Open Source AI — September 2, 2026
Today's three stories all land on the money side of Lucien's thesis.
In Plain English: Today's three stories all land on the money side of Lucien's thesis. A widely watched index of what AI companies actually charge per token hit a record low on Monday — prices have more than halved since summer — squeezing OpenAI and Anthropic's margins just as both are quietly prepping IPOs. Separately, Anthropic's new $35 billion cloud deal with an Nvidia-backed provider is drawing scrutiny for being financially circular: Nvidia leases the data center, a Nvidia-backed company runs it, and Anthropic foots the bill, with the real numbers undisclosed. And in China, Zhipu (Z.ai) reported revenue up almost 400%, nearly all of it from renting out its own open-weight models over the cloud — proof the open-weight gap is closing, but also that most of that usage still runs through a rented API, not a local machine. It's a day about strain in the rented-compute model, exactly where Lucien expects the pressure to show first.
AI Token Prices Hit a New Record Low, Squeezing OpenAI and Anthropic Right Before Their IPOs
A widely tracked index of average AI token prices fell to 97 cents per million tokens on Monday — less than half its summer peak — pushed down by cheaper Chinese open-weight models like Kimi K3, OpenAI's own price cuts, and falling production costs. Analysts describe the effect as "token deflation": revenue per query keeps shrinking while the compute commitments behind it stay fixed, a squeeze landing right as both OpenAI and Anthropic have filed confidentially for IPOs. For Lucien's thesis, this is close to the headline prediction — competition, including open-weight competition, pushing the price of a token down toward the point where the rented-API business model stops penciling out.
Yahoo Finance · Infrastructure & Economics · Sep 1, 2026
Anthropic's $35 Billion Cloud Deal Draws Scrutiny Over "Circular AI" Financing
Anthropic signed a $35 billion capacity deal with Nvidia-backed cloud provider Lambda, which will run a Texas data center that Nvidia itself is leasing from bitcoin miner Hut 8 — the same week a similar $45 billion Anthropic commitment surfaced with another Nvidia-backed provider, Nscale, in West Virginia. Analysts flagged the structure as "circular AI financing": Nvidia's capital effectively helps fund the infrastructure that buys Nvidia's own chips, while the underlying economics — what Lambda pays Nvidia, what Nvidia pays Hut 8 — stay undisclosed. For Lucien's thesis, this is precisely the kind of opaque, interlocking financing that makes it hard for outsiders to verify whether the buildout is genuinely profitable or just recirculating capital among a small group of players.
The Deep Dive · Infrastructure & Economics · Sep 1, 2026
China's Zhipu (Z.ai) Reports Revenue Up Almost 400%, Nearly All From Renting Out Its Open-Weight GLM Models
Zhipu AI, maker of the open-weight GLM model family, reported first-half 2026 revenue of about $142 million, up nearly 400% year-over-year, with its "open platform and API" business — cloud rental access to its own open-weight models — surging roughly 2,700% and pushing annualized revenue past $1.6 billion. Losses narrowed even as R&D spending rose by a third. For Lucien's thesis it cuts two ways: it's more evidence a Chinese open-weight lab is closing in on the frontier commercially, but it's also a reminder that most real-world use of that model today still runs through a rented cloud API rather than a local machine — model quality is arriving faster than the on-prem shift is.
South China Morning Post · Open-Weight Models · Aug 31, 2026
The Apple Angle (Standing Perspective)
Today's stories illustrate why this standing argument keeps pointing at Apple's hardware rather than Nvidia's. If token prices are cratering because the rental-API business is under margin pressure (today's first item), and multi-billion-dollar cloud deals now need increasingly circular financing just to keep GPUs full (today's second item), the case for skipping the rental relationship altogether only gets stronger. The core numbers haven't moved: a Mac Studio configured with 512GB of memory shared between CPU and GPU costs around $9,500 and can hold a trillion-parameter open-weight model entirely on one desktop machine. Matching that memory footprint with Nvidia's RTX Pro 6000 workstation cards takes five or 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, the argument goes that the competitive question stops being "who has the smartest model" and becomes "who can afford, and physically power, a machine to run it locally" — a race Apple's unified-memory design currently has few serious rivals in. Apple's MLX framework and Thunderbolt 5's RDMA networking already let multiple Mac Studios pool memory for even larger models, and Nvidia's own DGX Spark, a small unified-memory box of its own, is effectively a concession that this is the right shape for local inference — notably after Nvidia stripped NVLink pooling out of its consumer and workstation cards. The fair caveat still holds: a Mac Studio is built to serve one person running one model, not the thousands of concurrent users a data-center GPU cluster serves — so this looks like Apple leading a specific, valuable segment of AI compute, not replacing shared infrastructure outright.
Limited Edition Jonathan (Substack) · Perspective
Why This Matters
Today's three items sit mostly on the profitability half of Lucien's thesis rather than the model-capability half. A widely tracked token-price index hitting new record lows squeezes exactly the margin OpenAI and Anthropic need heading into their IPOs; Anthropic's own $35 billion cloud deal is drawing scrutiny for the kind of circular, opaque financing that makes independent profitability hard to verify; and even a fast-growing open-weight lab like Zhipu is, for now, monetizing mostly through a rented API rather than on-device or on-prem deployment. None of this alone proves the on-prem transition is imminent, but it sketches the exact pressure point the thesis expects: as token prices compress and the financing behind the infrastructure buildout gets more convoluted to sustain, the economic case for running open-weight models locally — the case the standing Apple Angle makes every day — gets a little stronger.
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.