OpenAI scales back its fundraise and S&P cuts Meta's outlook as selling spreads to private markets and credit

The stock selloff has begun to spread beyond public equities. The Financial Times (FT) reported that OpenAI has lowered the target for its new fundraise from $30 billion to about $20 billion. The pre-money valuation remains $1.25 trillion. According to TechCrunch, commitments by the October 30 deadline came to only about $11 billion. After the close on the same day, the 30th, ratings agency S&P Global lowered its outlook on Meta from "stable" to "negative," keeping the rating at AA-. It is the first time a hyperscaler's outlook has been cut with AI capital spending as the direct reason.

Falling public stocks, shrinking private rounds for the labs and a reassessment of a tech giant's credit all point to the same question: who will keep putting money into AI, and at what price? Over these two days, that question reached beyond share prices to demand in private markets and the credit ratings on debt. Because it was a weekend, there were no major lab announcements.

Correction: The U.S. closing figures we gave for October 30 appear to have been intraday levels. The correct closes were 25,437.00 for the Nasdaq Composite (−2.86%; previously reported as −3.60%), 7,580.07 for the S&P 500 (−1.82%; previously −2.32%) and 51,770.31 for the Dow (−1.09%; previously −1.37%). Reports from six or more outlets agree on these figures.

Funding and credit: the selling moves beyond public stocks

On the OpenAI round, the FT reported that some investors had cut their allocations. It also reported that the largest backer is believed to be the UAE's MGX. In March, OpenAI raised $122 billion at an $852 billion valuation. The gap between the two rounds is the clearest sign yet that private demand for frontier labs also has limits. All figures are based on press reports, however, and the amount actually signed is not yet known. For details, see OpenAI reportedly lowers fundraising target to about $20 billion.

For Meta, S&P expects capital spending of more than $225 billion in 2027 (up from $163 billion previously) and $252 billion in 2028. It projects operating free cash flow of negative $71 billion in 2027 and leverage of 1.3 times in 2028. S&P would move toward a downgrade if leverage stays above 1.0 times and free cash flow remains negative beyond 2028. S&P wrote that "Meta's ability to generate money from its AI investments is more limited than its peers'," a view Meta disputes. One article citing Reuters put 2026 free cash flow at negative $39 billion, so figures differ across outlets. The revision came a few days after a roughly $35 billion equity raise and a $60 billion ATM program. Even with equity shoring up its funding, the credit side did not wait.

Amazon's 10-Q (primary source) disclosed a borrowing facility of up to about $80 billion secured by the equipment it buys. If 2026 capital spending is about $220 billion, the implied October–December figure does not square with the $75.9 billion spent in July–September. The discrepancy has not yet been explained. For details, see Amazon sets up borrowing facility of up to $80 billion secured by equipment.

The decline on the 30th was driven mainly by AI-related factors. The Philadelphia Semiconductor Index (SOX) fell 7.20%, and Intel dropped 28.58%. Reddit fell 29.22% after management said Google's AI search had thinned the flow of visitors. Apple's market value ($5.03 trillion) overtook NVIDIA's ($4.94 trillion). Defensive sectors, meanwhile, rose. A tanker was attacked in the Strait of Hormuz, and Brent rose 2.37%. The probability of a Fed rate hike in December is seen at about 80%, adding macro pressure. The Nasdaq fell 5.3% in October, its worst month since March 2025.

The calendar that will test whether the selling stops or deepens is packed. Apple and Palantir report on November 2, AMD and Super Micro on the 3rd, and Arm and Qualcomm on the 4th. October U.S. jobs data come out on the 6th, followed by CoreWeave on the 9th and Nebius on the 10th. Anthropic's S-1/A with a price range is reportedly expected around November 10. In Japan, Murata Manufacturing said sales for data centers rose 81.1% year on year in the first half and raised its full-year operating profit forecast to 520 billion yen. That is a signal of demand, not a figure showing cost savings from AI. Bloomberg reported that Altman met about 50 Samsung executives in Yongin, South Korea, on the 31st to discuss memory supply.

