Open Source AI — September 12, 2026
Four stories today, and they all sit on the same fault line: money is still pouring into AI's most expensive, centralized layer even as cheaper alternatives keep getting stronger.
In Plain English: Four stories today, and they all sit on the same fault line: money is still pouring into AI's most expensive, centralized layer even as cheaper alternatives keep getting stronger. Nvidia is reportedly weighing putting up to $10 billion into Anthropic's IPO, which could value the company near $2 trillion. At the same time, chip startup Positron raised $875 million to build inference chips that use ordinary computer memory instead of Nvidia's premium memory, claiming it can be dramatically cheaper to run big models. Japan's Sakana AI released a system that automatically routes AI tasks to whichever open model can handle them, undercutting Anthropic's own pricing by up to 60%. And Kimi's maker, Moonshot AI, says it's on track for $2 billion a year in revenue by giving its models away and charging for everything built on top. Today's Apple Angle ties Positron's memory-first chip bet directly to the Mac Studio argument.
Nvidia Reportedly Weighing Up to $10 Billion Anchor Stake in Anthropic's IPO at Near-$2 Trillion Valuation
Bloomberg-sourced reporting says Nvidia is in early talks to anchor as much as $10 billion of Anthropic's planned IPO, which could raise up to $100 billion and value the Claude maker at roughly $2 trillion — more than double its $965 billion valuation from May. Anthropic's annualized revenue run rate topped $65 billion by the end of July and the company is forecasting $190–200 billion in 2028 revenue, while also diversifying its chip supply with more than a million of Amazon's Trainium2 chips on order and a new in-house chip-design team. The listing is reportedly being timed for before November's U.S. midterm elections, though talks remain preliminary. For Lucien's thesis, this is prediction #1's other side showing up at enormous scale: the smartest money in the industry is still betting billions that today's centralized, GPU-heavy AI buildout will pay off, even as reports elsewhere show the underlying economics under mounting strain.
Benzinga (via Bloomberg) · Infrastructure & Economics · Sep 11, 2026
Chip Startup Positron Raises $875M to Challenge Nvidia With Memory-First Inference Silicon
Positron AI closed an $875 million Series C at a $5 billion valuation — five times what it was worth in February — to bring its "Asimov" inference chip to market. Instead of Nvidia's expensive high-bandwidth memory, Asimov pairs its processors directly with up to 18.4 terabytes of ordinary, consumer-grade LPDDR5X memory per system, which Positron says lets a full server rack handle models as large as 32 trillion parameters and process up to 26 times more tokens per dollar than Nvidia's top Blackwell GB300 system, according to the company's own simulations. Asimov is due to tape out on TSMC's 3-nanometer process by year-end, with mass production targeted for the second half of 2027. For Lucien's thesis, this is predictions #1 and #2 meeting in silicon: a well-funded challenger is betting that ordinary memory, not Nvidia's premium hardware, is the cheaper path to running huge models — the same memory-first logic the standing Apple Angle below has been making about the Mac Studio for months, just aimed at the data center instead of the desktop.
SiliconANGLE · Infrastructure & Economics · Sep 10, 2026
Sakana AI's New "Fugu" Models Route Around Expensive Closed AI — at 40–60% Less Than Claude Sonnet 5
Japanese lab Sakana AI released Fugu Max and Fugu Ultra v2, a pair of "orchestrator" systems that don't answer questions themselves but instead learn to route each incoming task to whichever model in a pool — including open-weight models and Nvidia's Nemotron family — can handle it most cheaply and capably. Fugu Max prices out at $2 per million input tokens and $6 per million output tokens, which Sakana says is 40–60% cheaper than closed rivals including Claude Sonnet 5 and GPT-5.6 Terra, and even the open-weight Kimi K3, while Fugu Ultra v2 targets harder reasoning tasks. For Lucien's thesis, this is predictions #2 and #3 wired directly into a shipping product: rather than picking one closed frontier model and paying its price, Sakana routes traffic toward whichever open or specialized model is cheapest for the job, treating closed frontier pricing as something to engineer around rather than simply pay.
MarkTechPost · Open-Weight Models · Sep 10, 2026
Kimi-Maker Moonshot AI Targets $2 Billion in Revenue — Proof Open-Weight Models Can Still Make Money
Moonshot AI, maker of the open-weight Kimi model family, is targeting $2 billion in annualized revenue by year-end — roughly double its August run rate — on the strength of its K3 model, which OpenRouter data shows generating as many as 300 billion tokens a day across the ecosystem, even as usage has cooled slightly in recent months. That figure still trails OpenAI's roughly $40 billion and Anthropic's roughly $65 billion run rates, underlining that giving away model weights for free is a real but comparatively modest business next to closed frontier labs. The report also notes Anthropic has accused Moonshot of harvesting more than 23 million responses from Claude Opus to help train Kimi, a distillation dispute that has become its own flashpoint. For Lucien's thesis, this is a useful reality check on prediction #3: open-weight models are proving there's genuine commercial life to be built around free weights, just not yet at the scale of the closed labs they're pressuring on price.
TechCrunch · Open-Weight Models · Sep 11, 2026
The Apple Angle (Standing Perspective)
The clearest thread connecting to today's news is Positron's new $875 million bet: a chip built around ordinary memory rather than Nvidia's premium HBM, on the theory that memory capacity — not raw compute — is what actually gates how big a model you can run cheaply. That's the same logic behind this recurring argument that Apple, not Nvidia, may be quietly best positioned for a shift toward local AI. Apple's Mac Studio uses a unified memory architecture shared between CPU and GPU, and a maxed-out configuration with 512GB of that memory — around $9,500 — can hold an entire trillion-parameter open-weight model at once. Matching that capacity with Nvidia's RTX Pro 6000 workstation cards takes five to six of them, roughly $60,000–$75,000 combined, while drawing something like ten times the power.
As frontier-capable open-weight models keep shipping every few weeks — this week alone bringing Sakana's cost-optimized routing and Moonshot's revenue milestone — the competitive question keeps drifting away from "which model is smartest" and toward "who can afford the hardware to run a nearly-as-good one locally." Apple's MLX software and Thunderbolt 5's RDMA networking already let multiple Mac Studios pool memory for bigger jobs, extending that advantage further, while Nvidia has stripped NVLink from its consumer workstation cards and its own DGX Spark increasingly borrows Apple's unified-memory playbook rather than the other way around. The honest caveat still applies: a Mac Studio serves one person running one model at a time, while the data centers behind stories like Nvidia's prospective Anthropic investment are built for enormous concurrent demand — Apple looks well 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 sit on both sides of Lucien's thesis at once. Nvidia's reported willingness to anchor up to $10 billion of Anthropic's IPO at a $2 trillion valuation shows the smartest money in the industry still betting enormous sums on the current, centralized way of building AI — the opposite of what prediction #1 expects. But everything else pulls the other way: Positron raised $875 million specifically because it believes ordinary memory beats Nvidia's premium hardware on inference economics, Sakana AI shipped a product whose entire pitch is routing traffic away from expensive closed models and toward cheaper open ones, and Moonshot proved a free-weight model family can already support a multi-billion-dollar business, even if a smaller one than the closed labs. None of this settles the argument on its own, but it captures the exact tension the thesis is tracking: capital keeps flowing toward the current centralized buildout even as the tools, the economics, and now the memory-first chip architecture for a cheaper, more distributed alternative keep maturing in parallel — precisely the setup the standing Apple Angle argues Apple is already positioned to benefit from.
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.