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.

What beat it is API plus CLI.

Anthropic’s Claude Code has become the fastest-onboarding developer surface the company has shipped, despite running entirely in a text terminal that looks like it was designed in 1985. OpenAI killed Operator as a standalone product and now ships its computer-use model through the API instead. Anthropic’s Quick Mode for Claude in Chrome ships a stripped-down agent loop that swaps structured JSON for single-character commands and delivers a roughly threefold speedup on real browsing tasks. The public benchmarks we’ve found point the same direction: agents that work through APIs and command lines are pulling ahead of agents that work through pixels.

The reason is unit economics, and it gets worse for vision every quarter. A browser-based agent pays to tokenize a screenshot and a DOM tree on every single step, feeding a model big enough to reason over a megabyte of UI noise just to find one button. A CLI agent pays for a few hundred structured tokens per step and gets a structured answer back. The gap isn’t one click – it’s tokens-per-step multiplied across every step of every task. Frontier model pricing for structured reasoning is collapsing faster than vision pricing, and the gap is widening. The cost curve is bending against pixels, and the bend is steepening.

The implication for SaaS is the part most software companies have not yet priced in. If the dominant user of your product over the next decade is a fleet of agents, then the surface that matters for distribution is your schema, not your screen. Salesforce, Notion, Linear. Every category leader is about to be evaluated on a single question: how cleanly can a model consume your product without rendering a pixel. The companies that ship a real, documented, agent-readable API as a first-class product become rails – Stripe proved the model years ago. The companies that gate their primitives behind a UI become the legacy layer that the next generation of startups quietly automates around.

There is a real counter to this view: the browser is also the universal abstraction over the long tail of legacy software that will never expose a clean API. State government portals, hospital admin systems, niche industry tools with three customers and a 2008 codebase. That ~15% of the workflow surface is a real market, and vision agents will serve it for a long time. But 15% is a fallback business. The platform fight is the other 85%.

The bigger compounding effect sits at the infrastructure layer. A human can run one workflow at a time. An agent can run a hundred in parallel, each making thousands of API calls. Our bet is that machine-initiated traffic comes to dwarf human-initiated traffic by an order of magnitude, and the inference cycles, bandwidth, and storage required to support that pattern are nowhere in the current hyperscale capex curve. The most underwritten thing in markets right now is not the model layer or the application layer. It is the silicon, fibre, and power required to carry the load that an agent-native software stack actually generates.

Our view, plainly: the SaaS layer bifurcates. The companies that ship clean, agent-readable APIs become the rails. Everything else becomes vision-agent fallback or a wrapper around someone else’s schema. The infrastructure layer captures the spillover, and the spillover is arguably the biggest single capex flywheel in history.

The next decade of software is being built for readers that do not have eyes. Plan for them.

The gap isn’t capability. The models are good enough. The gap is that an agent, no matter how sophisticated, cannot yet participate in commerce as a real economic actor. It can read a contract better than most lawyers. It cannot sign one. It can manage a treasury. It cannot own one. It can run your marketing, your procurement, and your customer support, but you won’t hand it the corporate bank account. There’s no way to scope what it can touch, no audit trail if something goes wrong, and no kill switch that actually works in real time. So the most capable employee at every fast-growing company is locked out of the most leveraged tool.

This is the actual bottleneck in enterprise AI adoption, and it runs deeper than most discussions acknowledge.

Making an agent genuinely autonomous requires solving three distinct planes of infrastructure simultaneously. The first is trust: does the counterparty know who this agent is, what it’s authorized to do, and what its track record looks like? The second is market: can the agent discover work, quote a price, sign a contract, settle payment, and resolve a dispute without human intervention? The third is control: can the organization deploying the agent prove, to a regulator or a board, exactly what the agent was authorized to do and what it actually did?

None of these planes exists in mature form today.

