A trillion dollars of compute, bounded by the grid — the ceiling nobody can buy their way past

The capex forecasts keep re-rating upward. The megawatts do not re-rate on the same schedule. The defining compute fact of H2 2026 is that money is fast and power is slow.

Hyperscaler AI capex has been raised to $750 billion for 2026 and is set to cross $1 trillion in 2027, while US data-center demand climbs from 23 GW in 2023 to 42 GW in 2026. The upward revision is the tell: the forecasts are chasing demand, not settling, and a trillion-dollar 2027 is now the base case.

The constraint hides inside the capex number

A lab can commit capital in a quarter and order chips on a similar timeline. It cannot conjure a grid connection on that schedule. The electricity figure — demand nearly doubling in three years — is the load that has to be found, delivered, and interconnected, and it grows faster than the grid was built to accommodate. The dollars outrun the megawatts, and the megawatts are the binding constraint.

Follow the order books

You can see the constraint in who is suddenly getting rich. Eaton's electrical backlog jumped 48%, Caterpillar's power-generation revenue rose 41%, and GE Vernova's Q1 data-center orders exceeded its entire prior year. When the picks-and-shovels are transformers and turbines, the scarce input is power delivery, and the value accrues to whoever makes the gear that delivers it.

The lead times on that gear are now part of how fast any lab can bring compute online. A GPU order that arrives before the substation to power it is not compute — it is inventory. That inversion is why the compute story of 2026 is a story about grid interconnection queues and power-purchase agreements as much as about silicon.

Compute advantage becomes an energy-and-real-estate game

The strategic consequence is that the scarce asset is shifting from chips to power-ready land and grid contracts. Whoever holds those holds the actual bottleneck, which is why frontier capital is flowing toward energy deals and modular data-center designs as hard as toward accelerators. You cannot buy your way past a ceiling that is measured in years of interconnection lead time — you can only secure a place in the queue earlier than the next lab.

Help Net Security — Power shortages could slow AI data center expansion → · EnkiAI — AI data center power: grid limits reshape energy in 2026 →