Open Source AI — September 15, 2026
Four stories today, and together they sharpen both sides of the debate.
In Plain English: Four stories today, and together they sharpen both sides of the debate. On the skeptical side: Wall Street's bond market for data centers has exploded - $17 billion in real-estate debt backed by data centers since early 2025, more than triple the prior two years combined - and investors are discovering these buildings are far riskier collateral than typical offices or warehouses. Separately, S&P Global warned that the credit quality of the biggest cloud/AI companies is "gradually weakening" as spending outpaces returns, with hyperscalers increasingly propping up younger AI labs' financing too. On the open-weight side: US intelligence agencies accused six Chinese AI labs of systematically copying capabilities from Claude, GPT, Gemini and Grok, while China's Zhipu AI raised another $5 billion just two months after its last round - even as it admits its own revenue still can't cover its AI infrastructure costs. Today's Apple Angle ties the data-center financing worries directly to the case for owning simpler, local hardware instead.
US Intelligence Agencies Accuse Six Chinese AI Labs of Systematically Copying Claude, GPT, Gemini and Grok
The NSA, CISA, and FBI issued a joint advisory accusing DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI of "systematic extraction" of US frontier models' capabilities - using billions of tokens across millions of queries, routed through multiple accounts, proxies, and gray-market services, to generate synthetic training data for their own models. The advisory says DeepSeek alone distilled four versions of Claude, two of Gemini, five of ChatGPT, and Grok 4, while Moonshot distilled 18 different US models including Anthropic's own Fable 5. This escalates a narrower dispute touched on in this brief's September 12 edition, where Anthropic had accused Moonshot of harvesting millions of Claude responses to train Kimi - the new joint advisory elevates that fight to formal US government policy rather than one company's grievance. For Lucien's thesis, this is a complication worth sitting with: the open-weight capability gains behind prediction #2 are, by the agencies' own account, partly built on distilling the closed frontier models they're closing the gap with, and a serious government crackdown could slow that catch-up rather than the models simply out-innovating their rivals on their own.
Open-Weight Models · CyberScoop · Sep 8, 2026 (follow-up to Sep 12 coverage)
China's Zhipu AI Raises Another $5 Billion for Its Open-Weight GLM Models - Two Months After Its Last Round
Zhipu AI, maker of the open-weight GLM model family, raised roughly $5 billion in fresh funding, just two months after a $4 billion round in July, with proceeds earmarked for its next-generation GLM model, its own training framework, and more computing infrastructure. The company has shipped a new GLM version roughly every two months this year (5, 5.1, 5.2, 5.3), and reports first-half 2026 revenue grew almost 400% year-over-year to about 954 million yuan (roughly $134 million), with corporate API usage up more than 27-fold. Even so, the reporting notes plainly that "the domestic AI model industry...cannot cover enormous computing and R&D costs with commercial revenue alone" - meaning Zhipu, like its Chinese rivals, is still burning outside capital to keep an open model competitive. For Lucien's thesis, this is a genuinely two-sided data point: real capital is willing to bet billions on an open-weight lab, supporting prediction #2's momentum, but the admitted revenue gap is a reminder that "open" doesn't yet mean "self-sustaining" - someone is still subsidizing the compute race either way.
Open-Weight Models · Seoul Economic Daily · Sep 14, 2026
S&P Warns Hyperscaler Credit Quality Is "Gradually Weakening" as AI Capex Outpaces Returns
S&P Global cautioned that the credit quality of the largest cloud and AI companies is gradually weakening, pointing to capital expenditures rising faster than expected, increasingly opaque financing structures, and returns on the AI buildout that remain years away. The report estimates the top six hyperscalers - Amazon, Microsoft, Alphabet, Oracle, SpaceX, and Meta - are on track to spend more than $7 trillion on data centers and AI infrastructure through 2030, with Amazon alone raising about $100 billion in bond-market debt this year. S&P also flagged a "shadow lending market" forming underneath that spending: hyperscalers are using their own strong credit ratings to help riskier, less-established companies like Anthropic and OpenAI secure cheaper financing through leases, loan guarantees, and chip-purchase commitments - a structure that could transmit losses through the sector if a smaller player falters. For Lucien's thesis, this is prediction #1 from one of the industry's most conservative referees: a major credit-rating agency, not an outside skeptic, is now on record saying the debt funding this buildout is growing faster than the evidence it will pay off.
