$340 billion a year. That is the annualized pace of US large-computer imports as of March 2026, a figure that arrived quietly in a trade data release and did not get nearly the headline it deserved. Not because it is boring, but because it is almost too clarifying. The AI boom is not primarily a software story, and it never was. It is a logistics story, a real estate story, and a power story. The chips have to live somewhere, and getting them there costs far more than most people are tracking.
That number lands differently when you read it alongside two other things from the same week: Google and Blackstone announcing a large joint venture to build AI cloud infrastructure, and the quieter news that Microsoft’s custom Maia chips may soon run Anthropic’s models. Two different companies, two different supply-chain strategies, both arriving at the same conclusion. The commodity in shortest supply is not intelligence. It is compute, meaning physical, power-hungry, capital-intensive compute, and the scramble to secure it is where this cycle’s returns are actually being decided.
Google, Blackstone, and Where the Returns Are Priced
The Blackstone partnership is the more revealing of the two signals, precisely because Blackstone has no business being in an AI story. It is a real estate and private equity firm. The fact that one of the world’s largest alternative asset managers is now co-building AI cloud infrastructure with Google tells you where the returns are being priced, and it is not in the model. It is in the ground the model runs on: data center real estate, power capacity, cooling systems, and the unglamorous plumbing that makes everything upstream possible.
The division of labor in that venture is the argument in miniature. Blackstone brings capital and site acquisition expertise, which is to say the ability to find and finance land near power that can actually be delivered. Google brings engineering and hyperscaler relationships. Neither party would put it this way in a press release, but the subtext is that both believe the infrastructure layer is where value accretes, and both are moving before the next wave of demand makes moving prohibitively expensive. Site acquisition is not a business with a fast second act. Once the parcels near transmission capacity are spoken for, the people who did not move are not competing on price. They are waiting in line.
Microsoft’s Maia and the Custom Silicon Race
The Maia development is subtler and points the same direction. The premise of custom silicon is a trade: you give up generality and gain efficiency, surrendering the ability to run anything in exchange for running one thing extremely well at a fraction of the cost per token. That trade only makes sense at volume, which is itself a statement about how much volume these companies expect. Nobody designs a chip for a workload they think might plateau.
Amazon has Trainium. Google has TPUs. Now Microsoft is bidding to power a frontier lab’s workloads with its own silicon. Every major cloud provider is building its own inference chips, which means the hyperscaler contest of the next five years will not be settled on general-purpose GPUs alone. It will be settled on who owns the most efficient path from prompt to response, because at sufficient scale the cost per token is the product. A model that is marginally better but meaningfully more expensive to serve loses to a model that is adequate and cheap, and every provider building custom silicon is betting on exactly that outcome.
Set the two moves side by side and they are the same wager at different layers. Blackstone and Google are buying the physical floor. Microsoft is optimizing what happens on it. Both are acting on a conviction that the scarce input is not cleverness but throughput, and both are locking in position while locking in is still affordable.
CoreWeave, Nebius, and the Independent Infrastructure Question
This is where the CoreWeave versus Nebius debate comes in, and it is genuinely interesting to sit with, though not for the reason it usually gets raised. The framing itself is the tell. We are now at the stage where “which AI infrastructure stock” is a retail-accessible question, which reliably means the institutional money has already taken its position and is now looking for someone to sell to at a better price.
The smarter question is not which of them is the better buy. It is whether the independent infrastructure play survives the next five years at all while the hyperscalers build their own everything. An independent provider’s pitch rests on being faster and more specialized than a hyperscaler. That pitch works beautifully while the hyperscalers are supply-constrained and slow, and it gets much harder the moment they have their own chips, their own sites, and their own financing partners lined up for the next build cycle. The independents are not competing with a product. They are competing with a balance sheet that has decided to vertically integrate.
Bill Gates writes in Source Code that the circumstances shaping his career were “mostly out of my control,” acknowledging the unearned advantage of being born where and when he was, with access to machines at a moment when access to machines was the whole game. The honest read on this week’s infrastructure moves is similar. Google and Blackstone are not smarter than their competitors. They reached scale earlier, built relationships earlier, and those relationships are now compounding into positions that will be very difficult to replicate on any timeline that matters. Infrastructure moats do not announce themselves. They just get wider, one long-lead contract at a time.
What the IPO Chatter Actually Signals
Which brings us to the offering news, the one item most people are treating as the leading indicator. SpaceX and OpenAI are both reportedly preparing for enormous public offerings, and the analyst take is that mega-offerings cluster near market tops. As a historical pattern that is true and worth respecting. The more useful question is what kind of top, because the label covers two very different situations.
A valuation top driven by speculation reverses quickly and painfully, and the underlying business is largely unaffected once the multiple resets. A structural top, where the build-out has run ahead of near-term demand, takes years to resolve, because the capacity is already poured and financed and nobody can un-build a data center. Those are different problems with different time horizons and different survivors, and treating them as one story is how people end up right about the direction and wrong about everything that matters.
The market can stay irrational longer than anyone expects, and meanwhile the imports number keeps running. $340 billion a year in large computers is not the spending pattern of people hedging against a blip. It is the spending pattern of people underwriting a decade, and it is committed capital rather than announced intention. The offering chatter is noise about sentiment. The chip imports are signal about conviction, and the two are being read as though they were the same measurement.
What most observers miss is that the infrastructure wave and the frothy offering narrative are both true at once, and they are not in contradiction. The build-out is real, capital-intensive, and durable. The valuations placed on top of that build-out may or may not survive contact with reality. You can believe entirely in the foundation and still be skeptical of the price being charged to stand on it, and the next few years will probably reward exactly that combination.
The infrastructure is the argument, for the companies building it, for the capital flowing into it, and for every model that runs on top of it. The software gets all the headlines. The plumbing gets all the returns.
Leave a Reply