One week. That's how long it took for roughly half a trillion dollars to come off what investors think Anthropic is worth, and that was before the public got to read the filing. Now the public can read it. And it's long. It's Tuesday, October 27th, and this is AI Daily Briefing. On the rundown: Anthropic's S-1 goes public, Meta raises thirty-seven and a half billion dollars with Nvidia writing a three billion dollar check, OpenAI claims three more proofs and brings in outside referees this time, and Naver lines up nine billion dollars of compute without buying any of it. One tap on follow, and we'll be back in your ears before you know it.
Manya Saini and Echo Wang, reporting for Reuters:
Some current and prospective investors are valuing Anthropic at about $1.5 trillion, down from roughly $2 trillion a week ago, after the AI startup disclosed some financial details at a recent investor event, Reuters reported.
Anthropic made its US IPO filing public on Monday. We went through the leaked draft last Wednesday: sixteen point three billion in first-half revenue and a net loss of more than eighty billion. Reuters says the public filing confirms those earlier exclusives. The new part is the price. The market has now seen the numbers, and it's marking the company down. A quarter off in a week, and it's still worth more than the nine hundred sixty-five billion it raised at in May. Only in this market does that count as a haircut. The commitments number grew too. The draft reporting had about four hundred fifty billion in cloud deals. A CODEW roundup of the filing coverage puts total compute and infrastructure commitments at at least six hundred fifty-seven billion, stretching as far as 2045. That includes a hundred sixty-one billion Broadcom equipment lease, at least a hundred fourteen billion to Google, a hundred ten billion to Amazon, and up to a hundred twenty billion to SpaceX. The roundup says the ticker will be ANT, on Nasdaq, with reports pointing to the week of November 16th. And the Broadcom piece got its financing on Monday. Per the same roundup, Bank of America, Citi and Morgan Stanley arranged senior debt and Blackstone took the second lien, sixty-three billion through a special purpose vehicle that buys Broadcom accelerators and leases them to Anthropic. Broadcom backstops the senior lenders against the hardware losing value. That's the Nvidia backstop story from yesterday, wearing a different logo. Read the risk factors. The Verge pulled one line from the filing: competition for power and for data center sites with enough of it is intense, and the buildout leans on natural gas. The roundup adds that high-bandwidth memory shortages are listed as a key risk. Everything we've said for weeks about power and memory being the real bottleneck, Anthropic has now put in a document it can be sued over. Reuters also points out the fall IPO calendar has thinned out because of market volatility. Anthropic is going anyway. OpenAI filed confidentially in June and is expected to list by early 2027. Whoever prices second gets to use the first one as a comp.
Ambar Warrick at Investing.com:
Separately, Meta said it had upsized an offering of common stock to $25 billion from $22.5 billion. The offering, coupled with a depository share offering, will now raise up to $37.5 billion.
The common stock priced at six hundred twenty-five dollars a share. Alongside it, Nvidia and Meta put out a joint announcement dated October 26th: an expanded partnership on the Muse models, and Nvidia investing three billion dollars in the equity offering. So Nvidia buys three billion of Meta stock, and Meta uses the money to buy, per the release, more Vera Rubin systems, Vera CPUs, Spectrum-X networking and BlueField-4. The money doesn't go very far before it comes back. Some continuity. Last Thursday we covered Meta creating a non-voting share class to fund AI, trading in 2027. Investing.com says that's a separate sixty billion dollar offering, on top of today's raise. Meta has guided to between one hundred thirty-five and one hundred forty-five billion in capex this year. What's in the partnership for builders: joint work on training and inference performance for Muse, plus Nvidia's OpenShell for safe deployment and Sentry for monitoring agent behavior. Those last two are listed as things the companies will consider. Considering isn't shipping. And watch the framing. The release says Muse has delivered frontier-level benchmark performance. The backing for that is Meta's own evals plus one independent leaderboard, which we covered yesterday, with Muse Wonder Full tied for second. Huang's quote calls agents software that understands our goals. That's a press release, not a finding.
The Verge, writing up OpenAI's announcement late Monday:
What's different this time is the choice to have external experts that OpenAI describes as "shepherds" guiding the results in computer science and physics, while for the math problem it says that they advised it to release the AI output for researchers to work through their own explanations.
Three claims, posted to GitHub. Per the AI Lately roundup: in math, a proof of local smoothing for the wave equation in every dimension, which would also settle the Kakeya problem. In complexity theory, a faster algorithm for 3-SAT that breaks the Exponential Time Hypothesis. And in physics, a proof that a family of four-dimensional quantum field theories, including quantum chromodynamics, exist. Parts of each were checked in Lean. The ETH one is the one I'd watch. P versus NP stays open, the new algorithm is still exponential, just less so. But a lot of hardness results assume ETH holds. If this proof survives, a pile of lower bounds need re-checking. That's the big if. On October 14th we covered Scientific American finding OpenAI's Navier-Stokes proof wasn't what it first seemed. And on October 6th Terence Tao criticized OpenAI's last mass release of math documents, saying few experts had read even one proof in full. So this time there are outside reviewers, and per AI Lately, two of the ones on the complexity result first proposed ETH. That's the right reviewer for that claim. But the math result skipped the shepherds. OpenAI is handing the raw output to researchers and letting them write the explanation. That's a lot of unpaid refereeing. My position until the reviews come back: these are OpenAI's claims, partly machine-checked, and nobody has refereed them yet. AI Lately expects that to take weeks.
Ho-jung Lee at Digital Today, in Seoul:
Earlier, Naver agreed to attract a $1 billion equity investment from Nvidia and set a prerequisite of securing at least $9 billion of computing infrastructure excluding those funds. The contract meets that condition, allowing Naver to proceed with procedures for Nvidia's investment payment.
Naver disclosed the Brookfield contract today, Tuesday, Korean time. Brookfield funds a special purpose vehicle called Asset Co. with up to nine billion dollars for GPUs and data centers. Naver Cloud runs the service and sells compute to outside customers. The target is fifty-five megawatts in the first half of 2027 and two hundred by 2028, with the Gak Sejong data center as the hub and a one gigawatt global build as the long-term goal. And here's the clever part. Naver doesn't buy the hardware. Naver Cloud pays usage fees linked to its business revenue, and Naver guarantees those payments for up to eight years. Second special purpose vehicle of the episode. It's the same pattern. Anthropic with Broadcom, Naver with Brookfield. Somebody else owns the chips, and the AI company signs a long contract and a guarantee. Naver says the guarantee itself creates no actual cash outflow. Until revenue doesn't show up. Then the guarantee is exactly what turns into cash. I'm not saying it will, but an eight-year guarantee on GPUs is a bet that they're still earning in year six. Nvidia also gets technical cooperation through its DSX platform. Naver says details on the share issue to Nvidia will come later.
Reuters notes that Anthropic's own research has caught models in controlled tests doing code sabotage and fraud assistance. If you want that side of the story covered properly, try our sister show, AI Safety Daily. It covers alignment, model evaluations, emerging risks and governance every weekday: what changed, what the evidence says, and why it matters.
So: Anthropic's prospectus is out and its price is down, Nvidia's money takes a round trip through Meta, OpenAI has three proofs waiting on referees, and Naver has a nine billion dollar lease. Sources are in the show notes. This has been AI Daily Briefing, from Lantern Podcasts. See you tomorrow, Wednesday the 28th.