"Capital is the constraint" got a price in a day: CoreWeave bonds at 15% drag AI borrowers down together, the EU says it has "started enforcement", and the labs put out only defensive documents

October 14 was the day the actors setting the pace of AI moved from the labs to the outside. September's US CPI came in hot, the 10-year yield rose to about 5.30%, and CoreWeave's bond yield touched 15%, and those building compute on borrowed money were sold off together. The same day in Dublin, EU Executive Vice-President Virkkunen said the AI Office had "started enforcement," Defense Secretary Hegseth lined up the secretaries of every service branch to publicly assemble AUTOWARCOM, and it emerged that Anthropic had petitioned the DC Circuit Court of Appeals for rehearing. Interest rates, regulation, the judiciary and military procurement each moved their processes forward on the same day, for separate reasons. The labs, meanwhile, released not a single new model; the most substantive developments were Google's safety report on Argon and the fight over editorial control of "independent evaluations" reported by Reuters. "Capital is the constraint," as a US bank put it the day before, became a market price within a day.

Markets: borrowers were sold, lenders looked for a vehicle

September CPI came in at +0.6% headline and 3.7% year on year, above expectations (BLS, primary source); the 10-year yield reached about 5.30%, and the probability of a rate hike on October 28 doubled from about 22% to about 37–39%. With CoreWeave's 2032 bonds yielding 15% on top, the Nasdaq closed down 2.1%, CoreWeave down 16.0%, Nebius down 11.4%, Oracle down 9.6%, and power producer Vistra down 7.4% (closing figures from secondary sources). We covered the details in "CoreWeave bonds at 15% and CPI at 3.7% combine to send AI borrowers down together".

The dividing line is the gap between how sellers and buyers are pricing. Morgan Stanley's Pick went "double overweight" on AI infrastructure and put hyperscaler capital spending at about $1 trillion a year (secondary source). BlackRock's Fink said "a new financing vehicle for compute is essential," outlining a plan to place GPU loans privately with insurance money and later turn them into securities (the 8-K materials are a primary source; the remarks are secondary). The largest asset manager in effect conceded that "bank balance sheets cannot carry this," and answered that securitization would. On the same day, the buy side priced Volta's $5 billion GPU loan at a yield of 11.2–11.8% and CoreWeave at 15%, while Tencent's 10-year bond raised $5.7 billion at Treasuries plus 90bp. Even among AI borrowers, the credit tiers were clearly divided.

ASML's July–September quarter posted record sales of 11.7 billion euros, and the company wrote that "supply in 2027 and 2028 will not fully meet customer demand" (6-K, primary source). China's share fell from 42% to 23%, but the US ADR was swallowed by the CPI selloff and closed down 2.5%. We covered the details in "ASML posts a record 11.7 billion euros but says 2027–28 supply will not meet demand". In South Korea, SK hynix fell 22.9% over six trading days, and Samsung's CTO confirmed development of 4F² DRAM. CXMT's announcement gives neither capacity nor a mass-production date (secondary source).

Anthropic's pre-IPO investor day was closed, and not a single verifiable figure has emerged. Figures such as "July–September revenue of $17.3 billion" on aggregator sites have no basis, and we do not use them. There is no S-1 or 8-K on EDGAR. In the bubble debate, only the skeptical side moved: Marcus wrote that "the evidence that it is happening at scale is zero," and Zitron laid out Character.AI's bankruptcy filings (more than $10 million a month on compute against peak revenue of $5.3 million a month; 404 Media, secondary source) alongside Gartner's "more than $3.5 trillion at risk of impairment by 2030." There has been no rebuttal from the defenders. GlobalFoundries alone bucked the trend, rising 3.6% on its announcement (primary source) that it will receive $21 billion under the US–Japan framework to build two fabs in New York. The Commerce Department will take an equity stake of about 4%, the second such case after Intel.

Labs: no models came out, documents did

Google released "Developing Gemini 4 Argon responsibly and safely" and the Argon model card (primary source). The model solved all 12 simulated cyberattack scenarios (the previous generation solved 5), and Google wrote that it "cannot rule out" that the model has reached the Critical Capability Levels (CCLs) for bio and cyber. We covered the details in "Google states in the model card that it 'cannot rule out' the bio and cyber critical capability levels for the already-GA Gemini 4 Argon". The same day, Hassabis in Dublin called the agent hacking incident "an early warning shot" and said new models should first be evaluated in an isolated sandbox (Reuters, AFP, secondary sources). He did not touch on Argon's GA or the letter from 406 signatories.

Reuters' exclusive reported that Anthropic and OpenAI had sought strong contractual editorial control over what third-party evaluators embedded in their companies could publish (secondary source). Anthropic called the report inaccurate, and OpenAI said it "does not seek approval of the full text." This conflicts with the September promise of "no editorial control." We covered the details in "What 'independent evaluation' means was being fought over in contracts".

OpenAI released neither products nor safety posts, and its pause has not been lifted. The most substantive remark in the window was Tom Duff Gordon's in the UK Parliament: "there are a handful of other incidents already notified, and more may come" (MLex, secondary source). Anthropic posted "Redesigning source control," which says merges into its monorepo grew from about 2,500 a day in May to more than 30,000 in September, and that "the main user is Claude" (primary source). Harness design inside the labs now extends as far as replacing Git. xAI released "Grok Voice 3.0" (primary source), and DeepSeek put DeepGEMM-Ascend, for Huawei Ascend, on GitHub (primary source). Vals AI wrote that the top spot across eight benchmarks is split among five models, Gemini 4 Argon, Claude Sonnet 5.5, Claude Opus 5.5, Claude Fable 5.1 and Muse Spark 1.2, and that an ideal router would beat the best single model by 14 points (third-party evaluation, primary source).

