The grid was not built for what is coming.

A transmission corridor in West Texas carries 1.2 GW into a single campus, eight buildings across four million square feet. Every hour that campus produces what Jensen Huang named at GTC 2026: tokens. Every hour the lights are on, the economics hold. Every hour the facility sits dark, the opportunity cost compounds.

The grid was sized for industrial load. What is arriving is closer to a national GDP line item in a single postcode.

Last week, a grid operator in Tennessee approved a 500 MW connection labelled "light industrial." Nobody told the nuclear plant 40 km away.

That was the gap between facilities. This week is the gap beneath it. What the grid was actually sized for. Against what is arriving.

I have spent six editions on the gap between construction complete and permit to operate. The documents that disagree. The drawings that are already wrong. The standards that conflict. The facilities that do not see each other. Every edition has lived inside the gap. This week, it extends to the grid.

NVIDIA has stopped talking about data centres. It talks about AI factories. Specifically, 1 GW AI factories. At GTC 2026, Huang made the shift explicit: "Tokens are the new commodity. AI factories are the infrastructure that produces them."

The unit has changed. So has the scale.

NVIDIA and OpenAI announced a 10 GW deployment in September 2025. NVIDIA committed up to $100 billion progressively as each gigawatt is deployed. That is roughly £8bn ($10bn, €9.5bn) of NVIDIA hardware per GW. Civil, mechanical, electrical and power components wrap around it, on top. Lenovo joined the Gigawatt AI Factories Program at the same event as NVIDIA's enterprise counterpart.

This is not a speculative category. It is a named, priced, ecosystem-endorsed class of infrastructure.

Edition 1 put a number on delay. £79m ($100m, €95m) per day for a single 1 GW AI factory sitting dark. That figure came from a conservative model. $0.10/kWh sold compute. 1 GW sustained. 24 hours.

Six weeks later, that number reads differently.

Building on Huang's stage remarks, secondary industry coverage now projects $150bn in annual revenue per 1 GW AI factory at the high end. The driver is tokenised inference, not kWh rental. That is £324m ($411m, €392m) per day in gross token revenue potential per GW, at full utilisation.

Even at a fraction of that potential, Edition 1's £79m ($100m, €95m) per day is the floor. The ceiling sits several multiples above.

Every day of commissioning delay at a 1 GW AI factory prices in the tens of millions. Every week prices in the hundreds. A three-month slip is eight figures on the wrong side of the balance sheet. Before anyone has argued about compute demand.

In Texas, the first operational site of the Stargate programme at Abilene draws 1.2 GW and is scaling past 2 GW. Eight buildings, four million square feet. DPR Construction, Turner Construction and M.A. Mortenson on the build. Crusoe as site developer. Oracle Cloud Infrastructure runs the compute. NVIDIA Vera Rubin is the platform.

In France, Flamanville 3 is commissioning. In Singapore, the Energy Market Authority caps data centre power at 5% of national capacity. In South Korea, Korea Hydro and Nuclear Power delivers reactors on a six-year cadence. In the UK, the National Grid ESO connection queue exceeds 700 GW, and most of those applications will never be built.

The grid was sized for factories, warehouses, retail parks and homes. Not for single sites consuming the electricity of a mid-sized city, sustained, 24 hours a day, for decades. In most jurisdictions, the transmission and generation build-out required to serve the arriving demand has not started.

That is before anyone discusses whether the compute demand is real.

Critics of the Stargate programme raise three arguments. Speculative demand. Stranded-asset risk. The concentration of a single hyperscaler behind most of the committed capacity.

Those are real arguments. None of them change the constraint.

If only a fraction of the projected token economy materialises, the delivery question still dominates. Whoever can put a 1 GW AI factory into service first captures the revenue window. On schedule. With the grid connection, cooling, and power to match. Whoever slips loses the window to someone who did not.

In construction terms, this is not speculative. The Abilene site has been delivering against milestones since 2022, with its first buildings operational and subsequent phases under construction. On 22 April, Bentley Systems published the mechanics behind that pace, a piece I co-authored with Tomas Kellner. The discipline is reconciling work breakdown against location breakdown. At velocity. Under a shared reference architecture across every trade on site. That is public. The pace of delivery is the economic moat, regardless of which analyst currently doubts the demand profile.

At 1 GW, delay is the expensive opinion. Delivery is not.

Across Atlas, Vistergy tracks nuclear, data centres and LNG in one view. The arriving demand and the available supply do not appear in the same frame anywhere else in public.

A 1 GW AI factory in Texas sits within transmission range of generation assets built in a different decade. For a different economy. A hyperscale campus in Dublin draws from a grid supplied by LNG terminals that were not sized for compute load.

The questions that matter in the next 18 months are cross-sector:

These cannot be answered from a single-sector map. Atlas is the surface where the three layers connect, in public, without a paywall.

The NVIDIA investment case is now specific about the economic unit. Jensen called tokens the new commodity. That is not marketing. It is a pricing signal.

If the unit economics are what NVIDIA and its partners have publicly committed to, the infrastructure pace has to match. If it does not, the opportunity cost is not an opinion. It is a clock.

The grid was not built for what is coming. The teams delivering 1 GW facilities on time, on public record, are solving a problem others have not yet priced. Regulators. Transmission operators. Grid planners in most jurisdictions.

Edition 8: The gap is not closing. It is getting wider.

If you build, operate, invest in, or regulate infrastructure anywhere in the world, this is written for you. Subscribe to Still Dark.

This newsletter lives in the gap between digital delivery complete and permit to operate. That gap is where value dies, and where it can be recovered.

I also co-author The Vistergy Brief at vistergy.com/archive. Satellite and geospatial monitoring, facility lifecycle intelligence, and standards architecture across LNG, nuclear, data centres, utilities and construction. Subscribe to both for the full picture.

The podcast is live. Permit to Operate is at permit2operate.com, Apple Podcasts, Spotify, and YouTube.

Permit to Operate.

David Ayeni

AI has everyone talking. Not everyone has answers. At Gladly Connect Live, CX leaders from Condé Nast, Smith Optics, and more share exactly how they moved AI from pilot to production, the timeline, the systems, the QA loops. 13+ sessions built for the moment we're all in. For CX and ecommerce leaders. Atlanta, May 4–6. Space is limited, secure your spot now.

Register now