No One Knows What AI Infrastructure Is Worth

In the same July week, investors delivered three verdicts on AI infrastructure. They punished Tesla for spending too little — only $2.5 billion of a planned $25 billion in 2026 capex, with the stock down 18% for the year ^[1]. They punished Meta and Microsoft for spending too much, sending Meta down 12% and Microsoft down 20% despite ad revenue growing 33% and Azure accelerating 40% ^[2]. And they crashed memory-chip stocks on fears the whole sector had overbuilt: SanDisk fell 36%, Western Digital 35%, Micron 24% ^[3]. Three signals, arriving together, that cannot all be correct. The contradiction is not a disagreement about the right level of capital expenditure. It is something more fundamental: the market has lost the revenue model that would make any level of capex rational. When a company is punished for underbuilding and another is punished for overbuilding in the same week, the market is not calibrating. It is guessing. The natural explanation a reader reaches for is cash flow. Surely the companies weathering this are the ones funding AI infrastructure from existing profits rather than debt. The evidence does not cooperate. SpaceX generates $2.17 billion a month in AI revenue from its SpaceXAI unit and holds $22 billion in government launch contracts, yet its stock has fallen below its $135 IPO price to roughly $123 ^[4]. Meta's advertising revenue grew 33% and the stock is down 12% ^[2]. Microsoft's Azure grew 40% and Copilot reached a $37 billion annual run rate, and the stock is down 20% ^[2]. Non-AI revenue, it turns out, offers no protection from the repricing. What separates the companies holding ground from those being repriced toward junk is something narrower: customer diversification and the ability to pivot idle infrastructure into alternative revenue. The contrast between Oracle and Meta makes the point with unusual clarity. Oracle has a profitable database business. But it bet approximately half of its $638 billion remaining performance obligation on a single customer, OpenAI, through a $300 billion cloud deal, and it funded the buildout with $176.9 billion in long-term liabilities and plans to raise another $40 billion in debt and equity ^[5]. When OpenAI's IPO was delayed and questions about its financial stability surfaced, that risk transferred directly to Oracle's balance sheet. The stock has lost 65% of its value, falling to $121.50, and S&P downgraded its credit to BBB-, one notch above junk ^[5]. Meta faces the same overcapacity risk. Its $125 billion to $145 billion AI capex forecast nearly equals its entire projected 2026 EBITDA of $145 billion ^[6]. But Meta made a different move. In July it launched Meta Compute, a cloud unit designed to sell excess GPU capacity externally, and it is in talks to lease up to $10 billion of compute to Anthropic over two years ^[7]. Mark Zuckerberg made the logic explicit:

We haven't done that yet, because we think that we have a use for the compute. But obviously, if we get to a point where we feel that we have overbuilt, then that is an option that we have, and that is partially what gives us confidence in investing in building this out. — Pavel Durov

He said the quiet part out loud. The cloud unit exists because the company might have overbuilt, and that option is what gives confidence to keep building. Meta is turning a potential liability into a revenue stream before the liability materializes. The same pattern appears elsewhere, in quieter forms. Amazon secured $225 billion in AI chip revenue commitments for its custom Trainium accelerators, cutting its dependence on Nvidia and building a customer base that spans multiple model labs ^[8]. Alphabet funds its $180 billion to $190 billion in projected 2026 capex from search and advertising margins, with a Google Cloud backlog that has nearly doubled to exceed $460 billion ^[9][10]. Each has built an escape hatch: a way to monetize the infrastructure they are building even if the AI revenue model that justified it never fully arrives. The contradictory signals are a symptom of a deeper problem. The pricing signal for AI infrastructure depends on a revenue model nobody can see yet, and the only companies with a margin of safety are the ones who can turn what they have built into something else if the revenue never comes.