Open Source AI — September 9, 2026
Three stories today. A well-known Wall Street investor who runs a $20 billion fund said publicly he agrees with famous short-seller Michael Burry that the money pouring into AI data centers has become a bubble — though he thinks there's one more big rally before any crash.
In Plain English: Three stories today. A well-known Wall Street investor who runs a $20 billion fund said publicly he agrees with famous short-seller Michael Burry that the money pouring into AI data centers has become a bubble — though he thinks there's one more big rally before any crash. Separately, the French AI company Mistral raised $3.6 billion, one of the largest funding rounds ever for a European tech company, built explicitly around giving away its model's underlying code so governments and businesses can run it themselves instead of depending on a big U.S. cloud provider. And a widely-watched independent scorecard that ranks AI models updated its numbers: the best freely downloadable models are now only 9 points behind the very best paid, closed models on its 0-100 scale — the tightest that gap has been tracked. Today's Apple Angle explains why a shrinking capability gap matters even more once the hardware to run those models yourself keeps getting cheaper too.
'Black Swan' Investor Mark Spitznagel Says He Agrees With Michael Burry: A Historic AI-Driven Crash Is Coming
Mark Spitznagel, founder of the $20 billion hedge fund Universa Investments — best known for the outsized returns his fund posted in the 2008 and 2020 crashes — said publicly that he agrees with Michael Burry's view that hyperscaler AI infrastructure spending has become a bubble, predicting "one more really big, risk-on, insane, euphoric rally" before what he called a bigger crash than anyone alive has seen. Notably, he isn't calling the top yet: his own fund is reportedly still telling investors to stay bullish for now, treating this as a timing bet rather than an immediate short. For Lucien's thesis, this is another well-capitalized, historically prescient investor lending credibility to the profitability-skepticism side of the argument — not proof the capex won't pay off, but one more sign the doubt has moved well past AI-skeptic commentators and into serious institutional money.
Benzinga · Infrastructure & Economics · Sep 8, 2026
Mistral Raises $3.6B at a $24B Valuation, Betting Big on "Sovereign, Open-Weight" AI as the Alternative to U.S. Clouds
French AI lab Mistral closed a €3 billion (~$3.6B) round led by Samsung Electronics — with Nvidia, a16z, Salesforce Ventures, BlackRock and others participating — at a €21 billion (~$24B) valuation, one of the largest funding rounds in European tech history. The company is explicitly framing its open-weight models and a planned 1 GW of European "sovereign" compute capacity by 2030 as the alternative to depending on the big closed U.S. providers, and said plainly that "not being an American company" has helped its revenue with governments and corporations who want to keep control over their own data and infrastructure. For Lucien's thesis, this is close to a direct, dollar-denominated bet on prediction #3: serious institutional capital wagering that customers will increasingly pay for self-hostable, open-weight infrastructure they control rather than for continued API dependency on a handful of closed providers.
TechCrunch · Open-Weight Models · Sep 8, 2026
Independent Benchmark: Best Open-Weight Models Now Sit Just 9 Points Behind the Best Closed Ones
The widely-cited Artificial Analysis Intelligence Index, updated to version 4.3, shows open-weight leaders GLM-5.3 and Kimi K3 scoring 44 on its 0-100 composite scale, against a score of 53 for the current closed-model leaders, OpenAI's GPT-6 Astra and Anthropic's Claude Fable 5.1 — a 9-point spread, described as the tightest gap the index has tracked between what's freely downloadable and what commands premium API pricing. In a smaller but related detail, the index also found GPT-6 Astra completing a typical benchmark task for $3.26 on average versus $7.63 for Claude Fable 5.1, a reminder that even the closed frontier is under real price pressure. For Lucien's thesis, this is a concrete, continuously-tracked data point for prediction #2: the capability gap between free and paid frontier models keeps compressing, even as the closed leaders themselves keep shipping new versions to try to hold the lead.
Trending Topics (via Artificial Analysis) · Open-Weight Models · Sep 8, 2026
The Apple Angle (Standing Perspective)
Today's items line up almost too neatly with the standing argument here: as the gap between open-weight and closed models keeps narrowing to single digits on independent benchmarks, and as a company like Mistral raises billions specifically to sell customers the right to run their own models instead of renting someone else's, the question of what hardware actually runs those open weights only gets more important. That's where Apple's Mac Studio line keeps quietly sitting. A Mac Studio configured with up to 512GB of unified memory, shared between CPU and GPU, costs around $9,500 today and holds a trillion-parameter open-weight model — the exact kind of model Mistral, GLM, and Kimi keep shipping — entirely in memory, with no data-center lease and no per-token bill. Matching that 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 a single Mac Studio.
Today's other story, Mark Spitznagel echoing Michael Burry's bubble call, is a reminder of why that fixed-cost alternative matters beyond hobbyists: if a serious slice of Wall Street increasingly believes the money behind rented, centralized AI compute is overextended, then the appeal of a one-time hardware purchase that never re-prices grows right alongside the capability of the open-weight models it can run. Apple's MLX framework and 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 trains the best model" toward "who can afford to run a nearly-as-good model without a hyperscaler-sized bet behind it." The fair caveat still applies, as always: a Mac Studio serves one person running one model at a time, while the data centers behind today's economy-wide AI buildout are 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 three items split cleanly across both halves of Lucien's thesis, and reinforce each other. Mark Spitznagel adding his name and his $20 billion fund's credibility to Michael Burry's AI-bubble call keeps building the case that the profitability of current, centralized compute spending is a live and serious doubt, not a fringe one. Mistral's $3.6 billion raise, built explicitly around open-weight models and "sovereign," self-hosted compute, shows real institutional capital already betting that customers will pay to control their own AI infrastructure rather than rent someone else's indefinitely. And the Artificial Analysis Intelligence Index's latest update puts a hard, trackable number on how close the free, downloadable frontier has gotten to the paid one — just 9 points apart on its scale. None of this proves the transition to on-prem and on-device AI has already arrived, but each piece narrows the same gap the thesis has been tracking: as open models close in on closed ones and doubts about the economics behind the closed, rented model keep surfacing from increasingly serious sources, the case for owning rather than renting compute keeps getting easier to make — which is exactly the bet the standing Apple Angle argues Apple is already quietly positioned to win.
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.