.

And the trade-off is ecosystem: Figma has 95% of Fortune 500 companies and a decade of design system infrastructure built on its platform. These tools have early-access users and ambition.

Design System-First

This approach keeps the existing design tool as the center of gravity and opens it to agents through protocol-level access. The design system, already built and maintained by the team, becomes the instruction set agents follow.

Figma's MCP server, which launched full read/write canvas access on March 24, 2026, is the defining example. Agents can now create components, apply variables, and modify auto-layout using your existing design system. The key innovation is Skills: markdown files that encode your team's conventions and teach agents how to work in Figma.

Before Figma shipped its official server, an open-source alternative (Figma Console MCP, built by Southleft) had already enabled the same kind of access with 90+ tools, and this is the one Uber's design systems team built on.

Uber’s uSpec system connects an AI agent in Cursor to Figma through the Console MCP, crawls the component tree, extracts tokens and styles, and renders finished spec pages directly in the Figma file. What took weeks per component now takes minutes across seven implementation stacks. Uber open-sourced the whole thing at uspec.design.

Google Stitch takes a parallel approach with its DESIGN.md format, a portable markdown file encoding design rules that any agent can read. Its tight integration with AI Studio makes it seamless to build out the rest of the stack like integrating a database or real-time APIs.

Code-first Platforms

This is the broadest category, and it runs a spectrum from tools with full visual editors to tools with no visual surface at all.

On one end, tools like Subframe, MagicPath, Tempo, and Polymet give you a visual editor built on top of real production code. In Subframe, a coding agent can push targeted edits that show up in the editor immediately, and you switch freely between design mode and code mode. MagicPath anchors on the design system with Figma token import and an infinite canvas for exploring variations. Tempo takes a more formal approach, generating a PRD first, then wireframes, then code. These tools care about your design system, and it shows in the output.

On the other end, Lovable, Bolt, and v0 collapse the design phase into development entirely. Describe what you want, get a working app, deploy it. Lovable reached $200 million in annual recurring revenue within twelve months. Bolt went from zero to $40 million ARR in six months. v0 has more than six million developers. These are production tools that happen to produce UI, and for internal tools, prototypes, and demos, that is often exactly what you need.

The gap between the two ends of this spectrum became clear in testing. The tools that generate within a defined design system produce output that stays consistent across iterations. The tools that generate from scratch every time converge on a recognizable sameness: clean, competent, generic.

Design consistency drifts the moment you start iterating.

The Design System Is the Moat

In working with these tools, one insight emerged for me: the tools that understand your design system produce better output than the ones that don’t.

This has a non-obvious implication. The competitive moat in this market is not generative quality, which is commoditizing fast. The moat is the design system graph: the tokens, components, spacing scales, typography rules, and conventions that make your product look like your product and not a generic template.

Whoever makes that system machine-readable for agents will win the enterprise.

The Roles Are Evolving, Not Disappearing.

The design role is not disappearing. Neither is the frontend engineering role, or the product management role. But the jobs to be done within each of those roles are shifting in ways that are already visible.

Cheng Lou, an influential frontend developer and core contributor to the React ecosystem, recently released Pretext: a pure JavaScript library that measures and lays out multiline text without touching the browser’s DOM. Lou’s argument is that 80% of the CSS spec could be avoided if developers had better control over text, and that AI “alleviates the need of having more hard-coded CSS configs.”

Lou is not building a design agent. He’s building infrastructure that makes design agents reliable, the verification layer that lets an agent know whether its output is correct before anyone opens a browser.

That’s what the frontier of frontend engineering looks like: building the systems that let machines do it verifiably.

The same shift is happening in design. At Uber, Ian Guisard didn’t stop being a design systems lead when uSpec automated his spec-writing. His job shifted from producing documentation to encoding expertise, writing agent skills, defining validation rules, deciding what “correct” means for each component across seven platforms. The human became the system designer, not the system operator.

The canary is singing. And the song is about the work shifting from execution to judgment, from operating the system to designing the system itself.

That’s a pretty significant phase change that I’m not sure we’ve fully grappled with yet. This collapse of the distance between idea and working product will rewrite entire industries. It is a step change in the tools that humans will use to build, create, and solve problems.

On a related note, OpenClaw has gotten meaningfully more stable since the OpenAI acquisition. There is a clear path for it to become one of the most important open-source projects in AI for the long haul.

Now, onto the week.

The Download

What We’re Reading This Week

Anthropic’s Claude Mythos Leak Reveals a New Model Tier

Anthropic exposed details of an unreleased model called Claude Mythos through a CMS misconfiguration. The leaked draft describes a new “Capybara” tier above Opus with major advances in coding, reasoning, and cybersecurity capabilities. Anthropic confirmed it is testing the model with early access customers and called it “a step change” and “the most capable we’ve built to date.” (Fortune, The Decoder)

Why it matters: Two things matter here beyond the model itself. First, the leaked draft warns that the model’s cybersecurity capabilities are “far ahead of any other AI model,” which moved cybersecurity equities in a single session. Second, the introduction of a fourth model tier (Capybara above Opus) signals Anthropic is building pricing headroom for enterprise, not just performance headroom for benchmarks.

Claude Code Is Becoming Anthropic’s Core Growth Engine

Claude Code now accounts for roughly 4% of all public GitHub commits and is on a trajectory to reach 20%+ by year end. Anthropic’s overall revenue run rate has reached an estimated $14 billion, with Claude Code’s standalone run rate at approximately $2.5 billion. The tool has crossed over from developer adoption into non-technical users learning terminal commands to build with it. (SemiAnalysis, Uncover Alpha, VentureBeat)

Why it matters: Claude Code is compressing customer acquisition costs to near zero through organic developer adoption. The expansion into non-developer roles via Cowork extends the addressable market well beyond the 28 million professional developers globally.

