Lead storyThe Information
CoreWeave carried $35.4 billion in debt as of June 30: $13.7B in project loans, $10B in senior notes, $6.6B in convertible notes, and $5.1B in equipment and software financing. Its stock fell about 20% in a week, from nearly $80 to just over $61, while yields on its existing debt climbed above 15%. Blackstone debated a $2.3B chip-collateralized loan in early 2023; CEO Mike Intrator called the market reaction irrational, Jensen Huang publicly voiced confidence, Goldman Sachs upgraded the stock to buy, and Ben Horowitz called the business a 'Rube Goldberg financial machine.'
Why it matters
The debt-funded compute model is the weak link in the AI buildout. When borrowing costs push past 15%, the business lives or dies on a handful of concentrated customers: Microsoft, OpenAI, and Meta. If any of them walks, billions in project loans go with them. Founders and investors should treat neocloud capacity as repricing risk, not permanent infrastructure.
Story 02The Information
Nscale and Anthropic are in talks for a second compute deal worth roughly $45 billion, on top of the $45 billion six-year agreement struck in August for 460 MW of Nvidia Vera Rubin chips at Nscale's West Virginia campus. Separately, chip startup Tenstorrent raised $1.5 billion in Series E funding at a $6.2 billion valuation, led by Trace Capital and Innovation X.
Why it matters
The two-tier buildout is getting sharper: labs with locked-in demand keep signing bigger commitments while debt-funded hosts get punished, so capital flows toward demand certainty and away from leverage. Tenstorrent's raise shows there is still appetite for the picks-and-shovels layer, but only with a clear customer story attached.
Story 03OpenAI
OpenAI's new MultiLangBench asks models to build a SQLite-compatible database engine from scratch in each of 20 languages. GPT-6 Astra hit an 87.4% average test pass rate, up from 26.8% for GPT-5.6 Sol, with gains in every language: COBOL 10.9% to 72.1%, OCaml to 85.8%, Fortran to 83.5%, Visual Basic to 89.4%. Cheaper GPT-6 Luna reached 40.3% at about 78% lower API cost, and on a COBOL migration test with byte-for-byte behavior checks, GPT-6.1 Sol hit 86.6% at low reasoning effort and 97.9% at high effort. The eval setup is on GitHub.
Why it matters
Legacy modernization, a multi-year enterprise services business, is becoming a job models can do. COBOL still runs more than 40% of online banking and 95% of ATM transactions, so near-saturation migration scores mean banks and insurers can move from decade-long service contracts to model-assisted rewrites with expert review. The cheaper Luna tier keeps compressing prices for coding agents.
Story 04The Information
Huawei engineers are stationed at DeepSeek's offices, folding DeepSeek's requests into chip-design decisions for AI chips shipping next year and beyond, with the goal of DeepSeek fully training an advanced model on Chinese chips. DeepSeek plans to start pre-training on the Ascend 950DT as soon as the end of this year and is set to receive most of the next-gen Ascend 960DT supply (Q1 2027, then 970 in 2028, 980 in 2029). Epoch AI estimates Nvidia's B300 delivers roughly 7x the raw arithmetic performance of the Ascend 950, with Huawei staying 3-4 years behind through 2030; TileLang, an open-source kernel language, now handles most of DeepSeek's kernel code on both Nvidia and Huawei chips. DeepSeek is still seeking Blackwell chips via intermediaries who dismantle servers bought where sales are allowed.
Why it matters
China is building a full stack, model lab plus chips plus kernel software, instead of chasing Nvidia chip for chip. Export controls slow Chinese frontier training but do not stop it, because co-designed software can close part of the hardware gap. Nvidia's real moat narrows wherever CUDA lock-in loosens.
