A leading European fintech reportedly made self-hosted open-weight models the default for AI coding, cutting inference costs by 90%. Here's what the evidence actually shows and why it matters for every company running AI engineers.
For the past few years, the default way for a company to add artificial intelligence to its software engineers was simple: call a vendor's API. You paid per token, you got the best models available, and you never had to buy a single GPU. It was convenient, and it was expensive. Now a prominent European fintech is quietly rewriting the playbook, and the story has sent ripples through the AI industry.
Revolut, the London-based banking and payments company that serves tens of millions of customers, reportedly made a decisive move: it switched the default for its engineers' AI coding tools away from commercial, pay-per-use models like Anthropic's Claude and OpenAI's Codex, and toward models it runs itself on its own hardware. The reported result is a cut in inference costs of roughly 90 percent — about ten times cheaper — for a workforce of around 2,000 engineers. The switch relies on open-weight models from China's Z.ai, primarily GLM-5.1 and, according to one report, later GLM-5.3, running on NVIDIA B200 GPUs.
This is not a minor pricing tweak. It is a strategic signal. When a company the size and scrutiny of Revolut decides that a Chinese open-weight model is good enough to be the default for its engineers, it legitimizes an entire category of technology that was, until recently, treated as a risky side experiment. But before we celebrate or condemn the move, it is important to separate what the reports actually say from the more dramatic headlines that have grown up around them.
Why this matters: we are not talking about a research curiosity here. Revolut's reported move is a real, large-scale example of a company taking AI coding infrastructure in-house. It raises questions about cost, control, data security, and dependence on Chinese technology that will affect businesses for years to come.
The story broke across several reputable technology outlets on November 3 and 4, 2026. The Register and TechRadar both reported on the switch, and they attribute the core figures to David Abitbol, who leads Revolut's operations and infrastructure. The headline claim is striking: Revolut says its blended cost for AI-coding tokens dropped to about $0.30 per million tokens after moving to self-hosted GLM-5.1, roughly one-tenth of what it was paying to Anthropic and OpenAI.
"Self-hosting GLM, our blended cost of token is 30 cents per million, and so our invoice is one-tenth of what we were spending depending on Anthropic and OpenAI."
The deployment reportedly serves around 2,000 engineers and runs on 56 NVIDIA B200 GPUs. That is a serious piece of hardware, though a modest cluster compared to the giant arrays used to train frontier models. Inference for a coding assistant is a different workload entirely: it serves many short requests from many users throughout the working day, and the cost per token falls dramatically once those GPUs stay busy. Dividing the engineers by the GPUs gives roughly 36 engineers per GPU, which tells you how heavily the hardware is shared — and why utilization is the hidden variable behind the savings.
One report adds that Revolut later migrated from GLM-5.1 to GLM-5.3, which suggests the team is treating the stack as something it upgrades on its own schedule, rather than waiting for a vendor's roadmap. That detail matters as much as the price tag, and we will return to it.
But here is the crucial nuance that the most viral versions of the story have dropped. The reports say Revolut made self-hosted models its default and moved commercial services like Claude to a secondary, behind-the-scenes role. They do not clearly state that Revolut completely eliminated Claude. The original framing — "Revolut cuts out Claude entirely" — is a stronger claim than the evidence supports. Revolut reportedly stopped relying primarily on commercial models; it is not certain that every engineer lost access to Claude.
Let's lay out the figures the reports provide, and be honest about which parts are confirmed and which are our own arithmetic.
| Figure | Detail | Status |
|---|---|---|
| Inference cost cut | About 90%, or roughly 10x | Reported by The Register and TechRadar |
| Blended token cost | About $0.30 per million input/output tokens | Reported, attributed to Abitbol |
| Engineers served | About 2,000 | Reported |
| Hardware | 56 NVIDIA B200 GPUs | Reported |
| Model | GLM-5.1, later GLM-5.3 per one report | Reported; migration from a single report |
| Previous commercial spend | $1,000 to $10,000 a month for some engineers | Reported, applies to some engineers only |
| Coding time | Down 60% | Reported, sample of ~2,000 engineers |
| Task start to merge | Down 70% | Reported, sample of ~2,000 engineers |
If the 10x ratio holds, it implies a commercial blended rate near $3 per million tokens — though that specific figure is our calculation, not one the reports state. The 90% reduction and the 10x savings are the same claim expressed as a percentage and as a multiple; treat them as one data point, not two independent confirmations.
