Electricity is 1–6% of the price of a GPU-hour. The other 94–99% is how long you waited. The Time-to-Power Index is Algorithm AI's ongoing research on the real constraint in AI infrastructure — not cost, schedule.
Most industry commentary argues about the wrong 1–6% of the bill. We built the Time-to-Power Index to track what actually gates AI deployment: interconnection queues, equipment lead times, and the schedule risk that determines whether a data center exists at all — not just what it costs once it does.
Every piece we publish carries at least one finding that runs against our own commercial interest. That's the rule that separates a reference instrument from marketing, and it's why the Index is worth subscribing to even if you never do business with us.
Source: Orennia, 2026 US Behind-the-Meter Data Center Playbook (April 2026 data).
In 2024, power within two years was the industry's gold standard. That benchmark has already slipped to three years, and in some markets the wait now runs past six.
AEP Ohio's new cluster study sets a six-year minimum to connect. Data center projects it has already studied expect power in 2031 at the earliest, and its second cluster in 2033.
Texas tells the same story at larger scale. ERCOT's large-load queue reached 239 GW, up nearly 100 GW in a year, with about 140 GW more submitted but not yet counted. In 10 months it approved only 6 GW.
In Virginia, Dominion is preparing for 70 GW of data center demand. About 25 GW is expected by the end of 2031; the other 45 GW has no connection date at all.
Average time to build the limiting grid upgrade, in years
Source: Orennia, 2026 US Behind-the-Meter Data Center Playbook (April 2026 data), pp. 3, 13, 14. ERCOT, AEP Ohio and Dominion figures as reported in the playbook.
Algorithm AI builds compute that does not wait for that line to move.
Facing the queue, developers turned to building their own power. Natural gas is the listed fuel for 77% of announced behind-the-meter projects, and Texas and Pennsylvania lead because that is where the gas is.
The US has more than 300,000 miles of gas pipeline able to move over 160 Bcf/d, so proximity is rarely the problem. Capacity is. About half of that pipeline capacity is locked in long-term contracts with utilities, power generators and LNG exporters.
That makes available gas, not a nearby pipe, the new constraint. Up to half of 2025's contracted volumes roll off by 2030, but every large campus is now competing for the same headroom.
Source: Orennia, 2026 US Behind-the-Meter Data Center Playbook (April 2026 data), pp. 3, 5, 7, 8.
When the whole market crowds onto one workaround, the advantage goes to whoever needs neither the queue nor the pipeline.
Most behind-the-meter campuses announced so far sit in remote areas, away from the constraints of the major data center hubs. They are built for large-scale AI training. Orennia's own assessment is that few are well suited to inference, because they are far from population centers.
They also need room. A data center typically takes about one acre per MW of demand, and on-site gas generation adds another 0.2 to 0.4 acres per MW. Assembling that much buildable, unflooded, gently sloped land is its own multi-year project.
Source: Orennia, 2026 US Behind-the-Meter Data Center Playbook (April 2026 data), pp. 11, 19.
Inference is where AI meets customers, and customers live in cities. Algorithm AI puts compute where the work happens.
In Orennia's tracker, gas-fired projects account for roughly 48 GW of the 63 GW announced, across 26 projects. Solar is about 7 GW across 6 projects; wind about 2 GW across 4. Orennia also finds that the areas with cheap, plentiful gas closely match where the large campuses have been announced.
Most of the industry chose gas to get around the grid, and for big training campuses near a pipeline, that is a rational choice. We chose differently, for a different job: renewable-powered compute, sized to sit close to the people who use it.
One more caution: announced is not built. A second tracker, Cleanview, counts about 90 GW of planned self-generation across 59 projects, yet only about 2 GW was operating by mid-2026, across four projects.
Sources: Orennia, 2026 US Behind-the-Meter Data Center Playbook (April 2026 data), pp. 5, 10; per-fuel figures are approximate. Cleanview, Bypassing the Grid (2026).
Orennia Research, 2026 US Behind-the-Meter Data Center Playbook, data as of April 2026. Used with permission. orennia.com