What did Revolut change in its AI coding stack?
Revolut, the British fintech, has made self-hosted open-weight models the default for coding work by roughly 2,000 of its engineers. The main model is GLM-5.3 from Chinese developer Z.ai, served on 56 Nvidia B200 GPUs. The Register reported the details on November 3, 2026.
The move reverses the usual enterprise pattern. Many large companies rent frontier models from Anthropic, OpenAI or Google through per-token APIs. Revolut still uses those vendors for some tasks, but code generation now runs on hardware it controls.
David Abitbol, who leads AI operations and infrastructure at Revolut, described the migration at the AI Engineer conference in New York. The organisers published his talk on YouTube on October 29, 2026.
The path was gradual. Revolut deployed Gemini in early 2025, adopted Anthropic's Claude Code in early 2026, then moved to GLM-5.1 and later GLM-5.3. Each step shifted more everyday work away from metered APIs.
It is not a clean swap. A router sends image and interface queries to Gemini or GPT Luna, and keeps code generation on GLM. Abitbol described frontier models as the exception in the new setup, not the rule.
That hybrid design matters. Revolut did not conclude that open weights beat closed models at everything. It concluded that most coding work does not need the frontier, and that the frontier is too expensive to use by default.
How much did Revolut save on AI coding?
Revolut says its AI coding inference costs fell by about 90% in less than a year. The figure comes from Abitbol's talk and was reported by The Register.
The internal number is about $0.30 per million input and output tokens. The Register ties that figure to GLM-5.1, the version Revolut deployed before GLM-5.3. It is roughly one-tenth of what the company paid when engineers used OpenAI's Codex or Anthropic's Claude Code.
The starting point was expensive. Abitbol said engineers were spending $1,000 a month, and in some cases $10,000 a month, with little return to show for it. "We were highly skeptical about the cost of AI," he said.
For comparison, Z.ai's hosted API lists GLM-5.3 at $1.40 per million input tokens and $4.40 per million output tokens, according to Epoch AI's model profile. Self-hosting at $0.30 sounds impossible against those prices until you look at the assumptions. The $0.30 blends input and output, and it depends on keeping a shared GPU pool busy.
A rough calculation shows the size of the gap. If a workload were 70% input and 30% output, Z.ai's list prices would blend to about $2.30 per million tokens. That is roughly 7.7 times Revolut's figure, before any volume discount a large customer might negotiate. The more output-heavy the work, the wider that gap becomes.
This is an illustrative blend, not Revolut's actual token mix. It does explain why a large, steady workload can justify dedicated hardware. The break-even point depends on utilisation, which is why the 2,000-engineer figure matters as much as the price.
Revolut also reports productivity gains. Coding time improved by about 60%, and lead time improved by about 70%, across the 2,000-engineer rollout. Lead time is the elapsed time from a request to a delivered change, not the number of lines typed.
The gains did not come mainly from writing code faster. Abitbol said the bigger bottleneck was around the code: debugging, Jira tickets and Confluence pages. Cheap tokens made those chores worth automating.
"When you have unlimited tokens, you're free to do that," he said. Engineers no longer had to save capacity for the hardest task of the quarter.
The caveat is that every number here is Revolut's own. No outside auditor reviewed the costs, and the company has not published its measurement method.
Why are 56 Nvidia B200 GPUs enough for 2,000 engineers?
Fifty-six accelerators for 2,000 engineers sounds like heavy oversubscription. The explanation sits in how the model and the workload behave.
Z.ai released GLM-5.3 on August 14, 2026, and the weights are now public on Hugging Face. It is a mixture-of-experts model with 753 billion total parameters, according to its model card, so only part of the network runs per request.
Memory is still the constraint. The weights need roughly 380 gigabytes of GPU memory, and each active user needs another 6 to 26 gigabytes for context, depending on precision. That rules out giving every engineer a dedicated slice.
Instead, Revolut tuned the serving stack. It used NVFP4 quantization, disaggregated prefill and decode, and hyperparameter tuning. It also capped context at 256,000 tokens, even though the model supports one million.
The hardware has room for that setup. An Nvidia DGX B200 node carries eight B200 GPUs and 1,440 gigabytes of memory, according to Nvidia. Seven such nodes provide the 56 GPUs. One node alone can hold the 380-gigabyte weight set with space for context, so a node can host the weights without spreading them across the cluster.
That helps explain the utilisation. The weights fit inside a single node, so Revolut does not need all 56 GPUs just to hold one copy of the model. The extra nodes add throughput for many concurrent users rather than raw model capacity.
"Long context is a scam. You really need to be mindful of that," Abitbol said. In his view, a million-token window is not useful if serving it destroys latency and throughput.
The latency targets reflect how engineers actually work. Revolut aims for under three seconds to first token at a 256,000-token context, and about 200 tokens per second of generation. Engineers working after 5pm see throughput closer to 800 tokens per second.
"As soon as you exceed 5 seconds, my Slack inbox starts to blow up," Abitbol said. The team optimised for responsiveness rather than raw speed.
That pattern also explains the GPU maths. Engineers start parallel agent tasks and check back later, so demand is bursty. A shared pool with good scheduling can serve far more people than a per-seat allocation would suggest.
Together AI presented a GTC Berlin 2026 session titled "How Revolut Runs Production Agents on Open Models: European Infrastructure, at a Fraction of the Cost". The session was scheduled for October 22, 2026. The title emphasises European capacity and lower cost, two themes that match Abitbol's account.
Does cheaper inference mean worse code?
Revolut's own setup suggests the answer is no in most cases, but not everywhere. GLM-5.3 is a text-only model. Abitbol said it is not ideal for interface work, which is why the router sends vision queries elsewhere.