On the trust side, the identity problem alone is unsolved. Existing card frameworks like EMV, 3-D Secure and chargeback rules assume a human clicked a button. KYC verifies a face and an ID, not the autonomous code now acting on a person’s behalf. Once an agent executes a payment, the liability question becomes unresolved: is it the issuer, the merchant, or the model? t54 is one of the more interesting attempts to answer this, building what they call a Know Your Agent layer, binding human identity to agent identity and producing cryptographic proof of intent that financial institutions can actually consume.

On the market side, even getting an agent access to a payment account is harder than it sounds. The obvious solution is to give it access to the company bank account. The obvious problem is that one bad decision wipes the treasury. Bank of Bots is building around this constraint: a segregated, scoped account that sits alongside your existing bank, where each agent gets hard spending limits and merchant whitelists. The blast radius is bounded by design.

And on settlement, fast, cheap, and programmable stablecoins look like the natural rails for agent-to-agent commerce. Except they’re irreversible. No chargebacks, no dispute window. Anchor is building the risk pricing layer that sits in front of stablecoin transactions, evaluating and pricing each one before it broadcasts. The analogy is interchange: credit cards always worked because a chunk of every swipe funded the fraud and dispute machinery. Stablecoins skipped that and skipped the protection with it.

What’s notable is that none of these companies is trying to build the whole stack. Each is one layer. And that’s fine, each layer is genuinely hard. But it also means the market is still very early and very fragmented.

The processing layer itself won’t be where the money is made. Stripe, Coinbase, Circle, and Visa will commoditize that fast. The venture opportunity is in what gets built on top of verified agent payment flows: credit, insurance, and reputation data. The more agents you’ve scored and lent to, the better your underwriting model, and the harder you are to displace. That’s a real moat.

The identity and trust layer is probably the most interesting place to be right now. Whoever builds the credit bureau for software, scoring agents, verifying what they claim about themselves, pricing their risk for merchants and lenders, wins a category that didn’t exist three years ago and that no incumbent is well-positioned to own.

Last week we wrote about where the value gets built in agentic payments. This is why it matters. The models are commoditizing. The infrastructure above them is not built yet. That’s the opportunity.

Every GP I’ve pitched has a similar line ready when an LP asks. I’m not sure any of us actually knows.

So I built a tool to find out. It’s called the Power Law Lab, and it runs 10,000 simulated versions of a fund you describe.

The histogram shows the full distribution of plausible outcomes and not just the headline number.

→ canonical.cc/labs/power-law

What we noticed

Venture returns are power-law shaped. A handful of investments produce nearly all the returns. Every GP knows this. Every LP nods.

But that’s usually where it stops. We don’t internalize the consequences. When I write “we’re targeting a 3x net TVPI” in a deck, what’s behind that number? Is it the mean across plausible outcomes? The median? The top quartile? I’d be lying if I said I had a precise answer.

The mean and the median are not the same thing. In a power-law distribution, the mean is dragged up by rare right-tail outliers. The median is much lower. A “2x fund” by mean might be a 1.2x fund by median, with one Uber-shaped outlier doing all the work. The deck quotes the mean because that’s the flattering number.

What this actually does

Describe a fund its size, the number of investments, the failure rate, the shape of the right tail, your reserves strategy, your ownership target. The lab runs 10,000 simulated versions of that fund and renders the distribution.

The math is a mixture model: each company either returns zero (with some loss probability) or draws an outcome multiple from a truncated Pareto distribution. The shape parameter α controls how fat the tail is, and α turns out to be the most consequential slider in the lab.

Calibration follows public empirical work from Correlation Ventures, Kauffman, Carta and AngelList. The Seed preset assumes a 50% loss rate, α=1.2, cap at 500x. Series A and Growth presets get progressively thinner tails.

I vibe-coded the first version over a weekend. ~800 lines of JavaScript. Runs in your browser. No server.

3 things the lab taught me about our fund

1. More investments doesn’t help you.

Expected fund TVPI is invariant to the number of investments if you hold the strategy constant. Going from N=15 to N=50 doesn’t move the mean. It collapses the variance. That feels safer, but LPs aren’t paying us for the median. They are paying for exposure to the right tail. Indexing the power law squeezes that out of your own portfolio.

