Revolut Slashes AI Coding Costs 10x by Ditching Claude for Self-Hosted GLM

The Metered Commercial API Trap at Enterprise Scale

When enterprise engineering teams first adopt generative AI coding tools, commercial SaaS APIs appear economical. Paying a few cents per thousand tokens offers an effortless on-ramp. However, as adoption expands across thousands of software engineers, variable token billing morphs into a severe operational bottleneck.

Speaking at the AI Engineer Conference in New York City, David Abitbol, who leads AI operations and infrastructure at British fintech giant Revolut, detailed how the company confronted this exact scaling wall. With roughly 2,000 engineers actively building across Revolut banking systems, reliance on proprietary closed models like Anthropic Claude Code and OpenAI Codex produced unsustainable financial spikes.

Individual senior engineers routinely consumed between $1,000 and $10,000 worth of tokens each month. To contain spiraling budgets, management imposed rigid departmental quotas. Yet this financial defense mechanism created an unintended consequence: developers began hoarding their token allowances strictly for the most difficult algorithmic problems, abandoning AI for routine tasks like writing boilerplate tests, refactoring modules, or parsing documentation.

To break this dynamic, Revolut made the radical decision to dismantle its commercial API contracts and deploy an in-house inference cluster running open-weight models.

The Hardware Blueprint: 56 Nvidia B200 GPUs for 2,000 Engineers

Rather than leasing enormous hyperscaler server farms, Revolut calculated that modern high-performance hardware could support 2,000 engineers on a surprisingly compact physical footprint. The fintech firm provisioned 56 Nvidia B200 GPUs, organized into seven DGX B200 server nodes.

On this private silicon infrastructure, Revolut deployed open-weight GLM models (developed by Zhipu AI). By moving from pay-per-token cloud billing to depreciated hardware economics, the blended inference cost collapsed:

• Blended Token Cost: Dropped to approximately $0.30 per million tokens, representing a 90% (10x) reduction compared to commercial API tariffs.
• Fixed Cost Predictability: Engineering leadership eliminated end-of-month cloud billing shocks, locking in predictable capital expenses.
• Data Sovereignty and Compliance: Operating private nodes ensured sensitive financial transaction logic, proprietary banking algorithms, and customer KYC schemas never crossed public API boundaries.

This operational transition reflects a broader structural evolution across the tech landscape. As enterprise teams discover that commercial model markups frequently fund vendor margins rather than raw compute, self-hosting is becoming the standard for high-volume tasks. This trend complements developments like Fal.ai doubling its valuation to $16 billion on custom inference infrastructure, proving that specialized compute architectures consistently outperform generic public cloud wrappers, as seen in Nvidia deployment of AI digital twins for semiconductor fabs.

Psychology of Unlimited Tokens: Unlocking Non-Coding Productivity

The most striking outcome of Revolut infrastructure shift was behavioral. Once the company announced that internal tokens were unlimited and unmetered, engineering habits shifted across the organization.

Abitbol reported that developers stopped treating AI as a scarce commodity. Instead, engineers built automated pipelines that applied GLM to routine operational overhead:

• Jira Ticket Triage: Automated agents parse ambiguous user bug reports, cross-reference crash telemetry, and draft detailed acceptance criteria before human review.
• Confluence Documentation: Internal documentation and API integration guides are generated and kept in sync with GitHub pull requests automatically.
• Automated Test Generation: Unit, integration, and edge-case regression suites are populated across legacy microservices without manual developer fatigue.
• Continuous Incident Debugging: SRE teams use dedicated agents to parse millions of distributed log lines during production alerts.

The operational payoff was immediate. Revolut documented a 60% improvement in active coding time and a 70% improvement in overall feature lead time across the 2,000-person engineering organization.

Inference Tuning: Sub-3-Second TTFT and Why Long Context Fails

Deploying open-weight models inside a major financial institution required rigorous systems optimization. Out-of-the-box open-source inference stacks initially struggled to meet interactive developer expectations.

Revolut infrastructure team tackled two primary architectural bottlenecks:

1. Sub-3-Second Time-To-First-Token (TTFT)

In developer environments, interactive autocomplete and chat assistance become unusable if latency exceeds three seconds. Software engineers lose focus when waiting for a remote completion. Without purchasing additional hardware, Revolut re-engineered its model serving runtime using optimized vLLM and TensorRT-LLM execution engines, applying continuous batching, prefix caching, and custom memory management to drive TTFT under three seconds.

2. The Context Window Tradeoff

While commercial model providers aggressively market 1-million and 2-million token context windows, Abitbol noted that gigantic context windows are counterproductive for high-throughput coding clusters. Processing massive context windows consumes quadratically more compute, introduces severe latency delays, and frequently dilutes model attention.

Revolut standardized on a 256K token context window. According to Abitbol, 256K tokens provides ample space to ingest complete microservice repositories, dependency trees, and relevant database schemas without degrading inference speed or causing hardware memory thrashing.

Model Routing: GLM as the Workhorse, Frontier Models as Exceptions

Although Revolut decommissioned Claude Code and OpenAI Codex as default tools, the company did not eliminate commercial models entirely. Instead, infrastructure architects instituted a strict hybrid model routing architecture.

Under this setup, the self-hosted GLM cluster functions as the primary operational engine for over 95% of daily requests: code completions, syntax transformations, unit test scaffolding, and repository navigation. Commercial frontier models are retained on an exception list, accessible only through approval flags for specialized tasks:

• Complex Multi-Modal Layouts: Visual UI mockups and design-to-code translations that require advanced vision capabilities.
• Esoteric Formal Verification: High-risk mathematical proofs and cryptographic protocols requiring maximum model parameters.
• Cross-Departmental Semantic Governance: Queries that require deep semantic definitions, similar to the graph-based data architectures introduced in Snowflake Business Ontology and Horizon Context.

By routing routine volume to cheap private silicon and reserving frontier models for genuine edge cases, Revolut maintains elite developer capability while keeping infrastructure costs disciplined.

Technical Comparison: Proprietary Cloud APIs vs Revolut Self-Hosted GLM Stack

The table below summarizes the operational and financial divergence between metered cloud APIs and Revolut private cluster deployment:

Frequently Asked Questions