For decades, "large" meant 50 megawatts. If you operated a 50 MW data center, you were a serious player. That benchmark is now quaint. In 2026, the conversation has moved to gigawatts — facilities that don't just house servers but function as industrial-scale computing infrastructure, consuming as much electricity as a mid-sized city.
One gigawatt powers roughly 750,000 homes. We are building campuses that need that much power continuously, just to train and run AI models. The term "data center" no longer fits. These are AI factories.
Here's what that shift actually means across the four dimensions that matter most: compute, power, cooling, and cost.
The fundamental unit of AI infrastructure is the GPU rack, and the numbers have changed dramatically. Five years ago, a server rack drew 5–10 kilowatts. A single rack of NVIDIA's current-generation Blackwell chips draws over 120 kW. That's a 10-20x increase in power density for the same physical footprint.
This changes everything downstream. The facility design, the power distribution, the cooling infrastructure, the network architecture — all of it gets redesigned around this new density requirement.
At the cluster level, the shift is equally significant. Rather than thousands of independent servers, gigawatt-scale AI infrastructure operates as a single unified supercomputer. The GPUs aren't working independently — they're communicating constantly, sharing memory and computation across the entire cluster. This means the network connecting the chips (InfiniBand and high-speed Ethernet) becomes as critical as the chips themselves. A bottleneck in the fabric slows down the entire system, regardless of how powerful the individual GPUs are.
The limiting factor in AI infrastructure in 2026 isn't GPU supply or capital — it's electricity. The power grid simply wasn't built for this. Utilities that took decades to build out transmission infrastructure are being asked to provision gigawatts of new capacity in years.
The industry's response has been to stop waiting. Microsoft, Meta, Google, and Amazon are all pursuing some version of behind-the-meter power — building or contracting their own generation rather than depending on grid connections that may take years to provision.
Nuclear is getting the most attention. Small modular reactors (SMRs) represent a path to always-on, carbon-free power at the scale these facilities need. Several hyperscalers have signed long-term contracts with existing nuclear plants, and the first SMR projects specifically designed for AI infrastructure are in development.
In many jurisdictions, new data center permits now require the operator to demonstrate their own power solution rather than relying on the existing grid. Bring-your-own-power has gone from competitive advantage to regulatory requirement in some of the most desirable development markets.
You can't cool a jet engine with a desk fan. The same physics problem applies to AI racks drawing 120 kW each. Air cooling — which has served data centers for decades — hits a practical ceiling around 30–40 kW per rack. Everything above that requires liquid.
Direct-to-chip liquid cooling is now the standard for AI-class infrastructure. Cold plates attached directly to processors carry coolant to the heat source rather than relying on airflow to carry heat away. It's more efficient, more reliable, and increasingly cost-competitive with high-density air cooling systems.
The frontier is immersion cooling — submerging entire servers in non-conductive dielectric fluid. The economics are compelling at sufficient scale: better thermal performance, lower fan energy consumption, potentially longer hardware lifespan. Several large AI training clusters are already operating fully immersed.
One development worth watching: heat reuse. At gigawatt scale, waste heat becomes an asset. Some European facilities are capturing heat from AI workloads and piping it to nearby buildings for heating. What was an environmental liability is becoming an infrastructure offering.
Building at gigawatt scale is extraordinarily expensive. The rough math: facility shell and real estate runs around $11 million per megawatt. The fit-out — GPUs, high-speed networking, power distribution, cooling infrastructure — adds another $25 million per megawatt. A 1 GW campus therefore carries a price tag somewhere between $35 and $50 billion.
For context, that's comparable to a large semiconductor fab, a major airport expansion, or the GDP of a small country. These are not data centers in any traditional sense. They are industrial megaprojects.
The supply chain pressure is significant. Transformers that can handle the power loads required have a two-year backlog. Specialized cooling pumps and heat exchangers are similarly constrained. The facilities that break ground first will have a meaningful advantage over those that follow, simply because the equipment queue is long.
A few implications worth thinking through if you're not in the infrastructure business but you're building on AI:
Geography is shifting. AI training clusters are no longer being built near population centers for latency reasons. They're being built wherever cheap, stable electricity exists — which increasingly means remote areas with hydroelectric, nuclear, or wind resources. The map of AI infrastructure looks nothing like the map of traditional cloud data centers.
Power is the new competitive moat. Whoever controls the most affordable, reliable electricity will have a structural cost advantage in AI inference. This will matter more as inference costs become the primary determinant of AI product economics.
Efficiency metrics are evolving. The industry has moved from PUE (Power Usage Effectiveness) — a measure of total facility power vs. IT power — toward PCE (Power Compute Effectiveness), which measures intelligence produced per watt spent. As power costs dominate total cost of ownership, the pressure to improve this metric will only intensify.
