Revolut slashes AI coding costs by 90% by ditching Claude for self-hosted models

Moving away from commercial AI

Revolut, the London-based digital bank, is changing how it integrates artificial intelligence into its internal workflows. The company previously experimented with Google's Gemini and Anthropic's Claude to assist its developers. However, the bank found that heavy users were spending thousands of dollars every month on tokens without seeing a clear return on investment.

To address these rising expenses, Revolut decided to cut out Claude entirely for its coding workloads. The bank is now opting for self-hosted, open-weight models provided by China's Z.ai. This transition was driven by a need to eliminate the unpredictability of commercial billing and the limitations of external APIs.

Infrastructure Lead David Abitbol noted that the company was highly skeptical about the costs of AI from the beginning. Because of these concerns, the bank initially launched its AI initiatives with very strict budget limits. By moving to a self-hosted model, the company aims to create a more stable and predictable financial environment for its technology team.

The economics of self-hosting

The primary driver for this shift is the significant reduction in inference costs, which refers to the cost of running a model to generate an output. Revolut reported that switching to the Z.ai GLM-5.1 model slashed these costs by as much as 90% in less than a year. The new costs are approximately $0.30 per million input and output tokens, which is about one-tenth the price of similar commercial options.

While the per-token cost is much lower, self-hosting requires a massive upfront investment in physical hardware. Revolut now operates its own infrastructure, which includes 56 Nvidia B200 GPUs. The model itself requires more than 380GB of GPU memory, which does not include the additional 6 to 26GB of memory needed per user to maintain context retention.

The bank is trading variable operational costs for fixed infrastructure expenses. David Abitbol explained that the new system provides flat billing, which prevents the company from experiencing 'bill shock' at the end of the month. This strategy allows the bank to manage its own capacity and scale its AI usage based on internal needs rather than external pricing.

Boosting developer productivity

Beyond the financial savings, the move to self-hosted models has led to measurable gains in how the company's engineers work. Revolut reported a 70% improvement in lead time productivity and a 60% improvement in coding time productivity. These metrics suggest that the new system is helping developers complete tasks much faster than the previous commercial setup.

One major reason for this increase is the removal of 'token limits,' which are restrictions on how much data a model can process at once. By derestricting the environment, Revolut's developers can use AI much more freely. This allows for deeper integration into their coding workflows without having to constantly monitor usage limits or costs.

The bank also highlighted the benefits of using Z.ai's GLM models for data privacy and auditability. Because the models are hosted on internal hardware, the company has more control over how data is handled. This level of oversight is often a priority for financial institutions that must adhere to strict security standards while still utilizing modern technology.

The future of AI infrastructure

Revolut's move reflects a growing trend of companies looking to reduce their reliance on major commercial AI providers. While frontier models remain the exception for certain tasks, many companies are finding that less advanced models are sufficient for mundane work. This shift suggests that the economics of AI may improve for many businesses as they move toward self-hosting for routine operations.

However, the bank's experience also serves as a warning that deploying AI aimlessly is a poor strategy. The high costs of GPUs, electricity, and maintenance mean that the transition to self-hosting is not a simple shortcut to savings. It requires a sophisticated understanding of hardware requirements and the ability to manage complex infrastructure.

As more organizations follow Revolut's lead, the industry may see a divide between those who rent high-end capabilities and those who build private environments. For the reader, this means that the most efficient way to use AI might soon depend on the specific scale and type of work being performed. The goal is to find the right balance between model power and operational cost.