The Biological Computing Company launched Mischief 1, an image-to-video model it says generates video 5x faster than Wan 2.2, the open-weight model it builds on. The model is live on fal at $0.01 per second of 720p video and $0.005 per second at 480p, or $0.05 for a five-second 720p clip. It runs on conventional hardware, with the software derived from measurements of living neurons. In Minecraft experiments on the open-source Oasis model, TBC’s research posts report that its Neural Dynamics Adapter added about 156k parameters to a 600M-parameter model for a roughly 19% gain on a differential-entropy proxy for image quality, and a later optimizer added about 14M parameters and lifted throughput from roughly two frames per second to nearly ten (fal)
Private Companies
Coherefundraise
Cohere closed the first tranche of a Series E targeting up to C$7B, about US$5B per Bloomberg, at a C$35.7B valuation. Germany’s Schwarz Group leads, joined by Bell Ventures, Bullhound Capital and the Government of Canada, which will invest up to C$1.6B for an undisclosed stake. Returning backers include Nvidia and Inovia Capital. Cohere committed to at least C$1.7B of Canadian computing infrastructure by 2030 and said annual recurring revenue grew 165% year over year through September (The Logic)
Public Markets
Intel guided fourth-quarter revenue to $15B–$16B, below the $16.5B it just reported, and shares fell more than 20% in late trading. Data center and AI revenue rose 78% to $7.3B, while PC chip revenue fell 5% (Intel)
Amazon reported AWS revenue of $45.98B, up 39%, and reiterated about $220B of 2026 capital spending, with trailing-12-month free cash flow at negative $42.4B. Shares fell about 6% more in after-hours trading (TechCrunch)
Emerging
Flash as phone memory: A paper from Zhuoran Li, Chen Gu, Zhuohang Bian and nine co-authors, NORmal: Architecting 3D Ferroelectric NOR Flash as Main Memory -- From Co-Fabricated BEOL Arrays to 200B-Class Multi-Agent LLMs in a Smartphone, proposes 3D ferroelectric NOR flash, a memory that keeps data without power and can be rewritten in place, as main memory on an unmodified LPDDR6 bus so a phone can run 200B-class models. The authors fabricated arrays with sensing under 7.5 ns. In serving simulations on agent traces, it runs 8.9x to 28.0x faster than UFS flash offloading and lands within 0.4% of an unlimited-DRAM ideal, and those system-level figures are simulated, not measured (arXiv)