Yesterday our newsroom published a piece checking the numbers behind the AI data-centre slowdown. The short version: less has actually stopped than the headlines suggest. The most-repeated claim — that roughly half of America's 2026 capacity has been cancelled — did not survive contact with people who track construction directly. Most of what gets counted as "cancelled" was an announcement with no financing, no equipment order and no grid connection, which was never going to be built in 2026 anyway.

That article is a report. This one is not. Our newsroom counts things; I am going to argue with the conclusion most people draw from the count.

Because whichever way this goes — a balloon that pops, or a high-pressure pipe that splits — I do not think we come out the other side with cheap AI. I think we come out with AI that costs several times what it costs today, and with a lot of businesses discovering that the productivity tools they built into their workflows were never priced at what they cost to run.

The comforting story

The optimistic case has a good pedigree, and I want to state it properly rather than knock down a weak version of it.

In the late 1990s, telecom companies laid something like 80 million miles of fiber across the United States on the assumption that demand would be limitless. It wasn't, the bubble burst, and the survivors inherited an enormous amount of unlit "dark fiber." Bandwidth costs collapsed — by around 90% — and that collapse is arguably what made the 2000s possible. Cheap bandwidth gave us streaming, cloud computing, and every distributed application we now take for granted. The bust was painful for investors and wonderful for everyone who came after.

Apply that template to AI and you get a reassuring forecast: the buildout overshoots, the money gets burned, the assets get sold at cents on the dollar, and the rest of us end up running large models on somebody else's stranded capital. A lot of thoughtful people expect exactly this.

I don't. Three reasons.

1. Today's price is not a price

The first problem with predicting a fall is that we are not starting from a market price. We are starting from a customer-acquisition price.

You can watch this being corrected in real time, without any crash at all. On 1 June 2026, GitHub moved Copilot from flat seats to usage-based billing, with monthly AI credits metered against token consumption. The stated reason was straightforward: Copilot now drives agentic workflows that consume far more compute than autocomplete did. The headline seat price did not rise. What changed is that heavy users now pay for what they use instead of being carried by light ones.

That is what the end of a subsidy looks like from the inside. It does not announce itself as a price increase. It arrives as a change in billing model, a credit allotment, an overage rate — and the bill goes up for exactly the users who found the tool most valuable.

Meanwhile the entity setting the reference price for frontier inference is reported to be on course to lose around $14 billion this year. Whatever else is true, a number like that is not a stable basis for anyone's cost model. When it corrects, it corrects upward.

2. The glut asset melts

This is the part of the fiber analogy that I think simply does not transfer, and it is the heart of my disagreement.

Dark fiber had two properties that made it a gift to the future. It barely depreciated — glass in a conduit is still glass ten years later, and the expensive part of the network was the trench, not the strand. And once it was lit, the marginal cost of another gigabit over it was close to nothing.

Neither is true of AI compute.

On the first point, there is an active and unresolved argument about how fast these assets actually die. Michael Burry has argued that hyperscalers are depreciating Nvidia hardware over five or six years when the real economic life is closer to two or three, and estimates roughly $176 billion of understated depreciation across the industry between 2026 and 2028. You can disagree with his number — plenty of people do — but you cannot dismiss the direction, because the companies themselves have moved that way: Amazon shortened the useful life of a subset of its servers from six years to five in early 2025, explicitly citing the pace of AI development, and took a $700 million hit to operating income for the privilege.

A stranded fiber strand costs nothing to own. A stranded GPU cluster loses value every quarter whether or not anyone runs a token through it.

On the second point: there is no equivalent of "lit fiber" economics here. Every token has a real, recurring marginal cost — electricity, memory bandwidth, cooling, and the amortisation of a chip that is aging fast. You cannot serve the second billion tokens for free the way you can push the second gigabit down an installed strand. The thing that made bandwidth deflationary is precisely the thing inference does not have.

3. The inputs are already moving the wrong way

The third reason is the one I find hardest to argue against, because it is happening now, independently of any bust.

Memory first. Conventional DRAM contract prices rose 90–95% quarter-on-quarter in the first quarter of 2026, with a further 58–63% increase projected for the second. High-bandwidth memory capacity for 2026 is sold out, with manufacturers turning away orders, and suppliers have warned the imbalance may persist for years. This is not a speculative price that pops when sentiment turns. It is a manufacturing capacity constraint, and manufacturing capacity does not appear because a valuation fell.

Then electricity. In the PJM market, capacity prices went from $28.92 per megawatt-day in 2024/25 to $329.17 in 2026/27 — a factor of ten. Data centre demand was responsible for 63% of the increase in one of those auctions, amounting to $9.3 billion recovered from customers in higher rates. Households in that region are already paying for compute they will never use.

And permits. As our newsroom piece sets out, more than 300 data-centre bills were introduced in the first six weeks of 2026, New York became the first state to pause anything above 50 megawatts, and local moratoriums now number in the hundreds. Add transformer lead times measured in years, and the picture is clear enough: the input cost of a unit of AI capacity is rising, and the political cost of building more is rising faster.

What a burst actually does

So here is where I land. A crash in this market does not do what a crash in fiber did. It does not leave behind a cheap, durable, nearly-free-to-operate asset. It leaves behind buildings full of depreciating chips that still need power the grid does not have, memory that is still scarce, and permits that are harder to get than they were before.

What it does remove is the willingness to sell below cost. That willingness is the single largest reason AI looks affordable today, and it is a function of capital abundance, not of engineering. Consolidation follows: fewer providers, each with more pricing power, serving customers who by then have rebuilt their workflows around the tool and cannot easily walk away. That is not a recipe for deflation.

Balloon or pressure pipe, the failure mode is the same for buyers. The bill arrives.

Where I could be wrong

I would rather state this than have it pointed out to me.

  • Efficiency could outrun cost. If model efficiency keeps improving several-fold per year, the cost per useful task can fall even as the cost per unit of compute rises. This has been true so far, and it is the strongest counter to everything above.
  • The glut could be bigger than I think. If enough capacity lands at once and demand disappoints, GPU-hour prices could fall hard for a period — even if the underlying economics don't support it.
  • Cheap capital could return. The subsidy ends when the money stops. If it doesn't stop, it doesn't end.

What would change my mind is fairly specific: memory contract prices falling for two consecutive quarters, and at least one major provider raising published inference prices without losing volume — that would tell me the market has already repriced and I am arguing about something that has happened.

What we are doing about it

This is not an abstract position for us, and it is not specific to one product. Talivio runs more than thirty products across e-commerce and logistics, business CRM, communication and marketing, compliance and finance, and AI. The AI newsroom is one of them — the most visible one, because it is public and updates by the hour — but the assumption behind it is the same assumption behind the rest of the portfolio: token prices will not stay where they are, so nothing gets built as though they will.

Concretely, that means: no model identifier is ever written into application code, so any task can be moved to a different tier or provider without touching logic; a budget guard sits in front of every call and cannot be bypassed; cheap models handle the high-volume, low-stakes work and expensive ones are reserved for judgement; and cost per unit of output — per published article in the newsroom, per processed document or reconciled record elsewhere — is a tracked number, not an afterthought discovered on an invoice.

None of that is exotic. It is the ordinary discipline you would apply to any input whose price you did not control — and the mistake I see most often is treating AI as though it were the exception. It is a commodity input with a volatile, politically exposed, physically constrained supply chain, currently sold to you below what it costs to produce, and we build every product on that basis, not just the one that happens to write about it.

Build as though that ends. If I am wrong, you have merely built something efficient.