黄仁勋:模型推理速度翻倍等于收入翻倍
黄仁勋算了一笔账:不改硬件、只靠 kernel 优化把推理提速 2 倍,token 卖得掉的话收入就翻倍,适合做 AI 业务的人看看这笔账怎么算。
NVIDIA CEO 黄仁勋提出,通过软件或 kernel 优化把模型推理速度提升 2 倍,在不新增 GPU、电力和机房投入的情况下,AI 实验室的收入可以翻倍。其逻辑是推理收入等于服务 token 数乘以 token 单价,更快的 kernel 在同样硬件上每秒产出更多 token。前提条件是需求持续超过供给,每个新增 token 都能按现价卖出去。
Jensen Huang says doubling model speed (e.g. with software or kernel optimization) doubles an AI lab’s revenue, which holds only while demand outruns capacity.
Inference revenue is tokens served multiplied by the price of each token.
Faster kernels raise tokens per second on the same GPUs, the same power, and the same building.
If every extra token finds a buyer at today’s price, a 2x speedup does double revenue with zero new capex. https://t.co/uBi358ndjP