DeAI Daily Brief — 9 September 2026

Today in DeAI: Mistral raised €3 billion at a valuation above €21 billion to push sovereign open-weight AI, Chutes passed 34 trillion cumulative tokens on Bittensor, and Together AI's serverless deprecation deadline lands in five days.

Mistral raises €3B for sovereign open-weight AI

Mistral announced a €3 billion Series D at a post-money valuation above €21 billion on September 8, framed around making "sovereign, open-weight AI the technology frontier." The raise is aimed at scaling training and inference capacity under European control of compute and infrastructure. Why it matters: this is the largest single bet yet on the sovereign open-weight thesis — that governments and enterprises will pay for frontier-class models whose weights and infrastructure stay inside their jurisdiction. The valuation and the sovereignty framing are Mistral's own claims; what would verify the thesis is shipped weights and signed sovereign contracts, not a funding round. (Mistral AI) — background: open-weight vs open-source models

Chutes passes 34T cumulative tokens on Bittensor

Chutes, the serverless inference subnet on Bittensor (Subnet 64), has processed over 34 trillion tokens cumulatively with daily peaks exceeding 120 billion, according to Crypto Briefing's reporting of network figures. Companion network Surplus Intelligence reportedly grew volume roughly 50x in two months. Why it matters: onchain inference request share is one of the few public usage signals for decentralized networks, and Chutes is consolidating its lead — but these are provider- and outlet-reported figures DeAI has not independently verified, and "onchain requests" measures settlement, not necessarily total served volume. (Crypto Briefing) — background: decentralized AI inference networks

Together AI deprecation deadline: five days out

Together AI's removal of four serverless models — openai/gpt-oss-20b, google/gemma-4-31B-it, thinkingmachines/Inkling-Small, and intfloat/multilingual-e5-large-instruct — lands September 14, with Qwen3.5-9B and GLM-5.3-Flash as the recommended replacements. The same September 1 changelog cut H100 80GB dedicated-endpoint pricing from $5.49 to $3.99 per hour. Why it matters: if your stack calls any of the four deprecated models, the migration window closes Monday; the replacement recommendations are Together's own, so benchmark them against your workload before the swap. (Together AI changelog) — related: Together AI alternatives

Meta's Muse agent asks for your email, calendar, and payments

Meta launched Muse, a personal AI agent for adults that requests access to users' email, calendars, payments, and health services — TechCrunch frames it as the company's biggest consumer AI privacy test yet. Why it matters: a mainstream agent requesting payments and health-data access on day one sets the consumer baseline for what "assistant" means, and it is the centralized-data-access counterpoint to the private and local inference patterns this publication covers. (TechCrunch) — background: what is private AI inference

Watching tomorrow

Whether Mistral publishes a model roadmap alongside the raise — the round is priced on open weights shipping, and the next weights release is the first real test of the €21B number.