OpenAI 计算主管谈 AI 基础设施建设OpenAI Compute Chief on AI Infrastructure Buildout
The Takeaway: AI 对计算资源的需求远远超过供应，OpenAI 等公司正在推动人类历史上最大的基础设施建设之一。
OpenAI 工业计算负责人 Sachin Katti 分享了数据中心建设的现实：这些是巨大的超级计算机工厂，将电子转化为 token，需要大规模液冷系统来处理芯片产生的高温。电力是主要瓶颈，公司不仅连接电网，还投资新建发电和传输基础设施，以避免从现有供应中抽取资源。
Katti 强调核能作为密集清洁能源的重要性，并讨论了 Jalapeno 等自定义硅芯片，以优化 inference 效率，实现更多 tokens per watt。推理已与训练计算相当重要，许多“训练”阶段实际涉及 inference。
他指出，尽管物理世界建设速度较慢，但需求持续超出供应，AI 甚至开始帮助设计自身芯片。“Anytime you have thought you have enough compute, we can slow down. Always negatively surprises.” 这反映了行业对 scaling 的坚定信念。
OpenAI 工业计算负责人 Sachin Katti 分享了数据中心建设的现实：这些是巨大的超级计算机工厂，将电子转化为 token，需要大规模液冷系统来处理芯片产生的高温。电力是主要瓶颈，公司不仅连接电网，还投资新建发电和传输基础设施，以避免从现有供应中抽取资源。
Katti 强调核能作为密集清洁能源的重要性，并讨论了 Jalapeno 等自定义硅芯片，以优化 inference 效率，实现更多 tokens per watt。推理已与训练计算相当重要，许多“训练”阶段实际涉及 inference。
他指出，尽管物理世界建设速度较慢，但需求持续超出供应，AI 甚至开始帮助设计自身芯片。“Anytime you have thought you have enough compute, we can slow down. Always negatively surprises.” 这反映了行业对 scaling 的坚定信念。
The Takeaway: Demand for AI compute far outstrips supply, with OpenAI and others driving one of the largest infrastructure buildouts in human history.
OpenAI Head of Industrial Compute Sachin Katti details data centers as massive supercomputer factories turning electrons into tokens, requiring extensive liquid cooling for hot chips. Power is the core bottleneck; companies invest in new generation and transmission to avoid draining the grid.
Katti highlights nuclear as dense clean energy and custom silicon like Jalapeno for inference efficiency, maximizing tokens per watt. Inference now rivals training, with much "training" actually being inference.
Despite slow physical world scaling, demand always exceeds supply, and AI is helping design its own chips. "Anytime you have thought you have enough compute, we can slow down. Always negatively surprises." This underscores the industry's scaling conviction.
查看原文 →
OpenAI Head of Industrial Compute Sachin Katti details data centers as massive supercomputer factories turning electrons into tokens, requiring extensive liquid cooling for hot chips. Power is the core bottleneck; companies invest in new generation and transmission to avoid draining the grid.
Katti highlights nuclear as dense clean energy and custom silicon like Jalapeno for inference efficiency, maximizing tokens per watt. Inference now rivals training, with much "training" actually being inference.
Despite slow physical world scaling, demand always exceeds supply, and AI is helping design its own chips. "Anytime you have thought you have enough compute, we can slow down. Always negatively surprises." This underscores the industry's scaling conviction.