The Capex Cycle Behind AI: Depreciation as the Hidden Variable

Capital spending is reported as a single enormous number and consumed as a small one: a cost per accelerator-hour that ends up inside a price per million tokens. The conversion between them contains an accounting judgement that moves the answer by a factor of two, and almost nobody discussing the topic states which value they assumed.

Capex becomes a cost per hour

Buying a machine is not a cost in the period you buy it; it is an asset that turns into cost as it is used. For serving, the useful unit is one accelerator running for one hour:

H = C / (L * 8760 * A)  +  E  +  F

  C   capital cost attributable to one accelerator —
      the accelerator, its share of the host, networking,
      and the datacentre fit-out it needs
  L   assumed useful life, in years
  A   availability: the fraction of wall-clock hours the
      machine is actually able to serve
  E   energy, for the machine and for the cooling it needs
  F   facility, network and staffing per hour

8760 = hours in a year

Two things about that formula are worth pausing on. The first is that C is not the price of a chip. A serving machine needs power delivery, cooling, high-bandwidth interconnect to its neighbours, and floor space in a building that had to be built; the fully attributed figure is materially larger than the component, and comparisons that use the component price understate H.

The second is A, availability, which is distinct from utilisation. Availability is whether the machine can serve; utilisation is whether anyone bought the hour. Both divide into the same denominator eventually, and confusing them is how a cost estimate ends up optimistic twice.

Useful life is a judgement, not a fact

Now hold everything constant except L. All inputs are assumptions:

C = $40,000 attributable capital
A = 0.95    availability
E + F = $0.60 per hour

L = 3 years : 40000 / (3 * 8760 * 0.95) = $1.60  + 0.60 = $2.20
L = 4 years : 40000 / (4 * 8760 * 0.95) = $1.20  + 0.60 = $1.80
L = 6 years : 40000 / (6 * 8760 * 0.95) = $0.80  + 0.60 = $1.40

Same machine, same electricity, same building.
The cost of an hour differs by 57% between the
first row and the last.

Nothing physical changed between those rows. An accounting assumption about how long the asset remains useful moved the cost of goods, and therefore the reported gross margin, and therefore how the business looks. This is why the depreciation schedule is the single most informative line in a disclosure about this industry, and why a change to it — in either direction — is a substantive event rather than a technicality.

Lengthening assumed life is not automatically an accounting trick; hardware genuinely does keep working, and a machine displaced from frontier training can serve smaller models for years. Shortening it is not automatically conservatism either. The honest position is that the number is a forecast about a fast-moving hardware market, and forecasts are worth stating explicitly rather than embedding silently in a margin.

Economic life ends before physical life

The reason L is genuinely hard is that hardware stops being worth running long before it stops running. Suppose a new generation delivers twice the tokens per hour at the same energy and facility cost. The old machine still works perfectly. But its owner now compares:

Old machine, per million tokens served:
  (E + F) / T_old

New machine, per million tokens served:
  (E + F) / (2 * T_old)  +  new capital per hour / (2 * T_old)

The old machine is retired the moment its *cash* cost
per token exceeds the *total* cost per token of the new
one — capital already spent does not enter the decision.

That is a general result and it is counter-intuitive: sunk capital is irrelevant to whether to keep serving from old hardware. What keeps old machines in service is that their energy-and-facility cost per token is still competitive; what retires them is a generation efficient enough that its full cost undercuts the old machine’s running cost alone.

Between those states, displaced hardware does not vanish. It migrates to work that values it: smaller models, batch workloads, latency-tolerant traffic, cheaper regions. That migration is one of the quieter forces pushing down the price of capability that is no longer at the frontier.

Why fixed cost produces price wars

The most consequential thing about a capex-heavy cost structure is what it does to pricing behaviour in a downturn, and it is entirely predictable from the formula.

Once the machine exists, the capital term in H is sunk. The marginal cost of serving one more request on an idle machine is close to the energy it burns doing so. So any price above that marginal cost contributes something toward a fixed cost that must be paid regardless. A rational operator with idle capacity therefore sells it at prices far below full cost — not out of aggression, but because the alternative is earning nothing on that hour.

• This is why discount tiers exist at all. Batch, off-peak and spot pricing are ways to sell trough capacity above marginal cost without lowering the price paid by customers who need the peak.
• It is also why prices can fall faster than costs. A period of overcapacity produces prices that no operator can sustain indefinitely and that all of them are individually right to offer. Do not read such a price as evidence about long-run cost.
• And why capacity arrives in lumps. Buildings and power connections have multi-year lead times, so supply responds to demand slowly and then all at once. Industries with that shape have cycles; that is what a cycle is.

For a buyer, the practical reading is that today’s cheapest price is information about today’s capacity balance, not about what a sustainable price looks like. Build the budget on a price you believe covers full cost, and treat anything below it as temporary upside.

What to watch, and where it is disclosed

No figures are given here for spending totals, because any such number is stale within a quarter and is not derivable from first principles. These are the primary sources where the inputs live, and each one should be recorded with its date:

One habit makes all of this easier to use: keep the four inputs in one place with dates attached, and recompute H whenever any of them changes rather than carrying an old figure forward. The cost of an accelerator-hour is the number that connects a headline capital figure to a price you actually pay, and it is the only one of them that is stable enough to reason with for more than a quarter.

Everything in this cluster that touches capital intensity — whether serving is profitable, whether the investment is proportionate — comes back to these four inputs, and every argument about them is really an argument about L.