Compute Ordered Is Not Compute Installed

AI infrastructure is usually measured by what has been ordered: GPUs purchased, capex announced, gigawatts "in development". Investors, lenders and even markets increasingly treat these numbers as if they were capacity. They are not. Compute becomes capacity only when it sits in an energised, cooled, permitted building. Between the order and the power-on date lies a delivery gap, and in 2026 that gap is widening.

THE BANKS HAVE STARTED MEASURING THE GAP

In September 2026, the gap went from thesis to consensus. A widely shared Financial Times Alphaville piece brought together research from the largest banks, and each measures the same thing from a different angle.

An academic paper presented at the Brookings Papers on Economic Activity conference in September 2026 adds a longer view. Of a little over 500 GW of US data center projects announced, its central scenario sees about 183 GW operational by 2032, about 117 GW completed later, and about 227 GW never built. See what the paper means for lenders.

Sources: Jefferies research as reported in the press, Morgan Stanley research note of 21 September 2026 as reported by Blockspace and Yahoo Finance, Goldman Sachs Research (May 2026), Financial Times Alphaville (September 2026); Brookings Papers on Economic Activity conference draft (September 2026).

ORDERS MOVE AT THE SPEED OF CAPITAL

A GPU order can be placed in weeks. Capex commitments are announced in the hundreds of billions: Goldman Sachs estimates around $7.6 trillion of AI capex over 2026–2031, and Jefferies counts some 34 GW of hyperscaler capacity in development. Capital is not the constraint. It moves fast, and it is visible.

INSTALLATION MOVES AT THE SPEED OF PHYSICS AND PROCEDURE

Grid. In France, an RTE connection takes several years on a brownfield site that is already connected, and 3 to 5 years on a greenfield site. Queues are crowded with projects that will never be built, which complicates planning for everyone else.

Equipment. Major transformer manufacturers quote 48 to 60 months or more. Grid operators are reserving factory capacity years ahead, as the €350M Hitachi Energy–Enedis agreement shows, leaving less room for late private buyers.

Permits and community. A single permit can stall a campus: Oracle's New Mexico project waits on a gas pipeline right-of-way, and its loans have been reported trading around 90 cents. In France, several projects face legal challenges or local campaigns.

WHY THE GAP MATTERS

For an operator, ordered-but-uninstalled compute is capital that does not earn. Accounting depreciation usually starts only when equipment is placed in service, but economic obsolescence does not wait: each new chip generation erodes the value of the previous one, whether it is running or not. For a lender, the gap is the period in which interest accrues and debt service depends on a date nobody fully controls. For equity investors, it means announced capacity figures overstate what will actually generate revenue in a given year.

THE COUNTERARGUMENTS

Not every chip needs a new building. Part of the accelerators sold replace older GPUs in existing halls, and part are deployed outside North America. Chip-implied power figures therefore overstate the need for new capacity.

Construction may be undercounted. Some analysts using satellite imagery argue that figures on capacity under construction are underestimated, particularly for hyperscaler self-builds.

Reserved capacity is still revenue. Customers who reserve capacity pay for it, whether or not they use it fully. For a cloud provider, low utilisation is not necessarily a financial problem.

Power is being built around the grid. Onsite generation, fuel cells and other fast-power solutions are growing precisely to close the gap, which is why Morgan Stanley's shortfall shrinks from 57 GW to 33 GW once they are included.

None of these points closes the gap. They explain why its size is debated, not whether it exists.

WHERE THE UNPLUGGED DEMAND GOES: THE EUROPEAN ANGLE

Jefferies notes that the US build pipeline is already shifting toward regions with easier interconnection and permitting. The same logic applies across the Atlantic. Compute that cannot be powered in the US will look for markets where it can. France combines low-carbon nuclear baseload, a single transmission operator, a regulator-approved fast-track and 4.8 GW of pre-identified connection-ready sites. It has its own constraints on equipment and permits, but its timelines are legible. See grid data center France and how European queue reforms compare.

HOW TO READ AI INFRASTRUCTURE NUMBERS

If the answer to any of these is no, the capacity is ordered, not installed. That distinction is at the heart of our due diligence for lenders and funds. See also what lenders miss in AI data center due diligence.

FREQUENTLY ASKED QUESTIONS