AI power is being underwritten like infrastructure. CleanSpark closed $2.276B of senior secured notes, following its 20-year, 175 MW lease with an unnamed investment-grade tenant worth $6.6B. TeraWulf signed a similar 20-year, 401 MW lease with Anthropic worth $19B. This matters because the credit belongs to the tenant while the scarce asset is the energized site, which is why former bitcoin miners now borrow like utilities. We think the opportunity sits between power and compute: project-level financing, and capacity for tenants below investment grade that still need megawatts.

Compute futures hit a regulatory speed bump. The CFTC extended its review of CME’s proposed H100 and B200 rental index futures to November 9, citing novel or complex issues. A separate request for comment asks about cash-market liquidity, manipulation and customer protection. This matters because compute can’t be hedged or financed at scale until there’s a price people trust. We think the opening is index integrity: neutral, transaction-based benchmarks and verifiable settlement underneath any forward curve, GPU-backed loan or onchain compute market.

OpenAI is building a chip-design model with Synopsys. GPT-Synopsys will operate Synopsys’s EDA tools directly under a multi-year revenue-sharing deal, without training on customer designs. The same day, Synopsys signed a $1B-plus IP deal with Amazon that also covers its AI tools. This matters because a frontier lab now sees chip design as worth its own model, and chose to enter through the incumbent. We think a model trained to drive one vendor’s tools learns the tools, not the judgment of the designers using them. That leaves room for an independent layer built on real design decisions that works across every vendor’s flow.

Robots can do the work, just not cheaply enough. Anthropic estimates robots could perform 74% of physical job tasks, but only 0.3% are cost-competitive with human labor today. Training is getting cheaper: Magic-W0, a new world-action model, learns from human video, real-robot and simulation data and adapts to real tasks with little fine-tuning. This matters because physical AI’s constraint has moved from capability to unit economics. We think the nearer-term winners are cost-curve bets: cheaper hardware, structured environments, and the financing and integration layers that make deployment pencil out.

Stablecoins are becoming B2B payment rails. Allium counts $401-527B in stablecoin payments through August, up 42-63% year over year, with business-to-business the largest lane at $137-153B. Supply grew only 6%. This matters because volume, not supply, is now the story: stablecoins are settling real corporate flows. We think the next opportunity is credit on top of the rails: short-duration working capital for payment companies and their suppliers, financed onchain, where repayment data already sits in the transaction history.

We’ll share another edition next week.

Amazon blocked Meta’s shopping agent, and its rivals rushed to plug in. Meta launched Muse on September 8, a personal agent that shops, books, and pays from its own cloud computer, and it went straight to the top of the App Store. Amazon blocked it, but Walmart, Shopify, PayPal, and Expedia integrated. This matters because an agent never sees a sponsored listing, and Amazon made more than $68 billion from ads last year. We think this is the playbook in every category: incumbents block agents, challengers embrace them to take share, and incumbents come around later on worse terms. Most exposed are the businesses that monetize inertia, like subscriptions nobody cancels and insurance nobody reprices.

BlackRock says agents will pay in stablecoins. The biggest agent launch of the year pays with cards. BlackRock published The Machine-Native Economy this week, arguing that agents need programmable, always-on money and naming stablecoins, Ethereum, and Circle’s Arc as the likely rails. Meanwhile, Muse checks out through Stripe’s Link wallet, using a saved card or a single-use virtual one. This matters because consumer agents inherit the rails of the person they act for, and that person already has a card on file. We think stablecoins win where there’s no human account behind the transaction at all: agents paying APIs, other agents, and compute per call, across borders and in amounts too small for card economics.

950 Claude agents found a new CRISPR-like enzyme system in 21 hours. Anthropic’s new life sciences lab pointed Claude at a DNA database with a single prompt. Roughly 950 agents sorted more than 200,000 enzymes into 20 written reports, and one spotted a repeating DNA pattern beside an unusual enzyme that nobody had described, reminiscent of CRISPR. This matters because genome mining that used to take an expert weeks to months now takes a day of compute, and hypotheses are becoming abundant. We think the bottleneck in science is moving from generating ideas to verifying them, and the durable value sits with whoever owns the validation loop: wet labs, automated experimentation, and the taste to decide what’s worth testing.

Robots are generating data nobody can store. Felicis argues the physical world has no data stack. A single autonomous vehicle can produce terabytes a day of synchronized video, LiDAR, and telemetry, and most of it gets sampled down or deleted, because today’s tools were built for rows and columns. Foxglove, whose open MCAP format is the default logger in ROS 2, is the early leader. This matters because every fleet that scales becomes a data infrastructure customer, and every robot foundation model needs that data to improve. We think the pre-seed openings are at the edges: deciding on the robot what’s worth keeping, and making physical-world data portable enough to share and sell, because data that can’t move can’t compound.

DeFi built the right tools for the wrong assets. Multicoin’s new essay on DeFi 2.0 argues that order books, fixed-rate lending, and portfolio margin have existed onchain for years, but were built for volatile, perpetual cryptoassets that didn’t need them. Treasuries, equities, and credit have maturities, cash flows, and identifiable borrowers, which is exactly what those primitives were designed for. This matters because the RWA conversation has been stuck on tokenization, when the value is in what happens after an asset arrives. We think the best early bets are primitives whose market grows with the asset base, and fixed-term, underwritten credit for real borrowers is the one we care about most, because it’s what hardware and robotics credit needs to move onchain.

We’ll share another edition next week.

Three labs bet in the same week that teleoperation data is not the moat. Odyssey-3 controls arms, humanoids, cars, drones and video games from one pretrained world model with tens of hours of task-specific data, and its sim-trained driving policies traveled about 77% as far between safety interventions as policies trained on real footage. Reward AI’s OM-1 skips robot data entirely, learning from people wearing a capture glove and transferring zero-shot from tabletop arms to humanoids. Zeno-1 trains robots on each other, using four hours of closed-loop partner interaction to run decentralized multi-robot collaboration at 30Hz. This matters because much of the industry has spent three years assuming proprietary teleop data was a defensible asset, and three different technical routes just suggested otherwise. We think the moat moves down to deployment, meaning customer contracts, fleets in the field, and the service relationships that produce revenue rather than demos.

Hardware founders are multiplying. A founder can now describe a part and get usable CAD, have a board fabricated and assembled for <$100, buy a quadruped with a full SDK for $2,500, fine-tune an open-sourced policy instead of training one, and pull the first hundred units from an overseas supplier without leaving their desk. What previously took a year and $250K now takes a few weeks and a credit card. Prototyping was never what killed hardware companies; going from 10 units to 1,000 is, and that takes production tooling, safety certification, field service and maintenance, and working capital. We think cheap prototyping increases the number of companies that reach that wall rather than the number that clear it, which is precisely why we keep looking at the financing and insurance layers.

AI decisions are about to get too cheap to meter. Diogo Almeida, who worked on the RLHF research behind ChatGPT, came out of stealth with TypeSafe and a model called Jev that doesn’t generate text at all. It takes unstructured state and returns typed outputs with calibrated probabilities, at $0.042 per million input tokens with output free, in 70 to 500 milliseconds. The evals are self-reported and the reference answers come from frontier models, so discount the 200x claims accordingly. This matters because most production AI work isn’t conversation, it’s a fuzzy if-statement buried inside software, and nobody has been pricing it as its own primitive. We think the interesting question for founders is whether this is a category or a feature the labs absorb within a year.

