AI Chips Can Arrive Before the Power. The “Half Unused” Claim Needs Care.
![]() |
| Around 20% of planned global data-centre capacity could face grid delays by 2030, in the IEA’s 2025 analysis. This does not measure GPUs currently idle. Graphic: Routes, Bytes & Budgets, based on IEA, Energy and AI (2025), CC BY 4.0. Redrawn as a percentage bar. |
A graphics processor can be ordered, delivered and paid for without doing a single useful task. Somewhere between the purchase announcement and the working AI service, it still needs a building, cooling, networking and a dependable electricity connection.
That gap is becoming one of the most interesting parts of the AI boom for me. It also needs a more careful explanation than the claim that half the industry's chips are simply gathering dust in boxes.
Where the “half unused” argument comes from
In his September 22 analysis, Ed Zitron argues that more than half of Nvidia's AI GPUs have yet to become operational capacity. He builds that case from spending, infrastructure disclosures and estimates of installed computing capacity.
I regard it as an argument worth examining, not an audited count of stored processors. Turning capital expenditure or megawatts into a GPU inventory requires assumptions about hardware mix, costs and what has actually been installed. Buildings, CPUs, networking and other equipment complicate that calculation.
There is another distinction I want to preserve: a GPU waiting for installation, one installed in a building awaiting power and one running below full utilisation represent different problems. Calling all three “unused” hides information we need.
There is firmer evidence for the power bottleneck
The International Energy Agency's 2025 analysis estimated that grid constraints could delay around 20% of global data-centre capacity planned for construction by 2030.
That is a forward-looking estimate of construction capacity exposed to connection delays. It is not the percentage of purchased GPUs currently sitting idle, and it cannot be used to confirm the “half unused” claim. It does support the narrower point: access to electricity can limit how quickly new computing capacity becomes usable.
The IEA's report summary also describes lengthy transmission-development timelines and longer waits for key grid equipment. Ordering the processors is only one part of a project whose dependencies move at very different speeds.
Strong demand and delayed equipment can coexist
Microsoft said in its July 2026 earnings call that Azure demand continued to exceed available capacity and that additional capacity delivered during the quarter was quickly monetised.
That is Microsoft's account of its business, not an independent utilisation audit. Still, it illustrates why I do not see a simple contradiction here. Customers may be waiting for a particular service in one location while equipment elsewhere is waiting for a connection or commissioning.
“We bought more chips” and “we can serve more customers today” describe different stages of the same investment.
The waiting has a cost
My concern is the mismatch between the pace of construction and the pace of hardware development. A delayed machine may remain perfectly useful, but the economics can change before it starts earning money. Financing continues, and customers may gain other options.
That does not make every advance purchase irrational. Securing equipment ahead of a planned opening can be sensible. The question is whether the delivery schedule, electricity connection and customer commitments fit together—and how much room the plan has for delays.
It is the same reason I treated Amazon's proposed nuclear-power development as a project with substantial work ahead. Announced capacity and available electricity should not share the same date by default.
The questions I want companies to answer
I want clearer distinctions between equipment ordered, equipment delivered, capacity energised and capacity actually serving customers. A single enormous investment number cannot tell me all of that.
For anyone renting computing, the practical version is simpler: is the right hardware available in the region I need, on the date I need it, under terms that protect my project? A future campus does not finish today's job.
The physical limits of AI are worth taking seriously. I just want the evidence to stay as precise as the conclusion: power delays are documented; a universal count of chips in boxes remains much harder to establish.

Comments
Post a Comment