The Investor

The most informative prices in this file came from parties who declined to transact — a denied customer, a walked-away acquirer, a blocked listing. That inverts the usual evidence hierarchy: a completed primary round is now the weakest form of price discovery in AI, because the marginal buyer is frequently optimising for supply-chain position, deployment pressure or ownership rather than return. First proof: The Information reports that Gimlet Labs' $300M round, led by Andreessen Horowitz, priced across three tranches — $2.5B, $3B, and an undisclosed "significantly higher" level. The $3B headline won because most of the dollars landed there, which makes it a weighted mode rather than a clearing price, and the first money in sits roughly 17% below it. The prior mark implied by a 16x step-up was about $187M, on a three-year-old company that left stealth in October. The tranche schedule and the denied anchor customer are in the deep dive below.

Samsung led €3B into Mistral at more than €21B post-money, per Bloomberg, while Anthropic reportedly walked away from a purchase of Decart — anonymous sourcing, no company confirmation. Your primary comps are being printed by a buyer whose return function includes device distribution and supply-chain position; your exit comps were tested by a capability-per-dollar acquirer and declined. Reported figures for the same round diverge — the reconciliation, the implied prior mark and the reported walk-away price are in the deep dive below.

China's securities regulator has told banks informally that humanoid robotics IPOs now need recurring revenue, a credible path to narrowing losses, or evidence of real innovation, The Information reports, after category leader Unitree's volatile debut. Meanwhile PitchBook counts $18.6B of robotics and physical-AI funding in Q2 2026 against J.P. Morgan's $2B of worldwide humanoid sales for full-year 2025 — a quarterly funding figure against a full-year revenue base, so read the ratio as directional rather than precise. The marginal buyer of humanoid equity was withdrawn administratively rather than repriced, so exit dates move before prices do.

CISA, the FBI and the NSA jointly named six Chinese AI companies — DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI — for industrial-scale distillation of Claude, GPT, Gemini and Grok. Separately, Harvey and Baseten reported that rebuilding the orchestration harness around a model lifted their legal diligence benchmark from 23% to 62% across models, with in-harness post-training adding comparable gains again. If you are paying for proprietary weights, you are paying for something that is now both a federal procurement variable and a smaller share of end performance than most memos assume.

TSMC committed to High-NA EUV high-volume manufacturing in 2030 on scanners priced up to $400M each, with only TSMC and Samsung reported able to write that cheque. That retires an assumption embedded in many AI models — that compute gets cheaper because hardware gets cheaper. Cost-per-token deflation now has to come from the software layer: serving, quantization, scheduling and compilation. Half-field exposure on current masks forces field stitching or chiplet partitioning until larger 6×12-inch masks reach production in 2033.