Open Source AI — September 20, 2026
Two stories today, both about who actually owns the AI boom.
In Plain English: Two stories today, both about who actually owns the AI boom. A London research firm mapped $3.6 trillion in AI financing deals and found many of the same companies showing up as each other's investors, customers, and suppliers at once - a circular pattern that helped send CoreWeave's stock down after it announced more borrowing. Separately, 19 European countries formally launched a joint fund to build their own AI infrastructure - training, deployment tools, and industrial applications - so they stop permanently renting cloud AI from American hyperscalers. Today's Apple Angle argues a single $9,500 Mac sidesteps both problems at once: no debt to untangle, nothing to rent.
$3.6 Trillion in AI Financing Deals Are More Tangled Than They Look, Research Firm Finds
London-based Sona Asset Management mapped $3.6 trillion across 176 AI financing deals involving 202 companies and found roughly 120 of them "highly circular" - the same firms turning up simultaneously as each other's investors, customers, and suppliers within the AI buildout. The report landed the same week CoreWeave's stock fell more than 4% after it disclosed a new $3 billion convertible debt raise (with an option for $500 million more), a company that already carries $35.6 billion in debt against $35-39 billion in expected annual capex. Oracle, per the same analysis, carries $125.3 billion in borrowings plus $288 billion in off-balance-sheet data-center leases stretching 15 to 19 years. Prediction markets are still sanguine - Polymarket traders put just a 12% probability on an AI industry downturn by year-end - but the report's core point is structural: much of the buildout rests on the same pool of demand circulating through a small set of interlinked companies. For Lucien's thesis, this gives prediction #1 a name for the mechanism, not just the size: circular financing means a wobble at one company doesn't stay contained to that company.
Infrastructure & Economics · Benzinga · Sep 18, 2026
19 European Countries Launch Joint Fund to Build AI Infrastructure They Actually Own
The EU's IPCEI-AI program went live this month, with 19 member states - including Germany, France, Italy, Spain, Poland, and the Netherlands - coordinating public funding for a six-layer AI stack running from raw data processing up through sovereign foundation models, sector-specific models, and industrial applications. Germany is coordinating the effort and has committed €1 billion on its own, with individual projects eligible for grants up to €25 million; a second application deadline falls October 31, with a third in April 2027. The program's own framing is explicit that it isn't trying to out-spend American frontier-model labs - Europe invested roughly €8 billion in AI in 2024 versus $97 billion in the US - but instead to build federated, EU-governed infrastructure around the manufacturing and industrial data European companies already hold. The one acknowledged gap: chips still come from Nvidia, AMD, and Qualcomm, so hardware sovereignty remains incomplete even as the software and deployment layers move toward local ownership. For Lucien's thesis, this is prediction #3 playing out at government scale: an entire continent choosing to fund owned infrastructure specifically to stop renting it indefinitely from US hyperscalers.
On-Prem/On-Device Shift · Tech Times · Sep 18, 2026
The Apple Angle (Standing Perspective)
Today's two stories are really one story told at different scales: a $3.6 trillion financing web tangled enough that researchers had to draw a map of who owns what, and 19 governments deciding the fix is to stop renting and start owning their own AI infrastructure outright. Apple's hardware quietly offers the smallest possible version of that same move. A single Mac Studio, configured with 512GB of unified memory for around $9,500, holds an entire trillion-parameter open-weight model in memory on one machine - nothing financed, nothing leased, nothing that would show up on anyone's circular-financing map.
Reaching that same memory footprint on Nvidia's workstation cards takes five or six RTX Pro 6000s, roughly $60,000-$75,000 combined, at something like ten times the power draw. As frontier-capable open-weight models keep narrowing the gap with closed ones, the competitive question keeps sliding away from "who trained the smartest model" and toward "who can just afford to own the machine that runs a nearly-as-good one." Apple's MLX software and Thunderbolt 5's RDMA networking already let several Mac Studios pool memory for bigger jobs, Nvidia has pulled NVLink out of its own consumer workstation cards, and its DGX Spark leans on Apple's unified-memory idea more than the reverse. The honest caveat stays the same: a Mac Studio serves one person running one model at a time, while the debt-financed, multi-gigawatt facilities in today's first story are built to serve millions of concurrent users at once - Apple looks well-positioned to quietly own a valuable, low-friction slice of AI compute, not to replace the data center outright.
Limited Edition Jonathan (Substack) · Perspective
Why This Matters
Today's two items show the same instinct working at wildly different scales - a research firm untangling $3.6 trillion of circular AI financing, and 19 governments deciding the answer is to stop renting compute and build their own. Both sharpen prediction #1's central risk while reinforcing prediction #3's endpoint, and the standing Apple Angle argues a single owned machine solves the same problem this brief keeps tracking, just at the scale of one desk instead of one continent.
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.