The compute buildout

Frontier AI runs on an unprecedented concentration of capital, chips and electricity, and this thread follows the bill. The hardware base was Nvidia’s Hopper H100; the money base was Microsoft’s reported $10 billion into OpenAI and Amazon’s up-to-$4-billion into Anthropic, the deals that tied each lab to a hyperscaler’s cloud. Nvidia touched a $1 trillion valuation on the demand.

The scale then went vertical. The Blackwell architecture succeeded Hopper; a data-centre power crunch drove Microsoft to restart Three Mile Island; and the Stargate Project pledged $500 billion for US infrastructure — announced days before DeepSeek’s R1 suggested far less compute might be needed, briefly wiping a record sum off Nvidia.

The spending only accelerated. Nvidia passed $4 trillion and then $5 trillion; OpenAI signed a $300 billion Oracle cloud deal and raised a $110 billion round; and the labs began designing their own silicon, OpenAI unveiling its Jalapeno inference chip. The recurring worry the thread raises is circularity — chipmakers investing in the labs that buy their chips — and whether the buildout is matched by the revenue to justify it.