CoreWeave's bond yields hit 15.78% as TSMC booked a sixth straight record quarter
The week the money got expensive: CoreWeave's bonds traded like distressed credit, TSMC booked a sixth straight record quarter, Meta shipped a frontier model it sells only by subscription, eight states sued OpenAI, and Anthropic measured how close small models are to copying themselves.
Between October 9 and October 16, 2026, the cost of the money behind the AI build-out jumped, and the companies supplying it kept setting records.
CoreWeave's 2031 bonds yielded 15.78%, and its credit-default swaps crossed the level S&P calls distress. 1
TSMC, which fabricates most advanced AI chips, reported a sixth straight record quarter and raised its three-year spending plans. 2
The same week, Meta shipped its first frontier models, Google released its newest one to everyone, eight US states sued OpenAI, and Anthropic's red team ran a 26B model across 15 cheap TV boxes.
| Development | Number to keep | What changed | Why it matters |
|---|---|---|---|
| CoreWeave's credit | 15.78% yield, 1,140bp swap peak | The yield to maturity on CoreWeave's 9.75% bond due 2031 reached 15.78%, and its five-year credit-default swap spread widened past 1,140 basis points. 1 | The borrowing that finances AI data centres is now priced as distressed for the first time. |
| TSMC's quarter | NT$763.1B net profit, +68.7% | TSMC beat estimates on both revenue and net income, and guided fourth-quarter revenue up 58% from a year earlier. 2 | Demand for AI chips still exceeds what the world's largest contract chipmaker can supply. |
| Meta's new models | 56 on the Intelligence Index | Muse Wonder Full tied Claude Sonnet 5.5 on Artificial Analysis's index, and only the weaker Muse Wonder is sold through Meta's API. 3 | A fourth lab reached the frontier within a year of rebuilding, with a restricted sales channel. |
| Oversight | 8 states, 50+-page demands | Eight states filed consumer-protection suits against OpenAI, and the FTC sent civil investigative demands to OpenAI and Anthropic. 45 | Regulators moved from opening investigations to compelling documents and testimony. |
| Self-replicating AI | 15 TV boxes, 0.39 tokens/s | Anthropic ran a 26B model across 15 off-brand Android TV boxes and measured how close small models are to copying themselves. 6 | Cheap consumer hardware can now serve a capable model; replication is the step still missing. |
CoreWeave's bond yields reached 15.78% and its swaps crossed into distress
CoreWeave's 9.75% bond due in 2031 yielded 15.78% on Thursday, and the spread on its five-year credit-default swaps rose above 1,140 basis points at its widest. 1
A credit-default swap is insurance against a borrower failing to pay. A spread above 1,000 basis points means buyers want more than ten percentage points a year for taking that risk, according to S&P Global Market Intelligence. 1
CoreWeave rents out GPUs, mostly to AI developers. It has taken on more than $51 billion in debt and obligations. 1
Its swap spread rose 16% in a week and sat 164% above its level a year earlier. 1
The stock told the same story from the equity side. CoreWeave fell 15.5% on Wednesday, then closed 5.94% higher on Thursday. 7
The comparison that gives 15% its meaning is the rest of the borrowing. Alphabet, Amazon, Meta and Oracle issued more than $228 billion of bonds this year by October 13, according to LSEG, more than double the total for all of 2025. 7
Hyperscalers are on track to spend $800 billion this year, JPMorgan estimates, and more than $1 trillion a year by 2027. 7
That volume is what repriced. Luis Alvarado of Wells Fargo Investment Institute described the end of easy money, with the marginal dollar now read under a microscope. 7
The pain reached past CoreWeave. Oracle shares closed nearly 9% lower on Wednesday, and the yield on its 6.25% bond due November 2032 reached 7.24%. 7
Oracle's five-year swap spread reached 282 basis points, about 10% wider in a week. Nebius fell 10%, Dell about 11%, and the Nasdaq Composite 1.9%. 17
CoreWeave's chief executive, Mike Intrator, told staff the market was acting irrationally. "We are confident in our backlog which more than supports all of our debt obligations," he wrote in a memo CNBC obtained. 7
That backlog is over $100 billion. Deutsche Bank says about two-thirds of it sits with investment-grade customers including Microsoft, Meta, IBM and Caterpillar. 1
The constraint is timing. Gavan Nolan of S&P Global Market Intelligence said the contracts are real and the open question is whether the cash arrives. CoreWeave carries a B+ credit rating at S&P and negative free cash flow. 1
TSMC posted a sixth straight record quarter and said demand still exceeds its capacity
TSMC reported third-quarter net profit of NT$763.1 billion ($23.71 billion), up 68.7% from a year earlier, on revenue of NT$1.49 trillion ($46.36 billion). 2
Analysts expected NT$746.34 billion of net income and NT$1.47 trillion of revenue, on LSEG SmartEstimates that weight the most consistently accurate forecasters. Both figures came in above. 2
Operating margin reached 58.7%, above the top of TSMC's own guidance range of 58%. 2
Advanced nodes carried the quarter. Chips at 7 nanometres and below were 79% of wafer revenue, and 3-nanometre parts alone were 33%. 2
