Model Loyalty Is Expensive: What Revolut's 90 Percent Move Means for Your SME
In short
Revolut has reportedly cut its AI coding costs by around 90 percent by moving from OpenAI and Anthropic to a self-hosted Chinese model. For your SME, that means procurement discipline instead of model loyalty: portable prompts, clean data exit clauses, and a model-agnostic orchestration layer so you capture falling token prices.
The news: 90 percent lower AI costs with a self-hosted model
Alvin Foo reported on LinkedIn (2026) that Revolut cut its AI coding costs by roughly 90 percent after switching from OpenAI and Anthropic to the Chinese model Zhipu AI GLM-4.5. Running on 56 Nvidia B200 GPUs, inference costs are reported at around 0.30 US dollars per million tokens. The decisive shift is not the vendor but the mechanism: competitive advantage moves from the strongest model to the best performance at the lowest price at scale.
90 %
Reported cost reduction at Revolut after switching to Zhipu GLM-4.5
0.30 USD
Inference cost per million tokens on 56 Nvidia B200 GPUs as reported
Competitive advantage shifts from the strongest model to the best performance at the lowest price at scale.
Model loyalty as a cost trap
Token prices fall every quarter. If you are locked into a single vendor, you do not automatically capture those price reductions. A model-agnostic approach means evaluating performance and cost per task, not choosing one model as a matter of faith. The Swiss AI Podcast on DeepSeek, Apertus, and AI agents provides a practical framework for this shift.
What vendor lock-in costs in Switzerland
For Swiss SMEs, typical exit costs range from CHF 8,000 to CHF 120,000 per scenario, depending on whether you use APIs only (CHF 8,000 to CHF 20,000), a UI-bound platform (CHF 25,000 to CHF 60,000), or a proprietary agent platform (CHF 50,000 to CHF 120,000). Add 10 to 30 developer days for re-implementing prompts and logic.
The real lever: model agility and model routing
Costs fall when you use the right model for each task. Accenture notes that Chinese models can be valid for stable, clearly defined process steps with non-sensitive data; in highly regulated industries, stricter standards apply under Swiss data protection law or FINMA. The gap between Chinese and US models has shrunk from more than a year to one or two months. DeepSeek shows comparable performance at a fraction of the training cost.
A sovereign alternative for sensitive data
For data sovereignty, Apertus from ETH/EPFL on the Swisscom Swiss AI Platform is a valid Swiss option. It is performance-limited but useful where data must not leave the country or legal jurisdiction.
Practical framework: self-hosting, API, or outsourcing
Three paths lead to AI capacity. The comparison shows where you get the best price-performance ratio and the lowest lock-in risk.
- Self-hosting
- Pays off only at high, stable volume; requires in-house hardware and operational expertise. Unattractive for most SMEs.
- API usage
- Flexible to start, but often binds through platform and tooling logic. Switching costs arise from prompts and integration.
- AI outsourcing / model-agnostic orchestration
- Keeps options open, automatically captures price declines, reduces lock-in risk. Procurement as a managed service.
Procurement rules for model-agnostic AI
- Data exit clause: export formats and raw data must always be returnable.
- Notice period: keep it short and avoid automatic renewal lock-ins.
- Open standards: MCP and portable prompts, no proprietary locks.
- Semi-annual re-evaluation: actively compare models and tools, renegotiate prices.
- Avoid single-vendor commitments.
For most SMEs, model-agnostic orchestration is the economically sensible route. It encapsulates models behind an abstraction layer so you can route per task and switch vendors without rebuilding your processes. How midsize companies can anchor AI as a leadership topic is discussed in the Swiss AI Podcast on AI in the Mittelstand.
The simplest first step
Review every existing AI contract for exit costs and export options today. You will be surprised how rarely these are fixed in writing.
Conclusion: procurement discipline beats model loyalty
Revolut's 90 percent step is a signal, not an isolated case. Companies that buy AI as a commodity and place model selection behind an orchestration layer benefit permanently from falling token prices. The goal is not to predict the next model, but to ensure you are never dependent on a single one.
Frequently asked questions
- What exactly is model loyalty?
- Model loyalty means relying on the models of a single vendor over time instead of re-evaluating AI performance and price per task. This prevents you from automatically capturing falling token prices and increases lock-in risk.
- Is a self-hosted Chinese model sensible for my Swiss SME?
- For most SMEs, no. Self-hosting only pays off at high, stable volume and requires in-house hardware and operational expertise. For clearly defined, non-sensitive tasks a Chinese model can be valid; for revDSG or FINMA relevant data, stricter standards apply.
- What are typical switching costs for AI vendor lock-in?
- In Switzerland, typical exit costs range from CHF 8,000 to CHF 120,000 per scenario, depending on whether you use APIs only, a UI-bound platform, or a proprietary agent platform. Add 10 to 30 developer days for re-implementation.
- What is model routing and why does it reduce costs?
- Model routing means automatically selecting the right model for each task. It combines cheaper models for simple steps with stronger models for complex cases, so you never pay more than necessary.
- What Swiss alternative exists for sensitive data?
- Apertus from ETH/EPFL on the Swisscom Swiss AI Platform is a sovereign Swiss option. It is performance-limited but keeps data within the Swiss legal space and is valid for data sovereignty.
- How do I start a model-agnostic AI strategy?
- Start by reviewing your existing AI contracts and checking data exit clauses, notice periods, and export formats. Then move model selection into an orchestration layer and re-evaluate every six months.
Sources
Chris Jon Graf · AI Strategist & CEO
For over 15 years, Chris Jon Graf has built productive Swiss software platforms and advised executive teams on the strategic deployment of artificial intelligence. Would you like to evaluate this topic for your company?