Picture a coding assistant your company pays for by the token, with a monthly budget attached. You would probably ration it: save it for the hardest problem, go easy near the end of the quarter. That, by David Abitbol's account, is roughly where Revolut's engineers were until spring 2026. Then the company stopped paying per token for frontier models as the default, loaded open model weights onto GPUs it controls, and told engineers to use it freely.
Abitbol leads AI operations at Revolut. He gave the talk, "Shipping like a startup, regulated like a bank," on the main stage of AI Engineer NYC on October 13, 2026, and the conference has published the recording, a machine transcript and a short write-up. Unless stated otherwise, the numbers and judgments below are his. We found no written material from Revolut itself, so none of this can be verified outside the talk.
He opened with a chart tracking a sample of 2,000 engineers over the previous year on two measures: lead time, from picking up a feature to having it merged and ready to deploy, counted in days; and coding time, counted in hours. The first AI tool Revolut introduced was Gemini, which he described as the first time staff could use private company data with AI. He credited it with a 10% improvement in lead time and 20% in coding time, mostly because people stopped digging through Stack Overflow and documentation. Claude Code, rolled out next, made little difference at first. The reason was cost. Revolut was skeptical about AI spending after seeing some engineers burn through a thousand, even ten thousand, a month without a clear return, so Claude Code came with strict budget limits.
The change came around April 2026, when Revolut started self-hosting open models and settled on Z.ai's GLM, which he said the team is "a big fan of" for coding. With GLM 5.1, his chart showed coding time improving by 60% and lead time by 70%. His reading is that lead time moving more than coding time means writing code was never the main bottleneck. Once tokens no longer need rationing, people hand the model the inbox, Jira and Confluence work around the code too.
Keep one thing in mind. In April two things changed at once: the model, and the end of rationing. The conference write-up notes that the internal comparison doesn't isolate the model change from other causes, and Abitbol didn't say how "improvement" was calculated.
Open models are now the priority at Revolut, he said. Billing is flat, so there is no surprise invoice at the end of the month. Codex, Claude Code and Gemini are still in use, and he praised them, but they are kept as exceptions. He put two numbers on the economics: a blended cost of about $0.30 per million tokens for self-hosted GLM, and an invoice around a tenth of what Revolut had been paying Anthropic and OpenAI. The conference write-up adds a fair caveat: those figures reflect Revolut's own deployment and utilization, not a price any self-hosted model will deliver.
The hardware behind it, he said, is 56 NVIDIA B200 GPUs serving GLM 5.3 to roughly 2,000 engineers. For a sense of scale, GLM-5.1, which Z.ai released as open weights in April 2026, is listed on Hugging Face as a 754-billion-parameter, text-only model under the MIT license.
His main technical point was that dropping weights onto GPUs and starting inference gives you terrible performance; what you get depends on how you tune it. His working targets come from experience. Time to first token should stay under three seconds, because past five his Slack fills with complaints. Raw generation speed matters less: about 200 tokens per second is enough, since engineers now run several tasks in parallel and check back when they finish rather than watching the model think. After 5 p.m., when demand drops, he said engineers can get around 800 tokens per second.
To show what tuning bought, he showed a chart running from April to August with the GPU count held constant. According to the conference write-up's reading of the slide, requests per second rose 76% while time to first token fell 88%, from 14.26 seconds during migration to 1.68 seconds once things settled. His advice: resist adding GPUs until you have tuned what you have. He named a few levers. The model can serve 1 million tokens of context, but he said 256K is what's actually needed, and reserving more wastes GPU memory. On Blackwell-generation GPUs he runs FP4, which he said cuts the GPU footprint by about 40% against FP8 with no loss in precision, though he didn't describe how that was measured. For coding workloads the cache hit rate should be above 90%. He also mentioned lower-level options such as how to split the model across GPUs and speculative decoding.
GLM handles text only. His trade-off is to keep the smarter coding model and send images elsewhere, with a local router passing image work to a model such as Gemini. He said Revolut's UI designer gets good results from a vision-capable Flash model in the GLM family; the transcript garbles its name, and we couldn't pin down which one. The endpoints are OpenAI-compatible, including the Responses API, so engineers can use them from OpenCode, Codex and the Codex app.
He closed with a message he said was for the CIOs in the room, and this is where the "regulated like a bank" half of the title comes in. There is a stigma, he said, around Chinese open models, but Revolut isn't calling an inference endpoint in China; it takes the weights and runs them on a US-based service, which he considers fine from a data-sharing point of view. He described Revolut as an EU-regulated bank under the strictest compliance regime. The approach he described is to keep these coding tasks from counting as autonomous agents by always having a human in the loop, and to back that with a sandbox. Tools like Codex and Claude Code can already approve actions for you; you set rules for what they may not do, allowing grep but not gradle, for instance. That, he said, gives the regulator confidence this isn't an autonomous agent needing different supervision. The same sandbox guards against insiders misusing the tools ("you'd be surprised what you find in a large company," he said) and against prompt injection.
On data location, his view is that GPUs are fine anywhere within what he called the data equivalency framework, naming the US, the EU, the UK and Israel; that inference and GPU providers keep no data by default; and that you need strong OpenTelemetry tracing so you can dig through the logs after an incident. These are his reading of Revolut's own situation. He didn't name a regulator or a rule, and the conference write-up warns against treating his remarks as proof that any model, region or sandbox meets another organization's obligations.
If you manage the AI tooling budget at your company, the part worth taking from this talk isn't "switch to GLM." It's the metric Abitbol watches: lead time from ticket to merge, not how fast code gets typed. Start measuring that on your own team, then decide whether the answer is a bigger frontier-model budget or GPUs of your own. His math rests on about 2,000 engineers, 56 B200s and months of tuning, so if your team is much smaller, run your own numbers before you copy it.