Open Source AI — September 16, 2026
Three stories today, and two of them land on the same nerve.
In Plain English: Three stories today, and two of them land on the same nerve. MIT Technology Review published a detailed look at the math behind the AI spending boom - the industry needs a 2.7x jump in productivity gains by 2030, and $3.7 trillion a year in revenue by 2032, just to justify what's being spent, with one Wharton economist calling the risk "the largest misallocation of capital in history." Separately, Anthropic disclosed it has now locked in roughly $517 billion in future compute commitments - nearly triple what it had told investors before - and even that reportedly still isn't enough to meet demand. On the open-weight side, China's state-backed Shanghai AI Lab quietly released a 744-billion-parameter agentic model under a fully permissive MIT license, the kind of terms neither OpenAI nor Anthropic has ever offered for a flagship system. Today's Apple Angle weighs a much smaller, calmer hardware bet against both of those trillion-dollar wagers.
MIT Technology Review: What Must Happen for AI's Trillion-Dollar Gamble to Pay Off
A detailed MIT Technology Review analysis lays out just how large the gap is between AI infrastructure spending and AI revenue. Hyperscalers are on pace to spend over $750 billion this year, rising past $1.1 trillion annually by 2027 and up to $5 trillion cumulatively through 2029, while AI product revenue across the industry sits at only $150-200 billion today. To break even by 2030 the piece estimates the industry needs roughly a 2.7x jump in economy-wide productivity gains, and by 2032 needs about $3.7 trillion in annual AI revenue just to satisfy what investors are expecting. Wharton's Jessica Wachter is quoted warning that without those gains, "the current buildout will be the largest misallocation of capital in history," while former SEC chair Gary Gensler notes more plainly that "the spending does not have commensurate revenues yet." The piece also flags that Alphabet - one of the industry's strongest balance sheets - posted its first negative free cash flow since Google's 2004 IPO. For Lucien's thesis, this is prediction #1 laid out with unusually specific numbers: a mainstream, technically literate outlet is now publishing the exact productivity multiple the industry needs to hit, not just gesturing at general skepticism.
Infrastructure & Economics · MIT Technology Review · Sep 15, 2026
Anthropic Locks In $517 Billion in Compute Commitments - and Reports Say It's Still Not Enough
Anthropic disclosed that it has now secured roughly $517 billion in compute commitments extending mostly through the next decade, nearly triple the $180 billion figure it had previously given investors through 2029. The deals span Amazon and Google (a combined roughly $300 billion for about 11 gigawatts, blending custom chips with Nvidia GPUs), Microsoft (at least $30 billion for Azure capacity), SpaceX (up to $45 billion), AMD's newer MI450 chips, and neocloud providers including CoreWeave and Fluidstack. Despite the size of the number, reporting characterized it as still falling short: Nvidia's own figures imply roughly $40 billion in annual revenue opportunity per gigawatt on its upcoming Vera Rubin chips, and Anthropic's newly secured 14.8 gigawatts of capacity reportedly still trails what the company believes it will need. For Lucien's thesis, this is prediction #1's central tension condensed into a single company: one AI lab has now contracted for more future compute spending than most countries' entire GDP, a bet that inference revenue will eventually catch up to a sum of that scale.
Infrastructure & Economics · 24/7 Wall St. · Sep 13, 2026
China's Shanghai AI Lab Quietly Releases a 744B-Parameter Open Model Under an MIT License
Shanghai Artificial Intelligence Laboratory, a Chinese state-backed research institute, released Atria Dawn Preview without any product launch or press event: a 744-billion-parameter mixture-of-experts agentic model, with about 40 billion parameters active per token, built on top of Zhipu AI's open-weight GLM-5.2. The weights are posted openly under a permissive MIT license - notably more permissive than anything OpenAI or Anthropic has released for a flagship-scale model. Self-reported benchmark scores (not yet independently verified) show the model competitive on tool-use and web-browsing tasks, though it trails on coding benchmarks like SWE-bench Pro. For Lucien's thesis, this is prediction #2 in a fairly pure form: a frontier-scale, genuinely open and freely redistributable model appeared with almost no fanfare, continuing the pattern of open-weight capability catching up to the proprietary frontier - built, in this case, on another open model rather than a closed one.
Open-Weight Models · Startup Fortune · Sep 15, 2026
The Apple Angle (Standing Perspective)
Today's numbers make an unusually stark contrast for this recurring argument. Anthropic just contracted for roughly half a trillion dollars in future compute, and MIT Technology Review's new analysis says the entire hyperscaler buildout needs a 2.7x jump in economy-wide productivity by 2030 just to avoid being remembered as, in one economist's words, the largest misallocation of capital in history. Set against that, the case for a much smaller, calmer wager: a single Mac Studio, maxed out with 512GB of Apple's unified memory for around $9,500, can hold an entire trillion-parameter open-weight model - well north of the scale of today's freshly released Atria Dawn - in memory at once. Matching that on Nvidia's own workstation-grade hardware takes five to six RTX Pro 6000 cards, something like $60,000-$75,000 combined, while drawing on the order of ten times the electricity.
As permissively-licensed, frontier-scale open models keep landing every few weeks - Atria Dawn's MIT license this week, sitting atop GLM's own open release before it - the competitive question keeps sliding from "who trained the smartest model" toward "who can actually afford to own the hardware to run a nearly-as-good one locally, without underwriting a slice of the trillion-dollar bet MIT Tech Review just described." Apple's MLX software and Thunderbolt 5's RDMA networking already let multiple Mac Studios pool memory for larger jobs, Nvidia has stripped NVLink from its own consumer workstation cards, and Nvidia's DGX Spark increasingly borrows Apple's unified-memory approach rather than the reverse. The honest caveat still holds: a Mac Studio serves one person running one model at a time, while the gigawatts of capacity Anthropic just locked down are built for enormous concurrent demand across millions of users - this argues Apple may quietly own a valuable corner of AI compute, not that it replaces the data center outright.
Limited Edition Jonathan (Substack) · Perspective
Why This Matters
Today's three items sit almost entirely on one side of the ledger, which is itself worth noticing: MIT Technology Review's productivity-gap math and Anthropic's own half-trillion-dollar compute commitment both sharpen prediction #1's central risk - that the current buildout's scale now depends on returns nobody can yet demonstrate - while Atria Dawn's quiet, MIT-licensed 744-billion-parameter release shows prediction #2's capability-closing pattern continuing regardless, with a Chinese state lab now willing to give away frontier-scale weights outright. Taken together, the more capital-intensive and revenue-dependent the centralized bet becomes, the more attractive an open, ownable alternative looks by comparison - exactly the case the standing Apple Angle keeps making.
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.