The AI Arms Race Isn’t Slowing — It’s Getting Pricier

The New York Times, The Washington Post, Axios, and Business Insider contributed to this report.

Big Tech just slammed the accelerator on artificial intelligence spending — again. Google, Meta, Microsoft, and Amazon all told investors they’ll pour even more billions into chips, servers, and data centers, despite growing chatter that this could morph into a bubble.

The new bill: bigger than last quarter—again

• Microsoft shelled out ~$34.9B in capex last quarter (up 74% YoY). Roughly half went to GPUs/CPUs — “short-lived assets” — plus $11.1B on data-center leases. Management says demand keeps outrunning supply.
• Amazon spent $34.2B in Q3 and says ~$125B in capex this year, with even more in 2026 as it aims to double data-center capacity in two years. It just opened an $11B AI campus in Indiana (Project Rainier) and earmarked $5B for South Korea.
• Meta logged $19.4B in Q3 capex and lifted 2025 guidance to $70–$72B (up from $66–$72B). It’s building company-owned sites and renting space, and lined up $29B in project financing for a Louisiana campus. Investors weren’t thrilled: the stock fell ~11% as that spending swelled.
• Google spent ~$24B in Q3; 60% of that went to AI chips/servers, 40% to networking and data centers. It raised full-year capex to $91–$93B and got rewarded as Cloud revenue accelerated.
• Apple — usually frugal — signaled capex is heading higher next fiscal year.

Wall Street’s getting twitchy about how this is funded. As capex balloons, more firms are leaning on debt or off-balance-sheet structures. That’s fine if returns show up; risky if they don’t.

“Microsoft’s fortress balance sheet makes higher capex a bullish signal — demand is real,” says Gil Luria (D.A. Davidson). “But some companies are borrowing for demand that isn’t there yet.”

• Oracle is the poster child: its debt could double to $290B by FY2028 (Morgan Stanley). Luria warns of overbuild risk—the kind better financed with equity than debt. He sees similar dynamics at infrastructure players like CoreWeave and Crusoe.
• The treadmill effect: every new Nvidia generation raises the ante. To stay competitive, hyperscalers keep refreshing hardware—pushing out the date when spending can slow.

Big Tech can afford to splash cash — profits and cash flow are massive — but sentiment can turn if the payoff lags. Meta is the canary here: strong results, bigger spend, lower stock. Analysts say the “grace period” to show non-core AI returns is closing.

Meanwhile, companies want to protect credit ratings and avoid scaring shareholders with towering on-balance-sheet debt. Expect more structured financing, JVs, leases, and vendor prepayments — the kind of arrangements that are harder to track, but smoother on optics.

If AI drives real productivity and new revenue streams (from cloud, ads, enterprise software, and AI services), today’s debt looks smart. If not, Luria cautions, surging leverage tied to lackluster demand could seed systemic risks — his words, not ours — akin to prior credit bubbles.

The AI build-out is still a land-grab—only now the land is pricier, the checks are bigger, and the financing is under the microscope. The winners will be the ones who turn capex into cash flow fastest. The rest risk building very expensive foundations for a future that arrives late.