Why do agents lie? After the OpenAI-Hugging Face incident, Yoshua Bengio published the clearest account we’ve read of why misaligned behavior keeps recurring: safety goals are vague, task goals are sharp, and a capable optimizer finds the reading of the vague one that lets it win the sharp one. His uncomfortable corollary is that better monitoring may just select for agents that cheat without getting caught. This matters because it reframes agent safety as a permanent operating cost rather than a bug to be patched. We think it’s the strongest case yet for the verification stack, meaning evals, runtime monitors, permissioning, audit trails, and eventually insurance, because once agents take actions with financial consequences, somebody has to underwrite those consequences.

Mapping the AI buildout, site by site. Epoch AI expanded its data center research to 86 sites covering 44% of global AI compute and 53% of everything deployed this year, with satellite imagery, capacity in H100-equivalents, chip types, capital costs, and construction status for each. Chinese coverage sits at 9-31%, which is its own story. This matters because the largest infrastructure buildout in history has been financed almost entirely on private information, and site-level data on capacity and construction is the beginning of a real underwriting dataset. We think you can’t price collateral you can’t see, and the lenders, insurers, and infrastructure debt markets now moving into compute all need to see it.

We’ll share another edition next week.

In most industries that would be strange. Trucking companies don’t sell equity to buy trucks, and farmers don’t sell stakes in the farm to buy tractors. Machines that earn contracted revenue get bought with debt, because a lender can underwrite the cash flows and take the machine back if they stop. Robotics startups are funding exactly that kind of asset with the most expensive capital they’ll ever raise.

Equity is buying the fleet

A commercial robot costs $20-100K. Under a signed service or lease contract, it earns a few thousand dollars a month and pays for itself in 12-18 months. Land a customer for 50 units and you need millions in hardware before revenue arrives. Most startups can only fund that with another equity round, taking permanent dilution to cover an 18-month gap. Robotics startups raised $18.8B globally by June, more than in all of 2025, and a lot of it is buying hardware. One of our portfolio companies told us they’d put $20M into their fleet tomorrow if they could get it, but raising it as equity at their stage isn’t realistic.

In theory, there are three ways to fund this without equity:

1. Equipment financing covers the machines used to build robots, like CNC machines, which have deep resale markets.

2. Fleet financing lends against robots a startup deploys itself under a signed contract. The lender is repaid from the contract and can repossess the robots if that fails.

3. Customer financing does the most. A lender pays the startup upfront and the customer repays over its contract. The startup gets cash on delivery to buy its next batch of parts, and the credit risk moves to the customer, who is often a much bigger company.

The parts themselves are the hardest to lend against. Motors and gears get customized, built into prototypes, or used up in R&D, and their resale value drops fast once that happens.

Why nobody lends

Banks want years of operating history and profits. Equipment lessors want two years of financials, which is unrealistic for a seed-stage company. Venture debt usually starts at Series B, comes with warrants, and takes a lien on the whole company, so the founder pledges everything to finance one fleet.

Bank capital rules make it worse, and GPUs show how. Under Basel III, a loan repaid from a specific asset’s cash flows counts as “object finance” and gets slotted into preset risk buckets that punish anything unproven. So banks lend to private credit funds instead, and the funds lend against the GPUs at higher rates.

USD.AI was built for that gap. It takes stablecoin deposits and lends them against tokenized GPUs. Its largest loan grew from $620K in July 2025 to $98.1M this June, with $202M deployed and ~$30M of annualized revenue in Q2. Deposits went the other way, from $650M+ at its January token launch to ~$225M by August. On-chain capital is easy to attract. Origination is the business.

Insurance is moving faster

Corgi is an AI-native insurance carrier for startups. Instead of acting as a broker, it writes and prices policies itself, using company data rather than long application forms. It was founded in 2024, reported $40M of annualized revenue at its January Series A, and reportedly raised at a $4B valuation in July, tracking toward a $450M run rate by year end. The lesson is that a carrier underwriting from data can win the categories incumbents handle badly, and robots are one of them.

Boop is already there. It sells per-robot insurance to companies deploying robots and includes a black-box recorder that settles claims on evidence. It argues standard liability policies likely exclude autonomous failures. Boop started out pitching insurance for household humanoids, and its pivot to commercial fleets shows where the demand is.

The missing piece

Robotics operators need financing and insurance at the same time, for the same robots. Both are underwritten on the same data: what the robot is, where it works, how much it’s used, and who’s paying for it.

Other machine industries solved this at the point of sale. When you buy a car, the dealer offers you a loan and insurance at the same desk. GM started its own lender, GMAC, in 1919 because customers couldn’t buy cars they couldn’t finance. John Deere’s lending arm holds ~$49B of loans today.

Robotics doesn’t have this yet. Robot makers are young and don’t lend to their customers. We think a startup will fill that gap by offering financing and insurance right where robots are bought.

That’s where crypto comes in. Stablecoin holders want real yield, and short, contract-backed loans against working robots fit that well. But as USD.AI shows, capital isn’t the bottleneck. The hard part is originating good loans.

We’ve been thinking about this all year and recently met founders already doing that work. If you’re building credit or insurance for the physical economy, or you’re a robotics founder spending your seed round on a fleet, we’d like to talk.

Cities are software products. CitySim gives LLM agents beliefs, long-term goals, habits and spatial memory, then asks them to generate their own schedules. In a Tokyo setting, the system reproduced patterns in Japanese time-use and travel data, while forecasting Shibuya crowding and venue popularity better than prior agent systems. The paper demonstrates realistic behavior at 1,000 agents and system scalability to one million. Urban planning, site selection and infrastructure design can become iterative software workflows.

AI will find more bugs than teams can fix. OpenAI classified GPT-6 Astra as Critical for cybersecurity after testing showed sharp advances in vulnerability discovery and exploit development. That means security teams will receive more potential issues than they can assess and resolve manually. We think winning products will identify which reports are real and urgent, generate patches, test them against production dependencies, and safely deploy or roll them back. We also think winning companies will pair autonomous remediation agents with human ownership and clear operational controls.

NVIDIA is making custom chips work inside its ecosystem. It invested $3.5B in MediaTek and will let MediaTek offer customers a way to connect custom AI accelerators to NVIDIA’s NVLink systems. Model labs and hyperscalers increasingly want silicon built around their specific workloads. But a chip alone is not a data center: it needs memory, networking, software, cooling and operational support that work together. This is a bullish NVIDIA story. Custom chips expand, but NVIDIA can still own the systems that deploy them at scale. Startups need a clear workload-specific advantage, not another generic accelerator.

The next social network may be a world, not a feed. Diiverge turns a single image into a persistent point-and-click adventure. Choose an object, decide what happens, and the scene continues as video. Each choice creates a branch that remains available to the next visitor. This makes generative video feel less like a format for watching and more like an environment people can shape together. The interesting products will not simply generate infinite content. They will give people reasons to return to worlds they helped create, explore the paths other people made, and build a shared history inside them.

Agents need identity before they can become economic actors. Long-running agents will hold credentials, purchase compute, call services, move money and build reputations over time. That requires persistent identity, clear permissions and accountability when something goes wrong. The debate around self-sovereign agents is still speculative, but the infrastructure problem is already real. An agent acting for a person, an organization or itself needs different rights and limits. We think agent identity becomes a core layer of the agentic economy, making autonomy legible to businesses, financial institutions and the systems agents will increasingly operate.

We’ll share another edition next week.

The underlying assumption here is that progress in robotics is predicated on step-function model breakthroughs. But models and task reliability are already improving consistently and scaling behavior looks real - this isn’t the bottleneck!

