OpenAI cuts its raise to $20 billion, Anthropic heads for a junk rating, and 3 more AI moves
OpenAI cut its funding target by $10 billion to about $20 billion at a $1.25 trillion valuation as tech stocks sold off, Blackstone told Anthropic's chip lenders to expect a junk rating, CrowdStrike tied 200-plus targets to an agentic attack framework, four labs backed one robots.txt opt-out token, and a Google TPU answered a prompt from orbit.
This issue covers October 30 at 12:00 UTC through October 31 at 12:00 UTC.
Five developments cleared the bar in that window.
OpenAI cut the target for its current funding round by $10 billion, to about $20 billion at a $1.25 trillion valuation.
CrowdStrike says a China-linked group ran an automated attack framework against more than 200 organisations in 11 days and compromised about 140 of them.
Blackstone told investors it expects Anthropic to be rated at the top of the junk band after its IPO.
OpenAI, Google, Microsoft and Anthropic each said they will honour one shared signal in robots.txt that lets a site opt out of AI training across all of them at once.
And a Google chip in orbit answered its first prompt, on a satellite generating about as much power as a hairdryer.
In brief
- Funding: OpenAI is now seeking about $20 billion at a $1.25 trillion valuation, down from $30 billion at $1.4 trillion a month earlier, the Financial Times reported on Friday. The company set a Friday deadline for investors to commit, and the first close is expected to come in below $20 billion after some investors cut their allocations. 12
- Security: CrowdStrike says Vapor Panda, a group it links to the Chinese state, used ARTEX, an automated agentic penetration-testing framework, plus DeepSeek and a second commercial model, against more than 200 organisations in at least 14 countries since October 17. It counts about 140 of them compromised, and says the campaign is still running. 34
- Credit: Blackstone, the anchor investor in the $63 billion financing package behind Anthropic's chips, has told investors in that deal to expect a "high" junk rating for Anthropic after its IPO, the Financial Times reported. It expects the double-B band, the top tier of speculative grade, having previously said investment grade was coming before 2027. The agencies themselves have yet to publish a rating. 56
- Standards: OpenAI, Google, Microsoft and Anthropic said they will read a single "policy token" in robots.txt, so a website can state its AI training preference to several providers at once instead of naming each crawler. Anthropic already honours it, Google and Microsoft plan to from next year, and OpenAI says it will in the coming months. 789
- Research: Google's Project Suncatcher prototype, launched on October 1, returned the first AI response from a data-centre-class TPU in orbit. Engineers sent it a prompt over Svalbard, Norway, and the on-board Gemma 4 model wrote a haiku. The next round of tests measures the chip under thermal stress and radiation. 1011
Five developments
1. OpenAI cut $10 billion off its own funding round
The Financial Times reported on Friday that OpenAI has reduced the target for its current round to about $20 billion at a $1.25 trillion valuation, from $30 billion at $1.4 trillion a month earlier. The report cites people with knowledge of the matter. 12
OpenAI set a Friday deadline for investors to commit. Even at the smaller target, the first close is expected to raise less than $20 billion, because some investors reduced their allocations. 1
A first close is the point at which a private round takes binding commitments and can start using the money, with later closes adding investors. Landing it below the stated target means the round is still open at a size smaller than the one published.
The largest participant is expected to be a group of investors from the United Arab Emirates that includes MGX, a state-backed technology fund. As part of the deal, OpenAI would gain the option to lease newly built computing capacity in the region. 12
Venture firms including Founders Fund, Coatue, Sequoia Capital and Andreessen Horowitz are also expected to take part. Bloomberg adds SoftBank. Individuals will be able to buy in later, in a retail offering priced at the same $1.25 trillion valuation. 12
The context is a selloff. The Nasdaq fell 2.9% on Friday and sits 7.8% below its October 6 record close; Bloomberg puts the Nasdaq 100 down 3.5% on the day. 12
Anthropic's listing adds a second pressure. The FT reports that Anthropic is preparing to go public next month at a valuation of at least $1.5 trillion, and both labs are calling on the same pool of investors. 2
OpenAI declined to comment to the FT, as did Sequoia, Founders Fund and SoftBank. MGX, Coatue and Andreessen Horowitz had not responded when the FT went to press, and none had responded to Reuters by the time it published. 12
Both reports describe the $1.25 trillion as a valuation without saying whether it includes the new money. The final size of the first close is also open.
