Open Source AI — September 8, 2026
Three stories today, all about the scale and opacity of money and hardware behind closed, rented AI models — nothing about a new open model or new hardware for running one yourself.
In Plain English: Three stories today, all about the scale and opacity of money and hardware behind closed, rented AI models — nothing about a new open model or new hardware for running one yourself. Nvidia's CEO revealed that OpenAI's newest model was trained on more than 100,000 of Nvidia's top chips, and four times that many are already being wired up for whatever comes next. Separately, people considering buying into Anthropic's upcoming stock listing are reportedly demanding a very specific, hard-to-fake number: exactly how much real revenue the company earns per unit of computing power it buys, not just headline growth. And a little-known company called FluidStack, which builds AI data centers but doesn't own a single computer chip itself, just tripled in value to $18 billion almost entirely because it's building custom buildings for Anthropic. All three point the same direction: the compute committed to closed, centralized AI keeps compounding, and the people funding it are starting to demand harder proof it will pay for itself.
Nvidia's Jensen Huang: GPT-6 Astra Trained on 100,000+ GPUs, With 400,000 More "Coming Online Next"
Nvidia CEO Jensen Huang said on X that OpenAI's newest model, GPT-6 Astra, was trained on more than 100,000 Nvidia Grace Blackwell NVLink72 chips, declaring "AGI has arrived" — while also revealing that 400,000 additional GPUs, roughly quadruple that count, are already being deployed for whatever OpenAI trains next. OpenAI co-founder Greg Brockman was more measured about the AGI claim itself, calling the threshold "this blurry thing." For Lucien's thesis, the number that matters isn't the AGI talk, it's the hardware trajectory: each new closed frontier model appears to need not just more compute than the last one, but several multiples more, with still no public accounting of whether the resulting revenue covers it.
PC Gamer · Infrastructure & Economics · Sep 7, 2026
Anthropic's Would-Be IPO Investors Are Reportedly Demanding "Revenue Per Token" and "Revenue Per Gigawatt" Numbers
Ahead of a listing that could value the company near $965 billion, prospective Anthropic IPO investors are reportedly pushing for unit-economics disclosures well beyond the usual top-line growth figures — specifically, revenue earned per token served and per gigawatt of compute operated, numbers that would make it far harder to dress up margin pressure as growth. For Lucien's thesis, this is close to the profitability question showing up in its most concrete form yet: investors with real diligence and real money on the line don't appear satisfied with growth alone, and want the exact unit economics that would prove — or disprove — that scaling compute actually scales profit.
CryptoBriefing · Infrastructure & Economics · Sep 4, 2026
FluidStack, the Data-Center Builder That Owns Zero Chips, Triples to an $18 Billion Valuation on Anthropic's Buildout
FluidStack — a company that designs and operates AI data centers but deliberately owns no GPUs itself, leaving the actual silicon to its customers — raised $1.5 billion led by Jane Street Capital at an $18 billion valuation, roughly triple where it stood earlier this year, driven largely by its role as Anthropic's primary infrastructure partner on a $50 billion capacity commitment. Its revenue is projected to jump from about $66 million in 2024 to $660 million this year. For Lucien's thesis, FluidStack is a clean illustration of how much specialized capital is now stacked up behind a single AI lab's rented-compute bet — the same bet that, per today's other item, Anthropic's own prospective investors are now asking it to justify token by token.
Tech Times · Infrastructure & Economics · Sep 5, 2026
The Apple Angle (Standing Perspective)
Today's stories are all about scale on the centralized side — a single model's training run now consuming enough Nvidia hardware that its own supplier's CEO casually mentions "quadrupling" it for the next one, and an $18 billion infrastructure company whose entire business is building custom facilities for exactly one customer's chips. None of that changes the case this brief keeps making about the other side of the ledger. A Mac Studio configured with up to 512GB of unified memory, shared between CPU and GPU, costs roughly $9,500 today and holds a trillion-parameter open-weight model entirely in memory — no data-center lease, no dedicated infrastructure partner, no multiplying GPU count required. Reaching that same memory footprint with Nvidia's own RTX Pro 6000 workstation cards takes five or six of them, somewhere around $60,000–$75,000 combined, while drawing roughly ten times the power of the single Mac Studio.
The Anthropic IPO story sharpens why that gap matters right now: when the company at the leading edge of the rented-compute model is being asked by its own prospective investors for hard revenue-per-token and revenue-per-gigawatt numbers before they'll commit capital, that is effectively the market pricing in the same doubt this brief has tracked all along — that scaling compute doesn't automatically scale profit. Frontier-capable open-weight models keep shipping every few weeks regardless, and Apple's MLX framework plus Thunderbolt 5's RDMA networking already let multiple Mac Studios pool memory for even larger jobs, which is part of why the competitive question keeps drifting from "who has the best model" toward "who can afford to run it without a $50 billion infrastructure partner of their own." The fair caveat still applies: a Mac Studio serves one person running one model at a time, while the compute behind stories like today's GPT-6 Astra buildout is built to serve enormous concurrent demand — Apple looks positioned to win 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 three items all sit on the profitability half of Lucien's thesis rather than the model-capability half, but together they sketch an unusually clear picture of where the pressure is building. GPT-6 Astra's training run and its already-planned quadrupling show the pace at which compute committed to closed, rented models keeps compounding; FluidStack's tripled valuation shows how much specialized capital has stacked up to build the physical buildings behind that commitment, almost entirely reliant on one customer's contract; and Anthropic's own prospective IPO investors reportedly asking for revenue-per-token and revenue-per-gigawatt disclosures show that the people with the most at stake are no longer satisfied with growth alone as proof the model pays off. None of this proves the on-prem transition has arrived, but it sharpens the exact question the thesis keeps returning to — and the standing Apple Angle argues that whoever eventually needs a cheaper, fixed-cost alternative to that math will find Apple's Mac Studio already quietly waiting.
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.