Infrastructure & Economics · Axios · Sep 11, 2026
Wall Street's AI Data Center Boom Is Rewriting the Rules for Real Estate Debt
Data centers have become a fast-growing corner of the commercial mortgage bond market, with roughly $17 billion in data-center-backed CMBS issued since early 2025 - more than triple the total from the prior two years combined - and now about 8% of all new commercial property bond deals. Bond investors and analysts say these deals don't fit traditional real-estate underwriting: hyperscaler tenants often demand confidentiality about their identity and lease terms, a facility's value now depends more on cheap power and grid capacity than on location, and the specialized cooling and electrical systems built for one generation of AI chips may not suit the next generation, or a different tenant, if a lease doesn't renew. One bond manager, Ben Hunsaker of Beach Point Capital, put it plainly: "data centres are much more opaque" than the office and warehouse debt investors are used to pricing. For Lucien's thesis, this is prediction #1 showing up in the plumbing of the financial system itself: the same rapid chip evolution that makes open-weight models improve every few weeks is exactly what bond investors now say could make a data center's collateral value obsolete before its debt is repaid.
Infrastructure & Economics · Bloomberg (via Business Standard) · Sep 15, 2026
The Apple Angle (Standing Perspective)
Today's clearest thread back to this recurring argument is the CMBS story: bond investors are discovering that a data center's value can evaporate the moment its chips and cooling systems fall behind the next hardware generation, and that the tenant behind the lease is often a black box. That exact fragility - a large, illiquid, power-hungry asset that can go stale within a single chip cycle - is the risk this argument says a Mac Studio simply doesn't carry. Apple's unified memory architecture lets one Mac Studio, maxed out at 512GB for around $9,500, hold an entire trillion-parameter open-weight model in memory at once. Matching that on Nvidia hardware takes five to six RTX Pro 6000 workstation cards, roughly $60,000-$75,000 combined, drawing something like ten times the power - and, per S&P's warning above, increasingly financed through the same opaque debt structures now worrying bond investors.
As frontier-capable open-weight models keep shipping every few weeks - Zhipu's rapid GLM cadence being this week's example, distillation disputes notwithstanding - the competitive question keeps drifting from "which model is smartest" toward "who can afford to own, outright, the hardware to run a nearly-as-good one locally." 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 reverse. The honest caveat still holds: a Mac Studio serves one user running one model at a time, while the trillions in data-center debt now worrying S&P and CMBS investors are financing infrastructure built for enormous concurrent demand - Apple looks positioned to dominate a valuable slice of local, single-user AI compute, not to replace the data center outright.
Limited Edition Jonathan (Substack) · Perspective
Why This Matters
Today's four items sit on both sides of Lucien's thesis at once, with an unusually direct line to the plumbing of AI's financing. Two independent, conservative voices - S&P's credit analysts and the commercial mortgage bond market itself - are now flagging that AI infrastructure debt is growing opaque and outpacing proven returns, exactly the strain prediction #1 expects; meanwhile Zhipu's ability to raise $5 billion for open-weight GLM development shows real capital still backing the open-weight side too, even as its own admission that revenue can't cover costs, and the US government's distillation crackdown on six Chinese labs, both complicate how fast and how independently that side can actually win. None of it resolves the argument outright, but it captures the exact tension being tracked here: money keeps flowing into both an increasingly fragile centralized buildout and an unsettled open-weight race, which is precisely the instability the standing Apple Angle argues favors owning simpler, local hardware instead.
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.