Decision Guide For Engineering Leaders
Should Your Engineering Team Self-Host an Open-Weight Coding Model?
Revolut's 2,000 engineers say it depends on volume, not ideology.
It depends on engineer count, in-house GPU and ops skill, and data-residency needs, not on model quality alone. Revolut self-hosted an open-weight model at roughly 2,000 engineers and cut its frontier API bill by about 90%, but still keeps frontier tools on hand for its hardest problems. Smaller teams will usually find a frontier API cheaper to run.
Published
Key Takeaways
- At about 2,000 engineers, Revolut self-hosted Z.ai's GLM on 56 Nvidia B200 GPUs and cut its blended token cost to roughly a tenth of its prior Anthropic and OpenAI bill.
- Lead time (days from picking up a feature to merge-ready) improved more than coding time, around 70% versus 60%, which is evidence the real bottleneck was never pure coding.
- The gain tracked with removing token budgets, not the model switch alone: once tokens were no longer scarce, engineers pointed AI at email, tickets, and docs too.
- Self-hosting took real infrastructure work, not just downloading weights: tuning context length, quantization, and parallelism cut time-to-first-token by 88% on the same GPU count.
- Revolut still keeps Anthropic's Claude Code, OpenAI's Codex, and Google's Gemini on hand as the frontier exception for its hardest problems, not as the daily default.
The full guide
Tap any section to open it.
About this guide
This is a buyer's guide for engineering leaders deciding whether to self-host an open-weight coding model, like Z.ai's GLM, instead of paying per seat or token for a frontier API. It is built around Revolut's public numbers: a bank that moved most of its roughly 2,000 engineers off per-token billing, while keeping Anthropic's Claude Code, OpenAI's Codex, and Google's Gemini for its hardest problems. The lesson is not which model Revolut picked, but what changed once tokens stopped being rationed.
Self-Hosting vs. a Frontier API or Seat Plan: A Decision Checklist
| Factor | Leans self-hosted open-weight | Leans frontier API or seat plan |
|---|---|---|
| Team size and token volume | Hundreds to thousands of engineers generating steady, high daily token volume, as at Revolut's roughly 2,000-engineer scale. | A small or mid-size team where GPU and ops overhead would cost more than the API bill it replaces. |
| In-house GPU and ops skill | You already run or can hire for GPU infrastructure: serving-stack tuning, quantization, and capacity planning are real, ongoing work. | No dedicated ML infrastructure team, and no appetite to build one just to serve a coding model. |
| Data residency and compliance | Code or context cannot leave infrastructure you control, or a regulator expects you to name exactly where inference happens. | A vendor's standard data-handling terms already satisfy your compliance and data-residency requirements. |
| Fallback for the hardest problems | Budget for occasional frontier API usage stays available for the small share of tasks an open-weight model handles poorly. | Most of your workload is exactly the kind of hard, varied problem frontier models are tuned for, not routine coding. |
| Billing predictability vs. flexibility | You want flat, predictable infrastructure costs instead of a bill that moves with usage and model pricing changes. | You want to switch models or vendors quickly without re-tuning a serving stack, and usage is too uneven to size GPUs for. |
The Source CaseWhat Revolut Actually Did and Measured
David Abitbol, who runs AI operations at Revolut, described the rollout in a talk at AI Engineer's New York conference in late 2026. Revolut tracks two numbers across a sample of about 2,000 engineers: coding time in hours, and lead time in days from picking up a feature to having it merged and ready to deploy.
An early rollout of Google's Gemini with internal data access improved both metrics modestly, mostly by replacing documentation lookups. Anthropic's Claude Code went out next under strict per-engineer token budgets, after some engineers ran up large bills with little clear return, and Abitbol says it showed little measurable effect at first.
What changed the numbers was self-hosting an open-weight model, Z.ai's GLM, starting around April 2026, with no token limits at all. Coding time improved by roughly 60% and lead time by roughly 70%. Revolut now treats the self-hosted model as the default for everyday coding, with flat and predictable billing, and keeps frontier tools as the exception for the hardest problems.
At roughly 2,000 engineers, serving the model took 56 Nvidia B200 GPUs, aiming for a time-to-first-token under three seconds (complaints start past five). The blended cost came out to about $0.30 per million tokens, around a tenth of what Revolut had been paying Anthropic and OpenAI combined.
The Real LessonThe Lesson Isn't the Model, It's the Rationing
Abitbol's own read of the data is the most useful part: lead time improved more than coding time, which means coding was never the main bottleneck. Once nobody has to save tokens for the hardest ticket of the quarter, engineers start pointing AI at email, Jira, Confluence, and the rest of the work that surrounds a feature, not just the code itself.
That lesson holds regardless of which model a team runs. A team on a frontier API that budgets tokens per engineer, or caps usage to control cost, is recreating the exact rationing problem Revolut solved by removing limits. The fix is not always self-hosting; sometimes it is a cheaper or flatter pricing plan with the same frontier model.
The practical takeaway for any engineering leader is to measure lead time, not just coding time or lines generated, before deciding whether AI tooling is paying off, and to watch whether engineers are rationing their own usage because of cost or quota pressure.
This page focuses on the team-cost and rollout decision: whether to self-host at all, and how to roll it out once you decide. For the broader, model-agnostic question of what open weights, open source, and closed APIs actually mean and why labs and governments treat them differently, Nightschool AI's guide Open-Source vs Closed AI Models covers that ground in full.
