Open Source AI — September 10, 2026
Four stories today, and they all circle the same idea: nobody wants to depend on just one AI provider or one chipmaker anymore.
In Plain English: Four stories today, and they all circle the same idea: nobody wants to depend on just one AI provider or one chipmaker anymore. OpenAI's finance chief admitted the company slashed prices on its cheaper "Luna" model by 80% after usage jumped tenfold, and openly compared its price to a free Chinese model, saying OpenAI now undercuts it. China's DeepSeek quietly tested a faster, cheaper new model for a single day before pulling it again. The legal AI company Harvey raised $550 million and, notably, released its own free-to-download model so law firms rely less on OpenAI and Anthropic. And a major Chinese cloud company said it's building a 100,000-chip supercomputer using homegrown Chinese chips instead of Nvidia's. Today's Apple Angle connects the dots: everyone from OpenAI's rivals to Chinese hyperscalers is racing to control their own stack, which is exactly the appeal building for Apple's Mac Studio line.
OpenAI Cuts Luna Model Price 80%, Then Brags It's Now Cheaper Than a Free Chinese Model
OpenAI CFO Sarah Friar disclosed that enterprise revenue grew 32% month-over-month even as the company slashed prices on its lower-cost Luna model by 80%, a cut that drove roughly a tenfold increase in usage. Most notably, Friar directly compared Luna's price to Z.ai's open-weight GLM-5.3 model, saying "if you're deploying Luna and compare that to GLM 5.3, on a cloud layer, we are cheaper" — an explicit admission that a free, downloadable Chinese model is now the price benchmark a leading closed lab has to beat. For Lucien's thesis, this is a rare moment where a closed provider names its open-weight competition out loud: OpenAI isn't just facing abstract margin pressure, it is actively cutting prices to stay under a specific open-weight rival's cost, which is exactly the dynamic predictions #1 and #2 describe feeding each other.
93.3 The Drive · Infrastructure & Economics · Sep 8, 2026
DeepSeek Quietly Tests a New Low-Cost, Multimodal Architecture — Then Pulls It After One Day
DeepSeek launched a one-day-only beta of "V4.1 Flash," described as an interim model built on a new architecture with native multimodal support, aimed at "stronger performance and faster generation at a lower cost" than existing options. Developers could reach it through a special API identifier that was set to expire the very next day, at the same pricing as the existing V4 Flash tier and capped at 20 concurrent requests per account — a low-stakes, low-commitment way to stress-test a new design before any full release. For Lucien's thesis, this is a small but telling data point for prediction #2: open-weight labs are iterating on cost and efficiency fast enough to run quick, disposable public experiments, the kind of rapid trial-and-error closed labs rarely show this openly.
TechNode · Open-Weight Models · Sep 9, 2026
Legal AI Startup Harvey Raises $550M — and Releases Its Own Open-Weight Model So Law Firms Depend Less on OpenAI and Anthropic
Harvey, whose AI tools are used by 80% of Am Law 100 firms and five of the Fortune 10's in-house legal teams, raised $550 million at a $15.5 billion valuation co-led by Diffusion and Lightspeed. Alongside the raise, Harvey released what it calls its first post-trained open-weight model plus "Harvey LAB," a new legal-agent benchmark — a deliberate move to build and control its own model rather than route everything through OpenAI's or Anthropic's APIs. For Lucien's thesis, this is prediction #3 playing out one layer up the stack: a well-funded AI application company, not just a model lab or a government, choosing to own an open-weight model as insurance against depending entirely on commercial API providers.
Harvey (company blog) · On-Prem/On-Device Shift · Sep 9, 2026
China's JD Cloud to Build 100,000-GPU Cluster on Homegrown Moore Threads Chips, Skipping Nvidia Entirely
JD Cloud and chipmaker Moore Threads announced plans for a 100,000-GPU cluster — a tenfold jump from their earlier 10,000-GPU partnership, and described as the first cluster of its kind hosted by a major domestic Chinese cloud provider. The cluster will run entirely on Moore Threads' own "full-featured GPUs," built for large-model training, inference, and robotics workloads, notable because Moore Threads sits on the US Entity List and the whole project is explicitly about proving Chinese-designed chips can operate at hyperscale without Nvidia. For Lucien's thesis, this is a hardware-side echo of prediction #3: a major cloud provider making a deliberate, large-scale bet on chip independence rather than continued reliance on one dominant supplier — the same instinct toward self-hosted, self-controlled infrastructure the thesis is tracking on the model-weights side.
Data Center Dynamics · Infrastructure & Economics · Sep 9, 2026
The Apple Angle (Standing Perspective)
Today's four stories are all, in their own way, about escaping dependency — on one AI vendor's pricing, on one chipmaker's supply chain, on one company's API. OpenAI now measures itself against a free Chinese model's price; Harvey would rather own a model than rent one indefinitely; JD Cloud and Moore Threads want a 100,000-GPU cluster that owes nothing to Nvidia. That same instinct, pointed at the hardware layer, is where Apple's Mac Studio line keeps sitting quietly. A Mac Studio configured with up to 512GB of unified memory, shared between CPU and GPU, costs around $9,500 and can hold a trillion-parameter open-weight model — the kind DeepSeek, GLM, and Kimi keep shipping — entirely in memory, with no data-center lease, no per-token bill, and no dependency on any single chip supplier's roadmap. Reaching that same memory footprint with Nvidia's own RTX Pro 6000 workstation cards takes five or six of them, roughly $60,000–$75,000 combined, at something like ten times the power draw.
Frontier-capable open-weight models keep arriving every few weeks, and Apple's MLX framework plus Thunderbolt 5's RDMA networking already let multiple Mac Studios pool memory for larger jobs, which is a big part of why the competitive question keeps sliding from "who has the best model" toward "who can afford to run a nearly-as-good model without betting their whole business on one vendor." Notably, even a Chinese hyperscaler chasing chip independence and a $15-billion legal AI startup building its own model are, in their own contexts, chasing the same kind of independence Apple's architecture offers almost by default. The fair caveat still applies, as always: a Mac Studio serves one person running one model at a time, while the infrastructure behind stories like JD Cloud's cluster is built to serve enormous concurrent demand — Apple looks positioned to dominate a specific, high-value slice of AI compute, not to replace the data center outright.
Limited Edition Jonathan (Substack) · Perspective
Why This Matters
Today's four items all point toward the same instinct, showing up at every layer of the stack: OpenAI naming a free Chinese model as its price benchmark, DeepSeek casually test-driving a cheaper new architecture for a single day, Harvey choosing to own a model rather than keep renting one, and a Chinese cloud giant building a 100,000-GPU cluster specifically to avoid Nvidia dependency. None of these alone proves the shift to on-prem and on-device AI has arrived, but together they show the same behavior recurring across very different players — labs, application companies, and hyperscalers alike are all hedging against dependency on a single vendor, whether that vendor sells models or chips. That is precisely the environment in which the standing Apple Angle argues Apple's unified-memory hardware is quietly becoming the default answer for whoever wants that independence without building a data center of their own.
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.