Big Tech’s combined AI capex is on pace to hit nearly $700 billion in 2026, and nobody writing those checks can point to the top of the mountain. That’s the picture emerging from a Fortune analysis circulating on Hacker News, which tracks quarterly earnings from Alphagram, Amazon, Meta, and Microsoft. The four hyperscalers spent more than $130 billion on data centers and infrastructure in a single quarter, putting the full year on track to jump from about $410 billion in 2025 to roughly $700 billion this year. In 2024, that same number was just over $200 billion. The curve is not flattening.
Wall Street is openly disagreeing
The market reaction to last quarter’s reports was the most interesting tell. Meta dropped sharply after investors got a closer look at its spending plans. Microsoft slipped too. Alphabet and Amazon, by contrast, rose on strong cloud growth. Same buildout, opposite verdicts.
There’s a wrinkle behind the optimism. Hacker News commenters flagged a separate Fortune piece noting that roughly half of the “blowout AI profits” at Google and Amazon this quarter came from their stakes in Anthropic, not from operating businesses. So part of what’s powering the bull case is mark-to-market gains on a private investment, not revenue from selling AI compute. That’s worth holding in mind when the headline numbers look reassuring.
Where the $700 billion actually goes
This isn’t software spend. It’s a utility-scale industrial buildout:
- Chips. A single Nvidia GPU runs up to $40,000. An eight-GPU server costs hundreds of thousands. Hyperscale clusters of 100,000-plus GPUs run into the billions per site.
- Data centers. Meta’s Hyperion project in northeast Louisiana alone is a $27 billion bet, with some estimates putting it at millions of GPUs and consuming the power of a small city.
- Networking. The unsexy layer. Switches, fiber, network cards. Without it, the chips can’t talk to each other and the cluster is dead weight.
McKinsey projected last year that global AI capex needs to hit $6.7 trillion by 2030 to keep pace with demand. At today’s run rate, the industry is roughly on schedule.
The bear case isn’t going away
A growing chorus of analysts is calling overbuild. The argument: AI hardware depreciates fast, demand forecasts are still mostly faith-based, and the gap between training-cluster supply and paying-customer revenue has not closed. Fortune’s Shawn Tully has flagged the depreciation problem as the next shoe to drop, since today’s $40,000 GPU is tomorrow’s stranded asset if the next generation arrives sooner than the financing model assumes.
The counterpoint is that this race is now in its third year and the hyperscalers keep finding reasons to spend more, not less. Only Alphabet has explicitly told investors capex will rise again in 2027, but none of the four are signaling a pullback.
What it means for practitioners and operators
- If you build on AI APIs: compute supply will keep expanding, but pricing is unlikely to collapse soon. Hyperscalers need to recoup capex, and Anthropic and OpenAI need to fund their own training runs. Plan margins around stable, not falling, inference costs.
- If you run an AI startup: the SoftBank “Roze” IPO, reportedly targeting a $100 billion valuation around physical AI infrastructure, signals where late-stage money is heading. Software-only AI plays are competing for attention against picks-and-shovels stories with hard assets.
- If you’re an enterprise buyer: the divide on Wall Street matters because it shapes which vendors get the cheapest capital. Watch which hyperscalers can keep funding capex without diluting shareholders. That’s the one most likely to undercut on price 18 months out.
The climb isn’t over. What’s changing is that the bears now have receipts, and the bulls are increasingly leaning on accounting gains from private stakes rather than core operating profit. That tension is the story to watch through the rest of the year. Full breakdown is at the original source.