Tom Blomfield founded two of the UK's best-known fintechs, GoCardless and Monzo. Now he joins Anthropic with a role as atypical as it is telling: working on the compute team, meaning the processing power needed to train and run models like Claude. This is no random move.

Blomfield is neither a machine learning researcher nor a silicon designer. He comes from payments and financial infrastructure, fields where scaling complex systems reliably is a survival skill. Anthropic tapping him to tackle AI’s hardest problem — compute — says a lot about how the industry is evolving.

We are used to thinking of Large Language Models as a game played on algorithms and datasets. But without compute infrastructure capable of handling increasingly massive workloads, even the most brilliant model remains a lab experiment. GPUs, video memory, node interconnects, and energy efficiency have become as critical as pure research. When someone like Blomfield, who built platforms processing millions of transactions smoothly, is put in charge of this challenge, the signal is unmistakable: compute is no longer just a running cost — it is the strategic battleground.

What does this have to do with evaluating on-premise deployment? More than it might appear. Anthropic, like almost all frontier labs, runs on cloud infrastructure, but the problems it faces are the same as those confronting a company that decides to bring inference or fine-tuning in-house: GPU saturation, latency, management costs, and, not least, data sovereignty. Relying on the cloud may simplify operations but introduces dependencies and variables beyond control, such as availability of specific cards or fluctuating instance pricing.

Bringing models on-premise means taking back control of the whole stack, but it forces you to deal with the same resource scarcity that plagues the major labs. Anthropic’s move is a reminder that compute optimisation — from quantization to the choice of serving frameworks, down to VRAM configuration — is not a technical detail for sysadmins, but the prerequisite for extracting real value from LLMs without blowing up budgets.

Blomfield’s arrival also signals a shift in skills: expertise matured in managing high-reliability transactional systems is migrating toward AI, because the constraints are similar. It is no surprise, then, that even outside Silicon Valley, AI infrastructure roles are gaining weight. The lesson for those following self-hosted approaches is that the AI battle is not won with the biggest model alone, but with the most solid and efficient architecture.