First Sunday Rollout Report — the week's AI and tech news, summarized by someone who has to live with the consequences on Monday. The blog launched five days ago; the news graciously refused to slow down for me.
The headline: context windows hit filing-cabinet scale
Google shipped Gemini 3.5 Pro on July 17 with a 2-million-token context window and a Deep Think reasoning mode — about a month late after a reported rebuild, which by 2026 launch standards is basically on time. Two million tokens is a whole project binder in one prompt. My architect take: a bigger window is a bigger junk drawer unless you curate what goes in it. Retrieval, structure and SOPs still win; the window just changes how gracefully you fail.
The money: the picks-and-shovels quarter
TSMC reported a $40.2 billion Q2, up 36% year over year, with AI chips driving 61% of revenue. Meanwhile Thinking Machines Lab — Mira Murati's startup — raised $2 billion at a $12 billion valuation with no shipped product. I've had customers refuse a $200 change order with more due diligence than that. The infrastructure era continues: Microsoft is putting 3M's expanded-beam optics into Azure data centers, and NVIDIA launched a "Physical AI Initiative" with the Japanese government.
Quick hits
- Moonshot AI announced Kimi K3 — 2.8 trillion parameters, mixture-of-experts, up to 1M context, with open weights promised July 27. Remember when open weights meant a 7B you ran as a treat? The model treadmill now comes in trillion-parameter sizes.
- OpenAI bought Northslope, a forward-deployed-engineering firm. Translation: the model companies figured out that the last mile of enterprise AI is a person in your office — which is what I've been telling clients while standing in their office.
- Musk and Altman are fighting again, this time about Apple's lawsuit, and Google and Microsoft backed a rival agent protocol. The standards war has standards wars now.
2M tokens of context and the industry still can't remember last week. That's what this column is for.