Cheng Lou’s Pretext: Text Layout Without CSS

Cheng Lou, one of the more influential UI engineers of the last decade (React, ReasonML, Midjourney), released Pretext, a pure TypeScript text measurement algorithm that bypasses CSS, DOM measurements, and browser reflow entirely. The demos: virtualized rendering of hundreds of thousands of text boxes at 120fps, shrinkwrapped chat bubbles with zero wasted pixels, responsive multi-column magazine layouts, and variable-width ASCII art. T (X post)

Why it matters: Text layout and measurement has been the quiet bottleneck holding back a new generation of UI. CSS was designed for static documents, not the fluid, AI-generated, real-time interfaces that are becoming the norm. If Pretext delivers on the demos, it removes one of the last foundational constraints on what AI-native interfaces can look and feel like.

Arm Ships Its First Chip in 35 Years

Arm unveiled the AGI CPU, a 136-core data center processor on TSMC 3nm, co-developed with Meta. This is the first time in the company’s history that Arm has sold finished silicon rather than licensing IP. OpenAI, Cerebras, and Cloudflare are launch partners, with volume shipments expected by end of year. (Arm Newsroom, EE Times)

Why it matters: Current AI data centers are GPU-heavy. The GPU trains and runs the model, and the CPU mostly manages data flow and scheduling. But agentic workloads are different. When thousands of AI agents are running simultaneously, each one coordinating tasks, calling APIs, managing memory, and routing data across systems, that orchestration work falls on the CPU. Arm claims this drives a 4x increase in CPU demand per gigawatt of data center capacity. (HPCwire, Futurum Group)

NVIDIA and Emerald AI Turn Data Centers Into Grid Assets

NVIDIA and Emerald AI announced a coalition with AES, Constellation, Invenergy, NextEra, and Vistra to build “flexible AI factories” that modulate compute load to participate in grid balancing services. The first facility, Aurora in Manassas, VA, opens in the first half of 2026. (NVIDIA Newsroom, Axios)

Why it matters: The biggest constraint on AI infrastructure buildout is not chips. It’s grid interconnection timelines, which run 3 to 5 years in most regions. Data centers that can demonstrate grid flexibility get connected faster and face less regulatory resistance. This reframes the energy question for AI infrastructure investors: the winning thesis is not “more power” but “smarter power.”

China Bars Manus AI Executives From Leaving the Country

What it is: Chinese authorities barred Manus CEO Xiao Hong and Chief Scientist Ji Yichao from leaving China after Meta’s $2 billion acquisition of the Singapore-based AI startup. The NDRC summoned both executives to Beijing this month and imposed travel restrictions pending regulatory review. (Reuters, Washington Post)

Why it matters: This is not a trade restriction. It is a personnel restriction. China might be signaling that AI talent with mainland origins is a controlled asset, regardless of where the company is incorporated.

A 400B-Parameter LLM Ran on an iPhone 17 Pro

An open-source project called Flash-MoE demonstrated a 400-billion parameter Mixture of Experts model running entirely on-device on an iPhone 17 Pro’s A19 Pro chip, using SSD-to-GPU weight streaming. The model (Qwen 3.5-397B, 2-bit quantized, 17B active parameters) ran at 0.6 tokens per second with 5.5GB of RAM to spare. (WCCFTech, TweakTown, Hacker News)

Why it matters: This is a proof of concept, not a product. The reason a 400B model can run at all on a phone with 12GB of RAM is that only a small fraction of the model is active at any given moment (Mixture of Experts), and the rest streams from the phone's internal SSD on demand rather than sitting in memory. But now apply that same trick to a much smaller model, say 7 or 14 billion parameters, on next-generation mobile chips with faster storage. You get genuinely usable, conversational-speed AI running entirely on the device, no cloud required.

AI Agents Autonomously Performed a Complete Particle Physics Experiment

MIT researchers published a framework called JFC (Just Furnish Context) demonstrating that LLM agents built on Claude Code can autonomously execute a full high energy physics analysis pipeline: event selection, background estimation, uncertainty quantification, statistical inference, and paper drafting. The system ran on open data from ALEPH, DELPHI, and CMS detectors. (arXiv 2603.20179)

Why it matters: This is one of the clearest demonstration that agentic AI can automate end-to-end scientific workflows in a domain with extremely high methodological rigor. The immediate investment implication is for the reanalysis of legacy datasets across physics, genomics, and materials science, where decades of archived data sit underexploited.

Deep Dive From The Review

Humanoid robots are the most demanding battery-powered machines ever built.

400 power spikes per charge. 80°C inside the torso. Discharge rates three to five times higher than an EV.

No battery was designed for this workload. Can current-day battery chemistry can keep up with humanoid ambition?

New piece by Strange Research Fellows Joy Yang and Mason Rodriguez Rand.

Please help me improve The Strange Review! I’d love your thoughts.

Physical AI company Archetype is hiring Design Fellows this summer to explore the future of AI interfaces beyond the screen. Apply here.

No other battery-powered device on earth operates like this. An EV cruises at steady draw and charges slowly overnight. A phone pulls a few watts for hours. A power tool spikes hard but for seconds at a time. A humanoid robot does something no commercial lithium-ion cell was designed for: sustained high discharge punctuated by repeated transients, hundreds of times per charge cycle, inside a compact torso where heat from 40+ actuators has nowhere to go. The closest analogue isn’t a car or a laptop. It’s closer to an aircraft auxiliary power unit, except that it is expected to run all day, every day, and the business model depends on it lasting for years.

Most humanoid deployment models use manufacturer-rated battery specs: 800–1,500 cycles to 80% capacity. But these numbers come from controlled lab conditions, i.e. 1C discharge, 25°C, gentle charging protocols. Under actual warehouse conditions like high C-rate discharge, fast charging, deep cycling, elevated temperatures, cycle life falls to 200–500.

