用「每用户 GPU 小时」为 LLM 推理定价是否合理？
来源标题：Is "GPU-hour per user" a sane way to price LLM inference?
文章质疑以「每用户 GPU 小时」作为 LLM 推理定价指标是否合理。该指标把 GPU 使用时长分摊到用户身上，但推理成本还受模型大小、批处理效率、上下文长度等因素影响，单看 GPU 小时可能无法反映真实开销。
以下正文同步自 Hacker News 最新，版权归原站所有，已转换为易读排版。
We host and run open-weight models on our own GPUs, and sell that capacity the way broadband is sold: you pay for the window, never per unit. From one developer's day pass to a private AI factory in your data centre.
A flat time pass with unlimited tokens on open models, for developers and their coding agents. Self-serve, in minutes.
Open models for your whole workforce, billed per seat by time, never per token. For organisations from 60 seats, on quarterly or yearly agreements.
Infinite tokens for a finite time. An AI platform for scoped model engineering and serving: training clusters and inference reserved for your work on your own endpoint, <company>.infinitetokens.ai, for 3, 6, 9 or 12 months. Billed for the platform and the term, never per token.
An AI factory in your own data centre, on your hardware or ours, on your domain. We design it, bring it live and run it. Nothing leaves your walls.
Metered APIs bill every token, so every agent loop and every experiment becomes a cost decision. We sell capacity by time: the cost is agreed before you start, and your people and systems stop rationing their thinking.
Unlimited tokens or per-token pricing: how to choose
Every model we offer runs on GPUs we host and operate. We do not proxy your requests to another provider and we do not resell anyone else's capacity. Frontier proprietary models stay with their labs.
Any OpenAI- or Anthropic-compatible client, agent, IDE or script.
Keys, sessions, admission and queueing, routing.
Open-weight models, parallelism, speculative decoding, prefix cache.
Hosted and operated by us, or in your data centre.
MIRASTACK observability and root cause analysis
Every agreement is shaped to the workload: seats, sessions, capacity, term and location.
正文由 探How 从来源站点 RSS 同步，内容未经改写；遇到排版缺失或需要图片、视频时请以原文为准。