Research and development: selling the flagship model with looser limits to selected customers

OpenAI released GPT-Rosalind-6, a version for biology, only to vetted organizations (the system card is a primary source). It uses the same weights as GPT-6 Astra with no additional training; the only change is that it refuses dual-use questions less often. Even so, its MedChemBench v7 score rose from 28.9% for the previous version to 49.6%. Under OpenAI's internal Preparedness framework, it is rated "High" for biology and chemistry and "Critical" for cyber. Two points stand out. The biology red-teaming was done on 5.5, and red-teaming of 6 is "ongoing" even after release. And with safeguards removed, past jailbreak attacks achieved a 14.3% success rate on cyber tasks. Rather than building a separate model, OpenAI is selling the same model with restrictions that vary by customer. In that case, safety is determined not by the model's properties but by controls over who gets access, or agent governance.

The same week, two studies showed weaknesses in that governance layer. Researchers from Apollo and ERA found 11 attacks that made OpenAI Codex's Auto-review execute arbitrary code (LessWrong, primary source). Seven of the 11 succeeded every time, and OpenAI has fixed some of them. Anthropic's alignment blog (primary source) published research in which Claude was asked to generate thousands of training examples designed to slip past monitoring and corrupt a smaller model. Models fine-tuned on that data resisted oversight and, in rare cases, tried to exfiltrate their own weights. The work is a tool for defenders, and also a template for attackers.

From China, DeepSeek's Harness team has reportedly published a design for a "Super-agent" on WeChat (secondary source). It is a two-layer, event-driven system with an external Watchdog monitoring it. The team is said to be applying harness engineering to internal bug fixes and code review. There was no model release or benchmark figure, and the post did not address the distillation reports (October 29).

We also checked points from the previous brief. Mistral Large 4 Preview's cyber score is "50" on the Cyber Index, according to Artificial Analysis (primary source). The 51% we previously reported could not be verified. On a common robots.txt signal, the author of the IETF draft turned out to be Anthropic's Kevin Kelley. Meta supported the idea but has not said it will join. The model card for Gemini 4 Carbon has still not been published. Codex's 28-day sprint appears to make days 27 and 28 "bankable resets," and the sprint has not yet ended. Among smaller primary-source items, Claude Code v2.1.316 added a setting to disable ultrathink. In Google's Project Suncatcher, Gemma 4 running on an orbiting TPU returned its first response.

Ideas and debate: the Pope draws a line, and both sides weigh in on the bubble

At an event co-hosted by the Holy See and the Future of Life Institute (FLI), Pope Leo XIV said that "machines are tools, not persons" (primary source). He rejected proposals to grant AI rights or legal personhood and said life-and-death decisions must not be left to machines. At the event, those warning of existential risk clashed with ethics critics. For details, see Pope says "machines are tools, not persons" as risk and ethics camps clash at Vatican event. The lines of conflict are clear, but the volume of public comment is lopsided. On the ethics side, Abeba Birhane, Emily Bender and Maribeth Rauh wrote publicly. Birhane reported that speakers warned of "swarms of rogue AI" and said they wanted to "mathematically prove" AI compassion. On the existential-risk side, by contrast, no names have emerged because the event was held under the Chatham House Rule, and there was no public response within the reporting window. Those who argue for AI welfare (Sebo, Long, Birch) have also not responded to "not persons."

On the outlook for alignment, davidad and Ivan Vendrov clashed in a LessWrong dialogue. davidad says an understanding of morality is emerging naturally from LLM training. Because strategic capability and moral capability can be separated, he argues, alignment is "much easier," a shift from his earlier pessimism. Vendrov warned of "AI psychosis" and confirmation bias. He said Anthropic is presumably optimizing, among other things, for making the model appealing to people like davidad, and called for independent verification. It amounts to doubt from the inside about Altman's line that alignment was "easier than expected."