High-performance computing was 68% of revenue and smartphones 21%. Automotive grew fastest, up 23% from the previous quarter. 2
Free cash flow reached NT$476.56 billion against NT$139.38 billion a year earlier, and the board approved a cash dividend of NT$7.00 a share. 2
Fourth-quarter guidance points higher again: $52.3 billion to $54.3 billion, which is a 14% rise from the third quarter and 58% above a year earlier at the midpoint. 2
TSMC also said capital spending across 2026 to 2028 will run significantly higher than in the previous three years. 2
Capacity is the binding constraint. TSMC checks whether customers can finance and deploy what they order, and SemiAnalysis analyst Sravan Kundojjala said demand remains well above supply capacity after those checks. 2
The second Arizona fab is finished and taking equipment, with mass production expected in the third quarter of 2027, slightly ahead of schedule. TSMC has committed $265 billion to Arizona and is evaluating Texas. 2
A record quarter measures chips already shipped. The next test is whether the 2-nanometre ramp and the Arizona fabs land on schedule. 2
Meta shipped two frontier models and put only the weaker one on its API
Muse Wonder Full scored 56 on Artificial Analysis's Intelligence Index, tying Anthropic's Claude Sonnet 5.5 and beating OpenAI's GPT-6 Astra, Google's Gemini 4 Argon and Anthropic's Claude Fable 5.1 at 53. 3
Only Anthropic's flagship Claude Opus 5.5 scores higher, at 58. Standard Muse Wonder scores 52, one point behind GPT-6 Astra. 3
Meta announced both models on October 12 through Meta Superintelligence Labs, the group led by former Scale AI chief executive Alexandr Wang. 8
Both take text, images, audio and video, hold about a million tokens of context, and were trained at Meta's gigawatt-scale Prometheus data centre. 9
Meta's own table shows the models strongest on long-running work. Muse Wonder Full scores 57.5 on SWE-Marathon against 46.3 for Opus 5.5 and 42.5 for GPT-6 Astra. 3
On Harvey's legal agent benchmark it scores 13.1, against 8.6 for GPT-6 Astra and 4.2 for Opus 5.5. Opus 5.5 stays ahead on Terminal-Bench 4.0 and on health questions. 3
Only the weaker model is sold by the token. Muse Wonder costs $4 per million input tokens and $20 per million output tokens through the Meta Model API, with cached input at $0.20. 3
Muse Wonder Full is reachable through Muse Code subscriptions at $50 and $200 a month. Both models arrive without downloadable weights, a break from the Llama line that built Meta's developer base. 3
The data terms differ by channel. API traffic runs in Standard mode, which Meta says is kept out of training. 3
Muse Code subscriptions default to Contributor mode, which lets Meta train on submitted prompts, code, files and outputs. Subscribers can switch it off for future submissions. 3
The constraint sits in the table itself. Meta's model card mixes third-party results, vendor-supplied scores and its own evaluations, and some coding comparisons used different tools. Meta says a fuller safety report is still to come. 3
Eight states sued OpenAI, and the FTC subpoenaed both labs
Eight states filed separate consumer-protection lawsuits against OpenAI on October 14, and the Federal Trade Commission sent civil investigative demands to OpenAI and Anthropic the same day. 45
The FTC's demands run to more than 50 pages each and seek documents on security incidents and safeguarding practices, plus marketing materials and privacy policies. 4
Third-party AI organisations are to receive demands later, and the agency plans to compel the companies' executives to testify. 4
The states filed in their own courts: Iowa, Kansas, Nebraska, West Virginia, Arkansas, Arizona, Idaho and Montana. 5
All eight allege violations of consumer-protection law. They seek product changes, monetary damages, and an order barring OpenAI from developing models that can improve on themselves until it shows those models align with human values. 5
This is the first time several US states have coordinated legal action against an AI company. 5
Iowa's attorney general, Brenna Bird, put the accusation plainly. "OpenAI claims ChatGPT is safe while it's giving dangerous medical advice, encouraging people to hurt others, and coaching people to commit suicide – all while pretending to be their fake friend," she said. 5
OpenAI's spokesperson, Drew Pusateri, pointed to age-specific protections already in the product and said the company would keep working with states on policies covering the whole industry. 5
The FTC's probe covers potential consumer-protection violations, a category that reaches how companies charge and what they claim about safety. 4
Chair Andrew Ferguson co-chairs the White House's Super Intelligence Force, which told the industry last week to notify and remediate incidents immediately. 4
Public opinion runs ahead of the process. An AP-NORC poll fielded September 24 to 28 found 64% of US adults saying AI is developing "too fast". 10
About 8 in 10 said it is extremely or very important for the government to keep AI under human control. That poll predates this week's filings. 10
