Revolut Says Self-Hosted Open AI Models Cut Its Coding Costs 90%

Revolut spent heavily on OpenAI and Anthropic’s APIs to help its engineers code, then switched 2,000 of them to an open model it runs on its own GPUs. The company says the switch paid for itself fast.

David Abitbol, Revolut’s infrastructure lead, told an audience at GTC Berlin on October 22 how the fintech changed its setup. The internal coding assistant moved off metered frontier APIs and onto a self-hosted deployment of Zhipu AI’s open-weight GLM-5.3 model, running on 56 Nvidia B200 GPUs. The result, he said, was a 90% drop in the cost of running AI coding tools across the company. Inference is now priced at roughly 30 cents per million tokens.

That’s a striking number. The economics of the AI boom are under constant scrutiny, from Oracle’s debt-funded data centre buildout to Amazon reportedly cancelling cloud GPU orders. Revolut’s bet runs the opposite way: instead of treating frontier API spend as an unavoidable cost of doing AI at scale, it built the infrastructure to avoid it entirely.

The hardware isn’t cheap. A fully configured eight-GPU B200 server lists at around $390,700 from at least one systems integrator, according to pricing reviewed by FourWeekMBA. That means 56 GPUs represent something like $2.7 million in hardware before power, networking and the engineering time to run it. Revolut’s argument is that the maths still works once you’re running thousands of engineers through the system every day rather than paying per token to OpenAI or Anthropic.

Cost wasn’t the only number Abitbol cited. Time to first token on coding requests fell from more than 10 seconds to 1.6 seconds. Over the same period the system’s throughput rose from about 3 requests per second to more than 7.

The presentation was a GTC Berlin expo talk titled “How Revolut runs production agents on open models.” Abitbol framed it as a direct productivity story: a sample of the 2,000 engineers using the self-hosted assistant saw roughly a 60% improvement in coding time and a 70% cut in lead time after the April rollout.

Revolut isn’t claiming open models beat frontier ones at everything. Abitbol said image-related tasks still get routed to Gemini or similar hosted models, so the self-hosted path covers text and code generation only. That’s a meaningful caveat. It suggests the company’s actual position is narrower and more pragmatic than “open-weight models have caught up across the board.” The claim is smaller: for a specific, high-volume, internally-facing workload, a model like GLM-5.3 running on owned hardware got good enough, fast enough, and cheap enough to beat paying OpenAI or Anthropic by the token.

It is a real data point in a debate that’s mostly been theoretical. Plenty of AI infrastructure vendors have argued all year that open-weight models from labs like Zhipu AI, Alibaba’s Qwen team, and DeepSeek are closing the gap with closed frontier models. Their case is that the gap is shrinking fast enough to make self-hosting viable for serious production workloads. Revolut is one of the first companies of real scale to put a specific number on what that actually saves, rather than asserting it in the abstract.

Why Revolut, specifically, had the incentive to do this

Revolut isn’t a startup experimenting at the margins. The company reported $6 billion in 2025 revenue, up 46% year over year. Its valuation hit $115 billion after a secondary share sale earlier this year, up from $75 billion in late 2025. It’s pursuing an eventual IPO that bankers reportedly believe could value it as high as $200 billion. There are more than 75 million customers and a stated goal of reaching 100 million by mid-2027. Revolut runs engineering at a scale where even small per-engineer AI costs compound into real money fast.

Running 56 B200 GPUs and the staff to operate them only pencils out at a certain size, and Revolut has the volume.

Frankly, this is the kind of concrete infrastructure decision that’s more informative than another survey about enterprise AI adoption. Revolut didn’t self-host GLM-5.3 to make an ideological point about open-weight models or data sovereignty. It did it because the finance team ran the numbers on 2,000 engineers hitting metered APIs every day and decided owning the GPUs was cheaper.

Other large engineering organisations are watching their own AI coding tool bills climb. They are likely to run the same maths soon, especially if open-weight models keep closing the gap with GPT-5 and Claude on real coding benchmarks rather than just marketing slides.

Also read: Morgan Stanley CEO Says AI Buildout Could Hit $15 Trillion and We’re Barely Started • Apple guides margins lower again as memory chip costs keep climbing • Palantir posted its best growth quarter ever and the stock still dropped