Unplugged AI: the rising menace

On 25 November 2025, Satya Nadella was asked whether the industry was heading for a compute glut. His answer, quoted a great deal since and usually without its date, was not about surplus at all. It was about unplugged hardware:

The biggest issue we are now having is not a compute glut, but it’s power and it’s sort of the ability to get the builds done fast enough close to power. If you can’t do that, you may actually have a bunch of chips sitting in inventory that I can’t plug in. In fact, that is my problem today. It’s not a supply issue of chips. It’s actually the fact that I don’t have warm shells to plug into.

That is the chief executive of one of the largest buyers of AI accelerators in the world saying he owns chips he cannot switch on. Ten months on, a grid operator has published the size of the gap he was describing.

Not how much was bought, but how much is unplugged

Almost all the coverage of the AI build-out counts dollars. Dollars are the wrong unit, because a dollar of capital expenditure and a dollar of revenue-earning capacity are separated by a transformer, a substation and a queue.

The right question is how much of what has been bought is switched on. For most of the market that cannot be answered. No filing carries a line for accelerators owned but not installed, and any figure you see for it has been modelled rather than measured.

Texas is the exception, because ERCOT publishes every stage. In a deck given to the Texas Senate Committee on Business and Commerce on 1 April 2026, ERCOT’s chief executive put the interconnection queue at approximately 410 GW of large loads, about 87% of them data centres. In its operational report for June 2026, ERCOT states that of the 8,926 MW which have received approval to energise, it observed a monthly peak of 3,966 MW, which it describes as how much approved load it believes is now operational. That figure was down slightly on May.

So: roughly 410 GW asking, 8.9 GW approved, 4.0 GW drawing power. Under one per cent of the queue is doing any work at all, and the part that is fell month on month.

That needs a caveat, and the caveat is more interesting than the number. An interconnection queue is not a list of real projects: the same development gets filed in several places at once, and a request costs far less than a building. ERCOT says so itself, describing a restudy loop that extends timelines by years. Its regulator has gone further. From 2027, only large loads with an executed interconnection agreement will count in the load forecast at all.

A regulator has looked at the announcements and decided to stop counting them.

Which matters to the accounts, because depreciation begins when an asset is placed in service, not when it is paid for. Until then the cost is carried in construction in progress and appears nowhere in the income statement. That is the correct treatment, not a trick. But it means reported profits during a build-out are flattered by exactly the part of the build-out that is not yet working, and the correction arrives later, all at once.

The vendor is underwriting its own demand

NVIDIA’s quarterly filing for the period ended 26 July 2026 is the document to read.

Its supply and capacity commitments rose from $119bn to $279bn in a single quarter, against total future commitments of $366bn. It holds $99bn of equity investments with a further $25bn committed. In August it signed guarantees capped at $105bn on roughly 4.25 gigawatts of leases, on behalf of an affiliate of OpenAI, and memoranda of understanding with capital providers to mobilise, in its own words, more than $500 billion of third-party capital.

It has also committed to take its customers’ unsold capacity:

Under these agreements, AI clouds procure our data center infrastructure products and we commit to cloud service agreements, which the AI clouds can unilaterally stop providing to us and sell to third-party customers at more advantageous rates.

Those commitments run typically six years and totalled $36bn at the end of July. Read the sentence again for the asymmetry: the cloud can walk away and sell the capacity to someone else at a better price. NVIDIA cannot.

And it is extending credit, in a paragraph about payment terms:

In certain cases, for investment-grade customer purchases, we have and may in the future provide longer payment terms ranging from 90 days up to one year to assist customers with large data center builds depending on size.

Meanwhile inventories rose from $21.4bn at the end of January to $31.6bn at the end of July, with raw materials tripling to $11.3bn.

The comparison everyone reaches for, and where it fails

The fibre build-out of the late 1990s is the usual analogy, and it fits better than most. The capital was enormous, the demand forecast was a slogan, and the equipment makers ended up lending their customers the money to buy the equipment.

Lucent’s own filings record a provision for bad debts and customer financings of $2,249m in fiscal 2001 and $1,253m in fiscal 2002, against $505m in fiscal 2000. The fiscal 2002 filing is specific about the cause: provisions for three customer financings, including amounts due from One.Tel and Winstar, accounted for approximately 60% of the fiscal 2001 expense. Winstar filed for Chapter 11 in April 2001 and One.Tel went into voluntary administration about six weeks later.

Today’s instruments are not all the same thing, and most commentary lumps them together as “circular financing”, which is a phrase that saves people the trouble of saying which one they mean. Equity investment, a guarantee against a tenant default, an offtake for unsold capacity and extended payment terms are four different instruments with four different failure modes.

But the analogy is usually deployed to reassure, and read carefully it does the opposite.

Fibre was a 25-year asset. It was overbuilt against demand that arrived about a decade late, and when the demand finally came the glass in the ground was still good. Somebody made money on it, just not the people who laid it. Dark fibre was patient capital that had been mislabelled.

An accelerator is not patient. It is depreciated over five or six years, it is superseded in about two, and its successor is shipping while the racks it replaces are still waiting for a substation. Unlit fibre waits. An unpowered rack ages.

That asymmetry is the whole argument, and it points the other way from the comfortable version of the comparison. The dot-com overbuild left an asset the future could use. This one may leave one that is obsolete before the power arrives to run it.

What this means if you run an estate

Texas is one grid, and nobody publishes the equivalent nationally. Non-deployed chip counts, fleet utilisation and the split of NVIDIA revenue between resellers of capacity and consumers of it are all undisclosed, and anyone who gives you those numbers has estimated them.

So this changes very little for you immediately, which is the point worth making to anyone selling urgency. The capacity question is a supplier problem and a shareholder problem. It becomes yours in one specific way: pricing. A build-out funded from operating cash flow can be sold cheaply for a long time. A build-out funded from debt cannot, and reserved capacity bought at today’s rates against a multi-year term is a bet on which of those you are dealing with.

We would look hard at any contract that locks a term beyond the depreciation life of the hardware it runs on, and harder at any that prices as though the last four years of capacity growth were the normal case. That is not a prediction about a crash. It is the ordinary discipline of not signing a long contract at the top of a supply cycle.

Sources

Satya Nadella, BG2 podcast, 25 November 2025.
ERCOT, “ERCOT Update: Senate Committee on Business and Commerce”, 1 April 2026.
ERCOT, Monthly Operational Overview, June 2026, published 17 July 2026.
Public Utility Commission of Texas, Project 58480, effective 1 March 2026.
NVIDIA Corporation, Form 10-Q for the quarter ended 26 July 2026, filed 26 August 2026.
Lucent Technologies, Form 10-K for fiscal 2002, Exhibit 13.