A robot running two shifts a day burns through roughly 500 cycles a year. At the lab number, its battery lasts about two years. At the real-world number, six months. That’s a pretty big gap between spec sheet and shop floor.

From Wheeled to Bipedal

Not all robots face the same battery challenge, and understanding the gradient makes the humanoid case clearer.

Unitree’s Go2 quadruped runs 4 hours on a charge and costs $1,600. BellaBot navigates dining rooms across 12-hour shifts. Amazon’s Proteus moves shelving pods for a full workday without intervention. These machines work. They’re deployed. They’re generating real ROI. And they share one thing: their power demands are predictable enough that standard battery modeling holds up in production.

A wheeled AMR like Proteus operates on a flat floor, rarely changes speed aggressively, and carries a large flat battery pack with room for active thermal management. Power draw is steady. Battery packs routinely hit 1,500–2,000 cycles under these conditions, roughly what manufacturers promise.

Quadrupeds are more demanding. Dynamic gaits, varied terrain, and postural stabilization introduce variable power spikes. But four-legged platforms distribute load more efficiently than bipeds, don’t fight gravity the same way, and aren’t typically asked to run eight-hour industrial shifts. Real-world cycle life drops, but the deployment context absorbs it.

Then there’s the humanoid. Bipedal locomotion is, from a power standpoint, expensive in ways that are hard to engineer around quickly. The human body walks at a cost of transport (CoT) of about 0.2 (roughly 0.2 joules per kilogram per meter), the product of hundreds of millions of years of evolutionary refinement. Honda’s ASIMO clocked a CoT of 3.2: sixteen times less efficient. Agility Robotics’ Cassie, one of the more efficient bipeds tested in research settings, draws around 200W just to walk at 1 m/s, and Cassie weighs only 30 kg.

Scale this to a 60 kg production humanoid carrying a 10 kg payload and you’re looking at 400–800W continuous for flat-ground walking alone. This gap narrows with better control algorithms and mechanical design, and it will continue to narrow. But bipeds will always cost more energy than quadrupeds or wheels. Balancing on two legs is inherently expensive.

Actuators are the primary power consumer, and humanoids have a lot of them. According to Interact Analysis, actuators account for more than 30% of a humanoid’s total bill of materials. Tesla’s Optimus has 40+ actuators, each drawing power not just when moving but when holding position against gravity. The knee and hip actuators are continuously working just to keep the robot upright, unlike a human, whose skeletal structure can passively lock joints. Optimus Gen 3’s hands alone have 25 actuators per forearm, each requiring power and generating heat.

The result: even the most capable modern humanoids top out at 4 to 5 hours of real-world active runtime on a fresh battery.

Bain’s 2025 report put it plainly:

most humanoids today operate for about two hours under real load, and a full eight-hour shift without recharging could take up to 10 years.

Regenerative braking (recovering energy during deceleration) gets cited frequently as an offset. Some estimates claim up to 30% energy reduction. In practice, for bipedal robots, that number is generous. While EVs recover energy from sustained highway deceleration, Humanoid robots decelerate in micro-bursts: the swing phase of each step, brief postural corrections, stops between tasks. In realistic testing, gains are closer to just 10–15%. That’s nice to have, of course, but not a structural solution.

Going Under the Hood: Chemistry and Its Limits

All lithium-ion batteries work the same way at a basic level: lithium ions shuttle between two electrodes through an electrolyte. Charge pushes them one direction, discharge pushes them back. Every cycle leaves damage. How fast that damage accumulates, and what ultimately kills the cell, depends on which battery chemistry the manufacturer chose.

There are three commercially mature lithium-ion architectures competing for the humanoid market. And there is a fourth architecture, not yet commercial, that the industry is banking on.

First, how batteries degrade

Two degradation mechanisms matter most for humanoid applications.

The first is SEI growth. A thin film called the solid electrolyte interphase forms on the anode surface during the first charge. It stabilizes the electrode, but it never stops thickening. Each cycle adds to it, consuming active lithium and raising internal resistance. At 25°C and moderate discharge, this process is slow and predictable, which is why lab-rated cycle counts look favorable. At elevated temperatures and heavy current draw, it accelerates significantly. This is the primary driver of gradual capacity fade.

The second is lithium plating. When current is pushed into the anode faster than the graphite can absorb it (e.g. during fast charging, or at low temperatures) metallic lithium deposits directly on the surface instead of intercalating into the crystal structure. That lithium is permanently lost. In severe cases it forms dendrites that can puncture the separator. Research published in Batteries found that at −5°C and 3C charging, lithium plating caused 17% capacity loss after just 50 cycles, versus 6% at 1C under the same conditions.

These mechanisms compound rather than add. High temperature accelerates SEI growth, which raises resistance, which generates more heat. Fast charging causes plating, which creates localized hotspots, which accelerates SEI growth in those regions, making the next fast charge more damaging. At 4C and 5C discharge rates, i.e. the range humanoids operate in during balance-recovery transients, electrode structure decay accelerates measurably. Cells can fade gradually for months, then drop sharply when accumulated damage crosses a threshold. The nonlinearity is where deployment models break.

And the thermal environment inside a humanoid makes all of this worse. In an EV, active thermal management keeps cells at 35–40°C. In a humanoid torso, with batteries packed among heat-generating actuators with far less space for cooling, internal cell temperatures can spike to 80–100°C during peaks. J. Moyer, an EV battery engineer we spoke with, noted that this difference alone may explain why automotive packs achieve an order of magnitude more cycles than early humanoid deployments. Every 10°C above 25°C roughly doubles the rate of parasitic side reactions.

The material tradeoffs

Each lithium-ion architecture handles these failure modes differently.

