Coding agents have a power bill.

The paper's Thursday account of OpenAI turning tax corrections into Codex training loops treated Codex as an enterprise feedback system. The earlier Cerebras frame said the IPO was about revenue durability and customer concentration, not chip romance.

Futurum's S-1 teardown says Cerebras has an OpenAI master relationship agreement tied to 750 megawatts of inference capacity. [1] OpenAI's Codex changelog shows the demand side: coding agents moving through release channels, permissions, and enterprise workflows. [2]

The link matters because agentic coding is not only software. It consumes low-latency inference, servers, electricity, cooling, and capital contracts. MSM can file Cerebras as Nvidia competition. X can argue about whether coding agents are useful. The ledger joins them: if the agent becomes normal work, compute becomes the constraint that decides how normal it can be.

That makes the IPO more than a chip ticker. It is a bet that software labor will demand dedicated inference the way cloud workloads demanded data centers. The coding assistant becomes a power contract when every pull request asks for another run.

-- THEO KAPLAN, San Francisco