Who's Locking In America's AI Power Capacity, and What the Gigawatt Race Means Before 2028

The grid doesn’t move fast. That’s the structural fact driving every major AI infrastructure announcement in 2026.

Permitting a new high-voltage substation takes years. Expanding transmission capacity takes longer. The lead times on power delivery infrastructure are measured in the same units as AI compute demand forecasts, three to five years, and they don’t compress on demand. What this means practically is that the organizations that will have AI compute capacity in 2028 are, in large part, the organizations making infrastructure commitments now. Everyone else is waiting in a queue that isn’t moving.

TeraWulf’s acquisition of the Muskie Data Campus, disclosed via a Form 8-K filing on approximately May 26, is a move inside that constraint. A 285-acre site in Eastern Kentucky, master-planned for more than 1 gigawatt of capacity, acquired from a private equity seller that built it speculatively. The acquisition doesn’t deliver capacity today. It reserves the grid position.

The Power Constraint Thesis

The argument for aggressive infrastructure positioning rests on a simple supply-demand asymmetry. AI inference workloads are multiplying. Training workloads require sustained multi-gigawatt compute clusters. Both are power- intensive in ways that general-purpose data center demand isn’t.

The hyperscaler capital infrastructure pattern has made this visible at scale: Microsoft committed to 10.5 GW of renewable energy capacity through a Brookfield partnership in 2026. That’s not a corporate sustainability gesture. That’s a forecast that Microsoft’s compute demand will require 10.5 GW to serve. When organizations of that scale make commitments measured in gigawatts, smaller operators face a choice: get ahead of the queue or accept that premium grid-connected sites will be locked up by entities with larger balance sheets and longer investment horizons.

The private equity exit at Muskie is its own signal. Industrial Equity Partners built the site speculatively, assembled the land, developed the site infrastructure, and secured grid arrangements, then sold to an operator. That’s the standard infrastructure development playbook: develop at development economics, exit to an operator who prices in the operating risk. The fact that IEP found a buyer, and that the buyer is committing to a two-phase delivery through 2030, confirms that the speculative development thesis has market support.

Who Is Racing to Lock in Gigawatts

TeraWulf is one entry in a pattern that became visible across multiple registry items in the current cycle. The comparison is instructive.

The Epoch AI data center database update in mid-May documented the Carlisle campus at $35 billion and 1.1 GW as one of the largest declared AI infrastructure positions in the current build cycle. That’s a hyperscaler-scale investment. The NV Energy data center power deal covered in the same-day registry entry represents a different model: utility-side commitment to serve compute demand rather than operator-side acquisition of sites.

Hitachi and X LABS’ gigawatt-scale “Power as a Service” model represents a third approach: infrastructure delivered as a contracted service rather than an owned asset. Each model has different risk profiles for the enterprise teams and investors evaluating AI compute procurement options.

TeraWulf’s approach is operator ownership. Acquire the site, own the capacity, take the development and delivery risk, benefit from any upside if AI compute demand lands on schedule. The risk is the two-year-plus gap between acquisition and first delivery, a window during which energy costs, competing supply, and demand timing can all shift.

The Kentucky Grid Decision

Eastern Kentucky isn’t an obvious AI infrastructure hub. Understanding why TeraWulf chose this site requires looking at what the site offers structurally rather than geographically.

According to TeraWulf’s Form 8-K disclosure, Kentucky Power, an AEP subsidiary, is involved in grid access arrangements for the site, reportedly including a dedicated 345 kV substation connected to an existing 765 kV transmission network. That transmission infrastructure detail matters. A 345 kV substation connected to 765 kV backbone transmission isn’t a improvised solution. It’s heavy-voltage infrastructure designed to handle sustained bulk power delivery at the scale a gigawatt-class data campus requires.

The practical implication is that the site’s grid access appears to be purpose-built for this application, not repurposed from existing lighter- weight infrastructure. That’s a meaningful distinction for operators evaluating whether a campus can reliably deliver at nameplate capacity under sustained load.

What the 2028-2030 Delivery Window Means for Enterprise Teams

The Muskie Campus doesn’t exist as a capacity option until 2028 at the earliest. For enterprise teams planning multi-year AI infrastructure procurement, that timeline creates two distinct decision frames.

Teams planning for 2026-2027 deployments don’t factor this into current decisions. The capacity isn’t there. They need either existing hyperscaler regions, contracted colocation capacity, or existing purpose-built AI data center sites.

Teams planning for 2028-2030 enterprise AI compute capacity are operating in a market that doesn’t exist yet but is being shaped now. The organizations making commitments today, through long-term contracts, pre-lease agreements, or direct investment, are establishing their position in a capacity queue that will matter in two to four years. Enterprises that wait for 2027 to evaluate 2028 compute options may find the most favorable sites are already committed.

For infrastructure investors, the investment thesis is cleaner: the asset is a long-dated bet on AI compute demand materializing at scale in a location with genuine grid infrastructure. The bet works if demand arrives on schedule. It’s challenged if the AI compute buildout overshoots demand and capacity becomes a buyer’s market by 2028. Both scenarios are plausible. Neither is certain.

The Bet, Assessed

TeraWulf is not a hyperscaler. It’s an operator that built a track record in high-density power infrastructure through bitcoin mining, then pivoted to AI data center positioning as the economics shifted. The Muskie acquisition continues that pivot at a new scale.

The $6 billion AWS commitment Snowflake announced on the same day, covered in our Snowflake/Natoma brief, is a different kind of infrastructure bet. Snowflake is committing to cloud consumption. TeraWulf is committing to physical site delivery. Both bets rest on the same underlying forecast: AI compute demand reaches a scale, within the next five years, that justifies infrastructure investments that look aggressive by today’s standards.

The question worth tracking is pre-leasing velocity, not capacity announcements. Capacity announcements are cheap. Signed tenant commitments are expensive. Watch TeraWulf’s Q2 earnings call for any disclosure of committed tenants or pre-lease activity at Muskie. If TeraWulf exits 2026 with disclosed tenant commitments, the demand thesis is tracking. If the site enters 2027 without disclosed tenants, the gap between infrastructure positioning and confirmed demand is the risk story to watch.