Revolut Cuts Coding AI Costs 90% In-House
The switch to owned open-weight hardware shows high-volume enterprise AI economics may increasingly favor self-hosting over metered APIs.
Updated on Nov. 3, 2026 in Artificial Intelligence
Live Poll
Do you expect AI tools in your everyday life to get better or worse over the next year?
Revolut moved its internal coding assistant to a self-hosted deployment of Zhipu AI's open-weight GLM-5.3 model, reporting a 90% cut in the cost of running its coding AI tools. The deployment is in production use internally, not a publicly sold product, and was presented by infrastructure lead David Abitbol at GTC Berlin on October 22.
Why it matters
Revolut ran the numbers on 2,000 engineers hitting metered frontier APIs daily and concluded that owning GPUs was cheaper, a shift from per-token API spending toward self-hosted open-weight models for high-volume internal workloads. The move suggests that for steady, high-frequency coding use, the economics of hosted APIs are being reconsidered at scale.
The figures come from Revolut's own internal deployment and conference presentation rather than an independent benchmark, covering a 90% cost reduction, latency falling from over 10 seconds to 1.6 seconds, and throughput rising from about 3 to more than 7 requests per second on 56 Nvidia B200 GPUs. The source does not explain how the roughly 60% coding-time improvement across 2,000 engineers was measured.
The players
Revolut
A London-based fintech whose 2025 revenue reached $6 billion, up 46% year over year, and which reports 75 million customers.
Zhipu AI
A Chinese AI developer whose open-weight GLM-5.3 model powers Revolut's self-hosted coding assistant.
David Abitbol
Revolut's infrastructure lead, who presented the self-hosted coding assistant switch at GTC Berlin.
The details
Rather than paying a metered fee for every request to a frontier model's API, Revolut hosts an open-weight model, meaning one whose trained weights are publicly released, on its own hardware. The self-hosted deployment of Zhipu AI's GLM-5.3 model runs across 56 Nvidia B200 GPUs, the company's current high-end data-center accelerators, and the company says serving requests on owned hardware drives per-token inference cost down to roughly 30 cents per million tokens. The self-hosted path handles text and code generation only, while image-related tasks are still sent to Gemini or similar hosted models.
Timeline
April 2026: Revolut rolled out the self-hosted coding assistant to engineers.
October 22, 2026: Abitbol presented the switch at GTC Berlin.
The Tech Race
Revolut's move pits owned open-weight infrastructure against metered frontier APIs, a choice now being made by high-volume enterprise users rather than only by model developers. The switch illustrates how open-weight releases from developers such as Zhipu AI are giving large companies a credible alternative to hosted services from frontier providers.
Revolut's coding assistant is an internal tool, so the change reaches engineers at the company rather than outside customers, who will not see a new product from this switch. The self-hosted path currently covers only text and code generation, and the source gives no timeline or pricing for any wider availability.
The takeaway
For high-volume AI workloads, the economics may favor owning open-weight infrastructure over metered APIs, a trade-off other large enterprises will likely weigh. Watch whether Revolut publishes independent validation of the 60% coding-time figure and whether it extends the self-hosted path to image tasks now routed to Gemini.
Further reading
For broader context on how companies weigh hosted and self-hosted models, see the Artificial Intelligence section.
Source note: This article includes information reported by Startup Fortune.
Live Poll
Do you expect AI tools in your everyday life to get better or worse over the next year?
