Instinct wants a $10B valuation weeks after a $2.5B round
Two stories about what the AI race is actually buying: one company raising four times its own valuation to cover a compute bill it can't pass to users, and one chip vendor trying to make that bill smaller. Plus a paper showing what structural search finds when it stops reading sequences.
Instinct is negotiating to raise as much as $1 billion at a valuation of about $10 billion, according to The Information, with Sequoia and Benchmark in talks to lead a round that is not yet final and may still move on terms. That is roughly four times the $2.5 billion valuation the invite-only agent closed on in late August, when it raised $250 million to bring total funding to $350 million — we covered that round here: Viral AI agent Instinct raises $350M at a $2.5B valuation. The jump is being driven by supply, not by a new release: the same reporting has the company hitting a compute crunch as its user base pushes past 100,000, and Instinct reportedly wants to own its own chips and data centers eventually, because renting inference eats the margin it doesn't have — it still runs mostly on open-weight models rather than licensing from OpenAI or Anthropic, and it still doesn't charge users. A one-year-old product with no revenue and a fourfold markup in weeks is a bet that compute scarcity will keep rewarding whoever reserves capacity first; Perplexity and Character.AI both ran free-first under the same pressure before adding paid tiers, so "we don't want to charge" reads as a stage rather than a strategy — and Meta is building a direct rival, as we covered when Meta shipped Muse as a consumer agent inside a sealed VM.
Axelera AI has started shipping Europa, its second-generation inference accelerator, at 629 TOPS inside a 45W card, and it launches with validated servers from Dell and Supermicro already attached. The Eindhoven company sells the silicon three ways — bare chip for customers designing their own boards, a single-AIPU half-length Edge 232p card that is what ships in the Dell XE5 and Supermicro 111AD today, and a four-AIPU full-length Server 250p. The engineering pitch is that the card stops needing the host: eight second-generation AI cores on Samsung's 5nm process sit next to 16 RISC-V vector cores that handle pre- and post-processing on the chip, with an onboard video decoder so vision pipelines don't shuttle frames through system memory, backed by 128MB of L2 SRAM and 200GB/s of DRAM bandwidth. Axelera's own comparison charts claim 3x to 5x performance per dollar and 2x to 3x performance per watt on Llama 3 8B, Llama 3 70B and Llama 3.2 11B Vision; those are vendor figures built partly from competitors' published benchmark data, and the four-chip board does most of the winning. The target buyer is the regulated on-premises customer — banking, healthcare, legal, defense — that sovereign-AI money keeps circling, the same lane as Euclyd's €200 million memory-centric inference bet.
Researchers at the University of Miami's Sylvester Comprehensive Cancer Center used AI to compare the three-dimensional shapes of more than 214 million predicted proteins and pulled out a human protein that had been hiding in the dark proteome — TM184C, which looks like a G-protein-coupled receptor but doesn't behave like one. Published in Nature, the work searched by structure instead of sequence, which is why the family member survived decades of genomics: most of TM184C sits inside the cell, on vesicle membranes, and those vesicles ride microtubules out into the thin bridges cells grow toward each other to trade metabolites, vesicles and mitochondria. Knock TM184C down and the cell makes fewer of those conduits and its autophagy markers shift, which points at a role in how cells share resources under stress — including in glioblastoma, where bridged cells may be borrowing survival from neighbors. The detail that lands hardest: deleting the yeast equivalent broke those cells, and adding back the human protein fixed them, a function conserved across roughly a billion years.
What to watch: whether Instinct discloses a monetization plan alongside the round — the valuation is the headline, the price of compute is the story.
If a free AI assistant can't cover its own inference, is a $10 billion valuation a mistake or an early read on who owns the chips? Tell us in the comments.