The real bottleneck is economics - we need cheap, globally scalable robots that can be put in a warehouse, hospital, or field and drive top-line improvements or efficiency gains. Robotics doesn’t need a GPT-3 moment. It needs a Model T moment.

Cost is the entry point

Cheap robots help make robotic deployments economically feasible and accessible. Robotic deployments create data, data improves autonomy, and autonomy lowers cost and increases value. We think this loop is the most important one in robotics.

This flies in the face of traditional hardware economics, where companies treat cost as something to address only at scale. History teaches us to first build the most capable version, sell it at the top of the market, and use that to subsidize the ramp, driving BOM cost down later with economies of scale.

Unfortunately, that doesn’t work in robotics, because you need deployments to generate the data that improves autonomy, and you need better autonomy to increase value-per-robot and robot-per-teleoperator ratio.

In other words, a robot that starts expensive never gets enough deployments to become cheap. It gets stuck at the top of the loop.

Robo

Today, our portfolio company Robo comes out of stealth. Robo builds affordable, general-purpose robots and teleoperation infrastructure.

Their first robot is Robo-T. Designed and built here in Los Angeles, it holds 10 pounds and runs tasks across warehouses, hospitals, food manufacturing, and logistics environments.

Over the past three months I’ve seen Robo go from robotic-effectors to robot arms to a full mobile embodiment, and relentlessly cut deployment costs by owning the full stack to just $10/hour. They also built Roboport, their own teleoperation platform, rather than renting one. It turns demonstrations, rollouts, and interventions into improving autonomy, helping close the cost-autonomy loop.

Robo’s founders, Don and Kyle, are high-integrity, high-velocity builders with a rare blend of technical depth, product intuition, and customer obsession. They pivoted into robotics midway through our diligence, and seeing their speed and pragmatism helped get us over the line at pre-seed.

Robo is now looking for fleet partners. Similar to operators of Tesla fleets for Uber or Lyft - Robo are looking for people who buy 15+ Robo-Ts, deployed around the clock, and sell the hours. Those fleets ended up providing much of the liquidity on Uber’s supply side, and turned out to be very good businesses for the people running them. Robo handles delivery, setup, and maintenance; what you need is organization and the communication skills to teach operators a task.

If you’re building here, or want to run a fleet, reach out to them directly. And if you’re working on anything at the intersection of cheap embodiment and real deployment, come talk to us at Canonical.

Over the next few months I spent a lot of time with Somi and Sandeep as the vision for Synthefy came together. Later that year, Canonical had written the first check as the only investor in their pre-seed, and we started meeting every week.

The idea they landed on was simple, driven by a core thesis:

Most of the data that runs the economy lives in tables and time series: transactions, prices, sensor readings, inventory, customer histories. Predictions made from that data decide which payments get approved, how products get priced and stocked, which customers leave, and when machines fail.

And the way companies make those predictions has barely changed in a decade. Every workflow is a custom project: clean the data, engineer features, train a model, tune it, deploy it, maintain it forever. The Synthefy team lived this at Uber’s self-driving unit, where dozens of people spent 5-6 months building a single model. A production system can cost $500K-$1.5M to build and up to $750K a year to keep running.

The team believed pretraining would do for numbers what it did for language: one model, trained on huge amounts of structured data, that makes accurate predictions on a table it has never seen. Six months of ML work becomes an API call.

The path from there looked like most good pre-seed paths, which is to say ‘quiet’. First MVP in early 2024, pointed at defense and telecom workloads. Deutsche Telekom named them best AI startup.

Through 2025 Synthefy stayed heads down: deployments at enterprises like GoodRx, Schneider Electric, and a US Army contract, while the rest of the market has been chasing text, images & videos through LLMs.

Then the market turned, fast. In May, SAP agreed to acquire Prior Labs, an 18-month-old German lab that had raised €9M, and committed over €1B to build on it. In June, Nvidia bought Kumo for a reported $400M. Two of the largest companies on earth paid up for the category within 30 days of each other.

Weeks later, Synthefy shipped Nori. It ranks #1 for accuracy on a public benchmark of 96 datasets, ahead of Google’s TabFM and the TabPFN model SAP had just bought and remarkably, it does it with just 6M parameters, a tenth the size of comparable models. It crossed 440K downloads in its first 30 days, and recently surpassed 630K downloads.

In pilots, it beat forecasting systems customers had spent years tuning, with no custom training.

Today Synthefy announces its seed round, led by Wing with Haystack, Canonical, Samsung Next, plus angels including Srinivas Narayanan (ex-CTO at OpenAI), Aparna Chennapragada (CPO at Microsoft), and Manohar Paluri (VP of AI at Meta).

Congrats, Somi and team. The economy’s numbers finally have their foundation model!

PS: One note on sourcing, because people ask how a team like ours finds Somi and Sandeep. The best pre-seed deals rarely look like “deals.” They look like research: no deck, no go-to-market plan, sometimes no graduation date. Our job is to be in the lab while the conversation is still “could this be a company,” and to be useful.

1. companies building deep infrastructure at the bottom of the stack that is hard to replicate, e.g. NVIDIA, AWS, Cloudflare, Snowflake

2. companies building applications with incredible distribution at the top of the stack, e.g. Facebook, Instagram, OpenAI, Figma

Everything in the middle gets compressed.

I think that model breaks down in robotics - it’s just way too big and too broad for single winners! Instead, we’ll see winning robotics companies in mining, agriculture, logistics, deep sea, defense, manufacturing, healthcare, and dozens of other verticals. These companies will win because they operate in different environments with different safety requirements, regulations, buyers, economics, and failure modes. A winning underwater robot is likely to be irrelevant in a hospital setting. A robot designed to pick and pack fragile inventory stock will be irrelevant when maintaining a solar farm.

With this in mind, I think robotics will produce many large companies, rather than one dominant infrastructure or application layer.

I expect those companies to broadly fall into two categories:

1. vertically integrated companies that own a specific workflow end-to-end, and

2. companies that benefit from the modularization of the robotics stack.

1/ Vertically Integrated Companies

Think of these companies as taking Apple’s path in robotics: increasingly owning the hardware, software, models, deployment, and customer relationship around one valuable physical job. What’s most important for these companies is owning the operating loop.

Take, for example, a robotic mining company - they may start with third-party vehicles, NVIDIA compute, open models, and off-the-shelf sensors. But, so long as they own the operating loop, every hour teaches that company something a generic robotics company cannot know. They learn about how dust affects perception, how different routes change tire wear, which slopes create control problems, and which component failures produce downtime.

That knowledge improves the system. It improves the policies, maintenance schedules, sensor configurations, fleet-management software, and customer workflow. Over time, it also tells the company which hardware should be customized, built, or acquired. The company earns the right to build custom hardware because it owns the job.

I think that same pattern will appear everywhere physical work is valuable and technical. For example: an agricultural robotics company will learn which weeds emerge at each crop stage, how weather and light affect vision, which treatments work, and where its equipment loses performance. A subsea maintenance company learns about marine conditions, asset degradation, battery performance, and the specific equipment customers need inspected or repaired. A logistics company learns how a facility actually operates: where inventory sits, when workers intervene, what causes exceptions, and which tasks create the greatest bottlenecks.

This deployment data makes the system more useful and more economical over time.

This gives those companies a deep competitive advantage, enabling them to sell outcomes, i.e. tonnes moved or packages packed, and not just machines or hours. They may start with third-party components, but the operating loop gives them a growing advantage over generic robot OEMs.