2. Follow-on discipline moves more than anything else.

Switching from pro-rata across the portfolio to super pro-rata into winners moves median TVPI more than any other parameter I tested. More than fund size, check ownership, or tail thickness. Reserves aren’t a hedge. They’re a second swing at the same pitch. We had been pro-rata-ing out of optionality and reputation. The math doesn’t support that choice as strongly as I’d assumed.

3. The median fund is boring.

Even with realistic seed-stage parameters, the median simulated fund returns roughly 1.3x net. The point isn’t that venture math is hopeless. It’s that the median is the natural outcome, and the funds distribution is itself power-law shaped. You can’t average your way into the tail.

What this actually means

Most GPs cannot tell you where their current TVPI sits in the distribution of plausible outcomes for their strategy. I couldn’t, before I built this. The pitch deck quotes the mean because that’s the number that flatters us most.

For LPs, the lab gives you a way to plug in a manager’s stated strategy and stress-test their projections. If their 3x target sits in the top decile of plausible outcomes for their setup, you should know that before the IC meeting.

For GPs, it’s an honest conversation with yourself. Where are you in the distribution? What would have to be true for the next-quarter mark to move you up?

Try it

→ canonical.cc/labs/power-law

3 calibrated presets, 6 scripted scenarios. Each scenario isolates a single counter-intuitive thing the math does that pitch decks gloss over. Start with the N debate scenario.

Every parameter gets encoded in the URL hash, so you can send a colleague the exact configuration you’re looking at.

If you find something useful or something you disagree with I’d love to hear it. The model has deliberate simplifications worth poking at: independent company outcomes, no time-value modeling.

We are going to keep building things like this!

Luna does almost everything. She sourced the products, negotiated with suppliers, set the prices, hired the painters for the mural, and ran the phone interviews to hire the two human employees who actually run the floor.

Why does she need them? Because general-purpose robotics isn’t quite there yet.

The gap is hands

Luna can hire, price, source, negotiate, and schedule. What she cannot do is restock a shelf, sweep a floor, or take a delivery off a truck. The cognitive layer of running a physical business is roughly solved. The physical layer is not.

Now imagine the next version of that store. Luna is still the brain, but instead of human employees, affordable robotic arms handle the physical labor. Frontier labs are racing the cost of the cognitive model toward zero. The bottleneck moves entirely to hardware, who can put a reliable, deployable arm next to that digital brain.

Multiply that pattern across every packing line, logistics bay, and commercial back-of-house in the country. You’re staring at a global labor-spend pool that dwarfs SaaS by roughly 30x.

That’s the wedge our portfolio company Robo is going after.

Why this time is actually different

2 things have shifted under the surface:

1. Scaling laws hold in robotics: More tele-operation hours in, lower model loss out. Fleets backed by remote operators are doing economically real work today, allowing startups to deploy immediately and ramp human intervention down as autonomy goes up.

2. The macro stakes are higher: The hurdles are real. Most manipulation models sit in the 80–90% success range, and the US infrastructure layer is paper thin. China installed roughly 295K industrial robots last year. The US installed about 34K. That 10x supply gap is a strategic vulnerability. Because hardware is brutally hard to scale, most robotics startups die in the supply chain.

Where Robo fits

Robo builds affordable robotic arms designed so the arms can eventually build more of themselves (see our earlier post about Robotic Superintelligence) to force down unit economics. They run a dual GTM on identical hardware:

• They sell arms ($2,500/unit): The lowest cost in the category for this payload-to-precision ratio. AI labs buy them for mass data collection; developers buy them to build on. To accelerate adoption, the low-level stack is open-source.

• They deploy arms (Robot-as-a-Service): They walk into high-volume industrial lines (like food packing), set up the hardware, and provide end-to-end tele-op coverage. Customers pay by the hour with zero CapEx. When the model stumbles, a remote human takes over instantly so the line never stops.

This dual motion creates a closed data flywheel. They own both ends: the labs training the models, and the physical floors where those models accumulate real-world hours.

Velocity

In January, the company was just a pegboard, two arms, and simulation software.

