OpenAI算力负责人：我们建得不够快OpenAI Compute Chief: We Can't Build Fast Enough
The Takeaway：OpenAI最大的瓶颈不是需求，而是物理世界建算力的速度跟不上。
Sachin Katti是OpenAI工业算力负责人，背景包括斯坦福教授、多次创业、英特尔CTO。他强调，数据中心是把电子变成token的巨型工厂，必须液冷，因为芯片太热。OpenAI今年算力支出约500亿美元，整个行业约7000亿。他们不仅接电网，还投资发电、输电和变压器，甚至开始behind-the-meter自建电源（目前主要是燃气轮机），并强烈希望核电尽快落地。
Jalapeno是OpenAI的定制芯片，目标是最大化每瓦token数。因为他们知道自己的模型和负载，可以共设计硬件。AI已经在协助芯片设计，递归式“AI设计下一代AI所需系统”并不遥远。推理已占算力大头，但训练本身也大量用到推理（合成数据、后训练、test-time compute）。
对过建风险，Katti完全不担心：“每次我们以为算力够了可以放缓，结果都负面惊喜。我们最大的担忧仍是建得不够快。物理供应链、工厂、变压器加产能都太慢。”Stargate是整体算力策略的统称，组合微软、AWS、Google、Oracle、CoreWeave等，加上自建部分。他们还推出guaranteed capacity，让企业锁定token供应。
社区方面，数据中心给农村带来税收、就业和电网升级，水是闭环循环，净消耗极低。瓶颈无处不在：许可、燃气轮机、变压器、电工。轨道算力有趣，但要等发射和硬件经济性到位。
“Anytime you have thought you have enough compute, we can slow down. Always negatively surprises… Demand far outstrips compute supply today.”
Sachin Katti是OpenAI工业算力负责人，背景包括斯坦福教授、多次创业、英特尔CTO。他强调，数据中心是把电子变成token的巨型工厂，必须液冷，因为芯片太热。OpenAI今年算力支出约500亿美元，整个行业约7000亿。他们不仅接电网，还投资发电、输电和变压器，甚至开始behind-the-meter自建电源（目前主要是燃气轮机），并强烈希望核电尽快落地。
Jalapeno是OpenAI的定制芯片，目标是最大化每瓦token数。因为他们知道自己的模型和负载，可以共设计硬件。AI已经在协助芯片设计，递归式“AI设计下一代AI所需系统”并不遥远。推理已占算力大头，但训练本身也大量用到推理（合成数据、后训练、test-time compute）。
对过建风险，Katti完全不担心：“每次我们以为算力够了可以放缓，结果都负面惊喜。我们最大的担忧仍是建得不够快。物理供应链、工厂、变压器加产能都太慢。”Stargate是整体算力策略的统称，组合微软、AWS、Google、Oracle、CoreWeave等，加上自建部分。他们还推出guaranteed capacity，让企业锁定token供应。
社区方面，数据中心给农村带来税收、就业和电网升级，水是闭环循环，净消耗极低。瓶颈无处不在：许可、燃气轮机、变压器、电工。轨道算力有趣，但要等发射和硬件经济性到位。
“Anytime you have thought you have enough compute, we can slow down. Always negatively surprises… Demand far outstrips compute supply today.”
The Takeaway: OpenAI’s biggest constraint is not demand but the physical world’s inability to build compute fast enough.
Sachin Katti, OpenAI’s Head of Industrial Compute (Stanford professor, multi-time founder, former Intel CTO), describes data centers as giant factories turning electrons into tokens. They require liquid cooling everywhere because chips run extremely hot. OpenAI is spending roughly $50B on compute this year; the industry ~$700B. They connect to the grid but also fund new generation, transmission and substations, and are moving to behind-the-meter power (mainly gas turbines today) while urgently wanting nuclear.
Jalapeno, their custom silicon, optimizes tokens-per-watt by co-designing for known models and workloads. AI is already assisting chip design; full recursive “AI designs the systems for the next AI” is not far. Inference is already the majority of compute, and much of “training” is itself inference (synthetic data, post-training, test-time compute).
On overbuild risk Katti is unworried: “Anytime you have thought you have enough compute, we can slow down. Always negatively surprises… Demand far outstrips compute supply today.” The real risk is under-building. Stargate is the umbrella compute strategy spanning Microsoft, AWS, Google, Oracle, CoreWeave and self-build. They also launched guaranteed capacity so enterprises can lock in tokens of intelligence.
Data centers bring tax revenue, jobs and grid upgrades to rural areas; water use is a closed loop and minimal. Bottlenecks exist everywhere—permitting, turbines, transformers, electricians. Orbital compute is exciting to the engineer in him but needs better launch and hardware economics first.
查看原文 →
Sachin Katti, OpenAI’s Head of Industrial Compute (Stanford professor, multi-time founder, former Intel CTO), describes data centers as giant factories turning electrons into tokens. They require liquid cooling everywhere because chips run extremely hot. OpenAI is spending roughly $50B on compute this year; the industry ~$700B. They connect to the grid but also fund new generation, transmission and substations, and are moving to behind-the-meter power (mainly gas turbines today) while urgently wanting nuclear.
Jalapeno, their custom silicon, optimizes tokens-per-watt by co-designing for known models and workloads. AI is already assisting chip design; full recursive “AI designs the systems for the next AI” is not far. Inference is already the majority of compute, and much of “training” is itself inference (synthetic data, post-training, test-time compute).
On overbuild risk Katti is unworried: “Anytime you have thought you have enough compute, we can slow down. Always negatively surprises… Demand far outstrips compute supply today.” The real risk is under-building. Stargate is the umbrella compute strategy spanning Microsoft, AWS, Google, Oracle, CoreWeave and self-build. They also launched guaranteed capacity so enterprises can lock in tokens of intelligence.
Data centers bring tax revenue, jobs and grid upgrades to rural areas; water use is a closed loop and minimal. Bottlenecks exist everywhere—permitting, turbines, transformers, electricians. Orbital compute is exciting to the engineer in him but needs better launch and hardware economics first.