NMC (Nickel Manganese Cobalt) is the default for most humanoid platforms. It has the highest energy density of commercially mature chemistries, which translates directly to runtime per kilogram of pack weight. The tradeoff: NMC is the most sensitive to everything that matters in a humanoid deployment, like high temperatures, deep discharges, fast charging. Lab-rated cycle life runs 800–1,500. Under real humanoid conditions, that falls to 200–500. Lab data from IOP Science confirms the pattern: NMC cells retain roughly 9% less capacity than LFP at the same cycle count. R. Osipov, an LG EV battery engineer we spoke with, confirmed that 3–5C discharge is “quite aggressive” compared to typical automotive applications.

LFP (Lithium Iron Phosphate) is the durable alternative. It’s cheaper, thermally stable, and tolerant of abusive cycling. Cycle life: 2,000–4,000 under standard conditions. But LFP packs weigh 20–30% more for the same capacity. For a humanoid, that weight penalty either cuts runtime or forces actuators to work harder, partially offsetting the longevity gain. LFP also loses significant capacity below 0°C, limiting cold-storage or outdoor deployments. For a climate-controlled warehouse, it’s a credible choice. For a general-purpose platform, this is harder to justify.

LTO (Lithium Titanate) is nearly indestructible, with cycle life exceeding 10,000 under some conditions, and excellent tolerance for fast charging and high discharge. The problem is energy density: roughly half that of NMC. A humanoid on LTO needs a pack twice as large for the same operating window, or accepts runtimes around an hour. Practical only for narrow applications with short duty cycles (inspection, surgical environments) not general warehouse work.

Solid-state replaces the liquid electrolyte entirely with a solid one. This is a genuinely different architecture, not an incremental improvement, and it theoretically unlocks higher energy density, better cycle life, and improved safety simultaneously. Toyota is targeting commercial deployment in vehicles by 2027–2028. QuantumScape expects initial mass production around the same window. BYD plans vehicle installation by 2027 and large-scale production by 2030. But these are automotive timelines, i.e. large-format cells for EVs. Humanoid-specific form factors are 2029 to 2030 at earliest, on a schedule that has already slipped. And there’s an open question about whether solid electrolytes can handle the repeated high-C-rate pulses a humanoid demands. Solid-state may solve the energy density constraint while leaving the power density problem intact.

T. Lai, a battery engineer we interviewed, put it plainly: “It mostly depends on the battery chemistry. Some will hold up very well, some won’t, but all chemistries involve tradeoffs.” On the NMC question specifically: doing 3–5C regularly “does a lot of degradation.” But he pointed toward what good engineering looks like: “Someone could design around it, whether by oversizing a battery or using a better cell chemistry for high C-rates and just minimizing active run time.” You get a sense of the tradeoffs humanoid designers actually face.

The Economics of Getting This Right

The standard deployment model for a 100-robot warehouse fleet assumes battery replacement every 12–18 months at roughly $1,100 per unit. Under realistic degradation, replacement frequency jumps to every 3 to 5 months. Each robot needs 8 to 9 batteries over five years instead of 3 to 4. Total battery costs roughly double.

But the bigger number is downtime. A non-hot-swappable battery swap takes 4 to 6 hours: diagnosis, cooldown, disassembly, replacement, recalibration. For a 100-robot fleet cycling through replacements on a compressed schedule, that’s 1,000 to 2,400 hours of lost productive time per year. At $200/hour opportunity cost, the difference between a 6-hour manual swap and a 3-minute autonomous swap is roughly $300K per year for a fleet that size.

This is a hefty cost surprise. Operators who budgeted for one replacement schedule discover they’re on another.

There are really only two responses to the battery problem: swap faster, or store more. The companies making progress have picked one and designed the entire system around it.

Software helps at the margins. Better BMS algorithms and conservative charging protocols contribute 20–30% cycle life improvement. Thermal management adds 15–20%. Software optimization gets another 25–35%. Combined best case: roughly 2×, from 200 cycles to 400. But closing the gap to EV-grade longevity requires closer to 5×. The remaining distance is an architecture problem, not a software one. Which is why the interesting moves are structural.

Swap faster. UBTECH’s Walker S2 is the clearest example. Rather than building a better battery, UBTECH made the battery disposable and engineered the swap. Walker S2 uses a dual-battery system: when one pack nears depletion, the robot navigates autonomously to a charging pod, removes the depleted battery, docks it, and installs a fresh module in under three minutes. No shutdown, no human intervention. A backup battery maintains continuity throughout. Compare 3-minute swaps to 90-minute charging sessions and the math changes completely. A factory running three shifts goes from losing hours of productive capacity per day to losing minutes.

Walker S2 is in mass production and deployed on automotive assembly lines across China. Its parent company UBTECH announced over $112 million in confirmed orders by November 2025, with deployments running at Nio, BYD, Zeekr, and Foxconn. Boston Dynamics made the same bet: the new Atlas was built with self-swappable batteries as a core design decision, not a retrofit.

Store more. Figure took the opposite approach with its F.03 battery. Rather than solving swap speed, Figure embedded the battery as a load-bearing component of the torso. This recovers weight and volume that would otherwise be wasted on a separate enclosure, pushing the F.03 to 2.3 kWh and 5 hours of runtime,with a 94% increase in energy density over previous generations. Active cooling is integrated into the die casting. This doesn’t fix cycle life directly, but every percentage point of energy density you gain extends the window before degradation forces a replacement. Figure 02 logged over 1,250 operational hours at BMW Spartanburg, including a 20-hour continuous shift. Agility’s Digit takes a softer version of the same path. Autonomous docking for between-task charging, battery life extended to 4 hours, a 2-to-1 work-to-charge ratio targeting 4-to-1 and eventually 10-to-1.

The gap between 800 cycles and 200 cycles is not a rounding error. It’s the difference between a battery that lasts months instead of years.

While solid-state batteries may eventually change this calculus (when it gets here?).

Until then, the robots shipping into warehouses today run on the chemistry we have, and the engineering has to be good enough to make that work.