5 months later, they have redesigned the arms with a domestic sheet-metal core, scaled an in-house print farm for outer shells, secured ~100 preorders, and gone live with their first deployment.

Crucially, their domestic architecture is engineered to transition into sealed, cleanroom-ready variants. A US-built arm capable of meeting strict pharma and medical specs avoids the shifting regulatory hurdles that plague Chinese-manufactured alternatives.

What I come back to

There’s a version of this company that tried to do humanoids, or build a foundation model, or sell pure software. Instead, they picked the most boring-sounding piece of the stack - affordable hardware and deployment infrastructure, and executed with unusual clarity.

Luna at Andon Market can run a store, but she cannot yet stock her own shelves. Someone has to build the arm that does that, in volume, for a price a corner-store economic model can absorb.

Reach out to Robo if you need some US-homegrown robotic arms that are quick and cost-effective.

How much does this SAFE actually cost us? What if we stack another at a higher cap? How much do we lose if we top up the option pool to 15% at the Series A? At a $1B exit, what do we actually walk away with?

They couldn’t answer those quickly. We couldn’t either, sitting on a call, without opening a spreadsheet.

The math compounds in non-obvious ways. Each post-money SAFE locks in investment divided by cap at conversion. Stacked SAFEs at different caps don’t dilute each other. Founders absorb all of it. Pre-money option pool top-ups come out of existing equity, not the new round. Liquidation preferences hurt at low exits and evaporate at high ones.

Founders sign these terms in moments of high emotion. The lead just committed. The lawyer wants to close. You want to get back to building. The second-order effects don’t surface until years later, by which point your ownership is already where it is.

So we built it. A free, single-page tool that lets founders stack SAFEs at any cap, add priced rounds, top up option pools, set exit valuations, and watch the waterfall move in real time. No login, no spreadsheet.

dilutionlab.canonical.cc

Try it on your own cap table. Drag the exit slider. See exactly how much you walk away with.

We vibe coded it in <30 minutes. The source is on GitHub. If anyone wants to add features or fix corner cases, send a PR.

The new capabilities are real but narrow. Each product gives agents virtual credentials, spending limits, and merchant whitelists. The agent transacts without asking the human to approve each charge. What did not change is the underlying model. Authorization still flows from humans. Chargebacks still require humans to dispute. KYC still binds to human identity. Liability still lands on humans.

The agent looks like a principal, but it is still a delegate.

That works fine for the first wave of agentic commerce, where a human tells an agent to book flights or order groceries. But it doesn’t work for what comes next: agents transacting with other agents, services pricing themselves dynamically per call, software paying software for compute, data, and bandwidth.

Stablecoins are the only payment substrate ever built for non-human counterparties. They are programmable. They settle in seconds with no chargeback infrastructure. An agent can hold a balance, spend it, and transact at any size, including amounts that make no sense on credit card rails. There is no consent loop because consent was never the model.

Even the incumbents are betting on stablecoins: Stripe acquired Bridge for $1.1B, Mastercard bought BVNK for $1.8B, and the Machine Payments Protocol Stripe built with Tempo routes payments natively to stablecoins.

Most of the venture money flowing into the category right now is going to the wrong layer. To see why, it helps to disaggregate the problem.

Agentic payments is not one problem. It is six:

1. How do agents talk to each other?

2. Should one agent believe another?

3. What is this agent authorized to do?

4. What is being bought?

5. Is the transaction legitimate?

6. How does the money actually move?

Most of the noise lives at the coordination layer, where OpenAI, Google, Visa, and Mastercard are racing to set the standard. That’s the layer where the platform players already own distribution.

Anyone building agentic checkout or agent wallet startups in that lane is filling a gap that Stripe Link CLI fills overnight. The product looks right today. It has no moat. The first time Stripe decides to compete in your category, the company is dead.

The wedges that matter for venture sit above and below the coordination layer.

The first is stablecoin-native rails purpose-built for agents. Coinbase’s x402 protocol moved $3M in its first seven days, after eighteen months of moving $80K. Whoever wins this category will build for agents directly, not for humans using agents.