The honest caveat: One report explicitly notes that the savings figure may exclude costs such as cloud infrastructure, personnel, hardware, monitoring, and operations. The 10x applies to token pricing, and may not capture the full cost of running the cluster. Until Revolut publishes a complete breakdown, the sensible reading is that the direction is credible and the exact multiple is a company-reported figure.
To understand why this story is significant, you need to understand the difference between a "commercial API" and an "open-weight model." A commercial API is a model you access by paying a vendor — like Anthropic's Claude or OpenAI's Codex — per token. You get the latest, most powerful models, but you never see the model itself, you don't control when it changes, and your costs scale directly with how much you use it.
An open-weight model, by contrast, is a model whose "weights" — the trained parameters that make it intelligent — are released for anyone to download. You can install it on your own servers, run it on your own hardware, and keep your prompts and code inside your own network. The GLM family is developed by Z.ai, a Chinese AI company. Revolut downloaded GLM-5.1, installed it on its own B200 GPUs, and ran it.
This is the point where the story connects to a much wider shift in the industry. For years, the assumption was that only the biggest tech companies could afford to run their own models, and that doing so was risky. Revolut's decision suggests that open-weight coding models have matured enough to be treated as a production default by a heavily regulated financial firm — not a side project.
There is also a powerful control argument that has nothing to do with price. When you use a hosted model, it can change underneath you without notice — its behavior, its pricing, its availability. A self-hosted model does not do that. Revolut decides when to move from GLM-5.1 to a newer release, can test the change on its own tasks first, and can roll back if quality slips. Teams that have watched a hosted model shift behavior overnight understand exactly why this matters.
You might wonder what any of this has to do with you if you are not a fintech executive. The reality is that Revolut's move touches a trend that will shape your digital life in several concrete ways.
Imagine you are a small business owner. Right now, adding AI to your operations means paying per-use fees that eat into your margins every single time an employee uses an AI tool. If the "Revolut effect" drives open-weight models to become cheaper and better, those tools become affordable for teams of ten instead of just teams of ten thousand.
Imagine you are a software developer. Your daily tools — the assistants that write code, debug problems, and explain errors — are quietly changing. Whether your company uses a commercial model or a self-hosted one affects how fast your tools improve, whether they stay online, and how your proprietary code is handled.
Imagine you care about data privacy. When a company runs a model in-house, your data — your messages, your financial queries, your personal details — stays inside that company's network instead of being sent to a third-party vendor. That can mean stronger privacy, but it also means you have to trust that one company with everything.
Imagine you are worried about global technology dependence. Revolut is replacing an American model with a Chinese one. This is part of a larger story about who controls the AI infrastructure the modern economy runs on, and it raises genuine questions about data sovereignty, export rules, and supply chains that ordinary users feel the effects of without ever hearing the debate.
The move to self-hosted open-weight models is not limited to fintech. Several sectors are naturally positioned to benefit, and Revolut is just the most visible early adopter.
Financial services. Banks and insurers already run sensitive data through internal systems. Self-hosting an AI model means they keep trading data, customer records, and financial plans inside their own secure perimeter — a strong fit for an industry that is heavily regulated and deeply concerned about where their data goes.
Healthcare. Hospitals and research institutions handle some of the most sensitive data imaginable. Running a model on-premises lets them use AI for documentation, research, and diagnostics without patient data ever leaving the building — a configuration that aligns naturally with privacy laws like HIPAA.
Government and defense. These sectors have long preferred to control their own infrastructure for reasons of security and sovereignty. Self-hosted models fit neatly into air-gapped networks and classified environments where cloud access is simply not permitted.
Large enterprises and software companies. Any organization with thousands of engineers and high, predictable AI usage gets the biggest benefit from buying or renting hardware rather than paying per token. The more volume you push through a model, the cheaper self-hosting becomes compared to a commercial API.
