Dedicated Generation: The AI Build-Out Has Crossed a Threshold in Energy Infrastructure

1. The August Data Points

Data Center Knowledge reported in August 2026 that the AI infrastructure build-out has produced a new category of project: the dedicated-generation campus, in which the operator controls the power supply rather than purchasing from the grid. [Established — Data Center Knowledge, “New Data Center Developments, August 2026.” Tier 2.]

Three announcements bracket the month. Brookfield Asset Management and NextEra Energy are developing a 1.2 GW AI campus at the former DOE uranium enrichment facility in Paducah, Kentucky — a site chosen in part because its existing high-voltage infrastructure, which once served the enrichment process, can be repurposed for data centre load. [Established — Data Center Knowledge, August 2026. Tier 2.] The campus includes dedicated on-site generation; it does not rely on the regional grid as a primary power source.

Meta’s Hyperion campus in Louisiana is being scaled toward 5 gigawatts of capacity — a figure that, if realised, would make it among the largest single power consumers in the United States. Meta simultaneously announced a $9 billion investment in a 1 GW campus in Alberta, Canada, in a jurisdiction where hydroelectric power is available at scale and where the regulatory environment for large energy consumers is permissive. [Established — Data Center Knowledge, August 2026. Tier 2.]

The most structurally revealing data point is OpenAI’s Project Camellia in Georgia. Georgia Power’s utility filings documented a “3,200 MW customer commitment” months before OpenAI made any public announcement of the facility. [Established — Data Center Knowledge, August 2026, citing Georgia Power utility documentation. Tier 2 citing Tier 1.] The scale of the commitment — 3.2 gigawatts, comparable to three mid-sized nuclear plants — appeared in regulatory filings before it appeared in press releases. The utility knew before the public did, because the utility had to plan for it.

August 2026 · Dedicated-Generation Campus Announcements

Source: Data Center Knowledge, August 2026 · Georgia Power utility filings

2. The Scale Shift

The US Energy Information Administration tracks data centre electricity consumption as a distinct category. The figures describe a step-change rather than a trend: from approximately 23 gigawatts of data centre load in 2023 to approximately 42 gigawatts by mid-2026 — an increase of 83% over three years. [Established — US Energy Information Administration, data centre load tracking, 2023–2026. Tier 1.]

For comparison, the entire installed electricity generating capacity of the United Kingdom is approximately 76 GW. The United States is adding, in data centres alone, more than half of a mid-sized European nation’s total generation capacity in a three-year window.

US Data Centre Electricity Load (GW) — EIA, 2023–2026

Source: US Energy Information Administration · data centre load tracking 2023–2026

This is not, in itself, unprecedented. Industrial booms have always been electricity booms. The electrification of the US economy in the early twentieth century, the postwar manufacturing expansion, and the internet’s server-farm buildout all produced spikes in electricity demand that the grid eventually accommodated. The distinctive feature of the current phase is the concentration: the demand is not spread across millions of small consumers but is concentrated in a small number of very large facilities, each requiring hundreds of megawatts or multiple gigawatts, in specific geographic locations, on compressed timelines.

Grid operators are not built to accommodate that pattern. They are built for distributed load growth that can be managed through incremental transmission upgrades. A 3.2 GW facility appearing in a utility filing before it appears in the news cycle is not a grid-planning scenario any regional operator has modelled.

3. From Purchase to Production

The shift from power-purchase agreements to dedicated on-site generation is the structural inflection point. A power-purchase agreement (PPA) means the AI operator is a sophisticated consumer of electricity — it contracts for supply from generators, it negotiates terms, but the generation infrastructure belongs to someone else. Dedicated on-site generation means the operator is building and owning the generation asset. It has become, in a meaningful sense, an energy company.

This distinction matters for risk distribution. A PPA exposes the operator to electricity price volatility and to supply reliability risk from the grid. Dedicated generation internalises those risks: the operator controls its power supply but also owns the capital cost, the maintenance obligation, and the liability associated with operating a large generation facility.

It also matters for what happens to the grid. An operator that purchases from the grid adds load; the grid must supply it and plan for it. An operator with dedicated generation partially decouples from the grid — its reliability does not depend on the grid in the same way, and in some configurations it can sell surplus capacity back. But the grid still must accommodate the connection infrastructure and the backup arrangements. The grid does not become simpler when large consumers go to dedicated generation. It becomes more complex.

4. The Scarcity Inversion

The Navigator covered the silicon constraint in Sounding No. 4 (“The Loop,” 5 August 2026) and the debt-financing of GPU procurement in Sounding No. 19 (“The Chips Are Bought With Borrowed Money,” 22 August 2026). Both analyses identified the GPU as the primary bottleneck. That constraint has materially loosened: Nvidia’s production has scaled, alternative suppliers have emerged, and Broadcom’s $100 billion debt-financed procurement capacity has put GPUs at a pace that would have been implausible eighteen months ago.

The binding constraint has inverted. It is no longer silicon; it is electrons. Specifically, it is energised capacity at the voltage level, geographic location, and supply reliability that AI workloads require. The developers who secured large power contracts in 2024 and early 2025 — when grid operators were still negotiating in the old mode — have a significant advantage over those seeking capacity now, when utilities are queuing interconnection requests for years.

This inversion has implications for the Purser’s analysis of AI investment valuations. The circular financing loop documented in Sounding No. 4 assumed that the GPU was the scarce asset around which value accrued. If electricity is now the scarce asset, the value distribution in the AI build-out shifts toward generators, grid operators, and transmission companies — not toward chip vendors and their lending partners.