📄 中文摘要
运行自己的大型语言模型（LLM）在本地不仅更困难，而且比使用像Anthropic或OpenAI这样的云服务提供商要昂贵得多，甚至在忽略明显的服务器成本的情况下也是如此。许多团队预算5万美元购买单台服务器，却发现每月的电费高达800美元，未考虑冷却、维护和更新模型所需的实际时间。而使用Mac Mini集群的工作量则减少了99%。'我想要完全控制'的论点听起来不错，但当你意识到20万美元的服务器农场在不断消耗资金，而云API每千个令牌只收取0.005美元时，情况就变得不一样了。
离线大型语言模型的成本远超你的想象（真实的计算）
出处: Offline LLMs Cost More Than You Think (Here's the Real Math)
发布: 2026年3月1日
📄 English Summary
Offline LLMs Cost More Than You Think (Here's the Real Math)
Running your own large language model (LLM) on-premises is not only more challenging but also significantly more expensive than utilizing cloud providers like Anthropic or OpenAI, even when ignoring the obvious server costs. Many teams budgeted $50,000 for a single server, only to find that their monthly electricity bill for that machine alone was $800, not including cooling, maintenance, or the actual time required to keep the model updated. A Mac Mini cluster would have involved 99% less work. The argument for 'full control' sounds appealing until one realizes that a $200,000 server farm is draining resources while a cloud API charges just $0.005 per 1,000 tokens.