Compute

When people in AI talk about “compute”, they mean the sheer processing power a task demands: the specialised chips it runs on, and the hours or weeks of running them. Training a large model is one of the most compute-hungry things humans currently do, which is why the biggest models cost so much and take so long to build. Every answer a deployed model gives — its inference — uses compute too, just on a smaller scale.

How compute is measured

At the lowest level, compute is counted in operations — specifically floating-point operations, or FLOPs — because training and running a neural network is, underneath, an enormous amount of arithmetic. In practice people talk in more tangible units: how many chips, running for how long. A single frontier model can take tens of thousands of specialised chips running for weeks or months, which is why training runs are now measured in millions of dollars and megawatts of electricity.

Training versus running a model

It helps to separate two very different demands. Training is the one-off, colossal effort of building a model in the first place — the headline compute cost. Inference is what happens every time someone then uses it: smaller per request, but multiplied across millions of users it can end up dwarfing the training bill. That split is why so much recent effort goes into making inference cheaper, and why “the cost of running AI” has become a theme in its own right.

The chips, and the British angle

The workhorse has been the GPU, a processor originally designed for video-game graphics that turned out to be well suited to the maths behind neural networks. Nvidia dominates that market, which is part of why its chips are expensive and perpetually in short supply. That scarcity has created an opening, and a wave of companies — several of them British, such as Graphcore, Lumai, Fractile and OLIX — are designing new kinds of AI accelerator to do the job faster or more efficiently, while firms like Nscale build the data centres to house them.

Why compute became a national issue

Access to compute is no longer just an engineering detail; it is a matter of policy. The UK’s push for “sovereign” AI capacity — government-backed data centres and guaranteed access to chips — is really about one thing: making sure the country has enough compute to train and run AI without depending entirely on others. In an era where capability tracks closely with compute, who controls the chips increasingly shapes who can build frontier AI at all.