2/ Modular Platform Companies

The winners in this category will be the companies that make it dramatically easier for other people to build, deploy, and operate robots.

Rather than owning a single workflow, such as mining or agriculture, they will provide the shared infrastructure that many robotics companies rely on: fleet management, remote operations, simulation, data infrastructure, safety tooling, hardware interfaces, and marketplaces for task-specific models and components.

Their opportunity comes from the robotics stack becoming more modular. Hardware will get cheaper. Open models will become good enough for more tasks. Components will become more interoperable. And new platforms will let teams deploy robot fleets without rebuilding every layer of the stack from scratch.

A company should increasingly be able to combine a low-cost robot body from one vendor, a specialized gripper from another, an open model, and a shared fleet-management layer, then deploy the system into a real workflow. That is a much more accessible world than one where every robotics company has to build its own hardware, models, tooling, and operating software.

This is what open systems do well. They lower the cost of experimentation, bring in more builders, and speed up the entire market.

The “Android for robotics” opportunity is real.

But open source alone is not a business model. Open models and interoperable hardware will drive adoption, but the best platform companies will own a control point that compounds, e.g. the default fleet-management layer for heterogeneous robotics or managed data and simulation infrastructure. The key question for builders and investors alike is whether every new deployed fleet makes the platform better.

A company managing thousands of robots across different environments will learn which components work together, where deployments fail, and how to reduce the time and cost of getting a robot into production. Those insights improve the platform, helping attract more builders and fleets, helping further improve the platform.

This is a different type of operating loop from the vertically integrated company. The vertical company learns how to perform one specific job better than anyone else. The modular platform company learns how to help many different companies deploy robots more effectively.

Both can create enormous businesses.

At Canonical, we are particularly excited by vertically integrated companies with proprietary access to difficult physical workflows that improve with every deployment. We also want to meet teams building the modular platforms that help the broader robotics ecosystem build and deploy faster.

Zuck made the political case for open-weight AI, an NVIDIA-led coalition made the economic case, and Illia made the systems case. Meanwhile, 1,324 employees of frontier AI companies made the case for restraint.

What divides them is which danger comes first: a few companies owning everything, America falling behind China, AI finding software holes faster than anyone can patch them, or models improving themselves beyond anyone’s ability to follow.

Zuck’s case is about power: no single superintelligence can represent everyone’s values, so access must be distributed. Concentrated intelligence produces concentrated authority. That makes sense, but it also assumes equal compute, which remains unequal. Whoever can afford more inference, better memory, lower latency and more agents still has the stronger army. Weights can become open while power stays scarce.

Zuck then concedes that biological risks may require coordinated limits on how capable models are deployed. Open weights make those limits harder to enforce. Once the weights are released, anyone can run or modify them, and the original developer cannot take them back. His answer to concentrated power leaves this irreversibility problem unresolved.

NVIDIA’s case is that America will not win by producing a single frontier model. It wins by spreading American models globally before China does. Open weights lower costs, prevent lock-in and let organizations run AI on their own infrastructure.

The underlying incentive makes sense for NVIDIA. It wins when the model layer becomes cheap and compute remains scarce. Open weights push models toward commodity economics, while every fine-tune, evaluation and deployment still burns chips and energy. America gets broader distribution of American AI. NVIDIA gets a market where nobody upstream captures all the margin.

That incentive does not invalidate the economic case. It explains why NVIDIA is leading it.

The 1,324 frontier AI employees worry that models will soon improve AI research itself, creating a feedback loop that moves faster than labs can understand or control. The labs are asking for tools and governance that let society control the pace before that happens.

Their strongest objection to open weights is irreversibility. A hosted model can be monitored, restricted or withdrawn. Released weights cannot.

Anthropic makes this case harder to dismiss because its position is narrower than “open models are dangerous.” Amodei has called lower-capability open models a public good. He objects when downloadable weights become powerful enough to cause catastrophic harm and can no longer be monitored or recalled. This makes sense, but is also self-interested: restrictions on powerful open weights give frontier labs more time to extend their lead.

The open side still has the strongest empirical response. Open-weight models have been around for years, and no catastrophe has arrived. Open models have enabled independent research, local deployment, specialized models and competition. They have given startups an alternative to frontier-model APIs and governments a way to run sovereign AI.

That record matters. Restrictors have to explain what changes at the next capability threshold, where that threshold sits and why the record so far no longer applies. Their answer is simple: yesterday’s models could be safe to release while tomorrow’s are not.

Finally, Illia agrees that more capable models create new risks. He disagrees that restricting access to their weights solves them.

His point is that open vs. closed tells us who can access a model, not whether the system running it is secure. A hosted model can still access sensitive data, call the wrong tools or exploit vulnerable software. Securing AI therefore means securing the whole stack: the model, data, execution environment, hardware and surrounding software.

This is the strongest of the three open arguments. It is also incomplete. Better security can constrain models running inside systems we control. It cannot recall released weights or stop someone from running them elsewhere.

Illia’s position rests on a specific bet: security improves faster than dangerous models spread beyond its reach. If capability wins that race, Anthropic’s irreversibility objection stands. If security keeps pace, the open vs. closed distinction matters much less.

We think all four dangers are real. They just run on different clocks. Concentration is here now. The race with China plays out over years. Recursive self-improvement may arrive at any point, or not at all. The security gap is the one founders can build against today. That makes it the investable theme we’re tracking.

If securing AI is about the whole stack rather than the weights alone, the buildable surface is concrete: sovereign deployment, independent evaluation and red-teaming, agent security, verifiable inference and formal verification.

Some of this already has buyers and budgets. Companies need to know which model ran, what data it accessed and which actions it took. Verifiable inference and formal verification are earlier, more expensive and less proven.

We’re looking to fund companies selling security today with a credible path toward becoming verification infrastructure later. The bet is that security improves fast enough to constrain dangerous models before they spread beyond its reach.

That gap is where an early-stage investor earns the return.

Through this drawdown the speculators walked and the price broke, while the infrastructure kept getting laid and used. Measure the industry against the 2022 bottom instead of against last year, and the picture inverts. Ethereum transaction activity runs roughly thirteen times its 2022 level. Stablecoin supply has roughly doubled to around $300 billion and held there straight through the selloff. Value locked in DeFi is up more than 60%. Tokenized real-world assets sit at a record near $33 billion, led by tokenized Treasuries. Same bear-market price, an industry twice the size. Usage compounded the entire way down. Only the price disagreed.

The clearest tell is where the market already votes with real money. Crypto equities are up about 23% this year while the tokens are down about 36%. That split is not noise. The market is paying for the crypto it can underwrite and declining to pay for the crypto it can only speculate on. Miners sitting on signed power contracts are redirecting that energy into AI compute and getting valued as infrastructure. Lending protocols with real revenue are getting valued as businesses; one of them cleared roughly $900 million in fees over the past year. For the first time the tooling exists to see this plainly, because the fees and the buybacks settle on-chain where anyone can audit them. Cash flow earns a multiple. Narrative earns a chart.

There is a deeper shift underneath that split. For a decade tokens traded at a premium to the businesses beneath them. Today many trade at a discount to the same cash flows. The market has stopped paying up for the idea of a network and started paying for the economics of one. That is not the category dying. It is the category growing up.

Real institutions are building through the winter rather than waiting for the thaw. Federal stablecoin rules finalize this year and take effect in January. Stripe, Visa, BlackRock, and Coinbase are standing up new stablecoins behind that framework. Schwab and E*Trade are adding retail crypto rails to tens of millions of accounts. None of that is a wager on next quarter’s price. It is capacity being poured while the ground is still cheap, by counterparties who do not build on sentiment.