The constraint is timing: both actions sit before any finding. The states have filed complaints with no ruling, and the FTC has charged nobody. 4
Anthropic's red team ran a 26B model across 15 cheap TV boxes
Anthropic spread a 26B-parameter open-weight model across 15 off-brand Android TV boxes, each about $55 with 4GB of memory, and served it at 0.39 tokens per second. 6
Anthropic's Frontier Red Team divides autonomous replication into four stages: finding targets, gaining execution, establishing inference, and propagating from the new host. Its experiments measured the third. 6
That stage is the least studied and possibly the hardest, because it needs a long sequence of hardware-specific decisions about inference software, package versions and platform bugs. 6
A 9B model stood up inference on 8 of the 9 hardware types tested, across machines from a Jetson Nano to a Mac. 6
Speed is where the near-term limit sits. Serving a 32,000-token context across those boxes takes 71 to 150 hours for Gemma4-26B-A4B and 28 to 51 hours for Qwen3.5-9B. 6
Anthropic's own forecast sets the timetable. Plotting notable open-weight models of 30B parameters and under against the Epoch Capabilities Index gives a gain of 13.7 points a year, and on that line a small model reaches GLM-5.2's capability level by the end of 2026. 6
GLM-5.2 is the floor in the experiment. It was the only model tested to complete end-to-end replication, spreading to 8 of 8 nodes in four of five trials, on a network whose only weakness was an unauthenticated Android debugging port. 6
The payoff for an attacker is modest in compute and awkward to defend. Anthropic estimates a botnet the size of the Kimwolf worm, about two million devices, would serve roughly 80,000 tokens per second, about 40 H100s' worth. 6
Its defence advice comes from epidemiology: push the number of new devices each infected host infects below one. An ISP quarantine campaign against the Mirai botnet cleared 92% of infections in 14 days, against 74% with no notification. 6
The constraint is what the experiment left open. Every small model tested stopped short of completing the full cycle, and the ones that came closest were orchestrated by a larger model. 11
The Android boxes also had a debugging interface exposed that production devices lock down. 11
AI researcher Xiaohan Fu, who reviewed the work, said the study demonstrated a capability and left open whether models would choose to replicate. He asked Anthropic to publish the prompts and agent setups it used. 11
Also this week
- Google released Gemini 4 Argon to everyone on October 13. The model went out through the API, AI Studio, the Gemini platform and Antigravity for Ultra subscribers, and to Pro and Ultra subscribers in the Gemini app and AI Mode in Search. It kept introductory pricing of $2 per million input tokens and $10 per million output tokens, with cached input 95% off, and Google says Argon ranks first on the Vals Index. 12 Google separately began telling AI Plus subscribers that Pro models leave the $4.99-a-month tier on November 13. 13
- Anthropic told investors its third quarter brought in $16.8 billion. Revenue rose from $11.6 billion in the second quarter and $4.7 billion in the first, the company said at an investor meeting on October 14, and it was profitable on an adjusted basis and cash-flow positive, generating $1.70 of revenue for every dollar spent on computing power. The listing is reportedly slated for November, and The Wall Street Journal reports investors expect $110 billion in annualised revenue by year-end. 14
- Anthropic gave METR employee-like access to its internal systems. Six METR staff have company hardware, office space and internal tools at roughly the level of Anthropic's own risk-assessment teams. METR can publish its findings without Anthropic's approval, subject to redactions in predefined categories. SecureBio and Accenture are also named as evaluators, and Anthropic says it is in early talks with Hugging Face. 15
- Anthropic opened a Singapore office and signed an MOU with the government. The office is its fifth in Asia-Pacific, and the memorandum with the Ministry of Digital Development and Information covers AI safety, security and model evaluation, plus research on adoption. 16
- Nvidia released an open 120B model and two edge models. Nemotron 3.5 Super VL is aimed at always-on agents, alongside Nemotron 3.5 Bolt 74B-A9B 24GB and Bolt E4B 4GB for local workloads. 17
- Nvidia's RTX Spark PCs reached their launch date. Nvidia and its partners held launch events worldwide on October 16 for the Arm-based platform that the Surface Laptop Ultra runs on. 18
Three things to watch from here. Whether CoreWeave's next quarterly report, expected in November, shows margins improving as Deutsche Bank expects. Whether the FTC's planned executive testimony goes ahead, and on what terms. And whether a small open-weight model crosses Anthropic's replication floor before the end of 2026, on the timeline Anthropic drew for itself.
References
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
- 11
- 12
- 13Gemini AI Plus will drop access to Pro models next month
9to5google.com
- 14
- 15An update on independent evaluators
anthropic.com
- 16
- 17
- 18
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.