Concentration, Supply and the Compute Market

The market for AI compute has a structure unlike most technology inputs: a small number of suppliers at several layers, physical constraints that do not respond quickly to demand, and lead times measured in quarters. For an organisation buying rather than building, the practical consequence is that availability is frequently a harder constraint than price.

The layers and their constraints

Fabrication. Leading-edge logic is produced by a very small number of firms, and advanced packaging, which bonds high-bandwidth memory to the logic die, has been a specific bottleneck independent of wafer supply. Capacity here is added on multi-year timescales.

High-bandwidth memory. Supplied by a handful of manufacturers, and it is often the component that gates accelerator output rather than the processor itself, since bandwidth rather than compute is what the workload needs.

Accelerator design. Concentrated, with a dominant supplier in datacentre AI training and a software ecosystem that raises switching costs well above the hardware comparison. Alternatives exist and each requires porting effort proportionate to how much of the stack is written against vendor-specific tooling.

Datacentre capacity and power. The layer that has become binding most recently. Building shells is fast relative to securing grid interconnection, and in several markets the interconnection queue is years long. Power availability, not silicon, determines where and when capacity appears.

Cloud providers, who buy at the layers above and resell in units organisations can consume, with the pricing and commitment structures that follow from their own capacity risk.

What it means for a buyer

Secure capacity before optimising its price. For a scarce part, a commitment buys the ability to run at all, and a plan that assumes on-demand availability of the current generation is a plan with an unpriced risk.

Portability has option value. Writing against a portable layer rather than vendor-specific primitives costs something now and preserves the ability to move if availability or price shifts. Whether it is worth it depends on how much of the stack is affected and how much performance the abstraction costs.

Lead times drive planning horizons. Quarters for committed cloud capacity and longer for anything on-premises means capacity decisions precede the workloads that use them, which is why the exploratory demand class is the hardest to plan for.

Geography matters more than it used to. Power price and availability, latency requirements and data residency rules jointly constrain where compute can be, and the cheapest region is frequently not an option.

When it breaks

Scarcity is not permanent and is not uniform. Supply constraints ease as capacity is added and tighten as demand shifts, and the specific part that is scarce changes. Planning built on the current shortage generalises poorly.

Concentration is a systemic risk, not just a pricing one. A dependency on one supplier at any layer is an availability and continuity risk, and it is worth naming in a risk register rather than treated as a commercial detail.

Vertical integration changes the comparison. Providers designing their own accelerators for their own workloads shift the economics for their customers in ways that are hard to compare against a merchant part, since the pricing reflects internal cost and strategy rather than a market rate.

Efficiency does not reduce aggregate demand. Cheaper compute has historically expanded what is built rather than reducing what is bought, so improvements at any layer have not relieved the constraint at the others. Planning that assumes efficiency will resolve scarcity has been consistently wrong.