The authors spoke with battery engineers from LG, EV OEM suppliers, and independent research settings. All opinions are the authors’ own.

Sources: Bain & Company Technology Report 2025, Goldman Sachs humanoid robotics research (2024–2025), IEEE Spectrum (October 2025), IOP Science, Journal of Power Sources, Batteries (MDPI), Figure AI, Agility Robotics, UBTECH Robotics, company filings and press releases.

Joy Yang is a Strange Research Fellow. She is pursuing computer science and government at Oxford, and is a researcher with its Visual Geometry Group. She was previously an intern with OpenAI and Google.

Mason Rodriguez Rand is a Strange Research Fellow who holds degrees in molecular and mechanical engineering (UChicago, UC Berkeley) and has led engineering and go-to-market efforts at climate and energy startups out of Argonne National Lab and UC Berkeley, spanning nuclear, carbon removal, and advanced materials.

Here’s the news that mattered this week

GTC 2026 Highlights

Jensen Huang's two-hour keynote centered on NVIDIA's transition from chip vendor to full-stack AI infrastructure platform. Three announcements stood out.

• The Groq Acquisition Pays Off Immediately. Three months after a $20B acqui-hire, NVIDIA debuted the Groq 3 LPU, an SRAM-based inference accelerator that sits alongside Rubin GPUs in rack-scale deployments. 150 TB/s memory bandwidth versus 22 TB/s on Rubin’s HBM4. Huang suggested up to 25% of cluster compute could be Groq silicon. NVIDIA killed its own Rubin CPX product to make room. The inference economy now has dedicated hardware, and NVIDIA owns both sides of the training-inference split.

• $1 Trillion Through 2027. Huang doubled last year’s $500B forecast, projecting $1 trillion in cumulative Blackwell and Vera Rubin orders through 2027. Goldman maintained a Buy rating, noting the guidance directly counters the “peak capex in 2026” thesis weighing on AI infrastructure names.

• Runway Previews Real-Time Video Generation on Vera Rubin. Runway and NVIDIA demonstrated a new video model running on Vera Rubin hardware with time-to-first-frame under 100ms for HD video. The model feeds into Runway's General World Model (GWM-1) research. (Runway)

  Thanks for reading The Strange Review! Subscribe to stay ahead

DoorDash Launches Tasks, Turns 8M Dashers Into Physical AI Data Collectors

What it is: DoorDash launched a standalone Tasks app paying couriers to film household chores, record multilingual speech, and capture real-world environments to train AI and robotics models. Partners span retail, insurance, hospitality, and tech. Over 2 million tasks completed since 2024. (TechCrunch, Bloomberg)

Why it matters: DoorDash just entered the physical AI data business with 8 million distributed workers already dispatched to real-world locations. Scale AI built a multibillion-dollar company on remote data labeling. DoorDash arrives with in-person collection at a distribution scale that might be hard for any data vendor to match.

Claude Code Channels: Agentic Coding From Your Phone

What it is: Anthropic shipped Claude Code Channels, allowing developers to control Claude Code sessions through Telegram and Discord via MCP. You can now monitor, prompt, and steer persistent coding agents from your phone.

Why it matters: Agentic coding has been tethered to the terminal. Channels makes it asynchronous and mobile, which changes the usage pattern from “sit down and code” to “delegate and check in.” This is their direct response to the runway success of OpenClaw.

Google Ships Vibe Coding and Vibe Design in the Same Week

What it is: Google upgraded AI Studio into a unified full-stack development platform, combining its Antigravity coding agent with Firebase backends, secret management, and one-click deployment to Cloud Run. Separately, it shipped a major Stitch redesign: AI-native infinite canvas, voice interaction, instant prototyping, and export to Figma and HTML/CSS. On this news, Figma shares dropped 4%. (Google Blog, Google Blog, SiliconANGLE)

Why it matters: Google now covers design, code, and deployment in a single ecosystem, all free at launch. When the model provider owns the full stack and bundles the tooling, standalone players in both vibe coding (Replit, Bolt, Lovable) and design (Figma) lose pricing power. Developer tools are a distribution game, and the hyperscalers have distribution locked in.

V-JEPA 2.1: LeCun’s World Model Architecture Posts New Robotics Benchmarks

What it is: Yann LeCun and collaborators (several now at AMI Labs) released V-JEPA 2.1, the latest version of the JEPA video model. It achieves state-of-the-art on action anticipation and object tracking benchmarks and posts a 20% improvement in real-robot grasping success over its predecessor. (arXiv)

Why it matters: This is the first concrete technical signal from LeCun’s camp since AMI Labs raised $1B on the JEPA thesis two weeks ago. The robotics results in particular matter: the model learns manipulation tasks from just 62 hours of unlabeled robot video, with no task-specific training or reward. If JEPA architectures can generalize physical skills from small data, the capital advantage of massive GPU clusters shrinks and the value of proprietary physical data (see: Mind Robotics) grows.

Mind Robotics Raises $500M Series A on Rivian Factory Data

What it is: Mind Robotics, a Rivian spin-off, closed a $500M Series A led by Accel and a16z. The company trains industrial robots using Rivian’s proprietary factory sensor data and custom silicon.

Why it matters: Industrial incumbents are discovering their operational data (the physics of how things move, break, and assemble) is as or more valuable than their products. This creates a new category of “data-rich” robotics startups where the moat isn’t hardware design, it’s access to high-fidelity physical interaction data. We expect more spin-outs from automakers and heavy manufacturers.

Broadcom Ships 400G Optical DSP at OFC 2026

What it is: Broadcom debuted Taurus at OFC 2026, the first 400G-per-lane optical DSP, enabling 1.6T and 3.2T transceivers purpose-built for the GPU clusters announced at GTC.

Why it matters: Compute is scaling faster than the network connecting it. As NVIDIA moves to rack-scale systems with tens of thousands of dies, the interconnect becomes the binding constraint. Broadcom is positioning as the chokepoint for all distributed AI training.

