On February 5, the market sold off when three companies guided $490-520 billion in AI capex. Amazon fell 10%. The Nasdaq dropped 1.8%. Two days later, on February 7, the Nasdaq closed up 2.18%. Nvidia rose 7.92%. The same spending. The same companies. A completely different verdict. What changed wasn't confidence in AI. It was the realization that the half trillion wasn't really a bet on AI at all.
The Pre-Sold Infrastructure
On February 7, Sherwood News reported that Amazon, Google, and Microsoft had a collective $1.1 trillion in cloud computing backlog — contracted revenue from customers who have already signed multiyear deals. Microsoft alone held $625 billion in commitments.
Put that number next to the capex: $490-520 billion in spending to serve $1.1 trillion in contracted demand. The data centers aren't speculative. They're pre-sold. The infrastructure being built at the scale of national GDPs has customers waiting for it. The February 5 sell-off priced the capex as an AI wager. The February 7 rally priced it as a fulfillment operation.
But the backlog doesn't just explain the rally. It reveals something about the structure of the capex itself. These companies are building AI-capable data centers to serve cloud contracts that were largely negotiated before AI revenue materialized. The same infrastructure runs both workloads. Every GPU that trains a model can serve a cloud customer. Every data center built for AI also fulfills a contract. The AI capability is, financially, a free rider on infrastructure the cloud business already justified.
The Subsidy
The same day, Politico reported that Amazon's 2025 tax bill fell 87% year over year to $1.2 billion — while profits grew 45% to roughly $90 billion. Ninety billion in profit. An effective tax rate of approximately 1.3%. The mechanism: accelerated depreciation breaks in the GOP's "One Big Beautiful Bill," which let companies front-load deductions on capital expenditure.
Every server Amazon buys, every data center shell it builds, every piece of networking equipment it installs — all of it creates a depreciable asset that reduces taxable income. Amazon guided $200 billion in 2026 capex. The depreciation deductions from that spending will be even larger than those that produced a 1.3% rate on $90 billion in profit. The more Amazon spends on infrastructure, the less it pays in taxes. The capex doesn't just build AI capacity. It shields income.
This means the federal government is co-financing every data center through the tax code. The same Congress that will debate AI regulation is already subsidizing AI infrastructure through depreciation provisions. The political system hasn't decided what to do about AI. The tax code already has.
The Flywheel
Here is what the February 7 data, taken together, actually reveals: the capex race has no financial brake.
The cloud backlog means the spending generates revenue — $1.1 trillion in contracts that need infrastructure to be fulfilled. The depreciation breaks mean the spending reduces taxes — every dollar of capex comes back as a deduction. The Semiconductor Industry Association reported that chip sales hit $791.7 billion in 2025, up 25.6%, and projected $1 trillion in 2026 — meaning the spending creates demand that sustains its own supply chain. Jensen Huang went on CNBC and said the tech industry's "$660 billion capex buildout is justified, appropriate, and sustainable." Of course the man selling the picks thinks the gold rush is rational. But the three mechanisms behind his claim — contracts, tax shields, chip demand — form a cycle, not a line. Spend more, fulfill more contracts, claim more deductions, buy more chips, generate more demand, spend more.
The February 5 question was whether the returns on half a trillion would look like a software company's or an infrastructure company's. The February 7 answer was more unsettling: the spending doesn't need AI-specific returns to continue. The cloud business justifies the infrastructure. The tax code subsidizes it. The chip industry depends on it. Each mechanism reinforces the others. None of them require AI revenue to work.
SemiAnalysis reported that Claude Code already authors 4% of all public GitHub commits, on track for 20% by year's end. Goldman Sachs disclosed it is using Anthropic's AI agents to automate trades, transactions, and onboarding. AI usage is real and accelerating — but even if it weren't, the flywheel would still spin.
The Physical Brakes
If the financial logic says "more," what says "stop"?
On February 7, Wired reported that New York lawmakers proposed a three-year moratorium on data center development — making it at least the sixth state to introduce such legislation. Red and blue states alike were citing climate impact and rising energy prices. The same week, the Wall Street Journal reported that T-glass — ultrathin glass sheets critical to advanced chip packaging — was in short supply, dominated by a single Japanese company with no plans to expand capacity for months. In January, Oracle pushed back its Stargate data center completion from 2027 to 2028 due to labor and material shortages.
The financial flywheel has no brake: cloud contracts justify the spend, tax breaks subsidize it, chip demand sustains it. But financial logic doesn't build data centers. Land, power, water, glass, labor, and political will do. And each of those has real constraints that no amount of contracted revenue can bypass.
That's the tension February 7 exposed. The market rallied because the arithmetic works — $1.1 trillion in contracts, 1.3% tax rates, $1 trillion in chip demand. The arithmetic will keep working. The question is whether six state legislatures, one Japanese glass maker, and a finite electrical grid will let the arithmetic run.