Daily Brief

Review is the bottleneck once agents write most of the code, and OpenAI is buying capacity there rather than only improving the base model. It also puts OpenAI in direct acquisition competition with Meta for coding-agent talent.

Parallel agent orchestration is becoming a core product surface, and Meta wants it in-house instead of depending on third-party tools. Expect more small orchestration startups to be priced as acqui-hires.

Hyperscalers are moving capacity to jurisdictions that will approve it and can supply power. The Malaysia commitment fits a pattern of AI compute spreading across Southeast Asia.

Local opposition is now a measurable constraint on US compute buildout, not just a permitting nuisance. Developers will need to price in more siting risk, and more capacity will go where approval is easier.

This is hard transaction data from a large marketplace, and it shows agent-mediated commerce is near zero today. The "agents disintermediate marketplaces" thesis has no volume behind it yet.

A 65B open-weight model aimed at agentic coding would be cheap to self-host for teams building coding agents. It arrives alongside Kimi K3.1 Pro weights, so the open-weight field is getting crowded.