Story 05The Information
The U.S. Treasury sanctioned five Chinese entities connected to Linwei Ding: Shanghai Zhisuan Technology, the AI startup Ding founded in 2023, and four entities tied to MiraclePlus, the Beijing VC and incubator that began as Y Combinator's China arm before going independent in 2018. Ding was sentenced to prison last month after being convicted in January of stealing thousands of pages of Google's AI trade secrets. The sanctions block the entities' U.S. properties and bar U.S. banks and investors from transacting with them. Separately, DHS ordered federal civilian agencies to remove DeepSeek models from their IT systems.
Why it matters
Washington is moving from prosecuting individuals to sanctioning whole networks: startup, incubator, and backers together. Anyone whose capital or code touches these entities now carries secondary exposure, and compliance lists will likely widen to other programs that onboarded ex-lab engineers. Expect U.S.-China AI decoupling to accelerate from chips to people, code, and money.
Story 06The Information
Sam Altman told OpenAI employees about three proposals for governing AI development: an IAEA-like international body enforcing safety standards, which he appeared to support; the 'Demis plan,' merging all frontier labs into a single CERN-like entity; and the 'Eliezer plan,' in which one firm teaches itself to get smarter on its own and uses its economic might to 'melt everyone else's GPUs,' which Altman called obviously terrible and included only for completeness. The framing lands as the Trump administration builds its first system for policing 'superintelligence' developers. Altman also described 'pacing 2.0': within the next year OpenAI must pick a long-term governance path and adopt a 'speed limit' on self-improving AI.
Why it matters
AI governance stopped being abstract. The industry's most powerful CEO is shopping endgame designs inside his own company while the government builds its policing apparatus, and a formal 'speed limit' on self-improvement would become a regulatory ceiling every lab's roadmap must respect. Investors should price political risk into frontier-lab bets the same way they price compute risk.
Story 07The Information
Federal prosecutors say Abdulrahman Alfozan, the 35-year-old Saudi national arrested October 9 after OpenAI found a recording device in a meeting room, exchanged frequent messages with an unnamed member of the Saudi royal family, the latest just four days before his arrest. The royal tried to recruit him to a Kingdom AI initiative; Alfozan, who earned $385,000 a year with a $2.2M equity package, said he would not leave until his stock vested. He had booked one-way tickets for himself, his wife, and his son to Saudi Arabia on October 12, and was arrested three days before the flight. Prosecutors are seeking his detention.
Why it matters
The insider threat at frontier labs is now a court-documented pattern, not a hypothetical. Generous pay does not protect against recruitment. Expect labs to harden physical security and screening, and budget lines for internal monitoring of the kind Strider sells to become standard; for anyone hiring AI talent, access control is now a diligence item.
Story 08a16z
On The a16z Show, OpenAI's Karan Singhal said more than 300 million people use ChatGPT for health and wellness questions each week. OpenAI trains healthcare expertise directly into its general-purpose frontier models, works with hundreds of medical experts to evaluate safety and disagreement, and sees AI as a complement to doctors; the company is building integrations with Epic and AI copilots for clinicians. Singhal made the case that the cost of medical intelligence keeps falling, letting AI act as a safety net that surfaces insights from years of health data and navigates appointments and records.
Why it matters
Healthcare is the clearest consumer-scale AI battleground. OpenAI owns the front door with 300 million weekly health users, Epic owns the data pipes, and the falling cost of medical intelligence is the investability thesis for every AI health startup. Startups should not try to out-model the labs; the opening is agent UX and clinician workflows built on top.
Story 09SpaceX
SpaceX will post financial and operational results for Q3 2026 after the market close today, with management hosting a live audio-only webcast at 3:30 p.m. CT / 4:30 p.m. ET. The webcast and replay will be available at ir.spacex.com and livestreamed on X via @SpaceX, with pre-registration open and an investor Q&A platform taking questions ahead of time. This is the company's second quarterly report since its listing on Nasdaq as SPCX.
Why it matters
The first hard read on Starlink's revenue scale and the Starship spending pace under public-market scrutiny. Satellite and telecom investors will mine the webcast for capex and capacity signals, and the numbers set the tone for tomorrow's Tesla print.