The GPU Shortage That Defined a Decade of AI

The shortage of accelerators during the recent AI boom will be studied as a supply chain case for a long time. Demand arrived suddenly, supply was constrained by a handful of manufacturing steps with multi-year lead times, and the imbalance persisted for far longer than most forecasts expected.

Four bottlenecks, one tightest

• Advanced fabrication capacity takes years and enormous capital to bring online.
• High-bandwidth memory is produced by only a few manufacturers and is itself capacity constrained.
• Advanced packaging, the step that joins a processor die to its memory, was the tightest link of all.
• Substrates and power delivery components had their own shortages.

Prepayments, in-house designs, national compute policy

The scarcity reshaped the industry in ways that outlasted it. Companies signed multi-year prepayment agreements to secure allocation, which effectively turned capital into a supply chain asset. Cloud providers built their own accelerator designs. National governments began treating compute access as a strategic resource, comparable to spectrum or energy.

It also created a windfall for anyone holding inventory. Firms that had bought accelerators for other purposes found they were sitting on appreciating assets, and a secondary market emerged with prices well above list.

Quarterly buying turned into multi-year planning

Even as supply has improved, the experience changed expectations permanently. Procurement teams that once bought hardware on quarterly cycles now plan years ahead. The idea that compute is a commodity has been abandoned at every large organisation in the field.

An allocation was a relationship, not a purchase

During the tightest period, buying an accelerator was not a transaction but a relationship. Allocations were negotiated directly with the manufacturer, often with prepayment, multi-year commitments and an implicit understanding about future orders.

That structure favoured organisations with scale, capital and strategic importance to the supplier. Research groups and smaller companies were priced out entirely, which had a measurable effect on who could conduct frontier work and therefore on what got studied.

Renting time beat buying the scarce asset

A grey market emerged in which unneeded capacity was resold at multiples of list price. It was inefficient, opaque and, for many buyers, the only route to obtaining hardware at all.

Cloud rental became the practical alternative: instead of buying a scarce asset, an organisation rents time on someone else’s. That shift moved spending from capital budgets to operating budgets and made large-scale experiments possible for teams that could never have purchased the hardware.

The most lasting change may be internal. Organisations that went through the shortage now treat compute access as a strategic planning assumption rather than a procurement detail, and several have hired people whose job is to forecast and secure capacity. That is a new function in most companies, and it is not going away.