THE STORY AT A GLANCE

What you need to know

  • Revolut reports an AI coding bill about 90% lower after making self-hosted GLM the default.
  • Commercial models remain available; the story concerns internal engineering tools.
  • Productivity figures are observational and do not isolate the model change.
  • Evaluate accepted work, review effort and full service costs before copying the approach.

Self-hosted models become the default

Revolut has made self-hosted AI the default for roughly 2,000 software engineers, according to Dealroom’s account of a presentation by David Abitbol, who leads AI operations at the digital bank. The deployment uses GLM, an open-weight model from Z.ai. Revolut says its AI coding bill fell to roughly one-tenth of its previous spending with Anthropic and OpenAI.

The change leaves Claude Code and Codex available for exceptional work. Image-related tasks can take a separate route to a vision-capable model. This is a change in the default model strategy for internal engineering, rather than evidence that Revolut has abandoned commercial AI across its business.

The 90% figure needs a clear boundary

In the conference account, Abitbol describes around 56 Nvidia B200 GPUs serving the engineers and a blended price of about $0.30 per million tokens. He reports a 60% improvement in coding time and 70% in lead time. The accompanying slide instead labels the lead-time improvement as 60%, leaving that comparison unresolved.

These are self-reported observations. The presentation does not supply matched billing periods or a complete accounting of engineering costs. It also does not isolate the effect of changing models from the effect of broader access to AI. The reported bill reduction therefore cannot establish a 90% saving in total software-development costs.

Cheaper access can change the work engineers delegate

Dealroom reports that Revolut initially imposed strict individual budgets on commercial coding tools. Abitbol’s explanation is that engineers then reserved tokens for difficult coding problems. More predictable spending allowed wider use, including searching inboxes and working through Jira and Confluence.

Our reading is that this makes the rollout an experiment in access as well as model choice. A budget can influence which tasks people attempt to automate. Removing it changes the workload, so a before-and-after comparison may measure several changes together.

The infrastructure becomes part of the product

Abitbol describes a target of less than three seconds before the first token appears, alongside tuning to keep the service responsive. His production discussion also includes sandbox restrictions, human oversight and logs for investigating incidents. Running the weights internally still requires an operated service around them.

For another engineering team, our proposed assessment would put model quality and service reliability beside spending. A low token price has limited value if developers wait longer, repeat failed tasks or spend more time correcting the result. A team should be able to see those costs in the same review.

Compare the cost of accepted work

A useful pilot would begin with a representative set of real engineering tasks and a written definition of acceptable completion. Keep the review standard consistent across deployments. Record completion time, reviewer effort, defects, retries and any use of a fallback model, alongside infrastructure and staff costs.

Capacity deserves its own sensitivity check. A shared fleet may look economical when busy and expensive when demand falls. Include peak-hour responsiveness and the effort required for upgrades, support and incident recovery before treating the invoice as the full operating cost.

Revolut’s account offers a reason to test a different allocation of AI spending. The decision for another company should rest on its own accepted work, measured under realistic demand. A deployment that delivers useful results predictably has a stronger business case than one selected solely for a headline percentage.

FOLLOW THE EVIDENCE

Sources & further reading

This article draws on the source material and official product documentation below. Product capabilities can change; check each source for current details.

VT
Ventures Time Editorial

The Ventures Time editorial desk publishes original explainers and analysis on enterprise AI, data, security, and the business of technology. Articles distinguish practical editorial analysis from source material and link to primary documentation where relevant.

Our editorial standards