The model card claims strong coding and agentic results. It lists 34.6 on Terminal-Bench 3.0, 88.2 on FrontierSWE and 84.5 on CyberGym, a vulnerability-discovery benchmark. Those are vendor-published figures, not independent tests of Revolut's code.
The productivity story is more interesting than the benchmark story. Revolut's biggest gains came from delegating low-status maintenance work: bug hunts, ticket triage and documentation searches. That work was previously rationed because every token had a price.
Abitbol's framing is that the bottleneck for engineering is not necessarily coding. If that is right, cheaper inference helps most where teams are constrained by unglamorous queue work rather than by typing speed.
Revolut also kept guardrails. Air Street Press reported in June 2026 that the company uses a central model gateway and fallback chains, and that it right-sizes models to tasks. It said that approach has cut costs by up to eight times. The same report said human oversight remains mandatory for decisions that affect customers' lives.
None of that proves GLM-5.3 writes better code than Claude or GPT. It shows that a capable open-weight model plus routing and governance can cover most internal work at a lower price.
Revolut has not published its own model evaluations. The benchmarks on the model card measure public tasks, not the fintech's codebase. They do not show how often GLM-5.3's patches pass internal review, or how many defects reach production. Those numbers would be a stronger test than token price.
The quality question also has a lag. A cheaper model that produces slightly weaker code could still win on cost if review catches the errors early. It loses if reviewers spend more time fixing than they save. Revolut's 60% coding-time improvement suggests the balance is favourable for its internal work, but the company has not shown the defect data.
What does this mean for banks and other regulated firms?
The obvious objection to Chinese open-weight models is data. Abitbol addressed it directly. He said the main privacy and security risk comes from calling APIs hosted in China. Running the same weights on-premises or in a regional cloud avoids that exposure.
That argument depends on control of the serving stack. A firm that self-hosts can inspect the weights, pin a version and keep prompts inside its own network. A firm that calls a foreign API cannot make the same claim.
Security cuts both ways. NIST's Center for AI Standards and Innovation assessed GLM-5.3 on September 17, 2026. It called the model the most cyber-capable open-weight model it had tested, while noting it still trailed the US frontier by about four months.
For a bank, a model that is strong at finding vulnerabilities is useful for defensive work and risky in the wrong hands. That is an argument for tight access control, not necessarily for avoiding the model.
The economics also suit regulated budgets. "The billing is flat, we don't have any bill shock at the end of the month," Abitbol said. Finance teams can forecast a fixed cluster cost more easily than variable API spend.
There is a compliance cost that does not appear in the token price. Running a 56-GPU cluster requires specialised staff, monitoring and capacity planning. Smaller firms may find a managed open-model service cheaper than building that capability.
Documentation is another cost. A bank that runs its own model must show supervisors how the model was tested, which version served each request and how outputs were reviewed. A pinned weight set makes that easier than a vendor API that changes without notice, but it still requires an audit trail.
The trade-off is not unique to AI. Firms already decide whether to run their own databases or buy managed services. The new part is that a language model's weights can be copied, inspected and served on ordinary cloud hardware. That turns a vendor relationship into an infrastructure decision.
What are the limits of Revolut's cost claims?
The first limit is scope. Revolut's figures cover internal coding agents used by employees. AI Chat Daily reported that the deployment does not extend to customer banking systems. That is a much smaller blast radius than the headline might suggest.
The second is accounting. A blended $0.30 per million tokens is an inference cost. It does not include the capital or rental cost of 56 B200 GPUs, the electricity, the networking or the engineers who keep the stack running. The Register's report does not break those out.
The third is like-for-like comparison. Vendor list prices are not what large customers pay, and Revolut's pre-switch spending included behaviour that the new setup changed. Once tokens became cheap, engineers used them for tasks they had previously skipped.
The fourth is measurement. Coding-time and lead-time improvements of 60% and 70% are self-reported. Without a control group or a published definition, they are best treated as directional evidence.
The fifth is durability. Prices for frontier APIs can change, and Z.ai can change its model or licence. Revolut's advantage depends on maintaining a serving team that can re-qualify each new version.
None of these caveats makes the result fake. They do mean the 90% figure should not be copied into another company's business case without its own utilisation data.
Utilisation is the variable that decides the outcome. A cluster that runs at a steady load spreads its fixed cost across many requests. A cluster that sits half idle makes every token more expensive than the API it replaced. Revolut's 2,000 engineers give it a large, predictable base of users, which many smaller teams lack.
What should enterprises watch next?
The immediate test is replication. If other regulated firms publish similar numbers, the case for self-hosted coding models becomes stronger. If they stay quiet, Revolut may remain an outlier with unusual engineering depth.
Watch how Anthropic and OpenAI price enterprise coding. Their response to cheaper open-weight alternatives will shape whether the market splits into a frontier tier and a commodity tier.
Watch Z.ai's release cadence and licensing. Each new GLM version forces customers to re-test quality and safety. A custom licence also adds legal review that standard open-source licences avoid.
Watch regulators. CAISI's assessments of open-weight cyber capability are likely to influence how banks and critical-infrastructure operators justify deployments.
Watch the serving layer, too. Revolut's numbers depend on NVFP4, disaggregated prefill and a 256,000-token cap. Those are engineering choices, not model features. A team that installs the same weights without the same tuning will see different costs.
The cheapest test is a bounded pilot. Pick one team, measure its token spend and lead time for a quarter, then compare a routed open-weight setup with the frontier API it replaces. Include the cost of someone to run the cluster.
Revolut's experiment does not settle the open-versus-closed debate. It does show that the debate is now about arithmetic, not ideology.