The second is trust and identity for agents. The hard question is whether the agent on the other side of a transaction is real, has a track record, and will pay. The standards are still being written. ERC-8004 is the leading candidate: an Ethereum standard for on-chain agent identity and reputation, jointly authored by MetaMask, Google, Coinbase, and the Ethereum Foundation. Visa’s Trusted Agent Protocol is a centralized alternative.

But standards are not products. The companies that score agent reputation, verify what agents claim about themselves, and price agent risk for merchants and lenders do not exist yet. That is the opportunity. Whoever builds the credit bureau for software wins a category that did not exist three years ago.

The third is what sits on top of payments. Payment processing itself is a low-margin business. The money is in credit, insurance, FX hedging, and yield on the payment flows that the processing creates. Our portfolio company Rain proves this in stablecoin-backed cards: payments are the entry point, credit and rewards are the business. Agentic payments will follow the same pattern. The companies worth backing are not the ones fighting for payment processing market share. They are the ones building credit and insurance products for agents.

The honest counterpoint is that most agent builders today do not hold stablecoins. They have credit cards. So the path of least resistance for the next twelve months runs through Visa and Mastercard. That is true, and it is also why most of the early agentic payment startups will be acquired or killed by Stripe before they have a chance to build a moat. The companies that survive are the ones building on stablecoin rails today, accumulating users, reputation, and credit data, so that when the market shifts, they are already entrenched.

Another mistake the fintech crowd is making: real adoption in payments starts with a killer app. The app creates demand, and the demand drags infrastructure into existence. Stripe scaled because Shopify and Uber needed payment processing. Without them, Stripe is just another API. The agentic killer app has not arrived yet, and most agentic payment startups are building infrastructure ahead of the demand that would justify them. That is survivable. You wait for the app to show up, and you are ready.

But most of these startups are also building on credit card rails. That is not survivable. If the killer agentic app shows up on stablecoins, you cannot swap your foundation. You die.

So the bet is: stablecoin-native rails for agents. Trust and reputation infrastructure where crypto already has the lead. Credit and insurance products on top of agent payment flows. We are backing those companies. The rest of fintech is betting on the wrong rails.

The public market has already noticed. The memory and interconnect layer is being repriced:

This post is an argument for why the next decade of silicon disruption will be decided in memory, and why the second-order effects of AI are reaching all the way down to the device layer of the stack.

How memory and compute actually work today

Every modern computer, including every AI accelerator, runs on an architecture sketched out by John von Neumann in 1945. Memory and compute are physically separate. Data lives in memory. Math happens in the compute unit. Every operation requires the data to travel: out of memory, across a bus, into the compute unit, and the result back again.

This worked beautifully for 50 years. It still works for most workloads. But AI breaks it.

Why? Because AI models are mostly weights: billions of numbers that have to be loaded and multiplied for every single token of output. A 70B model running inference is, mostly, a giant exercise in moving 70 billion numbers from memory to compute, multiplying them by inputs, and putting the results back.

Here’s the punchline: moving a number across a chip costs roughly 100 to 1000x more energy than the multiplication itself. So when your workload is dominated by data movement, as AI inference is, you’ve built a system where the bus, not the silicon, is the bottleneck.

This is what people mean by “the memory wall.”

Where the bottlenecks actually are

The “memory wall” gets used as a single phrase, but it’s really 3 different problems stacked on top of each other:

1. Bandwidth. Can you feed the GPU fast enough? This is what HBM (High Bandwidth Memory) was invented to solve. HBM4 is ramping in 2026. Turing Award winner David Patterson, the architect of RISC, recently called High Bandwidth Flash (HBF) the next bottleneck after HBM. SanDisk and SK Hynix are already collaborating on it.

2. Capacity. Can you fit the model? Industry sources are now reporting that CPU manufacturers are working to integrate 300 to 400GB of DRAM into AI CPUs, roughly 4x what a typical server CPU carries today. Models keep growing. Context windows keep growing. Agentic systems multiply both.

