Moonshot AI

Frontier AI and the Kimi Delusion

The article analyzes market reaction to the release of Kimi K3, a 2.8-trillion-parameter open-weight model from Moonshot, and challenges the narrative that the model alone caused a broader sell-off. It notes a sharp market decline on July 17 — including drops in TSMC, SoftBank, and the Nasdaq — which traders labeled the "Kimi Moment." The author argues this attribution is incorrect on two structural grounds: (1) the visible catalyst (Kimi K3) is an attribution substitution for longer-running geopolitical damage to the chip supply chain, and (2) the bearish interpretation is inverted because cheaper open models are more likely to expand demand, push closed labs outward, and do not necessarily reduce hardware capex needs. The piece frames the market narrative as the "Kimi Delusion."

Alphabet developing efficient 'Frozen v2' AI server chip

Alphabet (Google) is developing an internal server chip called "Frozen v2" to accelerate its Gemini models by embedding parts of the model's architecture (the structure, not fixed weights) into silicon — a concept associated with Google/DeepMind chief scientist Jeff Dean. The design aims to reduce computation and data movement; engineers estimate it could serve roughly six to ten times more tokens per unit of power than current TPUs. Alphabet plans initial small-scale internal testing around 2028 and sees Frozen v2 as a specialized complement to, not a replacement for, its TPU lineup. The project is intended to relieve internal compute shortages that have forced Google Cloud to turn away business and incur large external costs. Reporting came via The Information/t3n; Google declined to confirm the report, saying it routinely researches hardware/software co-design.

The Best Model Loses

The newsletter discusses the democratizing power of AI for personal projects, experiments the author ran to test new models, and several industry developments. Key items: the federal government reviewed Anthropic’s Fable and OpenAI’s GPT-5.6 Sol; Thinking Machines released Inkling, a 975B-parameter open-weights model intended to drive revenue via a fine-tuning platform called Tinker, with estimated pretraining costs of $10M–$20M (pretraining compute only); Meta announced a Hyperion data center expansion to 5 gigawatts costing more than $50 billion and faces a market narrative of becoming a compute “landlord,” with reports Anthropic is in early talks to lease up to $10 billion of compute from Meta; SpaceX’s Colossus deal with Anthropic is cited as roughly $45 billion over three years with short termination windows; and MLB issued a mid-season ban on generative AI in dugouts. The piece mixes analysis of business models, infrastructure, and short-term market dynamics.