This is where the early-stage job actually gets done. Bull markets reward trading. Bear markets reward underwriting. The founders laying rails and shipping product right now are doing it into flat charts and thin attention, which is exactly what makes them cheap to back and exactly the kind of people worth backing. Anyone can fund usage after the price has confirmed it. The return comes from funding it while the price still argues.

So the discipline is simple. Ignore the quote and read the usage, because the usage is the business and the quote is only the mood. Sentiment collapses in a quarter and the commitment underneath it does not. When the price finally re-rates, it re-rates up to the usage that was there the whole time. The work is to already own the thing before it does.

In 2026, the largest sponsor of free frontier AI is the Chinese state.

Start with the facts. The open models that keep landing at or near the frontier are increasingly Chinese. DeepSeek, Qwen, and Moonshot’s Kimi now trade blows with the best American models on real benchmarks, ship under permissive licenses, and arrive within months of the closed frontier. The old story was that China was a cheap fast-follower, a year back at a tenth of the price. That story is out of date. The gap is now measured in months, and the strongest Chinese labs are confident enough to charge close to full price for a hosted version while still giving the weights away.

This is deliberate commoditization, sanctioned from the top. If you cannot win the frontier outright, the next best outcome is to make sure the frontier is not worth much. Open weights turn the core asset of OpenAI and Anthropic into a free good. They also set the default. Every developer who builds on a Chinese open model is one more person fluent in China’s tools, standards, and assumptions. The United States understood this logic perfectly when it gave GPS away and let the world build on top of it. Whoever owns the standard owns the leverage that comes with it. China is running that playbook one layer up, at the model.

There is a sharp irony underneath this. Export controls were meant to protect an American compute lead. What they actually did was force Chinese labs to get radically better at doing more with less. Starved of the newest chips, they leaned into efficiency: sparser models that wake only a fraction of their parameters per token, cheaper attention, aggressive compression. The same thing happened in lithography, where a constrained foundry reached advanced nodes on older equipment because it had no other option. Constraint is a forcing function. The controls meant to widen the moat are training the discipline that drains it.

This is the part worth sitting with. Open source was supposed to be insurance against concentrated power. Today the biggest force pushing free frontier intelligence into the world is an authoritarian state. We still believe open beats closed. We wrote that recently and we mean it. What changed is not the direction, it is who benefits.

The United States has no clean answer. Banning Chinese open models mostly binds American firms: weights spread by torrent, so you get costlier domestic AI while the rest of the world standardizes on Chinese models anyway. Winning the open game directly needs a lab both willing to open a frontier model and able to build one, and no American lab is both. Either path cedes the default.

For us, the investing implication follows directly. If the model layer is commoditizing and going open, and it is, then durable value does not sit in the weights. It sits in what open weights cannot commoditize: verifiable inference, private deployment, the sovereign and enterprise builds that cannot run on a foreign API, and the distribution that turns a free model into something people actually pay for. Open winning does not mean nobody captures value. It means value moves, and it moves to exactly the layers we have been backing.

The world’s default intelligence may end up open, cheap, and not American. That is not a prediction we love. It is one we would rather underwrite than ignore.

The reason is a collision. Demand for compute is going vertical at the exact moment the one trick the whole industry relied on has run out of room. Power, not chips, is now the binding constraint on AI: this year alone billions of dollars of datacenter capacity has been delayed or cancelled not for lack of funding or silicon, but because the grid cannot deliver electrons fast enough. And a large share of the energy we do deliver is spent not computing but shuttling data between memory and processor. But progress does not stop. It moves, and right now it is moving in three directions at once: new materials, new geometries, and new architectures.

New materials. There is a class of device that holds memory and does the computation in one place, the way a synapse does, instead of separating the two. The physics has been understood for roughly two decades and the performance ceiling is enormous, but none of these devices are in production. The reason: the materials involved have historically been impossible to build into a normal silicon chip without destroying the transistors underneath.

New geometries. Photonic chips do not need three-nanometer features. They need three-dimensional, multi-material, wavelength-scale structures that the world’s most expensive machines were simply never built to make. The unlock there is not smaller. It is a different shape entirely, and the tools optimized for shrinking transistors cannot produce it.

New architectures. Once you can co-locate memory and compute, you get to attack the von Neumann bottleneck directly rather than papering over it with faster interconnects and more cache. The prize is doing the core operation of a neural network in a single physical step instead of thousands of read-write shuffles. That is where the order-of-magnitude efficiency gains live. And when power is the constraint, efficiency is the only thing that matters.

In every one of these cases, the breakthrough is not the invention. It is the integration. The exotic material, the photonic structure, and the analog crossbar have existed in labs for years. What’s been missing is a way to manufacture them at scale without throwing away the trillion-dollar silicon supply chain. The companies we think win this decade are those who figure out how to build these devices on machines the fabs already own, inside an existing process. Complement the incumbent, do not try to replace it. The moat moved from what you can invent to what you can integrate, and that is a much less crowded, much more defensible place to stand.

This reframes what we look for. In diligence, we care less about a benchmark from a clean-room prototype than whether the process survives contact with a real foundry, which is where almost everyone stalls. It also reopens the map. The leading-edge logic race is lost outside Taiwan and a single town in the Netherlands, but the new wedges are early, unoptimized, and open to new entrants in a way the transistor race has not been for thirty years. That is why the best teams we see are not trying to out-TSMC TSMC. They are building what TSMC was never designed to build.

Why now? AI created both halves of the opportunity. It created the demand by making energy efficiency existential. And it supplied the tools to meet it: cheap GPU compute, better lasers, a consolidated supply chain. These approaches turned viable in the last twenty-four months, not the last twenty years. The constraint and the cure arrived together.

Arena announced a $100M annual revenue run rate, eight months after launching its evaluation product. It started as a research project at UC Berkeley, with one mission: to measure AI progress through real-world use. The company is barely a year into commercialization, but that research began back in 2023. By the time the revenue showed up, the hardest part was already done: earning the whole field’s trust as the neutral scoreboard for AI.

That gap is what we keep coming back to. Arena’s team figured out how to measure model utility before it was even recognized as a central problem by industry. The research was the moat, and the revenue is the market catching up to a bet placed long ago in a lab. We would rather get to know these people while they are still publishing than compete for them once every fund has the same warm intro.

This is why our focus on academics is so deliberate. At the frontier, in AI, robotics, energy, and cryptography, the durable edge tends to be technical depth and being early to a paradigm. Academics live there by definition. They often see the shape of the next S-curve before it has a name, and they carry conviction that is hard to reverse-engineer from a competitor’s launch.

It also shapes where we look. We think the best technical founders are spread far wider than the target-school shortlist suggests, so we put as much energy into the research groups at UIUC, UT Austin, and Georgia Tech as we do at Stanford or MIT. The work coming out of the less-trafficked programs is world-class, and the founders behind it are often the least contested. Getting to know them early is one of the most underrated edges we have found in venture.

So we try to treat the lab as the top of the funnel. We map research groups, we read the papers, and we build relationships years before anyone incorporates. Arena is a good reminder of why that patience compounds. The company was there for anyone paying attention, long before it printed a nine-figure run rate. The best companies are often visible on paper first. Our job is simply to be reading.

