AI Grid Capacity: 23 GW Is Not Data Centre Power

AI grid capacity has a seductive new number: 23 gigawatts. On 24 September, the US Department of Energy said a selected package of transmission upgrades could make more than that much additional electricity capacity available. You can almost hear the data centre sales decks being rewritten. Hold off. Moving power through a line is different from generating it, and neither guarantees that a particular site can take delivery. The distinction is where AI infrastructure plans meet the grid they actually have.

What the 23 GW announcement covers

The Department of Energy's SPARK announcement says it intends to help fund 31 grid-improvement projects across 26 states. The package totals $5.25 billion: $1.9 billion in federal funding and $3.35 billion in recipient cost-share. Recipients are expected to reconductor or rebuild more than 1,500 miles of transmission lines and install grid-enhancing technologies across nearly 21,000 miles. DOE says the combined work will make over 23 GW of additional electricity capacity available.

There is an important verb in that announcement: intends. These are selected projects, not completed upgrades. The programme page lists expected awards between October 2026 and January 2027. Selection is a credible step towards construction, but it is still a step. A capacity estimate in a funding announcement is not a commissioned connection.

The engineering is sensible. Reconductoring replaces existing wires with conductors that can carry more power. Grid-enhancing technologies can help operators measure conditions and use existing transmission assets more effectively. Both work on a problem the AI boom has exposed: you can have power somewhere on a system and still lack room to move it to the place asking for it. Upgrading an existing corridor can be less contentious than finding a new one, though that does not make the work automatic or immediate.

Notice what the release does not do. It does not assign 23 GW to AI data centres. It does not say 23 GW of new generation has been built. It does not promise that every selected line will serve a data centre corridor. Treat that national total as a proposed gain in the grid's ability to deliver power across particular routes, not as a pool of electrons waiting for a rack order. Geography is doing the hard work here.

A national total cannot sign a local connection agreement

A data centre needs much more than a large number on a federal press release. It needs a site, a feasible point of connection, available network capacity at that point, sufficient electricity supply over time, an agreed load profile, and terms for who pays when infrastructure has to be strengthened. Those are separate negotiations. One cannot be inferred from another.

Picture two locations. A transmission corridor in one state gains transfer capacity after an upgrade. A proposed campus in another state faces a local substation constraint. The first project is real, useful infrastructure; the second project's problem is still there. Even within the same state, a stronger high-voltage route does not automatically clear a downstream transformer, local network limit or queued connection study. The relevant question is whether the upgrade changes the binding constraint for the site in question.

That is also why "available capacity" needs careful handling. Transfer capability measures what the network can safely carry under specified conditions. It is not the same as a firm supply contract at a price the operator can live with. Weather, operating limits and competing demand can all affect what is usable at a particular time. AI grid capacity has to be tested at the point where a developer wants to connect, with its own supply arrangements on the table.

The International Energy Agency's energy-and-AI assessment makes a similar distinction in its discussion of connection queues. It calls for better disclosure of planned loads and closer work with system operators so grid investment matches projects that will actually materialise. That is a less photogenic answer than announcing another campus, but it is the one that stops everyone planning around the same scarce capacity.

The bill moves even when the bottleneck moves

There is another trap in reading the DOE number as free AI power: someone funds the network upgrade. The SPARK selections include $3.35 billion of recipient cost-share alongside the federal contribution. That tells you how these particular projects are funded; it does not settle how the cost of each eventual data centre connection will be recovered. A national grant and a local tariff are different parts of the ledger.

For a developer, the test is direct. If an existing transmission line is strengthened, does the project's connection charge fall, or does another piece of the network still need an expensive upgrade? If a utility builds ahead of a speculative queue, who bears the cost when the proposed loads arrive late or never arrive? If a campus takes power reliably at peak system hours, what generation and grid investment does that require? Those questions belong in the financial model before the plot of land becomes a glossy rendering.

The IEA warns that data centres are large, concentrated loads, often filled with servers gradually. They can initially seek connections larger than their actual near-term demand. If planners build around every application as though it were a fully operating site, the system risks investing too early; if they dismiss genuine projects, it risks building too late. The IEA says tariff design can help allocate the costs of new generation and grid upgrades fairly. That is policy, not a completed engineering fix.

We've covered the contract side in AI infrastructure's power procurement bottleneck. The federal announcement sharpens that argument rather than replacing it. Better wires create options. They do not sign the electricity purchase agreement, fix the delivery schedule or absorb the cost of the final connection on behalf of a campus.

Flexibility is a connection term, not a slogan

The serious opportunity is to make some AI demand less rigid. The IEA suggests that system operators could explore non-firm connections or reward data centres for reducing load during grid stress in exchange for faster access. It also describes batteries and other onsite assets as possible tools for supporting grid operation. Those are proposals and possibilities, not blanket claims that a data centre can safely switch itself off on command.

Different workloads tolerate different interruptions. Some training jobs may have room to move in time or between locations. A latency-sensitive service cannot simply disappear when the network gets crowded. A campus that wants a flexible connection has to specify which load can move, how quickly it can move, for how long, how it will be verified, and what backs up the load that cannot move. Otherwise "flexible" is just an adjective in a permit application.

We looked at that operating layer in dynamic energy control for AI infrastructure. The critical move now is contractual. Give the grid operator a credible curtailment envelope and give the developer a connection product it can price. The former protects reliability; the latter lets the project decide whether an earlier but conditional start beats waiting for fully firm service. Neither side should pretend the trade-off has vanished.

SPARK is worth watching because it attacks a physical limit that data centre builders cannot code their way around. The next useful update will be project-level: which routes actually gain transfer capability, when they enter service and which connection studies change as a result. Until then, 23 GW is a system-wide projection attached to selected upgrades. It is not 23 GW of data centre power ready for purchase. If someone uses it to sell you a site, ask for the connection agreement.