TSMC posts a record quarter; CoreWeave's CDS passes 1,000bp

AI² — AI Investor Sydney · Friday 11 December, The Mint. One widely cited forecast puts cumulative AI and data-centre capital spending at about US$11 trillion by 2029. One day, six evidence-led briefings on where that capital is creating and destroying value — no panels, no pitches. Every ticket includes a year of NOOPS Pro. A$995 until 30 October → Book your seat

TSMC reported a sixth straight quarter of record profit and says demand is still higher than its supply, in the same week CNBC counts more than US$1trn shed by the 20 largest AI and chip stocks and CoreWeave's five-year credit default swaps passed 1,000 basis points. Anthropic's third-quarter revenue is reported at US$16.8bn ahead of a planned November listing; separately, the FTC has sent it and OpenAI civil investigative demands. Further down: Meta and Microsoft push their own weights while a new benchmark puts three closed models in a tie at the top, Workday and Apple each make their case for AI, and the web turns out to be writing headers for agents already.

Chip stocks fall and CoreWeave's credit widens

Twenty of the most valuable global AI and semiconductor stocks have shed more than US$1trn in combined value so far this week, by CNBC's tally on Thursday. SK Hynix fell more than 10% in Seoul and Advantest almost 14% in Tokyo; in US trade, neoclouds were mixed, CNBC reported, with CoreWeave up nearly 5% in afternoon trading. CNBC and the analysts it quotes point to financing rather than demand as the concern: Dan Ives of Yorkville Ives argued that Treasury yields above 5% raise the cost of funding the build-out, but that "rates are a financing question, not a demand question", and named his own test — cuts to hyperscaler capex plans, cancelled contracts or falling GPU rental prices would change his view. These are accounts of sentiment rather than an established cause. The credit market is where the question is being priced most explicitly: spreads on CoreWeave's five-year credit default swaps rose above 1,140 basis points at one point on Thursday and last stood at 1,088bp, CNBC reports, up 33% in a week, and the yield on its 9.75% bond due 2031 reached 15.78%. Gavan Nolan of S&P Global Market Intelligence told CNBC that a spread of 1,000bp or above is generally a sign of credit distress; his framing is that nobody is questioning whether demand for CoreWeave's services is real, only whether the cash flows arrive in time to justify the capex. Chief executive Mike Intrator wrote in a memo obtained by CNBC that the backlog, which Deutsche Bank puts above US$100bn, "more than supports all of our debt obligations". Deutsche's Jim Reid noted that Oracle's five-year CDS also widened this week, to 288bp, and warned that this kind of concern can become self-fulfilling. Mark put the question in practical terms: how would a real hedge fund kick the tyres on a 15% yield?

TSMC reports record profit and more demand than supply

TSMC reported third-quarter net profit of NT$763.1bn, up 68.7% year on year and a record for the sixth straight quarter, on revenue of NT$1.49trn (US$46.36bn), CNBC reports, with both figures ahead of LSEG SmartEstimates. Operating margin was 58.7%, above the top of its guidance range, and TSMC guided fourth-quarter revenue to US$52.3–54.3bn. TSMC North America chief executive Sajiv Dalal gets at least ten calls a day about wafer capacity, CNBC reports, and TechInsights' Dan Hutcheson says all the big chipmakers are now testing at Intel in search of a second source. The company is also checking whether its customers can deploy what they order: TSMC is sending staff to inspect US data-centre construction and power availability as it plans further capacity, The Register reports from the earnings call. Chairman and chief executive C.C. Wei said TSMC factors cloud providers' capex, financing and power supply into its planning and still judges that "the demand is higher than the supply"; it expects 2026 capex of US$60–64bn. Wei also warned that "consumer and price-sensitive end market segments are being challenged" by rising component prices. Set beside the week's sell-off in AI hardware, the supplier at the narrowest point of the stack reports record profit and more demand than it can serve while its customers' financing is being questioned.

Labs heading for listings report growth and draw scrutiny

Anthropic generated US$16.8bn in revenue between July and September, up about 46% from US$11.5bn in the previous quarter, according to a Euronews report carried by Yahoo Finance that cites Bloomberg and people familiar with the matter. The figures were reportedly presented to investors weeks before a planned November IPO, and Anthropic reportedly told them it was profitable on an adjusted basis for a second straight quarter, with an 11% operating margin on the same basis and US$1.70 of revenue for each dollar spent on computing power. Revenue had more than doubled from the first quarter to the second, so 46% is a deceleration, and the Financial Times is cited as saying the result fell short of some investors' hopes. Anthropic has yet to publish its prospectus, so every figure here remains reported rather than disclosed. Anthropic and OpenAI are also reportedly under federal scrutiny: the US Federal Trade Commission sent civil investigative demands to Anthropic and OpenAI on 14 October over consumer harms linked to recent safety incidents, Ars Technica reports, citing an insider who spoke to Axios. The requests reportedly seek documents on the incidents, internal security measures and marketing that may have misled consumers; neither company responded to Ars. The same day, Reuters reported that a bipartisan group of eight states had filed separate suits against OpenAI under consumer-protection and public-nuisance law, seeking among other remedies an order barring it from developing self-improving models unless it can ensure they are aligned with human values. Anthropic, for its part, is opening itself to outside inspection: it says six METR staff now have "employee-like access", with company hardware, office space and internal tools roughly on par with those its own risk-assessment teams use. METR will publish its analysis of Anthropic's risk levels, incidents or practices; Anthropic cannot veto publication but may require narrow redactions, which METR is free to characterise publicly. Anthropic explicitly borrows the logic of bank supervision, where regulators place resident examiners inside the largest banks. John's caution is about the depth of the bench: very few people are capable of this work, they are likely to share similar biases, and METR also wrote the incident report after the Hugging Face incident at OpenAI.

