Ask ten people whether AI compute is a bubble and you'll get ten confident answers and no agreement, because the question is doing two jobs at once. It sounds like a question about demand. In practice it is almost always a question about asset value — specifically, whether the fleets being built today will be worth anything when the depreciation schedule catches up to them. That is a narrower and much more tractable question, and the industry keeps answering it with a single word when it needs two.

Two clocks, not one

The first clock is financial and contractual: a hyperscaler buys an accelerator, assigns it a useful life — commonly in the range of a handful of years — and expenses it across that window. That schedule is a policy choice, informed by tax treatment and investor expectation as much as by engineering. It says nothing at all about what the card can do on the day the schedule ends.

The second clock is technical. An accelerator becomes less useful when the workloads it can serve stop mattering, or when newer silicon makes it uneconomic to run for those workloads. That clock is set by memory capacity, interconnect, power efficiency, and — most of all — by whether the software stack still targets it well. It is not a fixed schedule. It is a moving target that different workloads cross at wildly different times.

Once you separate them, the shape of the market changes. A card sitting at the end of its accounting life is not a worthless card. It is a card whose owner has a capital-cost reason to stop holding it, sitting next to buyers who have a capability reason to want exactly that card — if the workloads it serves are still alive.

Why the two clocks drift apart

Three forces are pulling them apart in 2026.

  • Workloads are heterogeneous, and that is the whole story. Frontier training is the most obsolescence-sensitive workload in the world: it wants the newest interconnect, the largest memory pools, and the lowest power per FLOP, and its economics collapse on older silicon. Inference at the long tail is nearly the opposite: it is memory-capacity-bound, latency-tolerant of queueing, and frequently doesn't care about interconnect at all. A card that's a liability for one is often fine for the other, and the second market is far larger in unit terms.
  • Memory capacity ages slower than compute does. The generational step function in most accelerators is FLOPS and bandwidth; VRAM per card has advanced more sedately. A card with enough VRAM to hold a realistically-quantized model keeps serving that model long after its raw speed has been overtaken — see our VRAM arithmetic piece for why capacity, not compute, is the gate.
  • Utilization is the real driver, not age. A card at 20% utilization was mispriced on day one regardless of silicon generation; a card at 85% utilization is doing its job whether it's new or three generations old. Obsolescence shows up as an inability to reach utilization at a viable price, which is a market fact, not a hardware fact.

The residual-value problem nobody wants to price

This is where the bubble debate gets genuinely interesting, and where the industry's accounting habits become a problem. Because there is no transparent market for used accelerators at scale, residual value — the single most important input into whether the buildout pencils out — is not priced anywhere public. Ask prices in scattered listings, self-reported figures in earnings calls, and analyst estimates built on both are routinely used as if they were the same number, and they are not remotely the same number.

The financial consequence is subtle. If residual values are systematically overstated, then the effective cost of owning capacity is understated, which means the returns being reported across the industry are flattered by a quantity nobody has ever observed trade. If they are understated, the opposite. Either way, the market is not disagreeing about residual value — it is largely failing to observe it, which is a very different and more fixable state of affairs. We wrote about the collateral side of this problem separately; the point here is that the bubble question cannot be answered before the price question is.

What a real observability layer would change

Suppose for a moment that executed used-hardware prices were visible at reasonable granularity — by class, by age band, updated continuously, with volumes attached. What would actually change?

  • Depreciation schedules would stop being opinions. Boards would have an empirically grounded useful-life assumption rather than a defensible-sounding one, and the gap between the two would become visible and scandalous in the cases where it exists.
  • Price asymmetry would shrink. Today's gap between asking and paying is a tax on every transaction, borne mostly by the party with less information — typically the seller of surplus capacity, which in this cycle means the smaller operator.
  • A genuine second-life market could price itself. Right now the assumption is that old fleets get bought by sketchy intermediaries at whatever they'll pay. With prices visible, that becomes a normal, financeable secondary market — which is exactly the thing that keeps utilization high and total system cost low.

This is the part of the decentralized-compute thesis that stands up best under scrutiny, and it's worth being precise about why. It is not that decentralization makes GPUs faster or cheaper to build. It's that a network of independent operators, each pricing capacity against their own marginal cost, produces a continuous stream of executed prices where today there is only a vacuum of estimates. That is a modest claim, and it happens to be the input the entire buildout is missing.

Reading it wrong in both directions

The bears get one thing right and one thing wrong. They're right that infinite demand is not the same as profitable demand, and that capacity built for a forecast rather than a contract is a genuine risk. They're wrong when they treat the depreciation schedule as evidence about capability — as if an asset crossing an accounting boundary were the same event as it crossing a technical one.

The bulls make the mirror-image error. They're right that the underlying demand for compute has proven durable across several years and showed no sign of reversing. They're wrong when they treat durable demand as a guarantee of durable asset value — as if demand existing somewhere in the system meant every unit of capacity built to serve it would remain economic to run.

The truth is more boring and more useful: the demand is real, the assets are heterogeneous, the two clocks are badly out of phase, and the market that should reconcile them barely exists. That's not a bubble diagnosis or a bubble rebuttal. It's an infrastructure gap, and it's the kind of gap that gets filled by whoever shows up with real prices first — which is a much better thing to argue about than whether the AI trade is over.