Inference host Fal doubled its valuation to $16B with a $1.3B round, while people familiar with the financials say rivals Baseten, Fireworks AI and Modal run thin gross margins because compute is so expensive (TechCrunch, Nov 3).
Inference host Fal doubled its valuation to $16B with a $1.3B round, while people familiar with the financials say rivals Baseten, Fireworks AI and Modal run thin gross margins because compute is so expensive (TechCrunch, Nov 3).
Moonshot's Kimi K3.1 Pro is the open-weight model to watch today: its weights were promised for November 4, but I could not confirm they are out yet. Companies that rent out AI compute are raising huge sums even though their margins look thin, and power is now a bottleneck for new data centers. Meanwhile, a 27B open model keeps running on an ordinary laptop.
One item today, and it is a follow-up: the most-watched open-weight release of the week is still unconfirmed on the day it was promised. If the weights do ship, it tests whether near-frontier scores come with downloadable weights (thesis point 2), though the model's size (about 2.8T parameters, inferred from K3) means only well-funded operators could self-host it.
Moonshot released Kimi K3.1 Pro in its API on October 23 at $2 input / $10 output per million tokens, with a 1M-token context, and promised open weights by November 4. As of the latest sources I could read (October 23-26), no Hugging Face repository existed, and I could not verify whether the weights went live today. Moonshot's own benchmark numbers are vendor-reported; the third-party Artificial Analysis index, as relayed by a secondary blog, puts it at 55 versus 58 for Opus 5.5, at roughly $2.11 versus $5.98 per index task.
Codersera (secondary, relaying Moonshot and Artificial Analysis) · Open-weight model, Follow-up · October 26, 2026
Two items point the same way for the thesis: the businesses selling inference are growing fast but reportedly earn little per dollar of compute, and the physical supply of power is limiting how quickly new capacity can come online. Both support the view that the cost of serving tokens is stubborn rather than falling away.
Fal, which hosts AI image, video and audio models, raised $1.3 billion in a round led by Bessemer, doubling its April valuation of $8 billion (per Bloomberg, as cited by TechCrunch). Fal reportedly passed $800 million in annualized revenue earlier this year. TechCrunch also reports, citing unnamed people familiar with the numbers, that competitors Baseten, Fireworks AI and Modal have thin gross margins because buying or leasing compute is expensive.
TechCrunch · Infrastructure & funding · November 3, 2026
Some AI data centers are generating their own power with retired aircraft engines converted to burn natural gas. A Rystad Energy analyst says such units can have a data center running in 24 months, versus about seven years to connect 100 MW to the grid in Northern Virginia (IEA figure cited in the article). The trade-offs are lower efficiency and more pollution, and xAI is reported to have installed eight converted turbines for its Colossus 2 site in Memphis.
TechCrunch (citing The New York Times) · Infrastructure & power · November 3, 2026
Today's items show local inference getting easier in practice (a 27B model on a laptop, faster tooling) while also showing its limits: very long reasoning runs are slow on consumer hardware. The tooling side keeps moving toward making local and verifiable open-weight serving routine.
Simon Willison ran Qwen 3.8 27B (a 17GB Q4_K_M quantization) in LM Studio on his laptop against a prompt from a new "Weight-Space Programming" paper. It identified Pac-Man from raw weights at 72% confidence, but took 46 minutes and 43,506 reasoning tokens. The paper's broader claims, made with a proprietary model, are disputed by commentators Willison cites. The practical takeaway is that local hardware can run this class of task, but at a latency only a patient single user would accept.
Simon Willison's Weblog · On-device inference · November 3, 2026
The Ollama v0.40.15 release candidate adds support for the MiMo V2.6 model and loads bundled CUDA libraries in-process on Linux for its MLX backend. The v0.40.14 release shipped the same day with faster top-k sampling on MLX for large vocabularies, among other fixes. Small, steady releases like these are how new open-weight models reach desktop users within days.
Ollama releases (GitHub) · Local tooling · November 3, 2026
Per the MarkTechPost summary (I could not open the full page, so this rests on the feed text), Anthropic released FlashProve, an Apache 2.0 zkSNARK prototype that proves a reply came from a committed model without revealing the weights. The demo proves Llama 4 Maverick at 110 tokens per second on one H100, with 19 MB proofs. If it holds up, it could let operators of open-weight models prove which model they are serving, which matters for on-prem and third-party hosting. Treat the figures as vendor-reported.
MarkTechPost · Open-weight tooling · November 3, 2026
Some commentators argue Apple, not Nvidia, is best placed for the shift to on-prem and on-device AI. In one such essay, the author says a single Mac Studio with 512GB of unified memory (M3 Ultra, priced at about $9,500 as reported there) can hold very large open-weight models that would otherwise need five or six RTX Pro 6000 cards, which the author prices at $60,000-$75,000 with roughly ten times the power draw. As capable open-weight models keep shipping, the bottleneck moves from who has the best model to who has affordable hardware to run it. The counterpoint is that Mac Studios suit single-user, high-memory inference, while datacenter GPU clusters still serve many users at once. Note that the essay also claims the 512GB configuration has been pulled from sale; I cannot verify current availability, so treat the configuration and price as the source's own figures.
Limited Edition Jonathan (Substack) · Perspective
Inference sellers are growing fast on thin margins and are constrained by power, which keeps the cost of serving tokens high (thesis 1). Meanwhile, open-weight releases like Kimi K3.1 Pro and ever-easier local tooling keep pushing capable models toward hardware people own (thesis 2 and 3), even if the biggest ones still need serious infrastructure.
Keynotes, masterclasses, panels and board-room sessions.