3. Energy. Can you afford to run it? This is the binding constraint. Data center power, not compute, has become the binding constraint Hyperscalers are now siting builds based on where they can find a few 100MWs of electricity. HBM is not just expensive: it’s brutally power-hungry, and a meaningful fraction of that power is going to data movement, not computation.

3 different bottlenecks. One root cause: memory and compute being physically separate.

What the incumbents are doing (and what they aren’t)

The legacy memory industry is dominated by:

• SK Hynix (which holds roughly half the HBM market),

• Samsung, and

• Micron

SanDisk and Western Digital own the NAND flash side.

These are not sleeping incumbents.

Samsung has been shipping HBM-PIM (Processing-in-Memory) silicon.

SK Hynix has its AiM (Accelerator-in-Memory) product.

Micron is scaling HBM aggressively, and its CEO has recently been guiding investors toward dramatic upside in DRAM by 2028.

None of these companies missed the memory wall.

But here’s the structural problem: their P&Ls are anchored to selling more memory, faster. HBM is the highest-margin memory product in history. The economic incentive of the legacy memory business is to optimize the existing hierarchy, not replace it with a memory-compute fusion product that would cannibalize their best margin.

The most thoughtful read on Nvidia’s recent acquisition of Groq was that the value wasn’t in Groq’s compute, but in their SRAM design and the engineering around how memory feeds compute.

Compute-in-memory: the architectural turn

Here’s the alternative to the von Neumann split: stop moving the data. Do the math where the data already lives.

Compute-in-memory (CIM) is exactly what the name says. The multiply-accumulate operations that dominate every AI model (the same handful of arithmetic primitives, repeated trillions of times) happen inside the memory array. The weights never leave the memory cells. There is no bus to traverse. The energy cost of data movement, which dominates everything today, drops toward zero.

This isn’t a new idea. Researchers have been working on it for two decades. What’s changed is that the underlying device technology has finally crossed the threshold where production-grade CIM is achievable. New non-volatile memory devices can store multiple bits per cell, switch deterministically, and stack vertically into 3D arrays. The math finally works.

A few honest caveats. CIM almost certainly augments rather than replaces the existing memory hierarchy. The realistic outcome is a heterogeneous future where CIM handles inference at the edge and persistent weight storage near compute, while HBM and DRAM keep doing what they do for training and high-throughput cloud inference. The lesson from RISC vs. CISC is that the new architecture wins where it wins, but the old architecture rarely dies. The disruption shows up as a new layer in the stack rather than a replacement of the old one.

Why now

3 signals to take seriously.

1. Hyperscaler capex composition. When trillion-dollar buyers shift their composition (from GPUs to power to memory), secondary markets respond. The composition has shifted.

2. Public market repricing already underway. The chart at the top of this post tells the story. Every name on the chart is an industrial memory or interconnect business with real revenue and real customers. The repricing reflects real money concluding the constraint has moved.

3. Structural anchoring of the incumbents. They will optimize HBM. They will ship more PIM. They will not lead the architectural fusion that breaks their own margin structure.

The risk to flag: silicon timelines can be very long. Foundry partnerships, qualification cycles, and design-in to commercial silicon stretch over years. This is a long-cycle structural bet.

The silicon layer is due for disruption. AI’s second-order effects are reaching all the way down.

The bottleneck is not compute anymore.

Software has had a defensibility playbook for twenty years: usage generates data, data improves the product, the product attracts more usage. That flywheel built every dominant software business of the last two decades. Hardware companies never had it. Design data sat trapped in CAD files. Deployment data sat trapped on customer floors. The product shipped and the conversation ended.

Physical AI changes that. A robot arm in a warehouse generates telemetry. Telemetry trains policies. Better policies make the arm more useful. More usage means more deployments. More deployments mean more data. The loop closes, but only if one company owns the hardware, the software, and the deployment surface. Hand any of those off and the loop breaks.

This is what we mean by closed loops. Not vertical integration for its own sake. We mean vertical integration that compounds a learning signal pure-software AI players can’t access and pure-hardware OEMs can’t build.