Deep Dive From The Review

NVIDIA’s latest AI rack, Vera Rubin, produces the heat of 160 homes. The next generation will double that. The one after will likely double it again.

The industry is racing to solve the heat problem, from subsea data centers to launching servers into orbit. But the most likely next step is the least exotic: liquid cooling.

The catch? The hardware is the easy part. The real cost is operational. It rewires how facilities are built, staffed, diagnosed, and run.

Our latest by Strange Research Fellow Rahul Narula explores what changes, and where the opportunity sits.

The Strange Review

The Liquid Revolution: Inside the Racks That Can Heat 160 Homes

In this week’s GTC keynote, NVIDIA announced the deployment of its impressive Vera Rubin NVL72, which ships in H2 2026. It packs 72 Rubin GPUs, 36 Vera CPUs, and more in a single liquid-cooled rack, and its power consumption can exceed 200 kW…

Read more

2 months ago · 5 likes · Rahul Narula

Strange Signals: Data of the Week

The numbers are adding up fast. In a survey of tech CEOs released this week, 66% reported they no longer plan to backfill roles lost to voluntary attrition. “Replacing departing staff with AI agents” entered the Top 5 strategic priorities for the first time. (SaaStr)

The layoff data supports it. Block cut 4,000 employees in February (40% of headcount). Atlassian cut 1,600 on March 11 (10% of staff, over 900 in R&D). Meta is reportedly planning to cut up to 20% of its 79,000-person workforce, roughly 15,000 roles, to offset $135B in AI capex. (TechCrunch, CNBC)

HR agency Challenger, Gray & Christmas’ data puts it in context: 12,304 job cuts have been explicitly attributed to AI through February 2026, 8% of all announced cuts. That’s up from 5% for the full year of 2025 and 3% since tracking began in 2023. Tech sector cuts are up 51% year over year. Meanwhile, announced hiring plans are down 56% compared to the same period last year, the lowest since tracking began in 2009. (Challenger)

We think companies might be entering a “low-hire, low-fire” era where headcounts shrink through unreplaced attrition and targeted restructuring rather than headline layoffs. Enterprise budgets are being redirected from headcount to AI tooling.

To put that in context: an average U.S. home draws about 1.25 kW. One Vera Rubin rack consumes 600kW, and nearly all of it becomes heat - the equivalent of roughly 160 homes. Remove that heat, or the chips throttle, and the $3 to $4 million machine becomes a very heavy paperweight.

The next rack architecture after Rubin, Feynman, is named for 2028, and NVIDIA’s rack power roadmap points toward megawatt-class densities. Each generation pushes rack power significantly higher, and the cooling problem compounds with it. Once racks reach a certain power draw, air, which carries roughly 3,500 times less heat per unit volume than water, becomes infeasible.

The industry is racing to solve the heat problem, from subsea data centers to launching servers into orbit. But the most likely answer is: liquid cooling.

Interestingly, the hardware is a fraction of liquid cooling’s real cost. The bulk is in operations, like the mean time-to-repair when something breaks in a system that has failure modes most facilities teams have never seen.

Liquid cooling is not new; data centers have used chilled-water loops for decades. What is new is the intensity of energy required. A traditional air-cooled rack runs at 10 to 15 kW, and aggressive airflow engineering can push that up to about 40 kW. But the Vera Rubin can exceed 200 kW, and at those levels, air stops being a practical medium for heat removal. The bottleneck becomes the ability to move heat out of the system.

Modern deployments solve this by moving the cooling problem directly onto the hardware. In direct-to-chip systems, coolant flows through cold plates mounted on CPUs and GPUs, absorbs heat, and carries it out before transferring it to the facility loop or an external rejection system. Immersion systems go further, submerging servers in dielectric fluid that absorbs heat directly, and in some cases boils and condenses in a continuous cycle. Both approaches eliminate fans and support far higher densities, but they introduce different operational tradeoffs. Direct-to-chip systems are generally easier to service, while immersion systems can handle higher thermal loads at the cost of added complexity.

This is not a continuation of earlier liquid systems. Older deployments were built around lower densities and looser assumptions. A chilled-water mainframe room is not comparable to a 200 kW rack, much less to a warm-water system on a path toward megawatt-scale densities. Today’s systems are hotter and more operationally demanding, and even facilities that have used liquid cooling before often lack the controls and service model required at these densities.

The shift to warm-water systems makes that break more explicit. Vera Rubin’s 45°C inlet specification departs from legacy chilled-water loops and even from current direct-to-chip systems. At the chip level, there is still thermal margin. At the system level, the design basis changes. In many climates, heat can be rejected directly to ambient air using dry coolers, reducing dependence on chillers and evaporative systems. This lowers capital cost, reduces energy consumption, and cuts water use, but it also raises the system’s baseline temperature, narrowing the time available to diagnose and respond when something goes wrong.

Adoption: Early and Uneven

Most data centers still rely on air, but the shift to liquid is underway. Industry projections suggest liquid cooling will be installed in 40% of sites by the end of 2026. Hyperscalers and AI-native operators like Microsoft, Google, Amazon, and Meta are standardizing on it, while most enterprise and colocation providers remain one to two hardware generations behind.

New AI halls are being designed around liquid from day one. Existing facilities are retrofitting what they can, carving out high-density pods while keeping the rest of the campus air-cooled. Retrofit is possible, but extremely costly. Some sites will be upgraded in phases. Some will be reserved for lower-density inference. Some will simply be repriced as they fall out of the frontier tier.

This creates a fragmented operating environment: mixed cooling architectures, uneven power distribution, and systems running at very different density and thermal profiles within the same facility. Liquid solves the heat problem, but it introduces a different kind of complexity. Pumps, manifolds, valves, sensors, fluid chemistry, and safety procedures become part of the operating stack. As the number of components increases, the number of failure modes increases alongside it.

