Goldman: 中美 AI 竞争走向规模对效率
阅读原文： https://t.co/5NJcVMFmPQ
Goldman 拆了组数据：美国 2026 年 AI 资本开支约是中国 7 倍，但 DeepSeek V4.1 Flash 把 KV-cache 省了 437 倍，看效率派怎么打。
Goldman Sachs 估算 2026 年美国主要云厂商 AI 资本开支约 8060 亿美元，中国约为 1100 亿美元，绝对算力差距仍然很大。中国实验室靠效率竞争，DeepSeek V4.1 Flash 把每 token 的 KV-cache 需求相对 V1 降低了 437 倍。中国模型在 OpenRouter 上的 token 份额持续上升，并在智能体和编程等高价值工作负载中扩大占比。随着低价中国模型被广泛采用，混合 LLM token 价格不断下降，给全球推理定价带来压力。
阅读原文： https://t.co/5NJcVMFmPQ
阅读原文： x.com/Jadzo1_/status… J.D. @Jadzo1_ GS: US vs China AI (Scale vs Efficiency) → The US is still massively outspending China on AI infrastructure. Goldman estimates major US cloud-provider capex at roughly $806bn in 2026E vs ~$110bn for China, so the absolute compute gap remains huge. → China is competing through efficiency. DeepSeek V4.1 Flash has reduced KV-cache requirements per token by 437x vs V1, highlighting how aggressively Chinese labs are optimising inference economics. → That efficiency is translating into adoption. Chinese models have been gaining OpenRouter token share and repeatedly occupying leading positions in usage rankings. → The gains are showing up in more valuable workloads. China models are increasing share across agents and coding, not just general chatbot usage. → Pricing is becoming a competitive weapon. Blended LLM token prices have been falling as lower-cost Chinese models gain adoption, putting pressure on inference pricing globally. → The next battleground is agents and digital labour. Goldman sees virtual-worker use cases expanding from coding into finance, legal, accounting and broader white-collar workflows. 🔗 View Quoted Tweet 💬 1 🔄 0 ❤️ 0 👀 436 📊 1 ⚡
- Geek01:17原文
