What a large-scale AI investment has to carry commercially
A decision to invest in computing capacity is not a technology decision but a calculation with an unusually short useful life. An account of the figures that have to come together.
Investment in artificial intelligence is usually discussed in the language of technology: model sizes, training runs, performance benchmarks. For the question of whether an investment carries, that language is unsuitable. It describes what a system can do, not what it may be allowed to cost.
The calculation begins with useful life
Conventional industrial investments are written down over decades. A production hall stands for thirty years, a machine tool for fifteen. Computing capacity for AI training behaves differently: the hardware loses its advantage not through wear but through successor generations, and those appear at intervals considerably shorter than any customary depreciation period.
That shifts the entire calculation. An asset that must be amortised over three or four years rather than fifteen requires a return in that period several times larger than what suffices for long-lived capital goods. The question is not whether the technology works. The question is whether the return arrives quickly enough.
Utilisation is the second unknown
An asset carries itself only at high utilisation. With computing capacity, however, utilisation is not even: training runs come in bursts, inference less so, and the two load profiles place different demands on the same infrastructure.
Dimension for peak load and you pay, in normal operation, for capacity that stands idle. Dimension for the average and you cannot serve the peaks. Both are expensive, and the decision is taken before construction, at a point when actual demand is not yet known.
The price at which you can sell is not fixed
The third figure is the most volatile. Providers of computing power, and of the services built on it, operate in a market where capacity is being added briskly and where several participants are seeking share at the same time. In such conditions the achievable price is not a constant but the outcome of a competition whose course cannot be settled at the time of investment.
An investment calculation that works from today's prices and assumes a short payback therefore makes two assumptions that compound one another: it needs fast returns in a market whose prices may come under pressure.
What follows
None of this argues against the use of AI. It argues against treating the investment decision as a technology decision. Useful life, utilisation and achievable price are commercial figures, and they determine whether a venture carries — regardless of how well the system ultimately works.
For companies that do not operate infrastructure themselves, one thing follows above all: the interesting question is rarely how large a model is, but which specific process becomes cheaper or better than it is today — and whether that improvement can be quantified.