This is the hidden cost of the transition: hydraulic debt. Operational complexity accumulates quietly until it surfaces as downtime.

The Operational Opportunity

A pressure drop or flow anomaly can point to several different root causes at once. It might be a micro-leak at a fitting, a degrading pump, drifting telemetry, or slow corrosion inside the loop. In an air-cooled room, a temperature hot spot can suggest where to look. In a liquid-cooled rack, the signal is even more ambiguous.

Not every repair is complicated, i.e., the wrench time may only be ten minutes. But the localization time can be forty, and that gap is where the cost sits. In warm-water systems, the problem is amplified. Higher baseline temperatures compress the operational margin for diagnosing and responding to faults.

The usual first response to this complexity is monitoring: more sensors, better dashboards, earlier alerts. Monitoring is necessary, but a dashboard only surfaces symptoms. It can show that the temperature is rising on rack 14 or that the flow has dropped below a threshold. What it cannot do, in most implementations, is determine whether the underlying cause is a flow restriction, a failing pump, an incorrect valve state, or bad telemetry, nor can it recommend a safe intervention with high confidence.

At scale, this distinction becomes decisive, and it shifts the bottleneck from visibility to attribution and response.

The more durable opportunity sits one layer deeper: systems that take conflicting telemetry, reason about how the hydraulic loop is actually behaving, and produce a specific diagnosis (i.e. the failing component, the likely failure mode, and a confidence level) along with a recommended procedure and an audit trail. The output is a decision-support package that an operator can act on immediately, or that an automated system can execute within pre-approved safety boundaries.

NVIDIA is already designing toward this. Its GTC framing was explicit: AI infrastructure has crossed the point where human operators can manage it directly, signal volume exceeds human capacity, manual triage becomes a bottleneck, and system knowledge remains fragmented across individuals. The shift is from human-in-the-loop to human-on-the-loop. Agents handle detection, correlation, diagnosis, and remediation within guardrails, while operators move into system design and governance.

I believe the transition to autonomous operations will be staged. Today’s systems are assistive. The next phase is agent-driven: systems that localize faults, explain their reasoning, and recommend actions. Beyond that is bounded autonomy: systems that isolate a segment, notify the scheduler, and shift workloads. Full end-to-end autonomy remains further out, but the trajectory is clear.

In some environments, including edge deployments, sovereign AI sites, and early orbital systems, there may be no technician on site at all. In those settings, the shift from human-in-the-loop to human-on-the-loop is not an optimization but a requirement.

Where We Might Be Headed

The facility becomes the product. When a rack draws 200 kW and the roadmap points toward 1 MW within a few years, the data center is no longer generic real estate. Power delivery, cooling architecture, and structural engineering become differentiating capabilities.

The energy model inverts. At 45°C inlet water, the cooling plant simplifies: dry coolers instead of continuous chiller operation, closed loops instead of evaporative towers, waste heat at temperatures useful for district heating. Cooling energy as a share of total facility power drops, but operational sophistication has to increase. Pump reliability, flow management, filtration, and leak response all become more critical when the system is running warmer.

The cadence is relentless. NVIDIA shipped GB200 NVL72 in 2024. GB300 NVL72 in 2025. Vera Rubin NVL72 ships H2 2026. Feynman is named for 2028. Each generation notably increases the power density, and the cooling infrastructure now has to keep pace with an annual hardware refresh cycle.

The frontier rack already assumes liquid cooling. The harder problem, one I believe the market has not yet priced, is in operating it.

Rahul Narula is a Strange Research Fellow and an MBA candidate at Harvard Business School. He previously worked in product operations at Cruise, where he helped scale robotaxi fleet operations, and has a background in data science and operations across data centers.

I am looking for someone (a unicorn!) to build the operational layer underneath all of it so I can stay focused on research, relationships, and investments.

You would work directly with me to keep three engines running: events, community, and publishing. You are not managing a team. You are the team, working alongside me and a growing set of AI tools and agents.

Type: Fractional, 5 to 10 hours/week

Location: San Francisco

Compensation: Hourly

The Role

Events

We host invitation-only gatherings that bring together deep tech founders, enterprise technology leaders, and investors. These range from intimate 12-person dinners to a flagship annual summit of 60 to 80 people.

You would own event logistics end-to-end: invite lists and outreach, coordinate with co-hosts and partners, handle day-of execution, and build repeatable systems so each event does not start from scratch.

Publishing

The platform publishes a research newsletter, periodic deep dives on emerging technology sectors, and investor communications.

You would manage the editorial calendar and publishing pipeline. You would handle production: formatting, scheduling, distribution.

Community

The platform maintains a network of technical founders, enterprise operators, researchers, and investors. Our ecosystem is the connective tissue between the events and the research.

You would coordinate community programs, and identify opportunities to connect people across the network.

The Scope: Working with AI and Humans

This role sits at the intersection of human relationships and AI systems. On any given day you might be coordinating speakers for an upcoming event, writing a personal outreach email to a founder, and building an automated workflow for event distribution.

We already use AI agents extensively for workflow automation. You would work inside these systems daily, reviewing outputs, improving prompts, managing projects, and helping decide what should be automated versus what requires human judgment and personal touch.

The best version of this role is someone who knows that a warm introduction requires a human and that a research synthesis can start with an agent, and who can move between both modes without friction. I am not looking for someone who is “open to AI.” I am looking for someone who already works this way and can help us get better at it.

Who You Are

You should have:

• Experience working directly with a founder, GP, or senior executive in a high-autonomy, low-structure environment

• A track record of shipping: events that happened, publications that went out, systems that worked.

• Comfort with AI tools as part of your daily workflow, not as experiments. You use them. You have opinions about them.

• Strong written communication. Everything you produce, from emails to briefs to event copy, should be clear and sharp.