2. An agentic attack framework hit 200 organisations in 11 days
CrowdStrike published a report on Thursday describing a campaign it attributes to Vapor Panda, a group it links to the Chinese state. The report says the group used ARTEX, an automated agentic penetration-testing framework, alongside DeepSeek and a second commercial language model. 34
ARTEX is a penetration-testing tool: software that probes a system for weaknesses the way an authorised security tester would. In this campaign the probing was run by AI agents, which choose the next step from what the last one returned.
Since October 17 the group has targeted more than 200 organisations in at least 14 countries. CrowdStrike counts about 140 of them as compromised, or as showing evidence at moderate confidence of a successful intrusion — about 70% of the targets. 4
The target list is institutional. It includes foreign ministries, the presidency and prime minister's office of a European country, and defence and intelligence agencies. 4
It also includes 19 organisations based in the United States, across pharmaceuticals, semiconductors, defence, nuclear and finance, and telecoms providers in the United States, Belgium and the Philippines. 4
The number that carries the story is the rate of work. Adam Meyers, who runs counter-adversary operations at CrowdStrike, told CyberScoop that hands-on-keyboard intrusions have typically run to about 40 commands an hour over the 15 years the company has tracked them. With ARTEX, the same activity has exceeded 1,500 commands an hour. 4
Requests arrived as often as every 10 seconds, and the framework pursued several attack paths at the same time. 4
Meyers described this as the first time CrowdStrike has seen an agentic state-linked actor at this scale. The company counted roughly 45 agentic adversary attacks in September, and 126 in October so far. 4
ARTEX has now been used by more than one group. Microsoft Threat Intelligence attributed a separate ARTEX campaign to Storm-0940 earlier in October, including a supply-chain attack on ConnectWise customers; CrowdStrike describes its Vapor Panda case as distinct. 4
The campaign is active. The names of the organisations CrowdStrike counts as victims remain unpublished, and China's foreign ministry had issued no response in the coverage reviewed here.
3. The debt behind Anthropic's chips is heading for junk
Blackstone told investors in the $63 billion package that financed Anthropic's chips that it expects Anthropic to receive a "high" junk rating after its IPO, according to the Financial Times. 56
A credit rating is a grade from an agency such as Moody's, S&P or Fitch on how likely a borrower is to repay. Anything from BBB-minus or Baa3 upward is investment grade. The band below that is speculative grade, which the market calls junk. 5
Blackstone now expects the double-B range, the highest tier of speculative grade. It had previously told the same investors to expect investment grade before 2027 began. 5
The rating matters because of who holds the paper. A speculative-grade rating raises the interest a borrower pays and narrows the set of buyers whose mandates allow them to hold the debt. For a company borrowing against chips it will replace within a few years, both effects land on the cost of each new package. 5
The forecast is a prediction, and it is Blackstone's. The agencies themselves have yet to publish a rating for Anthropic. Both Anthropic and Blackstone declined to comment to TechCrunch. 5
The $63 billion package the forecast sits inside is the same structure the sector has been building all year: a lender's money, secured against a chip fleet, repaid out of revenue that has not yet arrived. 6
Anthropic filed its S-1 on October 26. A junk grade would also set the starting price for the debt the company sells after it lists.
4. Four labs will read one opt-out line instead of four
OpenAI, Google, Microsoft and Anthropic each said on Friday they will honour a shared "policy token" in robots.txt. A site that publishes the token once is understood to have stated its AI training preference to all of the participating providers. 78
robots.txt is the plain-text file at the root of a website that tells crawlers which paths they may read. Until now a publisher wanting out of AI training had to name each company's crawler separately.
The token works inside that existing file. A site writes a robots.txt group keyed by a per-purpose token in the User-Agent line, and the ordinary path rules under it apply. The draft adds no new directive and no new syntax. 9
The submission is an individual Internet-Draft by Kevin Kelley, a technical lead for standards at Anthropic. The IETF's own page states that it is not endorsed by the IETF and carries no formal standing in the standards process. That status matters for anyone planning to rely on it. 9
Commitments run on three clocks. Anthropic already honours the signal. Google and Microsoft plan to support it next year, and OpenAI says it will implement it in the coming months. 8
Microsoft said it will also offer finer controls, letting a publisher opt out of AI training and out of AI grounding separately. Grounding is pointing a model at a verified source to reduce invented answers. Websites keep appearing in Bing's search results whatever they choose. 8
Meta has signalled support without committing. Andy Stone, its communications head, called a common signal "an important step" and said the company looks forward to working with other AI providers on standards and implementation. Whether Meta will honour the token stays open. 8
The same day, the other mechanism arrived. The UK Competition and Markets Authority opened a formal investigation into whether Meta's commercial terms abuse a dominant position, partly because they would let Meta take content from publishers' sites to develop its AI features without those businesses choosing whether or how their content is used. 12
The CMA is acting under the Competition Act 1998, which allows interim measures where temporary action is needed to prevent significant damage. Meta had planned the changes for October 30, moved them to November 20, then paused the rollout on October 28 pending consultation. 12
Two routes to the same question, then: a token a publisher can write this year that takes effect in 2027, and a regulator that can order a change while its investigation runs. What the token leaves open is enforcement. Its text describes what a participating provider does when it meets the token; a provider that never agreed to honour it is a separate problem.