Last week a couple of days after my panel at SuperAI in Singapore on “Compute, Chips and the Cost of Intelligence“, I stood on a cleared lot in Johor, Malaysia, watching a future AI data center get staked out. Everyone on that panel was talking about GPUs. But what decides whether that lot in Johor becomes a working cluster is not just chips. It is whether you can get a gigawatt of power to the dirt, cool it, and find people who know how to run it. That gap is what this post is about.

It is a companion to our map at canonical.cc/labs/data-centers, which tracks 31 players across 7 layers. It sits next to our Semiconductor Stack and Decentralized AI maps. Body, brain, nervous system of the buildout.

The strange loop

Our silicon primer was predicated on this fact:

the technology breaking the chip industry is the same technology most likely to fix it.

AI demand is breaking the grid. A single campus now draws a gigawatt, a small city’s worth of power for one building. And AI is turning out to be the best tool we have for designing those power systems, tuning the cooling, and squeezing more compute from concrete we already poured.

A building, on a napkin

A data center is a building full of computers. Power comes in, heat comes out, compute happens in between. That is the whole thing. We went from server rooms, to the cloud, to the AI explosion, and the physics never changed. Only the scale did.

The number everyone quotes is PUE, Power Usage Effectiveness: total power divided by the power that reaches the chips. A PUE of 1.2 means you burn 0.2 watts on cooling for every watt of compute. The old server room ran near 2.0. Liquid-cooled AI halls are pushing toward 1.1. At gigawatt scale, that gap is a power plant’s worth of waste.

Now stack it up. From the rental product at the top to the grid at the bottom, every layer is its own market.

1. Neoclouds. The product most AI companies buy: GPUs by the hour, no enterprise baggage. CoreWeave, Crusoe, Lambda, Nebius. CoreWeave proved the category reaches public scale. It also exposed the category’s central risk, below.

2. Operations. Who runs the cluster once it is powered on. The quietest layer on the map, and we think the most mispriced. More below.

3. Compute silicon. The GPUs and accelerators. NVIDIA, AMD, Google TPU, Cerebras. Everyone fights here. We cover it in the silicon map, so we will skip it.

4. Networking. Past 100,000 GPUs, the bottleneck stops being compute and becomes moving data between chips. Arista, Broadcom, optical-I/O upstarts like Ayar Labs.

5. Cooling. Old racks drew 5 to 20 kilowatts; air handled it. AI racks draw 40 to 80 today, and NVIDIA is targeting a megawatt per rack on Rubin Ultra. Air cannot touch that. Liquid goes straight to the chip. Motivair, CoolIT, JetCool.

6. Power. The binding constraint. Constellation restarting Three Mile Island, Oklo and small reactors, plus a fast-growing layer of on-site generation and batteries that smooth a training run’s spiky draw.

7. Facilities. The buildings and land, mostly held in REITs. Equinix, Digital Realty, Vantage, QTS.

So where does it break?

Which layer can actually stop the buildout?

In silicon, the answer was two companies: ASML and TSMC. Here it is not a company. It is power.

About 2,300 gigawatts of generation sit stuck in US interconnection queues, more than the country’s entire installed capacity.

The average project waits five years to connect. Some data centers are quoted twelve.

US data-center demand is headed for roughly 76 gigawatts in 2026, up from 50 in 2024.

In Texas, large-load requests to one utility jumped sevenfold in a year.

Bessemer counts 190 gigawatts of hyperscale capacity already announced against 5-7 year queues.

“We need more compute” is something we are all hinding behind. The world is not just short on chips this quarter. It is short on energized, cooled, staffed megawatts. You can panic-order GPUs. You cannot panic-build a substation, and you cannot panic-train the people who run the room.

Where AI causes the squeeze

Power takes the direct hit. Cooling takes the next one, as rack density climbs an order of magnitude and liquid goes from exotic to mandatory.

Then the strangest one: utilization. The most expensive hardware ever built mostly sits idle. xAI’s fleet reportedly runs near 11%. Hyperscalers run 5-10%. We spent 25 years learning to share CPUs with VMs, containers, and schedulers. Almost none of that exists for GPUs yet. So we scream about a shortage while most of our GPUs sit parked.

Under all of it, people. Maybe a few hundred humans on earth have run a multi-thousand-GPU cluster end to end. The demand is the entire buildout. Construction is no better: the trade is short hundreds of thousands of workers, the average data-center worker is 53, and most operators cannot fill ops roles. The pain lands where the talent is thinnest. Not a coincidence.

Where AI relieves it

The same intelligence stressing the grid is being pointed back at the stack. AI permitting software is compressing the slowest, most bureaucratic step in the build (Paces, Lorica). AI runs cooling and power inside live halls (Phaidra, Corintis). A new layer makes GPU work power-flexible, so a cluster leans into cheap power and backs off when the grid is tight (Emerald AI, Verse). Construction is getting robots (Gecko Robotics, DroneDeploy).

The highest-leverage bet sounds the dullest. Build “VMware for GPUs,” real multi-tenancy and scheduling, and a 10 percent utilization problem becomes a 10x supply unlock. That is more new compute than most fabs add in years, and it ships in software.

Will any of this beat the power constraint this decade? No. Electrons obey physics and permitting, not roadmaps. But the software-shaped layers (permitting, orchestration, utilization, operations) compound in quarters, not half-decades. That is where venture lives.

Sovereign AI: the repatriation trend worth watching

For two decades the logic was centralize: ship compute to whoever had the cheapest power. That is reversing, fast. Countries now treat AI compute as critical infrastructure, like a grid or a port, and they want it on home soil under home law. Some call it geopatriation. I call it the most important capital-flows story in infrastructure right now.

The EU has put 20 billion euros behind AI Gigafactories. France is building a 1-gigawatt campus with the UAE worth tens of billions, while backing Mistral as its champion. Saudi Arabia stood up HUMAIN under its sovereign fund to build the whole stack. The UAE is breaking ground on a multi-gigawatt cluster in Abu Dhabi. Japan and Singapore are moving the same way. We see it at seed stage too, with founders pitching energy-first inference backbones built for Europe.

Repatriation does not shrink the stack, it widens it. Every sovereign build needs its own power, its own cooling, and its own operators, in a country that has never run a frontier cluster. The talent gap that is acute in Virginia becomes a wall in Riyadh or Johor. Many national buildouts means the power and operations layers get demanded everywhere at once. That is the picks-and-shovels case, multiplied by the flags on the map.

Why this could fail

I would be selling you a dream if I skipped the bear case. It is real and specific.

The sharpest risk is financial. Neoclouds buy GPUs with GPU-backed debt. If utilization or pricing softens, the collateral can depreciate faster than the loans amortize, and one demand wobble cascades through the most leveraged layer. CoreWeave’s bull case and bear case are the same fact. Add regulatory backlash (moratoriums, water fights, grid-reliability fights are already spreading) and the chance that hyperscalers re-bundle and fix the allocation problems that created the neocloud layer in the first place. And the deepest question stays open: does inference get effectively free as efficiency compounds, or does demand outrun supply forever, keeping power the permanent constraint? The thesis hinges on which way that breaks.

Our read: the bull case is already priced into the obvious layers, silicon and the public neoclouds. We would rather underwrite where the centralized option is structurally weak and demand is non-negotiable. Power. Liquid cooling. And the operators who turn a half-billion-dollar building full of idle silicon into a working cluster.

The exciting part

Today, standing up AI compute takes a hyperscaler’s balance sheet and a team that barely exists. Productize the operations and power layers and that capability gets unbundled: handed to the sovereign, the enterprise, the neocloud in a country you would not have guessed. Standing up a cluster starts to look less like building a refinery and more like provisioning a service.

