Is the $700 Billion AI Spending Wave a Bubble? What the Numbers Actually Show

Amazon, Google, Microsoft, and Meta are on track to spend somewhere north of $650–700 billion combined on AI infrastructure in 2026 — up from roughly $400 billion just a year earlier. That’s not a typo, and it’s not a one-quarter spike. It’s the fastest sustained increase in corporate capital spending in modern market history, and it has split Wall Street into two camps that are both looking at the same numbers and reaching opposite conclusions.

The case that this is justified spending, not a bubble

The bull case rests on one simple observation: the companies spending the money say they can’t build fast enough. Microsoft has disclosed an $80 billion backlog of Azure orders it cannot fulfill — not because of lack of demand, but because of power constraints on data centers. Google’s cloud backlog surged 55% in a single quarter to over $240 billion. Amazon’s executives have said publicly that they only commit to capacity when they’re already seeing the demand signals to justify it, and that as fast as they add capacity, they’re monetizing it.

There’s also a quieter data point that matters: efficiency is improving alongside the spending. Google says it cut the cost of serving its Gemini models by roughly three-quarters over the past year through model optimization. That’s the kind of gain that, if it continues, means the same infrastructure spend produces meaningfully more usable AI output over time — which undercuts the simplest version of the bubble argument.

The case for real concern

The bear case isn’t about whether AI is useful — it’s about the financial mechanics of how this is being funded. Free cash flow at the largest US tech spenders is on track to turn negative for the first time in roughly 35 years, as capital spending outpaces operating cash flow. Capex intensity — spending as a share of revenue — is now estimated near 34%, more than double the roughly 15% peak reached during the dot-com buildout of the late 1990s. Three of the four hyperscalers saw their stock sell off following recent earnings calls specifically because of capex guidance, even as revenue beat expectations.

The honest, uncomfortable middle ground is this: backlog and bookings data suggest real, currently-unmet demand — that’s a genuinely different picture from the dot-com era, where infrastructure was frequently built well ahead of any paying customer. But the sheer scale of spending, funded increasingly through debt rather than existing cash flow, means the bet only pays off if that demand keeps compounding for several more years, not just the next few quarters. A slowdown in enterprise AI adoption, or a plateau in what current-generation models can economically do, would leave a lot of very expensive, very specialized infrastructure without enough paying use to justify it.

What to actually watch going forward

Skip the “bubble or not” framing entirely and watch one number instead: whether backlog and bookings growth continues to outpace capex growth, or whether that relationship flips. As long as demand signals (backlogs, monetized capacity, cost-per-query efficiency) are growing faster than the spending itself, the aggressive build-out has a real demand floor under it. If capex keeps climbing while backlog growth flattens, that’s the signal the bears have been waiting for — and it would show up in earnings calls well before it shows up in a stock price crash.

Why this cycle is different from past infrastructure booms

It’s worth being specific about what actually separates this from the fiber-optic overbuild of the dot-com era, since the comparison gets made constantly. Telecom operators in the late 1990s built capacity based on projected internet growth that, in many cases, never materialized on the expected timeline — leaving fiber “dark” for years. The current AI buildout has a different constraint entirely: it’s power-limited, not demand-limited. Microsoft’s inability to fill Azure orders isn’t a sales problem, it’s a physical one — data centers need electricity, and grid capacity in many regions simply can’t be added as fast as chips can be installed. That’s a genuinely different risk profile. A demand-side bubble pops when customers stop showing up. A power-constrained buildout has a natural ceiling that forces discipline on spending regardless of how much capital is available, which is arguably a healthier failure mode than unconstrained overbuilding.

That said, power constraints cut both ways — they’re also why utility capex is rising sharply alongside tech capex, and why some analysts now watch electricity infrastructure spending as a leading indicator for the AI buildout’s real pace, separate from whatever the hyperscalers say on earnings calls.