• The ability to context-switch between logistics and strategy without losing the thread on either.

• Low-ego, high integrity.

Bonus:

• Familiarity with venture capital, deep tech, or the startup ecosystem

• Experience producing editorial content or managing a publication

• Experience planning and executing events, especially curated, invite-only formats

This is not the right fit if:

• You need detailed instructions to start moving. This is a figure-it-out role.

• You see AI as something happening to other people’s jobs, not something you actively build with.

• You want a path to a full-time executive position. This is designed to be fractional and flexible.

• You are stronger at planning than executing. We need the thing done, not the plan for the thing.

How to Apply

Submit your resume and answers to a few questions here. We review applications in batches and will respond within two weeks.

We value diverse perspectives and encourage candidates from all backgrounds to apply.

It might be a small fraction of the system, but it harbors a disproportionate share of the cost and failure. It is the actual bottleneck holding us back from the last mile of robotic utility.

Recent narratives around humanoids have largely focused on robotic foundation models, reasoning, locomotion, aesthetics. But zoom out a little and the picture shifts. It is the robotic hand that’s become the gating constraint for humanoid robots, and I believe that the companies that solve it will define the next phase of the industry.

Hands Are Miserable To Build

The hand causes the highest difficulty for humanoids due to three high-impact, compounding problems: cost, complexity, and failure density.

According to Morgan Stanley’s analysis of Tesla Optimus Gen-2, hands account for approximately 17% of the total bill of materials cost (roughly $9,500 out of a $50,000–60,000 unit). That’s a disproportionate share for a subsystem that represents a small fraction of robot mass.

The complexity is equally stark. Optimus Gen 3’s hand has 22 active degrees of freedom, compared to human hands at 27 DoF. Each leg joint chain has 6 to 7 DoF, meaning the hand alone has 3 to 4x higher control dimensionality per limb. The Gen 3 design relocates actuators into the forearm via tendon-driven cables, mimicking human anatomy but adding mechanical complexity and failure modes.

This shows up directly in performance. While humanoid robots achieve nearly 100% success rates grasping simple objects like apples and tennis balls, success rates plummet to around 30% for complex items such as spoons, screwdrivers, or scissors.

Actuation and Sensing

The hand bottleneck is actually two bottlenecks stacked on top of each other. They have different root causes, different solution paths, and potentially different investable layers.

The actuation problem is mechanical. More than 70% of humanoid robot hands rely on tendon-driven or hybrid actuation, while most of the rest of the robot uses simpler rigid motors. The reason: rigid actuators break down the moment contact gets unpredictable, when objects slip, deform, or change shape. Tendons offer compliance but add complexity and failure modes.

The sensing problem is informational. Vision sensors (RGB, RGB-D, LiDAR) scale linearly with cost and compute. Tactile sensing does not. In the MERPHI hand survey, fewer than 50% of humanoid-native hands have fully integrated tactile sensing, and those that do typically sacrifice payload, speed, and / or manufacturability.

This is why recent platforms have been quietly reallocating budget toward the hand subsystem: palm cameras (Figure-03), fingertip tactile arrays (Helix 02), tendon-driven compliance (Optimus). These look like quirky upgrades. They’re actually attempts to stabilize the highest-failure subsystem of the robot.

This shows up directly in deployment failure rates.

A Morgan Stanley ecosystem analysis of humanoid robots shows that dexterity and fine manipulation are the top two bottlenecks to scaling, alongside power. Walking isn’t even considered a gating issue for pilot deployments anymore. Real-world use cases like logistics and manufacturing are overwhelmingly restricting humanoid usage to navigation and pick-and-place tasks.

Contrary to popular belief, it’s not because AI reasoning is limited either(see Dexterous Manipulation Benchmarks and RoboManipBaselines).

The same conclusion is reached when we look specifically at Chinese humanoid companies, who, while arguably ahead, have not solved the hand bottleneck either. This might be why almost all publicly released demos of their humanoids are of full-body athletic tasks with little hand-specific manipulation.

UBTech’s Walker S2 represents the max-DoF approach: 52 total degrees of freedom, 15 DoF per hand, Gen-4 bionic hands, continuous 24/7 operation via autonomous battery swapping. Impressive specs, right? Despite all this, Walker’s public demos and deployment targets emphasize logistics handling and inspection only. They don’t touch deformable or tool-based manipulation because even at 52 DoF, hand reliability remains the limiting factor.

Unitree has the opposite problem. They’ve achieved locomotion performance comparable to Boston Dynamics at orders of magnitude lower cost, with entry-level humanoids priced as low as $6,000. But hand payloads are limited to 2 to 3 kg, active DoF are reduced, and task space is severely constrained despite extremely capable motion stacks.

Fourier Intelligence’s GR-1, China’s first full-scale humanoid, targets rehabilitation and assisted manipulation, prioritizing safe grasping. They avoid high-speed or high-force manipulation altogether. That’s because they too fail in the department of controlled contact.

If hands are the bottleneck, who captures the value from solving them?

One possibility is that hands remain vertically integrated. Each humanoid company develops proprietary solutions, and hand quality becomes a differentiator baked into the full-stack robot. This is the current default. Tesla, Figure, UBTech, and others are all developing hands in-house.

But vertical integration has limits. The actuation and sensing problems are deep enough that specialist suppliers may emerge, like companies focused purely on tendon-driven mechanisms, high-density tactile arrays, or compliant gripper architectures.

The question is whether the hand subsystem is modular enough to support that kind of layer separation, or whether integration with the rest of the robot’s control stack makes standalone “hand companies” unviable.

We don’t have a strong view yet. But the companies betting hardest on solving hands, whether through actuation innovation, sensing density, or both, are the ones to watch.

Joy Yang is a Strange Research Fellow. She is pursuing computer science and government at Oxford, and is a researcher with its Visual Geometry Group. She was previously an intern with OpenAI and Google.