Somewhere on that lot in Johor, rebar is going in for a building that will think. Whether it ever does comes down to the least glamorous layers on the map: the wire coming in, the heat going out, and the few people who can keep it alive at 3am. That is the layer we are building toward.

The flashy layers (silicon, the public neoclouds) are mostly priced. The middle (power, cooling, operations) is where we think the next infrastructure companies get built, and where AI is the relief, not the strain. We track all of it, layer by layer, with sources and caveats, on the full map.

Explore the map: Data Centers & Neoclouds at Canonical Labs →

Educational, not investment advice. Figures are point-in-time as of June 2026.

When the input commoditizes, value moves to whoever orchestrates it. That is happening in two directions at once, above the model and below it.

Above the model, the aggregator is becoming the intelligence layer. OpenRouter spent two years routing each request to the single best model. This month it launched Fusion, which fans a prompt out to a panel of models, has a judge reconcile their answers, and returns one synthesized result. The panel beats any single model: 69% on the DRACO research benchmark against 65% for the best solo model, and ahead of both GPT-5.5 and Opus 4.8. The routing layer stopped picking the winner and started manufacturing one.

Below the model, serving is becoming its own discipline. A cheap open model is only cheap if you serve it efficiently, and the optimization surface (batching, KV cache, speculative decoding) moves faster than any one team can track. vLLM is the floor, not the ceiling, and the frontier shifts week to week. The edge is no longer which model you run. It is how few GPUs you need to run it.

We think this is where the margin goes. Not to whoever trains the best model, but to whoever orchestrates models best on top and serves them cheapest underneath. We’ve said before that open weights commoditize the base layer and the moat moves up-stack. The model was the product for three years. Now it is the raw material, and the companies that matter are the ones that turn it into something cheaper, faster, or smarter than any single model could be alone.

We started paying attention anyway because of the talent density. Elite researchers and infrastructure veterans who could have shipped at any centralized lab were choosing to build here instead. They saw decentralization as a force multiplier, not a constraint.

At our AGM last year, we told our LPs that AI was concentrating power in a handful of labs, and that the answer was open, verifiable intelligence built on crypto rails. “Not your weights, not your brain.”

In the time since, four things changed. Governments started requiring AI to run inside their own borders, on infrastructure they control. The GPU shortage priced most startups out of compute. People grew wary of handing every prompt and document to a few opaque providers. And as AI agents began managing real money, they needed a way to prove the right model ran on the right data. Each of those is a reason to build outside the centralized stack.

And the results are no longer hypothetical. A network of 70-plus strangers trained Bittensor’s Covenant-72B on commodity hardware and beat LLaMA-2-70B on MMLU. x402, the HTTP-native agent payment rail, has settled over 161 million machine payments across roughly 69,000 active agents. Aethir booked more than $128 million in real revenue last year from 150-plus enterprise clients. Venice runs private inference for over 2 million users with zero data retention.

What makes decentralized AI different is structural. Three things stand out.

Openness is built in, not bolted on. Pluralis splits a model across many machines so no single party ever holds the full weights. You can use the model, but you can’t take it private and walk away with it.

Cryptography does the work that trust used to. You don’t have to take these systems at their word. EigenCloud ran the same inference 10,000 times and got an identical result every time, which is what lets a smart contract actually rely on the output. Venice encrypts your prompt on your own device, so neither Venice nor the GPU provider ever sees it.

And the people building this are not who you’d expect. NEAR’s founder co-authored the Transformer paper. Sentient’s co-founder invented the core technology behind 4G. This is serious technical talent choosing the harder path.

The obvious objection: OpenAI and Anthropic are running away with this, so why does any of it matter? Because the question was never who builds the smartest model. It is whether AI infrastructure ends up winner-take-all or multi-vendor. The centralized labs win on raw capability, and they will keep that lead. But capability is not the whole market. Governments that won’t run on American clouds, enterprises that can’t expose their data, and agents that need provable execution: none of that is a capability problem, and none of it is something a closed API solves well. Centralized AI compounds through data moats and switching costs. Decentralized AI compounds through networks and adoption. The first wins the short term. We are betting the second matters more over time, and regulation, GPU scarcity, and sovereign mandates are all pushing that way.

The bear case is real:

1. The centralized labs are extremely well capitalized, backed by the best investors and run by exceptional operators. That is an enormous head start.

2. Most decentralized AI projects are yet to find real organic demand. Many are propped up by token emissions rather than customers.

3. “Decentralized” often isn’t. Look closely, and many of the nodes for these networks are fairly centralized.

4. Several layers depend on trusted execution environments, which have been broken before.

5. A lot of this rests on research breakthroughs that have not landed yet, like privacy-preserving computation. By the time they arrive, the window of opportunity may have closed.

We don’t wave any of that away. Our job is to separate the projects solving a real bottleneck, where the centralized option genuinely can’t go, from the ones solving a problem that centralized AI already handles fine.

So we built the map we wanted to read. 35 projects across 9 layers, with the team, the traction, and our view on each. It is a companion to our Semiconductor Stack and Physical AI maps, and it is fully interactive: filter by layer, search by project, click into any one.

We said it a year ago and we will say it again. Still early, but not too early.

Explore the Decentralized AI map →

We think that bundling is the wrong abstraction. A person does not relearn the world every time they reason through a problem. Memory and reasoning are distinct faculties, and the most interesting work in AI is starting to separate them.

Two public results point the same way. DeepMind’s Mixture of a Million Experts treats knowledge as a vast store of tiny experts, retrieving only the handful a query needs and decoupling what a model knows from what it costs to run. From the other direction, Samsung’s Tiny Recursive Model treats reasoning as a small recursive loop: 7 million parameters scoring 45% on ARC-AGI-1, beating Gemini 2.5 Pro, DeepSeek R1, and o3-mini with less than 0.01% of their parameters. One unbundles memory. The other unbundles reasoning. Neither needed scale to win.

The metric that falls out of this is not parameter count but intelligence per FLOP, capability for every unit of compute and energy spent. The author of the TRM paper calls the belief that hard problems require billion-dollar foundation models a trap, and the results are starting to agree.

We think the next frontier is architectural. The teams that separate memory from reasoning will deliver the same intelligence at a fraction of the cost, and that efficiency is exactly what physical AI and on-device inference have been waiting for. Scale was never the point. It was a proxy for the things we had not yet learned to build directly.

27 years ago at Purdue, in ECE 362, I programmed my first chip: a Motorola 68HC12 with a few KBs of memory, hand-written assembly, bolted onto a racecar so we could control its speed remotely. I still remember the rush. Today the chips driving the AI boom cram 200+ billion transistors into one package, and half the world’s governments are fighting over them. Same starting material: sand. And that distance is what this post is about.

The strange loop at the center of this

Marinate on this for a second:

the technology currently breaking the chip industry is the same technology most likely to fix it.

AI eats chips faster than the world can make them. A single frontier model trains on tens of thousands of leading-edge GPUs. And AI is also the best tool we’ve ever had for designing chips, checking them, and squeezing more of them out of the factories we already have. That tension runs through everything below.

To see why, you have to know how a chip gets made. Most people don’t, and that’s fine; the industry hides behind acronyms. The structure underneath fits on a napkin. So that’s what we drew.

(This post is a companion to our full interactive map at canonical.cc/labs/semiconductor-silicon-stack, which tracks 20 incumbents and 40+ challengers across every layer)

The stack, on a napkin

A chip starts as quartz sand and ends as a die with billions of transistors. In between: six physical steps, plus one software pipeline running alongside. Each step is dominated by one or two companies you’ve probably never heard of.

