AI COMPUTE · MARKET STRUCTURE

Has AI computing become like Coca-Cola and Pepsi with a meal?

Dr Akanda Ashraf · 2024 · all articles · back to profile

Ordering a meal usually comes with a narrow drinks question: Coca-Cola or Pepsi? The choice feels free, but the shape of the market decided it long before the menu was printed. Accelerated computing for AI has arrived at a similar place. Whatever the workload — training a large model, running perception at the edge, serving inference to millions — the practical shortlist of vendors is short, and the defaults are strong.

What makes the analogy useful is not the number of suppliers but the role software plays. The moat is the toolchain: kernels, compilers, profilers, container images and the accumulated habits of every engineer who has ever debugged a training run. Hardware parity does not immediately buy adoption, because switching cost is measured in engineering months, not in dollars per FLOP.

For teams building on top, three consequences follow. First, portability is a design decision that has to be taken early — abstractions over the accelerator layer are cheap to add at the start and expensive to retrofit. Second, cost modelling should assume the price of compute is set by supply, not by your negotiating position. Third, model architecture and deployment target should be chosen together; a design that only performs on one vendor's stack quietly hands away future optionality.

None of this is an argument against the incumbents. Their tooling is genuinely good, and shipping matters more than purity. But it is worth knowing when a choice is being made for you, and pricing that into the plan.