Revolut moves 2,000 engineers to self-hosted GLM

Revolut has shifted approximately 2,000 engineers to self-hosted GLM models for coding tasks, effectively moving frontier models from Anthropic and OpenAI to an exception status. According to a report by Guillermo Flor on X, this transition was led by David Abitbol, who heads AI operations at the company. The most significant outcome of this infrastructure change is not merely cost savings, but a 70% improvement in engineering lead time, driven by the removal of token rationing that previously constrained developer workflows.

The infrastructure and cost breakdown

The technical foundation for this shift is a cluster of 56 Nvidia B200 GPUs, which Revolut uses to host the GLM models internally. Abitbol reports a blended cost of 30 cents per million tokens. This figure represents roughly one-tenth of the costs Revolut was previously paying to Anthropic and OpenAI for similar workloads. The financial delta is substantial enough to alter the economic calculus of AI adoption within the engineering organization. By internalizing the inference layer, Revolut has decoupled its AI usage from the variable pricing structures of external API providers, creating a predictable cost baseline for a large-scale engineering team.

The impact of unlimited access

The primary driver of the 70% lead time improvement is behavioral. When AI access was metered through expensive external APIs, engineers tended to reserve AI tools for complex, high-stakes tickets to conserve credits. As Flor notes, "the gains showed up when tokens stopped being rationed. Engineers quit saving AI for the hard tickets and pointed it at everything." This shift in usage patterns means that AI is now integrated into routine development tasks, from boilerplate generation to code review, rather than being treated as a specialized tool for difficult problems. The removal of financial friction has effectively democratized AI assistance across the entire engineering stack, allowing for more frequent and consistent interaction with the models.

Context in the AI coding landscape

This move aligns with a broader trend in the tech industry where large enterprises are increasingly evaluating self-hosted open-weight models as a viable alternative to proprietary frontier APIs. While models like GLM have historically been positioned as competitive alternatives to closed-source systems, their adoption at the scale of Revolut suggests that performance parity in coding tasks is now a practical reality for many use cases. The decision to push Claude Code and Codex to the exception list indicates that Revolut’s internal evaluation found the self-hosted solution sufficient for the majority of their coding needs. This is a significant data point for the broader market, as it suggests that the gap between open-weight and proprietary models in specific domains like software engineering is narrowing to the point where cost and control become the deciding factors.

What it means for developers

For developers at Revolut, the immediate change is the ability to use AI tools without the psychological and financial overhead of token consumption. This likely results in a more fluid development process, where AI is used as a constant companion rather than a resource to be managed. For other engineering teams, this case study offers a blueprint for evaluating self-hosting. The key metrics to consider are not just the raw cost per token, but the impact on developer behavior and throughput. If your team is currently limiting AI usage due to cost, this example suggests that the hidden cost of rationing—lost productivity and reduced adoption—may outweigh the savings from using a cheaper API. Teams should assess whether their workload can be handled by open-weight models and whether the infrastructure investment in GPUs is justified by the potential gains in developer velocity.

What to watch

• Model performance benchmarks: Further independent benchmarks comparing GLM’s coding capabilities against Claude and Codex will be crucial to validate Revolut’s internal findings.
• Adoption by other enterprises: Whether other large tech companies follow suit with similar self-hosting strategies for AI coding tools.
• Pricing responses from labs: How Anthropic and OpenAI respond to the loss of enterprise volume in coding-specific workloads, potentially through new pricing tiers or feature differentiation.
• Hardware availability: The scalability of this model depends on the availability of high-end GPUs like the Nvidia B200, which could become a bottleneck for other organizations looking to replicate this setup.