I saw something strange a few weeks ago while passing a construction site outside of Phoenix. It didn’t appear to be a warehouse because of the dust, the cranes, and the length of the perimeter fence. It appeared to be an airport. I was informed that it was a data center by a worker leaning on a Ford F-150. Only one.
Within a forty-mile radius, there were three more that fed the same hunger. In part because it contradicts the narrative we continue to tell ourselves about artificial intelligence, that image has stuck with me. Benchmark scores and chatbots continue to make headlines. Concrete, copper, and cooling pipes are receiving the real money—the kind that builds cities.
| Sector | AI Infrastructure / Data Center Hardware |
| Core Thesis | The biggest AI winners may not be model labs but the companies supplying compute, networking, and power |
| Estimated Global AI Infra Spend (2026) | Projected to cross $400 billion |
| Key Players Often Mentioned | Nvidia, Broadcom, TSMC, Supermicro, Vertiv, Coreweave |
| Average Cost of a Modern AI Server Rack | Roughly $3 million per fully loaded GB200 rack |
| Power Demand per Hyperscale Site | Often 500 MW to 1 GW, equivalent to a mid-size city |
| Major Buyers | Microsoft, Meta, Google, Amazon, Oracle |
| Reference Reading | Sequoia Capital’s Generative AI Act o1 |
| Risk Factors | Power grid limits, GPU supply, cooling constraints, geopolitical chip restrictions |
| Time Horizon for Thesis | 5 to 10 years |
The belief that the next trillion-dollar company won’t be a model lab at all is becoming more and more prevalent among those who actually move capital. Whoever sells the picks and shovels will win. Julien Bek of Sequoia recently made a similar claim, stating that a software company masquerading as a services company will be the next big thing. In terms of software, he’s probably correct. However, if you go one level below that, you discover an even more significant reality: without the servers, nothing would function. Every quarter, models become more affordable. The hardware beneath them doesn’t.
It’s difficult to ignore how uneven the spending has gotten. Nvidia, Broadcom, TSMC, Vertiv, and a few lesser-known suppliers that manufacture switchgear and immersion-cooling tanks in places like Iowa and Malaysia receive several dollars for every dollar a company pays OpenAI or Anthropic. This year, Microsoft alone is expected to invest nearly $100 billion in AI infrastructure.
Meta’s capital expenditures appear comparable. These budgets are no longer for software. They are more in line with what major oil companies or utilities used to spend on pipelines and rigs.
The bottleneck is what makes this particular moment unique. It is now possible to train a model with a smaller team. However, you are unable to create a fabrication plant that can produce three-nanometer wafers, a gigawatt of power, or 100,000 H200s. These things require political negotiations, years, and billions of dollars. Investors seem to think that durable monopolies are created by scarcity rather than intelligence, and I believe they are right. It is possible to swap out a model over the weekend. In Virginia, a power substation cannot.
Years ago, Tesla encountered similar skepticism, and the analogy seems helpful. Everybody was obsessed with the cars. It turned out that the true question was whether the factory could be built at all. Similar events are currently taking place, but on a much bigger scale. This spring, Coreweave went public. In just two years, Supermicro’s earnings have increased dramatically. Vertiv, a company that primarily sells uninteresting metal cabinets and liquid coolers, has quietly emerged as one of the decade’s top-performing industrial stocks. There are no tweets about it.
It’s still unclear if the money will be distributed among a dozen experts or if one company will take home the prize. Before the chips break, the grid might. China might catch up sooner than anticipated. However, I couldn’t stop thinking the same thing when I saw this build-out in person, next to that fence in Phoenix. The prompts are not written by the trillion-dollar corporation. The foundations are being poured by it.
