Primary sourceBlognone
View sourceEvent ·Multi-source
Revolut 自建 GPU 跑 GLM 替代 Claude Code,成本降 90%
英国银行 Revolut 工程师 David Abitbol 在 AI Engineer 大会上披露,公司把编程助手从 Claude Code 迁移到自建的 56 张 NVIDIA B200 GPU 上,运行智谱的 GLM-5.3(此前用过 GLM-5.1)。原因是 Claude Code 成本过高,曾达到每名工程师每月 1 万美元。自建方案成本为每百万 token 0.30 美元,约为 Claude Code 或 Codex 的十分之一,编写代码时间减少 60%。团队还通过 NVFP4 量化和将上下文窗口缩小到 256K,实现每秒 200 token 输出和 3 秒首 token 延迟,并用 model router 在 GLM 与 GPT、Claude 之间按任务自动切换。
Current read
Revolut 嫌 Claude Code 一人一月烧 1 万美元,直接买了 56 张 B200 自己跑 GLM-5.3,成本降到十分之一,还讲了量化调优的具体参数。
11 sources76° AI heatLast updated 2026/11/5 23:28:54
Evidence
11 sourcesRelated source 1rohanpaul_ai
View sourceRelated source 2elvis
View sourceRelated source 3VTV Công nghệ
View sourceRelated source 4CNBC Indonesia Tech
View sourceRelated source 5vLLM
View sourceRelated source 6IT之家
View sourceRelated source 7@atomic_chat_hq
View sourceRelated source 8Suhail
View sourceRelated source 9The Information
View sourceRelated source 10TestingCatalog
View sourceConflict check
Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.