The capex is the strategy

Ask the people running the largest companies in artificial intelligence what their strategy is, and you'll get a tour of mission statements: superintelligence, abundance, curing disease, the long arc of human flourishing. Listen to the keynotes long enough and you start to believe the strategy is a destination — a thing the spending is meant to get them to. It isn't. The spending is the strategy. The trillion dollars of concrete, silicon and power contracts is not a means to a plan you'll be told about later. It is the plan, fully expressed, sitting in plain sight on the balance sheet. Everything said into a microphone is downstream of it.

This is a power column, so let me be precise about what I mean. I'm not making the markets argument — that the numbers are too big, the revenue too thin, the loop too circular, and that it ends in a crash. That case exists and may even be right. But it's a valuation question, and valuation questions get settled by quarterly earnings. The question I care about is older and harder to undo: when a handful of firms commit capital on this scale, what they are really buying is not capacity. It is the right to set the terms for everyone downstream of them. The build-out is a claim on who decides.

Read the figures as a sentence, not a spreadsheet

Start with what's actually being committed in 2026, because the scale is the argument. The four big hyperscalers — Amazon, Microsoft, Alphabet and Meta — have guided to a combined capital expenditure in the region of $625 billion to $725 billion this year, depending on whose tally you trust, up something like 70 percent on 2025's already-record figure. Microsoft alone pointed analysts toward roughly $190 billion. Meta lifted its range past $145 billion. These are not research budgets. They are the kind of numbers that used to belong to nation-states building railways and grids.

Then look at the demand side, where the commitments are even more vertiginous. OpenAI's Sam Altman told the world in November that the company is looking at roughly $1.4 trillion in data-centre commitments over eight years — against an annualised revenue run rate he put at above $20 billion. Sit with that ratio for a second. Anthropic, not to be cornered, locked in up to a million of Google's TPUs, well over a gigawatt of compute capacity, in a deal analysts size in the tens of billions a year. Oracle's Stargate arrangement with OpenAI runs to $300 billion. Nvidia put $100 billion into the same orbit. Read as a spreadsheet, these are bets. Read as a sentence, they say something simpler: we intend to be the layer no one can route around.

Capacity is what they're buying. Power is what they're paying for. The two are not the same purchase.

Why capital, and why now

There's a reason the bet is being placed with the balance sheet rather than the slogan, and it's worth saying plainly because it's the whole game. In software, the historic moat was the model, the network, the brand — assets that are cheap to attack and expensive to defend, because a clever idea can route around them by Friday. Compute is the opposite kind of asset. A gigawatt of power, a multi-year silicon allocation, a substation, a water permit, a site with a grid connection that took four years to approve — these cannot be cloned by genius. They have to be bought, and increasingly they have to be bought before they exist. That's why the spending takes the form of forward commitments: not 'we have the capacity' but 'we have pre-committed the next decade of it, and you have not.'

This is the move I want to name. When you sign a multi-year, multi-gigawatt contract, you are not just securing your own supply. You are removing that supply from the market, raising the price of everyone else's ambition, and converting a fungible commodity — electricity, chips — into a structural advantage that compounds. The capex is simultaneously a moat and a message. The moat is the hardware. The message, addressed to every founder, every government, every smaller lab, is: the cost of entry is now a number you cannot raise. Decide accordingly.

The strongest counter-argument, which nearly moved me

I'll give the counter-case its best form, because it's serious and because I've been on the wrong side of exactly this kind of question before. I was too sanguine, years ago, about open source as a natural check on concentration; I assumed the cheap, the clonable and the clever would keep routing around the incumbents indefinitely. So I take the optimistic version seriously now. It goes like this: capital has lost these races before. The mainframe was supposed to be the moat, then the PC ate it. A sufficiently clever architectural breakthrough — a model that does more with radically less compute — could strand all this expensive hardware overnight and make the spenders look like men who built canals the week before the railway. Efficiency is the great equaliser, and efficiency is exactly what a roomful of brilliant researchers is paid to find.

That scenario is real. It's also, notice, the flattering one — the story in which genius beats capital, the underdog wins, and the gatekeepers are humbled by a better idea. It's the story the build-out's own narration is happy to keep you focused on, because it frames the spenders as merely placing a bet that could lose, rather than as parties already collecting rents while you wait to find out. But follow where the money is actually going. It isn't going into hedges against an efficiency breakthrough. It's going into owning the physical layer on the assumption that more compute will always be worth having — and crucially, into owning it whether or not any single model wins. The hyperscalers don't need OpenAI specifically to triumph. They need the era of expensive compute to continue, and they are spending precisely to ensure it does.

Where this leaves who's in charge

Here is the part the valuation debate keeps missing. Even if the bubble talk is correct, even if some of these commitments are written down in a spectacular reckoning, the power question resolves the other way. A crash redistributes assets; it doesn't redistribute them to you. When the dust settles on an over-build, the data centres don't evaporate. They get bought — at a discount — by whoever has the strongest balance sheet, which is to say by the same handful of firms, now with their position cheaper and more concentrated than before. The dot-com bust didn't democratise the internet's plumbing. It handed the survivors a decade of dominance at fire-sale prices. Concentration tends to survive its own excesses. That's not a bug in the bet. For the largest players, it's a feature.

And the legitimacy story is already being written to match. Watch how quickly the framing has shifted from 'we are building a product' to 'we are building national infrastructure' — to the language of sovereignty, security, the things states are supposed to guarantee. When Altman has to publicly insist OpenAI neither has nor wants a government backstop for its data centres, the very fact that the denial is necessary tells you the scale at which this is now being adjudicated. Infrastructure on this scale doesn't ask permission so much as it presents a fait accompli and dares the public to unwind it. By the time a regulator or a legislature has a view, the substations are pouring concrete and the power has been contracted through 2034.

The build-out doesn't ask who should be in charge. It answers the question by making itself too large and too load-bearing to put back in the box.

So when the spenders tell you the strategy is superintelligence, or abundance, or the future of work, take it as narration — sincere, perhaps, but narration. The strategy is the thing they are doing with their money, and the thing they are doing with their money is buying the right to set the terms: who gets compute, at what price, on whose conditions, and who has to ask. That right used to be diffuse, contestable, occasionally held by the clever and the small. It is being consolidated this year, in public, at a scale most governments can't match, by parties who would rather you debate the valuation than the verdict.

The bet, stated honestly, is that this time capital wins — not the argument about capital, but capital itself, the physical, contracted, irreversible kind. I used to think the better idea always eventually routed around the bigger balance sheet. I'm no longer sure that's the century we're in. The trillion dollars isn't reaching for a strategy. It is one. And it has already, quietly, decided a great deal of what the rest of us will be allowed to call our own choices.