Goldman's 98% AI Capex Surge: The Infrastructure Trade Gets Bigger-and More Dangerous

Big Tech's cash-flow squeeze is the story now

The bill is due now

Goldman says Big Tech capex is on track to consume 98 percent of operating cash flow this year. That leaves little room for buybacks, dividends, or mistakes. Bulls can call that commitment; bears can call it a shrinking margin of safety.

Earlier this year, the four largest U.S. tech companies were committed to more than $700 billion in capital expenditure, mostly for AI infrastructure. As that spending stays elevated, the near-term beneficiary remains the supply chain rather than the platform owners themselves.

Why the platform layer looks more fragile

The problem for the giants is not just the size of the spend. It is that revenue proof has not yet caught up. Goldman's broader framing is that the real question is whether AI commercialization can justify the scale of deployment. Meta's recent reaction shows how sensitive investors are to that gap: after it raised its capex target, its stock dropped 9 percent in a single session.

That is the tension in the market now. Investors are not paying for ambition alone. They are watching which companies can fund the build without straining financing or eroding flexibility.

AI infrastructure spend still favors the suppliers

Upward revisions keep the build-out visible

The cleaner beta still looks closer to the suppliers. Hyperscaler spending forecasts have been pushed up by nearly 60 percent in cumulative revisions, while EPS estimates for AI infrastructure stocks have also climbed sharply. That points to stronger order visibility going upstream, not just bigger long-term rhetoric.

The money is spreading across the stack

Hyperscaler budgets are not landing in one bottleneck. They are flowing across compute, networking, memory, power, and cooling. That broadens the set of companies that can monetize early in the build-out, before software economics are fully proven.

The macro backdrop also supports the idea that this is more than a narrative trade. In the first half of 2025, AI-related investment added nearly a full percentage point to U.S. real GDP growth. That does not prove end-market returns, but it does suggest the build-out is already showing up in economic activity.

Pricing power still leans toward scarce inputs

Tech giants have been locking in 3-5 year contracts to secure chip supply, while the DRAM market remains highly concentrated. When buyers are securing supply and sellers are few, pricing power tends to stay with vendors rather than return to the spenders.

That is why the upside path still looks wider upstream: revenue can convert faster, scarcity can support margins, and the cash comes in before platform owners have fully proved AI returns.

The risk: spending can rise faster than the payoff

The trade gets more dangerous at the platform level. Goldman notes that investors are already rotating away from AI names where operating earnings growth is under pressure and capex is debt-funded. In other words, the market is becoming more selective about who gets rewarded for spending.

What could break the setup

The clearest break point is not weak AI demand on paper. It is capex rising faster than visible monetization. Wall Street's consensus for hyperscaler 2026 spending is now $527 billion, and Goldman has warned that analyst estimates have consistently underestimated AI-related capex.

There is also a cost-risk layer. Goldman's research says next-generation data centers are becoming more complex and expensive as AI workloads push power density higher and system integration deepens. If costs stretch further, financing needs can arrive before income-statement benefits do.

What to watch next

• Constructive: infrastructure sellers with pricing power and visible orders.
• Selective: platform spenders, and only where monetization, financing, and shareholder-return capacity still look credible.

Meta remains the clearest live test of that boundary. The market did not reject the AI build-out; it rejected the idea that spending alone deserves a premium. For tech infrastructure, that means the opportunity is still expanding, but the tolerance for error is shrinking.