Revolut moves 2,000 engineers to self-hosted GLM, cutting coding token costs to about a tenth
- Revolut moved roughly 2,000 engineers onto self-hosted GLM for coding and pushed Claude Code and Codex to an exception list, according to David Abitbol, who leads AI ops there.
- The setup runs on 56 Nvidia B200s at a blended 30 cents per million tokens, roughly a tenth of what the company paid Anthropic and OpenAI, and engineering lead time improved 70%.
- The speedup appeared once tokens stopped being rationed: engineers quit reserving AI for hard tickets and pointed it at everything, which the poster frames as a pricing problem for labs heading to IPO.
- Hacker News commenters put the hardware and operating cost at roughly 2.5M to 3M dollars (power, cooling, updates included) and ask how many staff it takes to keep 56 B200s healthy, or whether the metal is rented instead.
- Commenters also argue self-hosting is driven by data control as much as price, citing financial rules that bar sending data to third parties outside national or EU jurisdiction, and note frontier-model access only gets expensive past the enterprise token-pricing threshold.
Hacker News 의견들
Honestly the difference is negligible for seasoned engineers. Maybe Claude is smarter if you're a junior, but if you know what you're doing GLM is enough. A few benchmark points less is nothing.
30 cents per million is wild. The number I want next is how many people it takes to keep 56 B200s healthy, that's usually where the savings quietly go.
Did the code get worse or just cheaper?
So that's maybe 2.5 to 3 million in hardware plus power, updates and cooling. If they were burning 500k to 1M a month on tokens it pays off in about a year, assuming the infra stays stable. Unless they're renting the metal, which changes the math entirely.
Unless you have very strict requirements, buying the hardware makes no sense. Most companies rent, and it's pretty hard to get the hardware in the first place.
If investors knew how powerful open models are, the AI company valuations would tank overnight.
You still need GPUs to run them. And with the news about OpenAI struggling to find investors, feels like it's already happening.
Open weight models are probably good enough for most software engineering now, but the goal posts will move. Frontier labs will find new areas that genuinely need the frontier.
Calling them frontier labs makes it sound like there's some secret sauce. There isn't, it's smart people on hard problems, and the Chinese open weight labs have those too. DeepSeek's inference architecture leads everyone. There's a real chance they end up being the frontier sooner than people expect.
We did the same thing on much weaker hardware and it's served us well. The bigger win is not having to worry whether inference providers can be trusted with our data long term. In my view the answer is absolutely not.
How do security conscious self-hosters sandbox their runtimes? Would love details on the setup people are running.
This really only matters once you pass 150 engineers, which pushes you into the enterprise bucket paying per token instead of the subsidized Claude Code plans. If you need 150 engineers on frontier accounts, maybe something else is wrong.
If you want zero data retention you have to go API only.
About 90% of the work we do goes to open source models. Ignore the cheap talk and define what you need precisely, then they follow instructions better than you'd expect over long timeframes. They're weak on vague prompts like fix all the problems.
The harness doesn't save you, you have to rework the prompts. Ambiguity and what we call the negative space really matter, you have to spell out and not y or z. With hardware work I have to state up front that the hardware is tested and fine, or the model will spend hours speculating about chip mask errors.