Revolut cuts AI coding costs 90% with self-hosted Z.ai models

Why Revolut switched from Claude to self-hosted AI

Revolut, the London-based digital bank, has replaced Anthropic’s Claude AI models with self-hosted, open-weight models from Z.ai, according to reporting by TechRadar. The switch reportedly reduced AI coding inference costs by 90%, a figure that reflects both the lower per-token cost of Z.ai’s GLM-5.1 model and the elimination of unpredictable cloud billing.

The company initially launched its AI tooling with strict budget limits due to skepticism about the cost of cloud-based AI. At the time, some heavy users of Anthropic’s Claude were spending thousands of dollars monthly on tokens without generating equivalent returns, according to internal observations. Revolut previously experimented with Google’s Gemini before settling on Claude, suggesting a broader evaluation of AI vendors.

Infrastructure and productivity gains

Revolut’s self-hosted model runs on 56 Nvidia B200 GPUs, each with over 380GB of memory. The hardware investment is substantial, but the shift to Z.ai’s GLM-5.1 model reduced inference costs to $0.30 per million input/output tokens, a fraction of Anthropic’s pricing.

Beyond cost, Revolut reports significant productivity improvements. Lead time productivity improved by 70%, while coding time productivity rose by 60%. The company attributes these gains to removing token limits, which allowed developers to use AI more freely. David Abitbol, Revolut’s Infrastructure Lead, stated that the billing is now flat, eliminating month-end bill shocks.

Trade-offs of self-hosting AI

While self-hosting slashed inference costs, it introduced new expenses: GPUs, infrastructure, maintenance, and electricity. For Revolut, the economics of open-source AI became favorable only after scaling usage and recouping hardware investments. Z.ai’s models were also noted for offering better data privacy and auditability, a consideration for regulated industries like banking.

The move reflects a broader trend of companies evaluating open-weight models as alternatives to proprietary AI vendors. However, self-hosting is not a universal solution—it requires technical expertise and capital expenditure that may be prohibitive for early-stage startups.

What this means

Lazyfounder analysis — our interpretation, not reported fact.

Revolut’s shift to self-hosted open-weight AI models is a signal for founders and operators weighing the trade-offs between cloud-based and self-hosted AI. The reported 90% cost reduction and productivity gains suggest that for high-volume use cases, self-hosting can be economically viable—provided the upfront investment in hardware and infrastructure is manageable.

For startups, this case underscores the importance of benchmarking AI costs not just in absolute terms but relative to outcomes like developer productivity. The move away from token-based pricing removes unpredictability, which matters for budget-conscious teams. However, the need for significant GPU capacity and maintenance means this approach may only make sense for companies with scale or those prioritizing data privacy and control.

The choice of Z.ai’s GLM-5.1 model also highlights the growing competitiveness of open-weight alternatives to proprietary models like Anthropic’s Claude or Google’s Gemini. Founders should test these models for cost efficiency, especially if their workloads involve high token volumes.

Key takeaways

• Self-hosted AI models can drastically reduce inference costs for high-volume users, but require upfront investment in hardware and infrastructure.
• Removing token-based billing can eliminate budget unpredictability, a critical factor for startups managing cash flow.
• Open-weight models like Z.ai’s GLM-5.1 are becoming cost-competitive with proprietary alternatives, warranting evaluation for cost-sensitive use cases.
• Productivity gains from AI adoption depend on usage policies—releasing strict budget limits post-self-hosting contributed to Revolut’s improvements.

Related on Lazyfounder

• Anthropic
• Google
• Nvidia
• More technology news