The frontier of AI has become a construction project. Epoch AI estimates the largest known AI data center, the Anthropic and Amazon site at New Carlisle, has computing power equivalent to about 700,000 H100 chips, drawing roughly 1.1 gigawatts and costing about 35 billion dollars. Microsoft's planned Fairwater campus in Wisconsin is projected at 5.2 million H100-equivalents by September 2027, nearly eight times larger.
Those figures move the contest off the chip and onto the grid. A gigawatt of facility power costs around 30 billion dollars to build and takes about two years of construction. The binding constraints are now electricity, permits, and steel, all of which move slower than a model release.
The capital math
GW_COST_USD_B = 30 # roughly $30B to build 1 GW of facility power
sites = {"New Carlisle": 1.1, "Fairwater (2027)": 8.6}
for name, gw in sites.items():
print(f"{name:18s}: ~{gw} GW -> ~${gw * GW_COST_USD_B:,.0f}B to build")The output explains why only a handful of companies are in this race. Spending tens of billions on a single building only pencils out if you expect to fill it, which assumes both demand and a power connection that may not exist yet.
What this concentrates
| Resource | Status | Who controls it |
|---|---|---|
| Chips | Improving in supply | Several vendors |
| Capital | Tens of billions per site | A few hyperscalers |
| Power | The hard limit | Utilities and regulators |
The cost of these bets ties directly to the scaling curve nobody wants to extrapolate: the bills are real, the payoff is a forecast. Watch the substation, not the launch event. The data center estimates come from Epoch AI.
Why a gigawatt is a different category
A gigawatt-scale AI campus is closer to an industrial project than a normal data center expansion. It needs land, transmission capacity, transformers, cooling, backup systems, fiber, and a construction schedule that can survive permitting delays. The model roadmap may move in months, but the physical plant moves in years. That mismatch is now part of AI strategy.
The scale also changes risk. A conventional cloud region can grow in phases as customers arrive. A frontier AI site requires a large commitment before the demand is fully proven, because the power and building shell have to be secured early. The buyer is making a bet on future model revenue, future inference volume, and future access to electricity all at once.
This is why the largest projects cluster around companies with enormous balance sheets. The advantage is not only that they can buy chips. They can sign power agreements, absorb construction delays, and finance capacity before it earns revenue. Smaller labs may still innovate, but renting the frontier increasingly means renting from someone who owns the grid connection.
The local effects are political
Gigawatt projects become local policy issues quickly. They compete for power with factories, homes, and other data centers. They create construction jobs, tax revenue, water concerns, and pressure on transmission planning. A project that looks like a model-capability decision from Silicon Valley can look like an energy-development decision to the county that hosts it.
That political layer can slow or redirect the race. Utilities need to decide how much generation and transmission to build for customers whose demand forecasts depend on uncertain AI revenue. Regulators need to decide who pays for upgrades if a campus is delayed or canceled. Communities need to decide whether the jobs and tax base justify the infrastructure burden.
For AI companies, community acceptance becomes part of execution risk. A site with cheaper power but slower permitting may lose to a site with stronger local support. The winning location is the one where capital, electricity, regulation, and schedule line up closely enough for the model roadmap.
What to measure next
The visible headline is H100-equivalent capacity, but the next set of useful metrics is more practical. How many megawatts are energized, not merely planned? What share of power is under firm contract? How much capacity is available for training versus inference? What is the expected utilization rate after the first year? Those details separate announced ambition from usable compute.
There is also an efficiency question. A larger campus is only better if it can feed chips, move data, and cool racks without wasting too much power. Power usage effectiveness, networking topology, and maintenance downtime all affect the real capacity customers experience. A chart of theoretical chips does not capture those losses.
The AI race now has a public-infrastructure dimension. Model releases still matter, but the next frontier may be decided by interconnection queues and construction schedules. The company that can turn capital into energized capacity fastest will have an advantage before a single training run starts.
The buyer's takeaway
Customers should read gigawatt announcements as capacity signals, not as product guarantees. A planned campus can lower future prices only if it is built, energized, filled with chips, and used at high utilization. Until then, buyers should ask providers where capacity is available today, which regions are constrained, and which workloads may be throttled during demand spikes.
The same questions matter for vendor risk. A provider with many smaller sites may be more resilient than one waiting on a single enormous project. A provider with firm power contracts may be safer than one with only a public target. The headline number is useful, but the delivery path is the story to watch.