Tuning Before ScalingSelf-Hosting Took Engineering, Not Just Weights
Revolut's infrastructure numbers are a useful caution against assuming self-hosting is simple once you have GPUs. From April to August 2026, with the GPU count held flat at 56, load rose from about 3 to over 7 requests per second while time-to-first-token fell from roughly 14 seconds to under 2, an 88% latency improvement with no new hardware.
Abitbol credits about 80% of that gain to serving configuration, not GPUs: capping context length well below the model's advertised maximum, running lower-precision quantization on current-generation GPU hardware, and tuning how the model's attention cache is split across devices. His summary line for teams considering this path is blunt: loading weights onto GPUs and calling it done gets terrible performance.
That tuning work is a real, ongoing ops cost most teams underestimate when they compare a self-hosted model's sticker price to an API bill.
Rolling It OutTrain the Habit Before You Touch Infrastructure
Revolut's gain tracked with removing token limits, not with swapping models. These are the steps that make that habit stick on any team, on any model.
Set a lead-time baseline first
Before changing any tooling or budget, measure how long a typical feature actually takes from pickup to merge-ready. Without this baseline you cannot tell whether a later change helped.
Remove quiet rationing
Check whether engineers are saving AI usage for the hardest ticket of the month because of a budget or quota. If so, that rationing is capping your results regardless of which model sits behind it.
Train on every ticket, not just hard ones
The habit that moved Revolut's numbers was pointing AI at email, tickets, and documentation, not only difficult coding problems. Teach that habit explicitly; it does not form on its own just because a limit was lifted.
Re-measure lead time, not just coding time
Track the same lead-time metric again after the change. If coding time improves but lead time does not, the bottleneck is still somewhere else in the process, and more tokens will not fix it.
Honest LimitsWhat Revolut's Case Doesn't Prove
Revolut's numbers describe one regulated bank's deployment at its own scale. Keep these limits in mind before treating this as a universal result.
One company's numbers, not a benchmark
The cost and latency figures in this guide describe Revolut's specific deployment, usage pattern, and negotiated terms. Treat them as evidence of what is possible, not as a price or performance guarantee for your own workload.
Frontier tools stayed in the mix
Revolut did not eliminate Anthropic's Claude Code, OpenAI's Codex, or Google's Gemini. It kept them as the deliberate exception for problems its self-hosted model handles less well, which is a cost most self-hosting pitches leave out.
Compliance details are company-specific
Revolut's reasoning about where it runs the model and how that satisfies its regulator reflects its own circumstances as an EU-regulated bank. It is not a general finding that any model, region, or setup clears another organization's obligations.
Teach the Habit Before You Touch Infrastructure
Revolut's gain came from removing token limits so engineers used AI on every ticket, not just the hard ones. That habit change comes before any self-hosting decision, and Nightschool AI's corporate training builds it across a whole engineering team.
Frequently Asked Questions
What is an open-weight coding model?
A model whose trained weights are downloadable and can run on infrastructure you control, unlike a closed model you can only reach through a provider's API. Z.ai's GLM, the model in Revolut's case, is one example; running it still takes real GPU infrastructure and serving-stack tuning.
Did Revolut stop using Claude Code, Codex, and Gemini entirely?
No. Revolut made a self-hosted open-weight model the default for everyday coding, but kept Anthropic's Claude Code, OpenAI's Codex, and Google's Gemini available as the exception for its hardest problems, where it reports those frontier tools still perform well.
How much did Revolut actually save by self-hosting?
Its own reported blended cost was about $0.30 per million tokens, around a tenth of what it had been paying Anthropic and OpenAI combined, serving roughly 2,000 engineers on 56 Nvidia B200 GPUs. That figure describes Revolut's deployment and usage, not a price every team should expect.
Is self-hosting actually cheaper once you count the engineering work?
Not automatically. Revolut reports that about 80% of its performance gain came from tuning its serving configuration, not from buying more GPUs, which is ongoing infrastructure work. A smaller team without that skill in-house may spend more reaching comparable performance than it would on a frontier API bill.
Is a self-hosted open-weight model like GLM a security risk because of where it was made?
Where a model's weights were trained and where you run them are two different questions. Revolut's own position, as a regulated bank, was that downloading weights and running them on infrastructure it controls in approved jurisdictions is a different data path than sending requests to a third party's hosted endpoint; organizations with different regulators should verify this against their own compliance requirements rather than assume it applies to them.
What should a smaller engineering team do instead of self-hosting?
Focus on the lesson rather than the infrastructure: find out whether engineers are rationing AI usage because of cost or quota limits, and if so, look at a flatter or cheaper plan on the frontier model you already use before considering the GPU and ops investment that self-hosting requires.
What should we measure to know if AI coding tools are actually helping?
Measure lead time, the days from picking up a feature to having it merged and ready to deploy, alongside coding time. Revolut's data showed lead time improving more than coding time, which is a sign that the surrounding work, not the typing, was the real bottleneck.
Is Nightschool AI affiliated with Anthropic, Z.ai, or Revolut?
No. Nightschool AI is an independent learning platform and is not affiliated with, endorsed by, or sponsored by Anthropic, Z.ai, Revolut, OpenAI, or Google. Claude, Codex, Gemini, and GLM are products of their respective companies, named here only to describe what they are and how they were used in this case.
Build the Habit, Then Decide on Infrastructure
The model and the hosting choice matter less than whether your team is still rationing tokens.