Open Source AI — September 23, 2026
Four stories today, all pointing the same direction.
In Plain English: Four stories today, all pointing the same direction. DeepSeek says it's betting $4.5 billion on Chinese-made Huawei chips instead of Nvidia just to keep training its biggest models, and Xiaomi quietly open-sourced a free model that now tops every other downloadable AI on a major independent ranking. Meanwhile California just made it noticeably more expensive to build the giant rented data centers this whole boom depends on. And legal-AI startup Harvey, whose profit margins went negative as usage exploded, swapped its paid subscription to a big AI lab for a fine-tuned open model and fixed the problem overnight - with several other startups doing the same. Today's Apple Angle connects the dots between cheaper open models and the hardware built to run them without a cloud bill.
DeepSeek Bets $4.5 Billion on Huawei Chips to Escape Nvidia Dependency
DeepSeek founder Liang Wenfeng told investors in a closed-door meeting that moving training off Nvidia is "one of the company's biggest strategic bets" and "must succeed," as US export controls push the lab toward Huawei's Ascend 950 series chips. DeepSeek is currently training a 2-trillion-parameter model and planning an 8-trillion-parameter successor, and estimates the job could take up to 200,000 Huawei chips versus roughly 50,000 Nvidia GPUs. For Lucien's thesis, this is a leading open-weight lab actively building its own supply chain around Nvidia rather than through it - prediction #3's shift away from the commercial GPU stack, driven by cost and access rather than preference.
Infrastructure & Economics · Android Headlines · Sep 23, 2026
Xiaomi Open-Sources MiMo-V2.6-Pro - Now the Top-Ranked Open-Weight Model on a Major Index
Xiaomi released MiMo-V2.6-Pro and Flash, omnimodal models handling text, image, video and audio with 1-million-token context windows, and Pro scored 46.32 on the Artificial Analysis Intelligence Index - ahead of Moonshot's Kimi K3 and Alibaba's Qwen3.8 Max, the best result yet for a freely downloadable model. On several agentic benchmarks it edges out Anthropic's Claude Opus 5 outright, though it still trails the newest closed flagships on aggregate. For Lucien's thesis, this is prediction #2's gap-closing pattern continuing on schedule - another open-weight release matching or beating a recent closed model on real, independently tracked benchmarks.
Open-Weight Models · SiliconANGLE · Sep 22, 2026
Harvey's Margins Went Negative -50% - So It Quietly Dropped Its Frontier-Lab Subscription
Legal-AI startup Harvey's gross margin collapsed from about 50% to roughly -50% by June, as a March agentic-AI update sent token usage up twentyfold while its flat per-seat pricing stayed put. Rather than raise prices, Harvey built "Harvey Tenet" by post-training Moonshot AI's open-weight Kimi K3, cutting the cost of a completed legal task from about $52 to $30 with no seat-price increase. Peers are moving the same way: Ramp is weighing its own model, Abridge is building on open weights, and Decagon already runs about 90% of its workload on open-source models. For Lucien's thesis, this is prediction #3 happening for spreadsheet reasons inside real, paying businesses, not open-source idealism.
On-Prem/On-Device Shift · THE DAILY BRIEF (beri.net) · Sep 22, 2026
California Signs Seven Laws Making Data Centers Pay Their Own Grid and Water Costs
Governor Newsom signed seven bills requiring data centers to cover their own grid-upgrade costs, meet clean-energy procurement standards, disclose water usage and fund related infrastructure, and lose blanket environmental-review exemptions in favor of local oversight. "We are ensuring that Californians remain in the driver's seat," Newsom said, "and that those profiting from data centers aren't doing so at our expense." For Lucien's thesis, this is a new, concrete cost and friction point layered directly onto the centralized buildout that prediction #1 says can't stay profitable as costs keep climbing - this time from state law, not just markets.
Infrastructure & Economics · Office of Gov. Gavin Newsom · Sep 21, 2026
The Apple Angle (Standing Perspective)
Every item above traces back to the same fork in the road: the rented, centralized way of running AI keeps adding cost and friction - a state now billing data centers for their own water and power, a leading Chinese lab spending billions just to route around Nvidia's export-controlled supply chain - while the models that can be owned outright keep getting better and cheaper, as Xiaomi and Harvey both showed today. Apple's answer sits on a desk rather than in a data center: configure a Mac Studio with its full 512GB of unified memory, about $9,500, and it holds an entire trillion-parameter open-weight model in memory, with no metering, no export license, and no state regulator involved.
Matching that on Nvidia's own workstation line takes five to six RTX Pro 6000 cards, roughly $60,000 to $75,000 combined, pulling something like ten times the power. As frontier-capable open-weight models keep shipping every few weeks - the same cadence that produced today's Xiaomi release - the competitive question keeps drifting away from who trained the best model and toward who can afford to own the box that runs a nearly-as-good one, locally. Apple's MLX framework and Thunderbolt 5's RDMA already let several Mac Studios pool memory into one larger machine, Nvidia has pulled NVLink from its own consumer workstation cards, and its DGX Spark borrows more from Apple's unified-memory approach than the reverse. The honest caveat still applies: a Mac Studio is built for one person running one model at a time, while the gigawatt-scale, now more heavily regulated facilities in today's California story exist to serve millions of concurrent users - Apple looks positioned to quietly own a valuable slice of AI compute, not replace the data center outright.
Limited Edition Jonathan (Substack) · Perspective
Why This Matters
Today's four items apply pressure from every angle at once - a leading open-weight lab paying billions to route around Nvidia, a state government adding real cost to centralized data centers, and two AI companies proving cheaper open models now work well enough to replace paid frontier access. That is prediction #1's cost pressure and prediction #2's capability convergence reinforcing prediction #3's shift toward ownership, all in a single day, with the standing Apple Angle pointing at the hardware built to catch it.
The Daily: Open Source AI - a recurring research brief for Lucien Engelen
Book Lucien for your next event
Keynotes, masterclasses, panels and board-room sessions.