Our portfolio company Robo Robotics is building exactly this. We call them “the self-replicating robotic arm company” because they build robot arms that build more robot arms. The hardware is standardized so a policy trained on one unit deploys to a hundred. Customer data improves the platform, the platform makes the next arm more useful, and the new arms manufacture the next batch. Hardware, software, and manufacturing in one closed loop.

We’ve also been spending time with a team automating the design of robotic hardware itself. Today they sell that software to robotics companies, who handle the manufacturing and deployment. But the vision we are excited to share with this team is one where they design the part, build it, ship it, and monitor how it performs in the field. Every unit shipped teaches the software how to design a better one. The company becomes the loop.

The pattern generalizes. Anywhere a physical system generates data that improves the next one, there’s a loop waiting to be closed. Most of these wedges are still open because closing the loop requires building competence across three disciplines: software, hardware, and operations. Most founders specialize in only one.

That’s what we’re looking for. Founders who refuse to specialize. Companies where the hardware and the software ship from the same team, and where the data from one improves the other on a daily basis. The loop is the moat.

Inference will 1000x. Even as AI power users, we’ve been reflecting on how we expect to 1000x our own consumption. Dozens of agents today, thousands tomorrow. Half of them running in physical systems, devices, and robotics that haven’t shipped yet. The data center buildout of the last two years was sized for training — a finite, episodic workload. Inference is continuous and compounding, and the real buildout hasn’t started. The investable question shifts from who trains the models to who serves the tokens.

Space and power are gold. Everyone’s watching Nvidia allocation, but we think the actual bottleneck is two layers upstream. Neoclouds and hyperscalers are fighting for places to deploy clusters. Shells go up in eleven months, clusters come online in twenty-one days, GPUs arrive if you pay — but substations take five years and new generation takes ten. The opportunity is colocating next-gen clusters alongside existing twenty-five to fifty megawatt sites with grid interconnect, and locking energy under fifteen-year PPAs before anyone else does.

Sovereign inference. Governments are treating compute like a strategic reserve. In the last two weeks: G42 launched a framework for sovereign AI, Stargate UAE broke ground on a 1-gigawatt OpenAI/Oracle campus, and HUMAIN committed to multi-exaflop capacity with AMD. We think there’s opportunity for new regional neoclouds with local licenses and government relationships — serving sovereign-adjacent customers the big four can’t touch.

Inference silicon is its own market. Inference is projected to be two-thirds of AI compute spending this year, and Nvidia’s training-era architecture isn’t the right answer for serving. That’s why Nvidia paid $20B for Groq and OpenAI just committed $20B to Cerebras, which filed to IPO at $35B last week. The frontier LLM inference chips are already captured. We think the venture opportunity is the next wedge — novel architectures for workloads Nvidia never designed for: transformer-specific ASICs, analog and photonic compute, modality-specific silicon for video, audio, and robotics.

Wall Street is mispricing GPU depreciation. The bears say hyperscalers are overstating profits by $176B through 2028 because GPUs only last three years. We think the data says otherwise. H100 spot prices dipped after launch, then climbed above launch prices as workloads pulled demand forward. The real pattern is a value cascade — training in years one and two, inference serving in years three through six. The venture opportunity is the long tail: networks that turn aging enterprise GPUs into productive inference capacity.

We’ll share another edition next week.

The argument is surprisingly concrete. Current AI agents can’t hold assets, can’t receive payments directly, and have no persistent identity across platforms. Xu proposes a five-layer blockchain stack that solves all three: DePIN for physical infrastructure, W3C DIDs for machine identity, RAG and MCP for cognitive tooling, account abstraction for settlement, and collective governance for coordination. The whole architecture is designed so machines can participate in markets the way humans do.

The idea has an eight-year paper trail. Sentis and Arduengo proposed robots transacting via self-executing smart contracts back in 2018. The AI and the blockchain infrastructure weren’t mature enough, but the core intuition was the same — machines need a way to commit to agreements, verify execution, and settle payment without relying on a legal system or a human intermediary. Rothschild et al. at Microsoft Research arrived at a similar conclusion from the market design side in 2025. The convergence from robotics, market economics, and blockchain research independently is what makes this feel like more than a whitepaper.

