Open Source AI — September 14, 2026
Three stories today, all circling the same question: how long can the current AI spending spree hold up?
In Plain English: Three stories today, all circling the same question: how long can the current AI spending spree hold up? SoftBank borrowed another $11.9 billion just to keep funding its stake in OpenAI - on top of $37 billion already raised this year - and investors sent its stock down 13% on the news. In South Korea, chipmakers Samsung and SK Hynix turned down their power utility's request to prepay years of electricity bills for new AI chip plants, saying they're not confident the chip boom will last long enough to justify it. And a coding-AI startup called Cognition released a new tool built on top of a free, open-weight Chinese model that now matches a leading closed AI model on real coding tasks at roughly a third of the cost - though the startup kept its own layer of improvements closed. Today's Apple Angle ties the Korean power story directly to the case for owning low-power hardware outright.
SoftBank Borrows Another $11.9 Billion to Keep Funding Its OpenAI Bet - and Its Stock Drops 13%
SoftBank secured an upsized two-year loan of $11.87 billion from a syndicate of roughly 20 banks to help finance its stake in OpenAI, on top of about $37 billion already raised this year through bond sales and other borrowing, plus a separate $10 billion margin loan against its OpenAI shares. Masayoshi Son's conglomerate is on track to have committed nearly $65 billion to OpenAI by October, and is reportedly weighing a further $10-20 billion junk bond sale on top of that. Investors did not take the news well: SoftBank shares fell 13%, their steepest drop since mid-July, as the scale of borrowing needed just to stay in the game became clearer. For Lucien's thesis, this is prediction #1 as a live stress test - the more debt one of the industry's biggest backers has to take on simply to keep funding its position, the more the whole structure depends on inference revenue eventually catching up to the capital already committed.
Infrastructure & Economics · Bloomberg (via Japan Times) · Sep 14, 2026
Samsung and SK Hynix Reject Korea's $18.7 Billion Power Prepayment Plan, Citing Doubts About How Long the Chip Boom Lasts
South Korea's state utility KEPCO asked Samsung Electronics and SK Hynix to prepay roughly 25 trillion won ($18.7 billion) - about five years of electricity bills in advance - to help fund the power grid upgrades their new AI chip plants will require. Both companies said no; a company official told reporters they were not convinced such a large upfront commitment was justified, citing "uncertainty over the long-term durability of semiconductor demand." That is a notable admission from two of the firms profiting most directly from the AI buildout: even they are unwilling to lock in a multi-year bet that today's chip demand holds up. For Lucien's thesis, this is prediction #1's skepticism angle arriving from an unexpected source - not an outside analyst, but the chipmakers themselves declining to commit capital on the assumption that current AI-driven demand is permanent.
Infrastructure & Economics · Bloomberg (via Business Recorder) · Sep 14, 2026
A Coding Startup Built Its New Model on a Free Chinese Model - Then Kept Its Own Improvements Closed
Cognition released SWE-2, a coding-agent model built by further training Moonshot AI's open-weight, 2.8-trillion-parameter Kimi K3 with additional reinforcement learning that the company says lifts scores by five to six points on several benchmarks. On FrontierCode 1.1, SWE-2 scores 50.0%, essentially matching the closed Fable 5.1 model's 50.9%, while running at roughly 64% lower cost, though it still trails OpenAI's GPT-6 Astra by a few points on the hardest tasks. Notably, SWE-2 itself has no public weights and no standalone API - it runs only inside Cognition's own Devin coding agent, even though the open-weight model underneath it is freely available to anyone. For Lucien's thesis, this is prediction #2 with a real complication attached: open-weight models are now capable enough to serve as the foundation for products that match closed frontier performance, but the companies building on top of them can still wall off the actual gains behind a closed layer of their own.
Open-Weight Models · MarkTechPost · Sep 12, 2026
The Apple Angle (Standing Perspective)
Today's clearest thread back to this recurring argument is the KEPCO story: two of the world's biggest chipmakers just declined to commit $18.7 billion toward power infrastructure because they aren't sure how long today's AI-driven demand will hold up. That same worry about locking in expensive, power-hungry capacity is exactly what sits behind the case that Apple, not Nvidia, may be quietly best positioned for a shift toward local AI. Apple's Mac Studio uses a unified memory architecture shared between CPU and GPU, and a maxed-out configuration with 512GB of that memory - around $9,500 - can hold an entire trillion-parameter open-weight model at once. Matching that capacity with Nvidia's RTX Pro 6000 workstation cards takes five to six of them, roughly $60,000-$75,000 combined, while drawing something like ten times the power - exactly the kind of long-lived, power-hungry commitment KEPCO's utility bill would have locked in for Samsung and SK Hynix's customers.
As frontier-capable open-weight models keep shipping every few weeks - this week's example being Cognition's Kimi K3-based coding agent - the competitive question keeps drifting away from "which model is smartest" and toward "who can afford to run a nearly-as-good one locally, without betting on demand or pricing staying put." Apple's MLX software and Thunderbolt 5's RDMA networking already let multiple Mac Studios pool memory for bigger jobs, while Nvidia has stripped NVLink from its consumer workstation cards and its own DGX Spark increasingly borrows Apple's unified-memory playbook rather than the other way around. The honest caveat still applies: a Mac Studio serves one person running one model at a time, while the multi-gigawatt data centers behind stories like SoftBank's OpenAI financing are built for enormous concurrent demand. Apple looks well 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 sit on both sides of Lucien's thesis at once, and in an unusually pointed way. SoftBank's fresh $11.9 billion loan - and the 13% stock drop that followed - shows one of the AI industry's biggest backers taking on real financial strain just to keep its centralized bet funded, exactly the pressure point prediction #1 expects to eventually bite. Samsung and SK Hynix's rejection of KEPCO's prepayment plan is the same worry from a different angle: the chipmakers profiting most from the boom are themselves unwilling to lock in a multi-year, power-hungry commitment on the assumption that today's demand is permanent. Meanwhile Cognition's SWE-2 shows open-weight capability continuing to close the gap with closed frontier models, even as the company's decision to wall off its own improvements is a reminder that "open" and "independent" aren't automatically the same thing. None of this settles the argument on its own, but it captures the exact tension the thesis is tracking: capital keeps flowing toward - and straining under - the current centralized buildout, even as the underlying case for cheaper, owned, local compute keeps getting stronger, which is precisely the setup the standing Apple Angle argues Apple is already positioned to benefit from.
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.