Follow the sand.

1. Raw materials. First you purify sand into polysilicon that is 99.9999999% pure. Nine nines. Take a billion atoms; only one is allowed to be the wrong kind. Two companies, Wacker and Hemlock, dominate, not because the chemistry is secret but because nobody else wants to run a billion-dollar chemical plant that needs a decade of contracts to pay for itself.

2. Wafers. The polysilicon gets grown into a single perfect crystal, then sliced into mirror-polished discs. Two Japanese companies, Shin-Etsu and SUMCO, make most of the world’s supply.

3. Lithography. The famous one. You print the circuit pattern onto the wafer with light, and the features are so small that ordinary light is too fat to draw them. So the machine makes its own: it fires a laser at droplets of molten tin, 50,000x/second, and collects the extreme ultraviolet flash with mirrors so smooth that if you stretched one to the size of a country, the tallest bump would be about a millimeter. One company on Earth makes these machines: ASML, in the Netherlands. The newest model runs $380 million. There is no second source. None.

4. Deposition and etch. Now you build the chip up like a layer cake, hundreds of layers, depositing material and etching it away. Applied Materials, Lam Research, and Tokyo Electron split this market. A healthy three-horse race, though each horse specializes in different tools.

5. Inspection and test. After every step, you hunt for defects a few nanometers wide. Miss one and the chip is garbage. KLA owns this.

6. The fab. TSMC in Taiwan puts the whole symphony together at scale. Samsung is a credible second on some nodes. Intel is rebuilding. At the true leading edge, though? TSMC or nothing.

And running alongside all of it: design software, called EDA. Before anyone touches silicon, the chip lives entirely in software, where it’s designed, simulated, and torture-tested. Three companies (Synopsys, Cadence, Siemens) hold about 75% of that market, with tools whose bones date to the 1990s.

Notice who’s missing: Nvidia. It doesn’t make chips. Neither do AMD or Apple. They’re “fabless”: they design in the EDA tools and hand TSMC the blueprint. The companies above are the ones even trillion-dollar giants depend on.

So where does it break?

Now ask the question a curious person should ask: which of these layers could actually stop the world?

Only two. ASML and TSMC.

If either one stops shipping, the AI industry grinds to a halt within months. Every other layer has at least two serious competitors keeping each other roughly honest.

“Semiconductor shortage” is a vague phrase hiding a precise reality. The world isn’t short on sand, wafers, or etch tools. It’s short on EUV exposures and leading-edge fab slots, which funnel through exactly one Dutch company and one Taiwanese one. And TSMC’s grip goes beyond printing silicon: the advanced packaging that stitches finished dies into a working accelerator also runs mostly through it. Hold that thought.

Where AI causes the squeeze

Now pour the AI boom into that structure and watch where it pinches.

The fab layer takes the direct hit. Everyone who matters (hyperscalers, frontier labs, now entire countries) is fighting over the same TSMC capacity. And you can’t panic-build a fab. Demand moves in months. Fabs move in half-decades and cost tens of billions.

The pressure flows downhill to lithography. More fabs means more EUV machines, and ASML can only assemble so many a year. Each machine has hundreds of thousands of parts, from a supply chain that is itself backed up.

Meanwhile the design side has its own problem: people. AI chips are enormous, and verification, the work of proving a design is correct before you bet $100M+ on manufacturing it, already eats up to 70% of engineering hours on a project. You can’t hire your way out. Training a chip designer takes a decade.

And the chips themselves changed shape. A modern AI accelerator isn’t really one chip. It’s several dies stitched together on an interposer (that’s the CoWoS packaging you’ve maybe heard of), and that stitching capacity is scarcer than wafers.

The pain shows up exactly where competition is thinnest. Not a coincidence.

Where AI relieves it

Here’s the other half of the loop, and the part I find more fun.

On the software side: the people behind Google’s AlphaChip, which used reinforcement learning to lay out chips better than human engineers, left to start Ricursive Intelligence. Salt the claims in this category appropriately. But the direction matters: if AI compresses design cycles from years to months, every design team on Earth just got bigger overnight.

On the hardware side, AI money is funding swings at the chokepoints themselves, which nobody dared for decades. Substrate aims to make chip-printing light with particle accelerators instead of ASML’s tin droplets.

Those are 2 of many. Agentic verification copilots, AI inspection layers squeezing 30% more throughput from existing fabs, free-electron-laser EUV with CHIPS Act backing: the full roster, with traction numbers and caveats, lives on the map.

Will any of them dethrone ASML or TSMC this decade? Almost certainly not. Watch the software layer instead: software ships in months, and the incumbents’ AI offerings are bolt-ons, not rebuilds.

Why hasn’t a 30-year-old monopoly been disrupted already? Rational fear. When a manufacturing run costs $100M+, nobody volunteers to be first to trust an unproven tool. That’s the real moat: challengers can’t just be better, they have to be provably, boringly, run-after-run better.

But if AI draws blood anywhere first, it will be EDA.

The part that should excite you

Follow that thread one more step and you get to the prediction I’d actually bet on. Today, designing a serious chip costs so much (tens of millions in tools and talent before you even pay for manufacturing) that custom silicon is a rich company’s game. “Fabless” mostly means Nvidia, AMD, Apple, Qualcomm.

We’ve seen what happens when that kind of cost collapses. Renting a server from AWS turned “starting a software company” from a $5M proposition into a weekend project, and we got millions of software companies. If AI-native design tools cut the cost of a credible chip by 10x, you don’t get slightly more chip companies. You get thousands of fabless startups designing silicon for problems too small for Nvidia to bother with.

And the demand is already lining up. Physical AI is pushing intelligence into robots, cars, drones, wearables, and factory floors, and most of those chips don’t need the bleeding edge. They need to be cheap, low-power, and exactly right for one job. (We map that whole demand wave separately on our Physical AI & Robotics tracker)

A world of thousands of designs running on mature nodes is also the world where Atomic Semi’s many-small-fabs bet stops sounding crazy. More designers, more specialized chips, more places to make them.

Somewhere on a campus right now, a junior is flashing their first assembly program onto a microcontroller and feeling the same rush I felt with that racecar. The stack they inherit is being rebuilt today.

The hardware chokepoints are decade-long bets you underwrite like fusion. We track all of it, layer by layer, with sourcing and known caveats, on the full interactive map: Semiconductor Stack Disruptors

Educational, not investment advice. Data is point-in-time as of May 2026.

That framing is breaking, and two things changed at once. World models now learn “what happens next” from internet-scale video, giving robots physics priors without hand-coded simulators or armies of teleoperators. And a thin layer of real demonstrations now bootstraps enormous synthetic scale. NVIDIA’s Manipulation-Augmented dataset turns 10 human teleoperated demos into 1,000 domain-randomized examples. The real world becomes the seed, not the substrate.

The data confirms the shift. Robotics went from 1,145 datasets on Hugging Face in 2024 to 26,991 in 2025, climbing from rank 44 to 1, with synthetic generation as the primary driver. And this week, NVIDIA launched Cosmos 3, an open model that collapses vision reasoning, world generation, and action prediction into a single system. Cosmos runs as a physics-grounded simulator that predicts approaches, evaluates them in a closed loop, and converges on behavior without real-world risk.

We think the moat is moving. Not to whoever owns the best model or the most teleoperators, but to whoever builds the best learned simulator. Reality becomes the verification step, and the data flywheel that used to live on customer floors now runs in software. The teams that own the simulator own what comes out of it.