The compute buildout now has named addresses and confirmed cost figures. That’s new.
Epoch AI’s updated Frontier Data Centers database identifies the Anthropic-Amazon New Carlisle facility as the largest currently tracked AI data center in its data collection, at a tracked cost of $35B and 1.1 GW of capacity. Per Epoch AI’s independent tracking methodology, this is their classification, not a figure independently verified from Anthropic or Amazon’s facility operator disclosures. That distinction matters: Epoch AI is T1 for AI infrastructure tracking and its data collection methodology is named, but the underlying figures derive from their database, not from facility operator filings.
Then there’s Fairwater. Epoch AI projects Microsoft’s Fairwater Wisconsin facility to reach approximately 8x the scale of New Carlisle by 2027. That’s a projection, not a confirmed build spec, and it should be read as Epoch’s modeling output rather than a Microsoft commitment. Eight times $35B would imply a $280B facility. Eight times 1.1 GW would imply 8.8 GW. Even if the projection is directionally right and substantially overstates the final scale, the magnitude suggests Fairwater is intended to be a qualitatively different class of infrastructure than anything currently operational.
$280B in a single facility would exceed the GDP of most countries.
Epoch AI Compute Trajectory (EPOCH-VERIFIED)
On the macro compute figures: per Epoch AI’s updated data, AI chip performance per dollar is improving at 37% annually, and global AI computing capacity is doubling approximately every seven months. Both figures carry Epoch AI verification status, these are the organization’s independent research findings, not vendor-stated benchmarks. That’s the relevant distinction for investors and infrastructure planners using these figures as baseline assumptions.
Why the facility-level data matters now: The prior Epoch AI compute coverage established the trajectory at the macro level. This update names specific facilities, which unlocks a different kind of analysis. Infrastructure investors, energy planners, and competitive intelligence teams can now map the geographic distribution of committed AI compute. New Carlisle is in Indiana; Fairwater is in Wisconsin. Both are in PJM territory, the same grid market where wholesale power prices have been rising in direct response to AI data center demand.
That’s not coincidence. The two largest tracked AI facilities are both drawing from the same grid market that’s already reporting price pressure. The energy constraint story and the infrastructure scale story are the same story viewed from different angles.
The 37% annual chip performance improvement and seven-month compute doubling time also reframe the Fairwater projection. If compute capacity doubles every seven months, a facility that’s 8x more powerful than today’s largest isn’t a 2027 outlier, it’s roughly consistent with two doubling cycles from current New Carlisle scale. The math checks out directionally, which is either reassuring about the projection’s plausibility or alarming about the pace of infrastructure concentration, depending on whether you’re an infrastructure investor or a grid operator.
Analysis
New Carlisle and Fairwater are both located in PJM grid territory, the same wholesale power market that has been reporting price increases directly attributed to AI data center load growth. The infrastructure scale story and the energy constraint story are converging geographically. Infrastructure investors need to assess grid transmission capacity alongside facility specifications.
What to Watch
What to watch
Epoch AI updates its Frontier Data Centers database on a rolling basis. The specific signals to track are whether New Carlisle’s tracked cost and capacity figures are revised as Amazon/Anthropic make further disclosures, and whether Fairwater’s 2027 projection gets updated as Microsoft publishes facility specifications. The gap between projection and disclosed spec is where investment uncertainty lives.
The real story is what named facilities do to the investment thesis. When the compute buildout was a macro trend, it was a sector call. Now that it has specific addresses, cost figures, and grid locations, it’s a project-level underwriting question. Infrastructure investors who’ve been betting on AI compute as a category now need to assess whether their capital is upstream or downstream of New Carlisle and Fairwater in the value chain. Watch the next round of PJM grid capacity filings for the first hard data on whether transmission infrastructure is keeping pace with these facility projections.