Meta and Microsoft push their own weights

Meta released the weights of Muse Spark Open on Thursday, describing it as "the most capable US open-weight model to date": 816bn total and 42bn active parameters, a one-million-token context window, and, Meta says, the first model in its class to run on a desktop, using two Apple M5 Ultra Mac Studios with up to 512GB of unified memory each. MacRumors notes that the 512GB Mac Studio has not yet been released, so "desktop" means two very expensive workstations; the licence carries "limited commercial terms", and the capability claim is Meta's own and awaits independent testing. Microsoft is building at frontier scale too: Microsoft AI says it has completed pre-training MAI-Base-2, a mixture-of-experts foundation model with 80bn active and 2.8trn total parameters, trained on 16,000 GB200 GPUs over 38 days. It claims the model outperforms every open-weight model it tested on held-out evaluations and that its recipe is 4.2–12.3 times more pre-training-FLOPs-efficient than DeepSeek-V4-Pro-Base. Every figure is Microsoft's, the comparisons are against open base models rather than proprietary frontier checkpoints, and the model is still in post-training. A new benchmark offers one outside reading of where open weights stand on long-horizon agentic work: on Artificial Analysis's AA-3D, in which models act as agents building interactive 3D scenes that a recruited panel compares blind, Claude Sonnet 5.5 is first at 1194 Elo but statistically tied with Claude Opus 5.5 and GPT-6 Astra, with Anthropic's Fable fourth in a lower band. The best-placed open-weight models, Kimi K3 and MiMo-V2.6-Pro, are 11th and 12th, more than 200 points behind. The more useful finding is behavioural: across the 19 models that can view images, how early an agent first looks at its own scene tracks Elo closely (Spearman 0.82).

Workday and Apple each make their case for AI

Workday told financial analysts that automating work such as HR tasks could expand its market "beyond the software budgets into the labor and services spend", The Register reports, with chief AI officer Joel Hellermark saying customers are "spending tens of millions of dollars on outsourcing" and claiming self-service agents had raised resolution rates for common tasks by 25%, a figure that is Workday's own and unaudited. Chief executive Aneel Bhusri said agent charges will grow with consumption rather than the number of licensed employees, and Workday has opened "headless" access to its HR and finance systems through Model Context Protocol tools. That is a seat-based vendor saying the seat is the wrong unit, and its pitch for labour budgets points at outsourcing providers as the counterparty, since any labour spend it captures would come out of their revenue. Apple is making a different argument. Apple closed above a US$5trn market capitalisation for the first time on Wednesday 14 October, at US$342.96 a share, MacRumors reports, the second company to close there after Nvidia. Separately, in an interview with TIME, marketing chief Greg Joswiak said Apple Intelligence is aimed at "everyday people" rather than Silicon Valley's "lunatic fringe", and that Apple doesn't "feel threatened" by competitors; nothing in either report connects the interview to the milestone. Apple is also preparing to bring basic Apple Intelligence features to China in iOS 27.3, MacRumors reports citing Bloomberg. John is less persuaded by the stance; his read was that it "might explain why Siri continues to largely be useless".

Agents get their own headers; mathematicians get a fund

Signals Corps scanned the HTTP response headers of 217m page responses, 10% of Common Crawl's September 2026 crawl, and found 7,959 sites sending a header written for an AI system. Most were pointers, 6,989 of them to an llms.txt file; 30 sites sent injection attempts, and 22 of those were pranks. In the firm's October tests, Claude Opus, Sonnet and Haiku running in Claude Code found its test headers and did not act on them by default, but obeyed in 7 of 9 runs when the user's prompt told them to follow the site's directions, so resistance depends on how much authority the prompt grants a website. The volume is tiny; the channel matters because headers are invisible to a human reader. In mathematics, where results can be checked and capability claims are hard to inflate, the field is getting help to adapt: the Simons Foundation, Anthropic and OpenAI will commit US$15m in combined funding and resources to help the mathematics community navigate AI, with Simons managing and awarding the money and the two labs adding AI tools and computing resources. Derek Thompson writes that OpenAI has published hundreds of mathematical findings in recent months, many of which had resisted the world's best mathematicians for decades or centuries. Two of the labs whose models drove that disruption are co-funding the adjustment, which is support and also a seat at the table where norms for judging AI-assisted proofs will be set, even with an independent foundation holding the purse.

Read all 14 signals →