Lead storyThe Information
Anthropic filed its preliminary prospectus, known as an S-1, on Monday, giving a first public look at its business in the first half of this year, plus new governance details and more than 90 pages of risk factors. An investor roadshow is expected in about two weeks, with the listing expected in mid-November, ahead of Thanksgiving. The deal is expected to value Anthropic at $1.5 trillion or more, though the filing itself did not disclose expected share pricing.
Why it matters
The most valuable private AI company is about to get a public price tag, and that price will reset how every other AI company is judged. A $1.5 trillion listing would put Anthropic in the top tier of public companies on day one, and it opens the IPO window for the rest of the AI field. Watch the roadshow for the first hard numbers on revenue and costs, because those will set the terms for every AI funding round that follows.
Story 02The Information
Meta plans to seek as much as $95 billion through equity sales to pay for AI infrastructure and computing capacity. About $35 billion would come in underwritten public offerings: about $22.5 billion in Class A shares and about $12.5 billion in depositary shares representing preferred stock. The other $60 billion would come through at-the-market stock sales once trading starts in new nonvoting Class C shares. The Class C structure lets Meta raise capital without diluting chief executive Mark Zuckerberg's voting control: Class A shares carry one vote each and Class B shares carry 10 votes each, mostly held by Zuckerberg-controlled entities. Nvidia, AMD, and Broadcom have indicated interest in buying up to $5 billion in aggregate of the Class A sale, at the same price as other buyers.
Why it matters
Meta is paying for the AI race by selling pieces of itself instead of borrowing, and it is doing it without giving up control. That is a very different funding model from the debt-heavy neoclouds, and it is much harder to break when interest rates rise. Chip suppliers buying in is the other tell: the companies selling Meta its hardware now want a stake in Meta itself, which ties the whole supply chain together.
Story 03The Information
As AI assistants like Instinct and Meta's Muse gain traction in shopping, Stripe has taken a leading position in how agents pay, often through single-use virtual cards issued via its Link digital wallet. The merchant accepts the virtual card like any other card, while Stripe separately charges the shopper's own card, so one purchase becomes two transactions that are harder to track end to end and harder to resolve when something goes wrong. Stripe could also switch an agent's payment to a different card network than the shopper's original card, diverting transactions and fees. More than a quarter of Muse's 10 million users are shopping with the app, according to Meta chief executive Mark Zuckerberg, with an average purchase size above $200. Meta had to work with Stripe in the week after Muse launched to raise credit limits because shopping volumes were so high. Visa and Mastercard are pushing token-based alternatives tied to shoppers' existing cards, called Visa Intelligent Commerce and Mastercard Agent Pay, and Bank of America plus five other banks published proposed industry guidelines last month favoring payment methods that can be audited.
Why it matters
Whoever controls the payment step controls the fee and the data, and right now that is Stripe, not the card networks. If agents do the shopping, the shopper never sees a card, so the company that issues the agent's card wins by default. Visa and Mastercard are fighting back with tokens that keep their cards in the loop, but the banks' push for auditable payments shows this is now a standards fight, not just a product fight. Any business selling through agents should assume the checkout layer will change hands at least once.
Story 04The Information
Tenants in new data center deals are asking for late fees upwards of $5 million per day of delay, up from about $1 million a day for a gigawatt-scale facility this spring. Brokers Marsh and Aon are now designing bespoke delay-in-delivery coverage, and bankers and bondholders are demanding more insurance before they fund projects. A 1-gigawatt shell delayed from Christmas to the end of February could owe $300 million in penalties while also missing out on rent, and a facility plus its chips and servers costs around $50 billion to $60 billion per gigawatt. S&P Global puts data center insurance premiums at around $10 billion in 2026, already twice the global aviation insurance market, and Swiss Re estimates they could approach $24 billion by 2030. Hut 8 has bought 19 months of delay-in-startup insurance, and one report found a single 45-minute outage could wipe out 40% of a facility's annual cash flow under the traditional promise of 99.999% uptime, which allows only about five minutes of downtime a year.
Why it matters
Being late is now a priced financial risk, not just an embarrassment, and that changes who can build. Developers who can insure their delays get funded; developers who cannot pay more or do not get built at all. For anyone renting AI computing power, your provider's delivery risk is now your risk too, so ask what happens to your capacity if their building slips by a quarter.
Story 0520VC
Microsoft chief executive Satya Nadella joined Harry Stebbings on 20VC for a 26-minute episode released Monday. The conversation covers whether every AI model should be assumed compromised, how to stop agents from causing harm, and whether a model's chain of thought, the reasoning it shows for its answers, can be trusted. It also covers open versus closed models over the next two years, why companies need to own their own intelligence, whether every company becomes AI-native, whether the world can support $100 billion training runs, and whether AI will transform sales, finance, and marketing the way it changed coding.
Why it matters
The head of one of the world's biggest AI sellers is telling buyers not to bet everything on any single model, including his. The winning move he describes is building your own loop of data and feedback, so your company's knowledge improves with use and cannot be copied by switching providers. That is a direct warning to startups built entirely on top of one lab's model: your moat has to live in your own data, not in someone else's model.
Story 06The Information
Florida drone-component supplier Unusual Machines closed its acquisition of Upgrade Energy last week for $52 million in cash and stock. Upgrade's founder Matt Barnard built the company entirely on his own money, with no venture capital or debt. The deal strengthens Unusual Machines as a US-made component supplier to the Pentagon's Drone Dominance program, worth $1.1 billion in contracts. In the latest stage last month, Unusual Machines supplied components to four of the nine companies that placed, including Perennial Autonomy and Neros Technologies, and Upgrade supplies both of those firms as well. Its Torrance, California factory has about 50 workers packaging cells from suppliers including Molicel and Amprius Technologies, producing 400 battery packs a day. Unusual Machines shares are up 21% since the deal was announced in May and up 32% over the last year.
Why it matters
The defense drone supply chain is consolidating around American-made parts, and small bootstrapped suppliers are becoming acquisition targets. A $52 million deal is small next to today's AI numbers, but it shows where real defense money is going: not to the drone brands, but to the batteries and components inside them. For hardware founders, the lesson is that a boring, working component with a defense customer is worth more than a flashy product without one.