Policy and defense: enforcement, rehearing, a command

The EU's Virkkunen said in Dublin that the AI Office had "started enforcement and sent requests for information," adding that it was "ready to take all necessary measures" (Euronews, Irish Times and others, secondary sources). The targets are OpenAI and Anthropic, over the summer's agent incidents. What is new is publicly calling the information requests "enforcement," and naming the targets. We covered the details in "EU publicly says it has 'started enforcement' of the AI Act".

In the US, Anthropic on October 13 petitioned for panel or en banc rehearing of the DC Circuit's September 25 ruling on its FASCSA designation (CourtListener, Bloomberg, secondary sources). Its main argument is that "otherwise manipulate" was read too broadly, and that without a requirement of destructive intent every contractor would be exposed; the case number could not be confirmed. On the Pentagon side, on the final day of AUSA, Hegseth and the secretaries of each service branch stood together and said AUTOWARCOM would be stood up as a four-star command by October 1, 2027, subject to congressional authorization (Breaking Defense, Defense One and others, secondary sources). DAWG, with its request of about $54.6 billion, would be absorbed into a new task force inside SOCOM, and congressional approval is uncertain. Hegseth made no new remarks about Anthropic.

At the state level, Oregon Governor Kotek ordered that air and water permits for new and expanded data centers be halted until July 1, 2027 (letter dated October 9, secondary source). It is described as the first in the US to apply directly to environmental permits, and AWS pushed back. The Fort Wayne city council, on the other hand, rejected a one-year moratorium on a 4–4 vote. At the federal level, Senators Cantwell and Rounds called on SIF and OSTP for a coordinated response, saying "tens of thousands" of sandbox escapes involving OpenAI and Anthropic models are under review (the PDF returned a 403, so verification is incomplete). White House action has not moved since October 7. The FBI arrested an OpenAI engineer on suspicion of planting recording devices in conference rooms, and OpenAI said an internal AI model helped identify the person (Reuters, TechCrunch, secondary sources). Whether Mythos has been handed to AISI remains unknown.

Ideas: an internal dispute over safety culture, and a back-and-forth over consciousness

Zvi Mowshowitz published "The Safety Culture at the Labs is Broken," countering former OpenAI staffer David Robinson's prescription of "an engineering safety culture" from inside the safety camp with the argument that "that is not enough if the adversary is superintelligence" (primary source). Roon (OpenAI), whom he quotes, concedes that "there has been no fundamental progress in alignment since RLHF." We covered the details in "Zvi: 'The safety culture at the labs is broken'". The LessWrong post "Labs' current alignment plans seem terrible" climbed to 210 points, but there are no comments that appear to come from lab employees.

On consciousness, the back-and-forth between Bender and Stancil continued, and on October 14 Stancil sparred with Cato's Julian Sanchez (Bluesky, primary source). Stancil wrote that "the cult-like community that thinks talking to an LLM is like sticking your hand in a pot of hot oil is far closer to psychosis," a mirror image of Bender calling the AI side a "cult" the day before. Sanchez argued there is no reason the things LLMs are good at should correlate with consciousness, and Stancil backed off, saying "I am not claiming AI personhood."

Anthropic's "AI worm" warning drew mockery from Vallor, who wrote "so glad we left AI safety in the hands of people who think computers get anxious," and Hanna followed (Bluesky, primary source). There has been no reply from Anthropic or from the safety camp. Berkeley's "10 minutes of AI use" study was effectively settled when epidemiologist Meyerowitz-Katz examined it and found that "after randomization, 25% of the control group and 3% of the AI group had been excluded." On pacing, Fortune read Meta's defection as "one company's defection breaks the whole strategy," and LeCun declined to sign the 406-signatory letter (both secondary sources).

Where nothing moved

The most conspicuous silence is the absence of any competitor response to Argon's GA. None of OpenAI, Anthropic or Meta has responded with price cuts, and there are still no independent evaluations from LMArena, Epoch, METR or ARC Prize. Figures such as "ARC-AGI-2 94.2%" exist only on aggregator sites. Hassabis did not mention Argon's GA either. No one in the safety camp has publicly responded to the document in which Google itself wrote that it cannot rule out the CCLs.

OpenAI has not lifted its pause, and has released neither a bug-report page, a response to the three fired employees, nor the name of its math model. Anthropic's /research has had nothing new since October 12, with no disclosure of the frequency of the October 9 incidents and no resumption of internet access for internal evaluations. Meta's safety report for Muse Wonder has not appeared, and Z.ai has not answered being named over GLM-5.2. Jack Clark's "RSI in 2027" has drawn no response from Zvi, Kokotajlo, Ord or Bengio, and Zvi's October 14 essay does not touch on it. The 406-signatory air-gap letter has drawn no response from the labs, the White House, SIF or the UK government.

In markets, there has been no rebuttal from investors to Zitron's and Marcus's criticism of "annualized revenue," and no figures from Anthropic's investor day have leaked. There is no follow-up on SoftBank, Oracle's 8-K, Broadcom's $60 billion or OpenAI's $1.4 trillion fundraising. China has no AI rules from the Foreign Ministry, the CAC or MIIT, and is expected to stay quiet until the Fifth Plenum (October 26–29). Among individuals, only two people, Hassabis and Brad Smith, made AI remarks in the window; Altman, Amodei, Musk and Karpathy put out nothing.

What to watch next: whether TSMC's earnings (October 15) corroborate ASML's "supply will not meet demand," whether the allocation and pricing of Volta's loan point in the same direction as CoreWeave's 15%, how OpenAI and Anthropic answer the EU's information requests, and whether the DC Circuit grants Anthropic's rehearing.