Revolut Cuts AI Coding Costs by 90% After Switching to Self-Hosted Zhipu AI GLM Models

Fintech Giant Slashes AI Expenses by Moving Away from Proprietary APIs

European fintech giant Revolut has reduced its artificial intelligence coding costs by 90% after transitioning its engineering workforce from commercial OpenAI and Anthropic APIs to self-hosted open-weight models developed by Chinese AI lab Zhipu AI (Z.ai). Shared by Revolut’s Head of AI Operations David Abitbol, the infrastructure overhaul replaces unpredictable, metered software bills with dedicated internal compute. By hosting the models directly on its own servers, Revolut lowered its blended inference cost to approximately $0.30 per million tokens—just one-tenth of what it previously spent on third-party coding assistants like Claude Code and OpenAI Codex.

Hardware Architecture, Latency Tuning, and Hybrid Model Routing

Rather than simply buying more hardware, Revolut achieved dramatic performance and cost efficiency through aggressive inference optimization on a fixed cluster:

  • Dedicated Nvidia B200 Cluster: Revolut runs Zhipu AI’s open-weight GLM architecture (moving from initial deployments to newer GLM iterations) across 56 Nvidia B200 GPUs.
  • 88% Faster Response Times: Through techniques like speculative decoding, prefill-decode disaggregation, and cache optimization, time-to-first-token dropped from over 10 seconds to roughly 1.6 seconds—even as internal request volume surged by 76%.
  • Hybrid Multi-Model Routing: Because GLM specializes in text and code generation, Revolut routes multimodal image tasks to hosted models like Google Gemini while reserving frontier APIs like Claude and OpenAI strictly as rare exceptions.
  • Regulated Sandbox Governance: To meet strict banking compliance standards, the weights are hosted entirely on Western infrastructure inside tightly controlled sandboxes with mandatory human-in-the-loop approval rules.

Removing Token Caps Unlocks Major Engineering Productivity

Eliminating per-engineer spending limits fundamentally transformed how Revolut’s roughly 2,000 software engineers utilize artificial intelligence in their daily workflows. Under previous commercial API pricing—which generated bills between $1,000 and $10,000 per engineer monthly—developers rationed their token budgets exclusively for complex code architecture. With unlimited internal access, engineers now delegate routine operational bottlenecks such as clearing Jira tickets, updating Confluence documentation, and debugging. Across a tracked sample of 2,000 engineers, coding time fell by 60% while overall task lead time dropped by 70%.

Challenging the Economics of Frontier Enterprise AI Subscriptions

Revolut’s transition highlights a growing inflection point in enterprise artificial intelligence economics as organizations scale from experimentation to mass developer deployment. As agentic coding workflows consume massive volumes of tokens through iterative reasoning and repository searches, per-token API pricing models rapidly become cost-prohibitive at scale. By demonstrating that open-weight Chinese models deployed on sovereign hardware can match or outperform proprietary assistants on unit economics and latency, Revolut provides a compelling blueprint for large enterprises seeking predictable AI operating expenses.

Setting a New Benchmark for Self-Hosted Enterprise AI

Ultimately, Revolut’s shift underscores how rapidly open-weight AI models and modern GPU architectures are eroding the moat of closed, proprietary API providers. By combining rigorous inference engineering with strict regulatory sandboxing, the fintech leader proved that organizations operating at scale can achieve faster response times, deeper developer adoption, and dramatic cost reductions without sacrificing data control.

Share

WhatsApp Channel