The Picks and Shovels of the AI Gold Rush
During the California Gold Rush of the 1850s, the people who got richest weren't the miners. They were the ones selling picks, shovels, jeans, and lodging to the miners. Replace "miners" with "AI companies" and you've got a pretty good summary of what's happening in 2026.
NVIDIA's market cap crossed $3 trillion in 2024. By mid-2025 it briefly touched $4 trillion. A company that spent most of its existence making graphics cards for video games is now more valuable than any company in history except maybe Apple. If you'd invested $10,000 in NVIDIA stock in 2019 and held until 2026, you'd have somewhere north of $300,000.
This isn't a bubble, exactly — NVIDIA's revenue is real, growing at triple-digit rates year over year, driven by insatiable demand for the H100 and B200 GPUs that power essentially every major AI training run. But it does raise a question: who else is benefiting from the AI boom that nobody's talking about?
Why AI Needs Entirely New Data Centers
A traditional data center rack might draw 10-15 kilowatts of power. An AI training cluster with NVIDIA H100s can easily draw 40-60 kilowatts per rack. The latest B200 ("Blackwell") systems push that even higher — we're talking about individual racks consuming as much electricity as a small apartment building.
Just running more power to these facilities is a logistical challenge. In Northern Virginia — the world's largest data center market — the local utility has warned that power demand from data centers could double in the next five years. In some markets, data center developers are being told they can't build because the grid literally can't supply enough electricity.
Then there's cooling. Traditional data centers use air conditioning. It works fine up to about 20 kilowatts per rack. Above that, air simply can't move heat away fast enough. You need liquid cooling — running water or specialized coolant directly through the server racks to carry heat away at the source.
The Liquid Cooling Revolution
A few years ago, liquid cooling was a niche technology used mainly in supercomputing centers. Today it's becoming standard for any facility hosting AI training workloads. The shift is happening fast because it has to — you literally cannot cool a dense GPU cluster with air alone.
Companies like Vertiv, Schneider Electric, and a handful of smaller players that specialize in direct-to-chip cooling have seen their order books explode. A liquid cooling system for a large AI data center can cost tens of millions of dollars. The market for data center cooling is projected to roughly triple between 2024 and 2030.
There's an interesting environmental dimension here too. Liquid cooling is actually more energy-efficient than air cooling for these dense workloads — it can reduce the power used for cooling by 30-40%. When you're running a facility that draws as much power as a small city, that's meaningful. Some AI companies are now designing data centers around liquid cooling from the ground up, which allows for even greater efficiency gains.
The NVIDIA Story in More Detail
NVIDIA's dominance isn't just about making fast chips. It's about a three-decade head start in building the software ecosystem around those chips.
CUDA — NVIDIA's parallel computing platform — was released in 2006. At the time, most people thought it was interesting but niche. What NVIDIA understood, and what almost everyone else missed, was that if you could make it easy for developers to write programs that run on thousands of GPU cores simultaneously, you'd create an ecosystem that would be extraordinarily hard to compete with.
That's exactly what happened. When deep learning took off in the 2010s, researchers discovered that GPUs were perfect for the matrix math that neural networks require. And because CUDA was already mature and well-documented, everyone built on NVIDIA's stack. By the time competitors like AMD, Intel, and various AI chip startups realized what was happening, NVIDIA's moat was already dug.
Today, even if someone builds a chip that's technically faster than an H100 — and several companies claim to have done so — they still have to convince developers to rewrite their entire software stack for a new platform. That's a much harder sell than "buy our faster chip."
The Supply Chain Squeeze
Here's something most people don't appreciate: advanced AI chips don't just appear. The supply chain is incredibly complex and has a single point of failure.
NVIDIA designs the chips. TSMC in Taiwan manufactures them using advanced processes that no other company can match at volume. The chips are then packaged using technology from companies like Amkor and ASE Group. Specialized memory (HBM, or High Bandwidth Memory) comes from SK Hynix and Samsung. The whole assembly goes onto a circuit board, gets integrated into a server by companies like Dell or Supermicro, and gets installed in a data center.
Every link in that chain is capacity-constrained. TSMC can only make so many advanced chips. SK Hynix can only produce so much HBM memory. The advanced packaging step — connecting the GPU die to the memory — is particularly tight, and it's controlled by a tiny number of suppliers.
This is why, despite massive demand, NVIDIA can't just 10x their production overnight. It's also why governments are suddenly very interested in semiconductor supply chains. The CHIPS Act in the US and similar initiatives in Europe and Japan are direct responses to the realization that AI hardware supply chains are both critically important and dangerously concentrated.
What This Means For The Rest of Us
You're not going to build your own data center. But the infrastructure boom affects you in ways you might not notice.
The cost of running AI inference — every time you ask ChatGPT a question, every time Copilot autocompletes a line of code — is falling fast as more efficient infrastructure comes online. This is why AI companies can offer increasingly capable models for free or cheap. They're passing on infrastructure efficiency gains to users, at least for now.
The energy question is going to become more prominent. AI data centers are already major electricity consumers, and as models get larger and more widely used, the power demands will keep growing. Some of the largest tech companies are now among the biggest buyers of renewable energy in the world, largely because their AI operations need so much power. Nuclear is back on the table too — Microsoft recently signed a deal to restart a reactor at Three Mile Island specifically to power AI data centers.
The jobs story is more complicated than "AI will replace everyone." Right now, the AI boom is creating enormous demand for electrical engineers, HVAC technicians, construction workers, and project managers — people who build and maintain the physical infrastructure that makes AI possible. These aren't the jobs that make headlines, but there are a lot of them and they pay well.
I don't know where the AI industry will be in five years. But I'm reasonably confident that the companies building the infrastructure underneath it — the chips, the cooling, the power, the buildings — will do fine regardless of which AI company "wins."
