What separates this buildout from other buildouts in the past doesn’t come from size. It comes from capacity being bought before it even exists. And I’ve compiled the evidence, stacked in order of contractual weight.
Exhibit one: Oracle. Oracle has reached US$553 billion in RPO (Remaining Performance Obligations), up 325% year over year, which is revenue contractually committed but not yet delivered.35 For clarification, RPO is not a forecast, pipeline, or management aspiration. It’s an accounting category governed by ASC606 reserved for signed contracts.36 Why I bring up Oracle’s RPO is because there is half a trillion dollars of committed demand sitting on the books of a company that has yet to construct the buildings to serve it. The primary constraint disclosed in Oracle’s filings is not sales but physical infrastructure constraints like concrete, transformers, and time.37
Exhibit two: Microsoft. Microsoft has disclosed a backlog of roughly US$80 billion in Azure demand that it is currently unable to serve.38 Despite describing building at the quickest rate they’ve ever built, Microsoft is being forced to take computing power away from their own internal products to preserve capacity for their paying cloud customers.39 Think about that for a bit. One of the most valuable companies on earth is rationing compute away from their own products that they have been selling for the past twenty years, because external customers have contracted more computing infrastructure than what currently exists today.
Exhibit three: Amazon. AWS sits at a revenue run rate of about US$142 billion and consistently reports that its capacity is monetized as quickly as it’s being built.40 The whopping 3.9 GW added in 2025 did not create any form of buffer for Amazon; it was all absorbed and used on arrival.
Exhibit four: CoreWeave. The biggest neocloud reports a revenue backlog of US$99.4 billion against 3.5 contracted GW.41 I will go more in depth with CoreWeave’s balance sheet later. But for now, know that their backlog is made of signed take-or-pay commitments from the largest technology companies in the world.42
There are softer signals that round out the exhibits. Meta, with a spending order of US$130 billion a year, reports that it frequently gets regular inbound offers for their computing power.43 Offers so frequent that it created enough pressure for Meta to launch its own external compute business within months.44 We also have evidence that requires no disclosure at all because it is structural. But the existence of neoclouds, like Nebius and CoreWeave, companies that rent GPU capacity at a premium, shouldn’t even exist. In a market where hyperscaler supply met demand, these companies would not be as relevant as they are today. The existence of neoclouds reflects the fact that hyperscalers spending US$700 billion a year just isn’t enough. The existence of the middleman is itself the demand signal.45
But a contract is only as good as the balance sheet behind it, and Oracle’s half-a-trillion-dollar backlog is owed by a company that isn’t even making a tenth of that in cumulative revenue.46 Calling that “pre-sold” is a bit of a stretch. A commitment from a buyer who cannot yet pay is speculative capacity disguised as contracted capacity. If the entire order book looked like that, I would not be writing this. This is just a piece of the picture. By using who stands behind these contracted commitments, I’ve graded the demand into three tiers.
Starting from the first tier that doesn’t announce itself as demand, hyperscalers are buying from themselves. Microsoft is rationing compute away from its own products. AWS is monetizing fresh capacity on arrival. This is demand in its purest form. No external counterparty, no customer default risk, and no question of whether the demand is real because the hyperscaler is self-aware of the capacity that it needs. The customer is the infrastructure owner. The second tier would be contracted demand to buyers who could clear the bill multiple times over: Meta being accountable for 21% of CoreWeave’s backlog is a good example.47 Essentially, a promise from someone who keeps their promises. Finally, we have our venture-stage labs, OpenAI, which is the loudest of them, signing for capacity that reflects their ambition for tomorrow over what is actually possible today. A demand signal, but it shouldn’t be confused with demand security like the hyperscalers I mentioned above. It reflects their ambition but not what they can actually pay.48
Therefore, the honest version of the argument is narrower than the headline claim: the buildout is pre-sold to the extent that the buyers are real. The problem with the market is not that they are pricing the third tier skeptically. They should. The real problem is that they price the self-funded and investment-grade majority of the demand at tier three discounts. Not a single person has rewarded these companies for trying to build the capacity they need.
Plot Part I’s supply figure with this section’s demand evidence on the same chart and you’ll get a strange picture: the fastest capacity expansion in corporate history, building capacity flat out, constantly behind its order books. Every quarter the builders add reactors’ worth of power, every quarter their backlog grows faster. Just think about it for a second. As I said before, this is all publicly disclosed information. Everyone has seen this chart. Which begs the question this essay exists to answer: Why can’t it read it?