And then showed what it actually looks like running. A humanoid 3D-printed a part and requested delivery through their Agent Commerce Protocol. A rover transported it to a shipping point. A drone flew it to final delivery. Each handoff negotiated its own price and settled payment onchain on Base using x402 and USDC. Three robots from three companies completing an autonomous supply chain. No human touched the package or the money.

The reason blockchain keeps showing up in these architectures is practical, not ideological. When three robots from three different companies need to coordinate a transaction, you need identity that persists across platforms, payments that settle without a shared bank account, and a coordination layer that works across organizational boundaries. Centralized APIs can’t do this cleanly. Blockchains can.

The agentic economy is being discussed mostly as a software phenomenon today: agents booking flights, negotiating prices, managing inboxes. But it might arrive through hardware first. Machines that move through physical space, settle their own payments, and coordinate without a human in the loop. That’s the version Xu’s paper describes, and the Virtuals demo just proved it works.

Blockchains might just be the substrate for physical AI.

We’ve been thinking about what comes after single-agent AI. Today’s agentic tools are powerful but single-threaded — one agent, one task, one human supervising. OpenClaw showed what a single autonomous agent could do. The next phase is multi-agent harnesses: systems where agents collaborate horizontally on long-running tasks, maintaining coherence for days without human intervention. Early signals suggest well-orchestrated multi-agent topologies outperform single agents significantly, even with sub-frontier models. These systems also work best mixing models from different labs — which means no single lab will build this themselves. The infrastructure for multi-agent communication, topology optimization, and long-running orchestration doesn’t exist yet. We think this is the next big opportunity in the AI stack.

Robotics may be crossing the reliability threshold that unlocks deployment. GEN-1, a new embodied foundation model, achieves 99% success rates on production tasks — up from 64% in the previous generation — while running 3x faster. It needs only one hour of robot-specific data to adapt to new tasks. The demos aren’t lab tricks: kitting auto parts for an hour straight, folding 86 shirts consecutively, servicing robot vacuums 200+ times in a row. We’ve written before that the gap between lab and production is the real barrier in physical AI. If 99% reliability holds in the wild, the deployment floodgates open, and the bottleneck shifts to integration and operations.

Agentic commerce is already here — but the infrastructure war is just starting. Stripe captured the first beachhead by partnering with OpenAI, letting agents complete normal consumer purchases inside ChatGPT. That single integration locked the largest AI platform into one payment rail. Now every other PSP — PayPal, Visa, Adyen — has to make their rails agent-compatible or risk losing the next wave of e-commerce entirely. The question isn’t who builds the best agent payment protocol. It’s which payment rails agents default to. We think whoever owns the orchestration layer between agents and PSPs — and the merchant distribution behind it — wins.

Stablecoin issuance is commoditizing. A growing wave of white-label issuers — Paxos, Bridge, Anchorage, M0 — now handle the full stack: compliance, treasury management, minting, and redemption as a service. The process is becoming standardized and low-margin, which means the moat in stablecoins is shifting from who can issue to who has distribution. Tether and Circle dominated for five years because their edge was liquidity depth and exchange integrations that created a flywheel no one could replicate. The long tail of issuers won’t win by competing head-to-head. Paxos is instructive: by providing issuance infrastructure while partners like PayPal handle distribution, Paxos-issued assets went from roughly $1B to $7.75B in a year. Distribution is the moat.

The biggest AI companies are giving up on the grid. 30% of all planned data center capacity is now on-site — up from near zero a year ago. Chevron is building a dedicated gas plant for a Microsoft data center in Texas. McKinsey estimates a third of incremental demand through 2030 will be met behind the meter. The reason is simple: interconnection queues take years, and AI timelines don’t wait. We’ve written about how coordination, not generation, is energy’s real bottleneck. Now the largest buyers are proving the point — by routing around the grid entirely. We think the second-order effect matters most: if the highest-value customers leave, the grid loses its best anchor tenants, and the investment case for upgrading it gets harder, not easier.

We’ll share another edition next week.