Big Tech Is Hitting the Brakes on AI Spending — Wall Street Wants Receipts

The gold rush enters its receipts phase

The AI spending slowdown is no longer a vibe — it's a market event. This week, Microsoft and Meta both signaled they're recalibrating infrastructure plans set during a period of near-unlimited optimism. Anthropic, preparing one of the most anticipated IPOs in years, is reportedly facing investor pushback on its valuation as revenue growth decelerates. And in the financial world, major banks are drawing a hard line against letting third-party AI agents touch customer money. The common thread: capital is getting disciplined, and the companies that avoided the arms race are suddenly looking prescient.

What changed

Microsoft is preparing the infrastructure to run AI cheaper rather than just bigger: its Maia 300 AI accelerator enters production deployment in 2027 as a primary inference engine across Microsoft products, with multi-gigawatt-scale deployment aimed at improving efficiency and cutting internal AI costs. That's a pivot from capacity-at-any-cost to cost-per-token.

Meta is recalibrating too, and the financing side is getting creative — CoreWeave is exploring insurance-industry capital as traditional financing struggles to keep pace with data center build costs, while Singapore-based DayOne expanded its Malaysian data-center financing package to $7.2 billion. When the builders start hunting for exotic capital, the easy money phase is over.

Meanwhile the banks' stance on AI agents is telling: institutions that move trillions won't let third-party agents near customer funds. Trust, not capability, is the gating factor — which rhymes with everything we know about rogue agent incidents this month.

The winner: the company that didn't play

Apple, long mocked for sitting out the LLM race, is the unlikely beneficiary. Its AI-related spending is projected at roughly $11 to $13 billion for 2026 — a fraction of the hundred-billion-dollar budgets at Microsoft, Google, Amazon, and Meta. Its market value briefly touched $5 trillion in July, only the second company ever to do so, and Apple shares are up about 25% in 2026, the best performance among the Magnificent Seven. Tesla, Microsoft, and Meta have all declined.

Being late looks like genius when the bill comes due. Apple's bet — modest model spending, AI as a feature layer across devices people already own — is exactly the shape Wall Street now wants to see.

The buildout isn't stopping. It's being audited

Don't mistake discipline for retreat. Demand is real: ASML reported 11.7 billion euros in quarterly revenue on the back of chip-equipment demand. Oracle's cloud infrastructure revenue jumped 121% year over year to $7.4 billion, with a record $664 billion in remaining performance obligations — contracted future revenue signaling years of committed AI compute demand. The data center bottleneck hasn't moved either: power, chillers, and permits remain the binding constraints, as we detailed in our look at the buildout's real bottlenecks.

What changed is the question being asked. Not 'how fast can we build' but 'what's the return.' The venture market shows the same concentration: fewer, larger deals — Ramp's $1.85B at a $60B valuation, Modal's $750M at $15.75B. The era of blank-check AI infrastructure spending is over. The era of having to prove it works has begun — and that's healthier for everyone except the companies that already spent the money. The discipline also changes who gets funded: infrastructure plays with contracted revenue keep raising, while speculative capacity without signed customers faces real scrutiny for the first time since 2023.