Looking at the short term, the next one to two years, we expect more large engineering organizations to test self-hosting, using Revolut as a reference point in budget meetings. Hybrid routing will likely become standard: cheap open models handling routine edits, while expensive closed models tackle the hardest problems. Commercial vendors will probably cut or restructure their enterprise pricing to compete.
In the mid-term, three to five years out, we may see independent audits of total AI cost become common, since the reports flag costs that self-hosting can hide. The open-weight release cycle will likely quicken, with model makers competing on release speed and coding quality to win over enterprise buyers. A mature market for the supporting software — the tools that make running your own model easy — will grow around this infrastructure.
Longer term, over ten years, self-hosting may become so streamlined that even medium-sized companies run their own customized models as easily as they use a spreadsheet today. AI could become a truly foundational utility, running inside the organizations that use it rather than flowing from a handful of central providers.
The Revolut case sits at the intersection of several parallel technological advances. Open-weight models have improved dramatically: the GLM family, along with others from China and elsewhere, now writes code that is genuinely competitive with the most expensive commercial models. At the same time, inference software — the layer that serves a model efficiently to many users — has matured, making it possible to run large models on modest hardware.
Hardware is another key driver. NVIDIA's B200 GPUs, used in the reported Revolut deployment, are among the most powerful AI accelerators available, and their increasing availability outside the biggest data centers makes self-hosting feasible for more organizations. Meanwhile, companies are building "inference-specific" chips that could eventually make running models even cheaper and more energy-efficient.
The trend is clear: the gap between open and closed models keeps narrowing, while the cost of running models keeps falling. That combination is exactly what makes a story like Revolut's possible.
Like any major technological shift, this move carries both promise and risk. A balanced view requires looking at both.
The positives. First, cost. Even if the full picture is murkier than the headline, self-hosting can meaningfully reduce the cost of AI for organizations with enough usage to justify the hardware. Second, control. A company decides its own upgrade schedule, its own quality standards, and its own availability — no vendor can change the rules overnight. Third, data sovereignty. Sensitive information stays inside the organization, which matters enormously for regulated industries and for anyone who values privacy.
The risks. First, hidden cost. Buying 56 GPUs is only the beginning. Someone has to patch drivers, tune the software, watch latency, and answer the pager when the cluster stalls. Those costs may sit outside the reported savings. Second, security. When you host the model, you own the guardrails around it — prompt-injection defenses, secret scanning, and output review become your responsibility. A self-hosted model with no safeguards can be manipulated into leaking data or producing harmful output. Third, dependence on China. Sourcing your core AI infrastructure from a Chinese provider raises questions about export controls, potential future restrictions, and geopolitical risk that a European company must weigh carefully.
There is also a subtle risk worth naming: the productivity gains — 60% less coding time, 70% faster merges — may reflect behavior as much as model quality. When a tool becomes cheap, people use it more, and that extra usage can look like a productivity boost even if the model itself is not dramatically better. The reports do not fully isolate these effects.
Revolut's reported switch to self-hosted Chinese open-weight models is a genuinely significant moment, but it deserves a careful, unsensational reading. The direction is credible: open-weight models are now good enough for a major fintech to run them as its default, and the cost savings on token pricing are real. But the exact 10x figure is company-reported, the full cost is unclear, and the claim that Revolut "cut out Claude entirely" goes beyond what the reports actually support.
What Revolut really demonstrates is that the era of unquestioned dependence on a handful of American AI vendors is ending. Whether you are a giant fintech or a small startup, you now have a genuine choice about who powers your AI — and that choice is good for everyone.
What do you think? Is self-hosting an open-weight model the right call for your organization, or is the convenience of a commercial API still worth the premium? Would you feel comfortable running a Chinese model with your company's data? Share your thoughts below — this is exactly the kind of debate that will shape the future of AI.
The bottom line: The most important lesson from Revolut is not the headline number — it is the method. Measure your own token costs, run an open-weight model on a small slice of real work, price the full stack, and keep a commercial model available as a fallback. Copy the method, not just the headline.