The buyers started making their own

Every large cloud operator is now shipping an accelerator it designed itself. That is not a claim about beating the incumbent on raw performance, and treating it as one misses what the programme is for.

Coverage of the shift describes custom designs moving from internal experiments to production volume, with each of the major operators running its own part alongside purchased hardware.

The advantage is not the die, it is the stack. An operator that controls the silicon, the compiler, the framework and the workload can specialise in ways a general-purpose part cannot, because it knows exactly which operations matter.

That specialisation pays at scale and nowhere else. A meaningful efficiency gain on inference silicon is worth the cost of a design programme only if you are running enormous volumes of the same class of model, which describes a very short list of companies.

Inference is where the shift concentrates for a straightforward reason. Training runs are heterogeneous and change with each generation; inference is repetitive, predictable and enormous, which is the profile a fixed-function design handles best.

The commercial effect is felt in the negotiation rather than the benchmark. A buyer with a credible internal alternative is a different counterparty, even when it continues buying most of its capacity externally.

Software remains the moat that keeps the incumbent's position durable. A custom part is only useful if the frameworks target it, and the operators have had to invest heavily in compilers to make their own silicon usable by their own customers.

For customers renting the capacity, the practical question is portability. Committing a workload to an accelerator that exists inside one provider is a lock-in decision, and the pricing that makes it attractive is the reason it is being offered.

The most likely outcome is not displacement. It is a market where the general-purpose part carries training and the frontier, and specialised silicon carries the steady, high-volume inference underneath it.

That is a normal maturation for a computing market and an uncomfortable one for a supplier whose margins were set when there was no alternative.