The four hyperscalers are spending more on compute infrastructure this year than the entire economy of Belgium produces — and they're just getting started.
The Summary
- Meta, Amazon, Microsoft, and Alphabet are projected to spend $750 billion on capex in 2026, more than half of what all S&P 500 companies spent combined in 2025
- This translates to building out the physical infrastructure that makes the agent economy possible: the data centers, chips, power, and cooling systems where AI actually runs
- The scale dwarfs national GDPs and represents a fundamental bet that compute demand will justify returns measured in decades, not quarters
The Signal
The $750 billion figure is a floor, not a ceiling. Each of these companies is racing to build data centers that can handle the training runs for frontier models and the inference loads for millions of agents running simultaneously. A single hyperscale data center now costs between $10 billion and $20 billion when you account for land acquisition, construction, power infrastructure, networking gear, and the GPU clusters themselves.
The breakdown matters. Land costs have exploded in regions with available power capacity. A site near a nuclear plant or with access to renewable baseload power commands a premium that didn't exist three years ago. Then comes the buildout: cooling systems that can handle the heat density of H100 or B200 clusters, networking that moves petabytes between racks without bottlenecking, redundant power systems that keep inference running when the grid hiccups.
"These aren't traditional tech investments. They're infrastructure plays with 20-year depreciation schedules competing against quarterly earnings pressure."
The chip costs alone tell the story. Nvidia's H100 GPUs run about $30,000 to $40,000 per unit. A competitive training cluster needs at least 10,000 of them. That's $300 million to $400 million in silicon before you've powered them on. The next-generation B200s will cost more, draw more power, and require even more sophisticated cooling. Multiply that across dozens of data centers going live every year.
Here's what separates this capex wave from previous infrastructure buildouts:
- Unlike telecom fiber overbuilds in the 1990s, there's no clear ceiling on compute demand
- Training runs for GPT-5 class models require clusters these data centers are only now coming online to support
- Inference costs for agent workloads don't compress the way serving web pages did
The companies writing these checks aren't building for current demand. They're building for a world where every knowledge worker has multiple AI agents, where every app has an agentic layer, where inference happens continuously rather than on-demand. Microsoft's bet on OpenAI only makes sense if Azure can deliver the compute. Google's Gemini strategy requires data centers that don't exist yet. Meta's Llama approach means building capacity their own products won't fully utilize for years.
The capital intensity creates a moat. Startups can rent GPU clusters from these hyperscalers, but they can't match the economics of owning the full stack. That shapes the entire agent economy. The companies that control inference at scale control which agents get built, how they're priced, and who can afford to run them.
The Implication
Watch where the next wave of data centers gets built. Power availability, not proximity to markets, is the new location variable. Companies securing long-term energy contracts today are positioning for the inference economy of 2030.
For anyone building agent companies, this capex wave means two things. First, the hyperscalers are your landlords whether you like it or not. Second, they're investing ahead of demand, which means inference pricing has room to compress. Build for a world where compute costs fall but never disappear.