The AI chip gold rush¶

The scramble for AI accelerators resembles the historical gold rushes in every particular except that instead of prospectors with pickaxes there are data centres with procurement budgets, and instead of gold they are mining computational capability measured in FLOPs per watt. Like all gold rushes, it involves tremendous excitement, questionable economics, and the nagging suspicion that the people selling shovels are making more reliable profits than those digging.

The question occupying economists, investors, and anyone forced to explain their hardware budget is whether this is genuine demand for computation to solve real problems or speculative positioning on the assumption that whoever controls the most AI hardware controls the future. The answer appears to be yes, which is unhelpful for decision-making but accurately reflects the state of things.

Real demand, with speculation on top¶

Some of the demand is unquestionably real. Training frontier models consumes tens of thousands of accelerators for months, and serving millions of users consumes more; these are genuine computational requirements that cannot be wished smaller. On top of that sits the speculative layer: infrastructure announced before the applications that would justify it are identified, on the logic that AI is important, therefore vast AI infrastructure is needed, therefore GPUs must be acquired urgently. Cloud providers build ahead of customer demand that may or may not arrive at predicted scale, which makes them speculators with very expensive chips, and startups acquire clusters before acquiring customers, which investors are invited to read as strategic positioning. Whether scarcity is entirely a production fact or partly a pricing instrument depends on whether you ask the supplier’s CFO or the buyer’s, and the merchants of Ankh-Morpork have known both kinds of shortage.

The economics of expensive shovels¶

The capital costs are staggering: clusters run to hundreds of millions of euros before networking, cooling, power and staff, which concentrates frontier AI capability among large tech companies, well-funded startups, and wealthy states. The depreciation schedule is brutal, since each hardware generation outclasses the last within a couple of years, and utilisation is everything: an idle accelerator is capital quietly evaporating. The economics work only if the resulting capability earns its keep through revenue, cost savings, or strategic value, and for a visible share of buyers the honest justification is that rivals are buying too.

Bubble indicators¶

Several classic markers are present: prices detached from fundamentals, fear of missing out as an investment thesis, and the assumption that current growth continues indefinitely, which is the traditional belief that trees grow to the sky. The awkward complication is that transformative platforms often look exactly like bubbles during their growth phase; the internet bubble was both a genuine revolution and a graveyard of spectacular failures, and telling the two apart in real time was hard then and is hard now. The practical reading is that current demand contains both genuine need and speculative excess, in proportions that will be known only afterwards. The rush will continue until it does not. The shovel sellers will have banked their profits either way, which is how gold rushes have always worked.

The clerk’s brief¶

February 2026: Half the money, a fraction of the chips¶

Deloitte’s 2026 outlook, published February 2026, projects chip sales of 975 billion dollars for the year, with AI semiconductors around 500 billion dollars: more than half the industry’s revenue from under 0.2 per cent of its unit volume. The clerks invite his Lordship to admire the shape of that ratio, in which nearly all the money rides on almost none of the objects, and to recall where similar ratios have been observed before, mostly in the vicinity of tulips.

December 2025: The rush reaches the corner shop¶

Reporting by CNBC in December 2025 described Nvidia’s pivot to LPDDR memory as a seismic shift, making the AI supplier a memory customer on the scale of a major smartphone maker, with memory prices expected to rise 30 per cent in the quarter and further in early 2026, and consumer brands warning of sizeable retail price rises. The gold rush has begun to set the price of bread, or at least of telephones. The clerks note that when a boom starts taxing bystanders, the politics of the boom change.

October 2025: The money goes round in a circle¶

The American Prospect’s analysis of AI vendor financing, published October 2025, traced the loop by which the chip monopolist invests in AI startups whose main cost is computing, purchased from clouds running the monopolist’s chips, returning the money as revenue that would not otherwise exist, and counted commitments including 850 billion dollars of data centre plans by a single model maker that does not have the money to pay for them. The clerks do not use the word ouroboros lightly, and note only that a snake eating its own tail reports excellent throughput.

The claims register¶

The register’s standing question, real demand or speculation, has not been answered; it has been quantified. As of early 2026 the money is real and enormous, the concentration is measured, and growth is decelerating from the vertical: industry analysis in September 2025 already put infrastructure growth at roughly a quarter of its 2022 to 2024 pace, with spending expected to peak as a share of data centre investment. The circular financing entered the file in October 2025 and has not left it. The clerks’ standing assessment: the claims are being staked faster than the assays are coming back, the shovel counter remains the only till that always rings, and the register exists so that, afterwards, nobody can say the pattern went unrecorded.