Who Owns the Compute Owns the Story

Strip the mission statements, the keynote theatrics and the founder mythology away from the biggest companies in artificial intelligence, and ask a flat question: what do they actually own that a rival couldn't replicate by Friday? Not the models — those leak, get distilled, get cloned within weeks of release. Not the talent — researchers move between labs the way consultants move between clients, and they take the good ideas with them. What's left, increasingly, is the compute: the chips, the packaging capacity that turns those chips into systems, the data centers that house them, and the power contracts that keep them running. The compute is the company now. Everything else is narration.

I want to be precise about what I'm claiming, because this is the kind of thesis that sounds like a slogan and dies as one. I'm not saying ideas don't matter, or that a better architecture can't humiliate a fleet of expensive accelerators. I'm saying that when you stop reading the press releases and start reading the contracts, the durable asset — the one that's expensive, slow to build, and impossible to fake — is physical. It is silicon, square footage and electricity. And it is being locked up, right now, by a very small number of buyers.

Follow the packaging, not the press release

Start at the chokepoint nobody puts on a slide. The constraint in AI hardware in 2026 is not lithography — it's advanced packaging, the step that fuses a GPU die to its high-bandwidth memory. TSMC's CoWoS lines are the bottleneck, and the company is racing to widen it, scaling from roughly 35,000 wafers a month at the end of 2024 toward a projected 130,000 a month by the close of 2026. That is an extraordinary build-out. It is also already spoken for.

Nvidia has reportedly booked over 60 percent of TSMC's total CoWoS capacity for 2025 and 2026, with industry trackers putting its 2026 allocation near 595,000 wafers — the lion's share earmarked for CoWoS-L, the variant that feeds its next-generation Vera Rubin systems. Reserve the majority of the world's most advanced packaging, for years out, and you have done something more durable than ship a great model. You have decided who else gets to compete. AMD, Google's own silicon team, the second-tier ASIC vendors — they aren't fighting Nvidia on benchmarks first. They're fighting over what's left of the wafers.

You don't need to win the argument about whose model is smartest if you've already bought the only factory that can build the hardware to run it.

This is what a moat looks like when it's made of concrete and not of vibes. It isn't a patent or a brand or a network effect. It's a multi-year purchase order against a supply line that takes years and billions to expand. You don't need to win the argument about whose model is smartest if you've already bought the only factory that can build the hardware to run it.

The strongest counter-argument — which nearly changed my mind

I'll concede the best version of the case against me, because it is genuinely strong and it is the scenario I find myself rooting for. Ideas and talent have routed around capital before. The whole history of computing is littered with incumbents who owned the expensive thing — the mainframe, the fab, the proprietary stack — right up until a cleverer, cheaper architecture made the expensive thing irrelevant. A breakthrough in model efficiency, in sparsity, in training method can, in principle, make a great deal of very expensive silicon suddenly unnecessary. We have already seen tremors of this: efficiency jumps that let a small team do with a modest cluster what was supposed to require a nation-state's worth of GPUs.

That's real. And it is precisely the story the narration is built to keep you watching, because it is the flattering one — the David-and-Goliath frame where genius beats balance sheet. I was too sanguine about a version of this argument once before; I genuinely believed open weights and open research would act as a structural check on concentration, that you couldn't fence off intelligence the way you fence off oil. I was wrong about the timeline, at least. The models did commoditize. The compute to train and serve them at the frontier did not.

Where the money is actually committed

Here is the test I trust more than any keynote: not what companies say is the future, but what they are signing legally binding, multi-year, balance-sheet-bending contracts to secure. On that measure, the verdict is not close.

The four largest hyperscalers — Amazon, Alphabet, Microsoft and Meta — are guiding toward roughly 700 billion dollars in combined capital expenditure for 2026, close to double what they spent the year before. Amazon alone has pointed to around 200 billion dollars; Alphabet to 175–185 billion, up from 91 billion in 2025; Meta to 115–135 billion; Microsoft to 110–120 billion. These are not research budgets. They are construction and procurement budgets — buildings, transformers, transmission and the chips to fill them. People do not commit a trillion dollars over two years to a thing they believe a clever algorithm is about to make obsolete.

• Compute: Nvidia's commitment to invest up to 100 billion dollars in OpenAI as at least 10 gigawatts of its systems are deployed — with the first phase targeted for the second half of 2026 on the Vera Rubin platform.
• Real estate: OpenAI, Oracle and SoftBank's Stargate program, a 500 billion-dollar build-out, with the Abilene flagship scaling toward 1.2 gigawatts and more than 5 gigawatts of capacity across sites under development.
• Power: at least 13 announced nuclear deals committing roughly 9.8 gigawatts to AI load — including Microsoft's 20-year contract to restart Three Mile Island Unit 1, Amazon's investment in X-energy's reactors alongside its Susquehanna campus, and Meta's multi-gigawatt agreements with Constellation and a roster of new reactor developers.

Read that list as a strategist rather than a fan. The bets are not being placed on whose model wins. They are being placed on owning the inputs to any model that could win: the silicon, the square footage, the electrons. That is a wager, made with balance sheets rather than slogans, that this time capital is the durable thing.

Power that hides behind a thermostat

The most underrated line in all of this is the boring one: more than 60 percent of hyperscaler spending now goes to power and the physical plant, not to the chips themselves. That reframes the whole industry. The binding constraint on frontier AI is shifting from how clever you are to how many gigawatts you've contracted, on what terms, and for how long. When companies start signing 20-year power-purchase agreements and resurrecting nuclear plants, they are not buying electricity. They are buying the right to keep computing at a scale their competitors can't match — and locking that right away from everyone who showed up to the auction late.

This is the part the mission statements are designed to obscure. A keynote about democratizing intelligence plays a great deal better than a slide showing that you have quietly cornered a region's spare power and several years of the world's advanced packaging. But the second slide is the company. The first one is the narration over it.

What it means for who's actually in charge

If the durable asset is compute, then power in this industry sits with whoever controls the bottlenecks: the foundry that allocates packaging, the handful of buyers rich enough to reserve it for years, the utilities and reactor developers writing 20-year offtake terms, the governments deciding whose grid gets the next gigawatt. The model labs we treat as the protagonists are, increasingly, tenants — brilliant ones, but tenants — on infrastructure they rent from someone with a harder asset.

I could still be wrong, and I'll tell you exactly how I'd know. If a genuine efficiency revolution arrives — if the frontier can be reached on a tenth of the silicon and a tenth of the power — then the contracts I've been pointing to become stranded assets, and capital will have bought itself a very expensive monument to the last paradigm. That is the scenario worth watching for, and I'm watching. But notice that even in that world, the question that decides everything is still a compute question: how much you need, and who controls it. The narration changes. The subject of the sentence does not.

So when the next launch arrives wrapped in a story about intelligence and benevolence and the future of work, do what I do. Skip to the back. Find the line about wafers, gigawatts and the length of the contract. That's not the footnote. That's the company. Everything else is narration.