Report: Revolut moves 2,000 engineers to self-hosted GLM, cutting AI coding costs to a tenth

What the report claims

A post on X by Guillermo Flor, which reached the Hacker News front page, says Revolut has moved roughly 2,000 of its engineers onto a self-hosted installation of GLM for day-to-day coding work, with Anthropic's Claude Code and OpenAI's Codex kept around only as exceptions for cases the in-house model can't handle. Flor credited Oliver Molander for surfacing the item.

The post attributes its figures to David Abitbol, described as the person running AI operations at Revolut. The claimed numbers: the deployment runs on 56 Nvidia B200 GPUs, blends out to about 30 cents per million tokens, and costs roughly one-tenth of what the company was previously paying Anthropic and OpenAI. It also claims engineering lead time improved by 70 percent.

Worth stating plainly: this is a single social media report relayed through Hacker News. No detailed write-up from Revolut or Abitbol accompanies it in the available material, so the figures should be read as reported claims rather than audited facts.

The detail that does the work

The most consequential part of the claim is not the price per token but the behavior change around it. According to the post, the productivity gains only appeared once engineers stopped treating AI usage as something to spend carefully. When tokens were billed per unit by external vendors, engineers saved the tools for the hardest tickets. Once the marginal cost of a token fell toward the cost of running hardware the company already owned, they pointed the assistant at everything.

That framing matters for anyone evaluating coding assistants. It suggests per-token API pricing acts as a hidden throttle on adoption: teams under-use a helpful tool because every request carries a visible price. Owning the infrastructure turns that variable cost into a largely sunk one, and usage expands to fill it. Under that reading, a meaningful share of the claimed lead-time improvement could come from volume of use rather than from GLM being stronger than the frontier models it replaced.

Why it matters

If the report is accurate, it sketches a template other large engineering organizations could copy: run an open-weight model, GLM being Zhipu AI's openly released family with a strong coding focus, on your own GPUs for the bulk of daily work, and keep frontier APIs as an escalation path for rare hard cases. The frontier labs' product then becomes a premium tier rather than default infrastructure, which is exactly the pricing problem the original post flags as those companies head toward public listings.

The unit economics are the heart of the story. Fifty-six B200 GPUs serving 2,000 engineers is a modest footprint by datacenter standards, and owned hardware converts a per-token bill into amortized capital plus power. The logic cuts both ways, though: self-hosting pays off only at sustained high utilization, and organizations without Revolut's scale or platform engineering muscle will not land the same blended rate.

Two caveats are worth holding onto. First, everything here traces back to one social media account; a figure like a blended 30 cents per million tokens depends heavily on how utilization and capital costs are accounted for. Second, coding assistant quality varies by task and model version, so good enough for everything is a claim each organization would need to test against its own workload. Even with those caveats, as a signal of where enterprise AI spending is heading, open weights by default and frontier models by exception, this is one of the clearer data points to surface publicly.