Revolut says it cut its AI coding costs by about 90% by self-hosting a Chinese model
This Week's Briefing
Full episode out now, find us wherever you get your podcasts. Upside Podcast - European venture, every week.
Worth a scroll.
Read the episode notes: Upside 115 - EU Inc's 28th Regime, the US Midterms, and Anthropic's IPO
Key Takeaways
- Revolut’s engineers had to ration AI tokens
- 56 Nvidia B200 GPUs serve around 2,000 engineers
- Without token limits, engineers used AI for everyday tasks
The Bigger Picture
- The figures are Revolut’s own and haven’t been audited
- Frontier models still handle some of the work
- In SuperSeed’s tests, GLM wasn’t the best value for any workload
Why it matters
“More on AI costs, the US midterms and EU Inc in the full episode”