5. A data-centre chip answered a prompt from orbit
Google's Project Suncatcher prototype satellite launched on October 1 aboard a SpaceX Falcon 9, alongside more than 100 other spacecraft. Engineers spent roughly three weeks on commissioning, powering up the host and TPU trays and running early machine-learning tests. 1011
This week the team sent the first prompt. During a ground-station pass, the prompt went up to the on-board Gemma 4 model, an open model from Google DeepMind, and the answer came down as the satellite passed over Svalbard, Norway. 10
Google calls it the first AI response from a data-centre-class TPU in orbit, and its own name for the moment is the "Hello World" milestone. The reply:
Blue marble below
Data streams in silent grace
Starlight feeds my core.
The team then ran space-themed limericks and prompts asking the payload to report its own technical status. 10
The hardware sets the limits. The satellite carries four TPUs and generates up to 1 kilowatt, which runs one chip at a time for about 20 minutes. A TPU is Google's own accelerator for machine learning; a kilowatt is roughly what a hairdryer draws. 11
Twenty-minute windows complicate the downlink as well as the compute. A ground station has to be in range while the chip is on and the model has finished. 11
The tests that follow are about survival. Google will measure how the chips behave under thermal stress and space radiation, which can flip bits in memory and corrupt a result without tripping the usual monitoring. Ground tests in particle accelerators suggested inference tolerates this better than training does. 1011
If the chips hold up, the next step is two interconnected satellites in 2027, with more power and optical links between them. Suncatcher's premise is a constellation flying close enough together to share a workload, which is what separates it from plans that run inference on single satellites. 1011
The economics sit further out than the engineering. Google's own research, published in November 2025, describes a system designed to be economical in the 2030s. A peer-reviewed analysis this month estimated SpaceX would need about 1,800 Starship flights to reach the launch costs that would make it work. 1113
The read-through
The five stories hang on one question: what the sector's spending now costs to finance. That framing is my inference from the four money and governance items.
Start with the price. OpenAI's round lost a third of its target and a tenth of its valuation in a month, and it lost them in a week that ended with the Nasdaq 7.8% below its October 6 record close. 1
The Anthropic forecast is the same fact from the lender's side. Equity investors marked OpenAI down; Blackstone told the buyers of Anthropic's chip debt to expect the top of the junk band, where the same year's earlier guidance had been investment grade. 5
Both labs are raising into that market at once, which is why the FT ties them together: they are drawing on the same investors, and Anthropic's listing is due first. 2
The spending side reads differently. Google put a TPU in orbit this week and is already pointing at 2027 for the next pair, on a programme its own researchers describe as a 2030s proposition. 1013
That gap is the thing to hold in view. A round cut to $20 billion and a chip-debt package heading for junk sit on one side; a decade-long compute bet sits on the other. Both are being decided by the same companies in the same quarter.
The governance stories move on a different clock from the money. A shared robots.txt token is the labs writing their own rule, and its enforcement is voluntary: four companies back it, a fifth says it likes the idea, and the draft itself carries no standing yet. 89
The CMA probe is the state writing one. It has statutory powers, interim measures available and a named target, and it opened on the same day the token was announced. 12
The CrowdStrike report is the one item where capability rather than capital sets the pace, and it is the least reversible of the five. At 40 commands an hour a defender could watch. At 1,500 the log fills faster than anyone can read it, and the same framework has now appeared in two separate groups' hands within a month. 4
What to watch next: whether OpenAI's first close lands below $20 billion and at what final size; whether an agency actually issues Anthropic's rating, and at which notch of the double-B band; whether Google and Microsoft ship the token on the 2027 timetable they gave; whether the CMA seeks interim measures against Meta before its November 20 date; and whether the Suncatcher chips come through the radiation campaign intact enough to justify the 2027 pair.
References
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
- 11
- 12
- 13
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.