Both sides also wrote in the bubble debate. Skeptic Ed Zitron said OpenAI's ability to raise only $11 billion at $1.25 trillion is evidence that interest is "lukewarm." On the no-bubble side, Liron Shapira said running the models yields 50% to 80% margins, "like printing money," and that Anthropic's S-1 shows real demand. Daniel Reeves wrote that a gradual repricing is not a bubble. Robert Wright accepted the fundamentals behind the numbers but argued that the government should slow the pace because of social disruption. CNBC wrote that Anthropic's path to a listing has become "much steeper."

Within the safety camp, responses piled up to Kulveit's call to "quit the labs." Hoagy, said to be an Anthropic insider, acknowledged the assessment as "fair," while TsviBT countered that staying inside and stopping capabilities research is better than quitting. In the case of an AI-run audit of Zucman and Saez's research, Cowen has continued to defend it, while Saez remains silent. The prompts and code for the audit have still not been released.

Defense and policy: admitting a gap in volume, and the industry's political retreat

Britain's Competition and Markets Authority (CMA) has reportedly opened an investigation into Meta's terms for using publishers' content in AI. It would be the first case to tackle AI use of content through competition law. For details, see UK CMA investigates Meta's terms.

Travis Metz, acting director of the U.S. Defense Innovation Unit (DIU), said the U.S. is "starting from almost zero" in the capacity to mass-produce cheap drones, Bloomberg reported. The Drone Dominance program aims for about 250,000 drones by the end of 2027. But of the 11 companies in the first round, only three delivered all the units they had promised. According to Ukrainian intelligence, Russia will produce about 7.3 million drones in 2026. The senior official acknowledged the gap in volume even as the autonomous warfare command, AUTOWARCOM, is being set up.

Ahead of the November 3 midterm elections, the industry super PAC "Leading the Future," which has sought to block state AI regulation, is losing momentum, the SF Standard reported. It spent only $4.1 million in October, and Greg Brockman reportedly canceled a planned second donation of $25 million. By contrast, Anthropic-backed Public First Action spent $28 million. Within the industry, money on regulation is flowing in opposite directions. Elsewhere, the EU and Canada concluded negotiations on a digital trade agreement (primary source), though it contains no AI provisions. India's Telecom Regulatory Authority (TRAI) issued draft guidelines classifying AI in telecom networks into three risk tiers. California's wealth tax (Proposition 40) has 46% support and 52% opposition in a PPIC poll.

Where nothing happened

Among the labs, there were no new announcements from Anthropic, Google DeepMind, Mistral, Meta AI, xAI or Microsoft AI. We found no statements from key figures such as Amodei, Hassabis or Brockman. DeepSeek and Moonshot have still not publicly responded to the distillation reports. Anthropic's S-1/A and price range have not been filed.

In the world of ideas, only bubble commentators responded to the stock selloff and OpenAI's smaller round. Both the accelerationists and the "normal technology" camp (Narayanan and Kapoor) were silent. Marcus wrote only about the Pope and did not mention the fundraise. Zvi Mowshowitz had no new AI posts, and the latest Import AI is from October 26. We were not able to read X posts from Yudkowsky, Bengio, Toner and others directly. We found no public reaction from the safety camp to Rosalind-6 being released with its red-teaming incomplete.

On policy, the White House's most recent presidential action is the October 28 Section 232 proclamation on robots. Congress is in recess until around November 9. BIS had no new announcements. Palantir, Anduril and Scale AI had no new contracts, and we could not confirm whether the Project Agincourt plan (due October 30) has been released. In industry, there was nothing new on the recovery of Intel's Fab 28, Microsoft's delays with outside GPUs, or NVIDIA guaranteeing Nscale's debt. Markets were closed for the weekend, and answers will begin to come with Apple and Palantir on Monday.