The Numbers Behind Anthropic's IPO, AI Bubble Questions, Meta's Spending Continues

Summary

• Anthropic’s strong operating business does not remove the cost of staying at the frontier. Thompson says subscriptions are profitable and the API is more profitable, but training-related R&D exceeds serving costs and increased more during the first half despite exploding usage. Competition determines spending whose pace is “out of their control.” He expects a bubble and painful reset, while explicitly declining to identify this IPO as the top.
• The headline loss and headline margins require different accounting adjustments. Thompson attributes the reported $80 billion first-half 2026 loss to repricing Amazon’s loan liability, which could be repaid in shares: “It’s not money out the door.” Conversely, about half of revenue comes through cloud platforms, whose roughly 30% cut—with estimates around 37% also cited—is classified as sales and marketing. He thinks that treatment overstates gross margins, although the underlying business still looks excellent.
• The hyperscalers could capture more of AI’s upside by financing the industry’s funding gap. Google, Amazon and Microsoft have the greatest capacity across free cash flow, debt and equity: the “girders of the bridge” until customer revenue catches spending. Existing shareholders fund equity issuance through dilution, and distressed labs could eventually accept worse terms. The hyperscalers that help finance the bridge could also gain power to develop competing chips and press Nvidia’s prices.
• The demand question is about timing, even for an AI believer. Craig Moffett’s test is whether final customer payments can support the entire value chain in 2027–2028, excluding transactions between suppliers. Thompson says current Anthropic revenue comes nowhere close to industry capex, but that infrastructure serves demand two to five years ahead. “Multiplying humans” into new businesses could create that demand; the difficulty is that such transformations usually take longer.
• Frontier labs need customer attachment to escape competition based solely on model quality and price. Microsoft and Palantir want customers to keep models replaceable, while Thompson argues the labs eventually need to own applications, workflows and interfaces. Cloud workspaces and accumulated habits make switching harder than changing an API endpoint: “Lock in is an important part of making money.”
• Meta’s Muse strategy makes heavy training expenditure integral to the investment case. The discussion cites up to $95 billion of stock offerings and shares down roughly 10% over a week; Sharp says dilution is a reason the shares should be down. Thompson says companies like Meta may be able to get away with dilution. Meta must control Muse’s underlying model, so the coherent objection is that “Meta doesn’t actually need Muse.” He also sees an indirect threat to Netflix: AI funding pressure pushes Meta to maximize Reels engagement and advertising.
• Government power remains a constraint on private AI and communications infrastructure. Thompson’s argument is conditional: if AI becomes as powerful as Anthropic predicts, governments will assert authority because “they have the guns.” Both hosts describe the Ukraine Starlink episode as opaque; Thompson guesses Musk acted in alignment with Washington, while Sharp guesses Washington requested the action. Neither presents those guesses as established facts.

Deep dive

1. Anthropic’s profitable products sit inside an expensive race

• Asked whether Anthropic’s IPO could mark AI’s peak as Coinbase’s listing did for crypto, Thompson answers, “I don’t know.” He considers a bubble, painful collapse and eventual recovery inevitable, but cannot identify the timing. His concern is investment arriving faster than useful ideas can be absorbed, rather than a belief that the technology lacks value.

• Sharp’s reading of the S-1 is that Anthropic looks like “a rock-solid business.” Thompson agrees about the operating core: subscriptions are profitable, and the API is more profitable still. That distinction matters because an attractive business serving existing models can coexist with a much harder economic proposition in developing their successors.

• Sharp notes that subscriptions also create distribution advantages. Having already paid, customers use Claude, develop habits and advocate for it inside their employers. Thompson thinks paying for both competing services is worthwhile, but doubts many customers will. The subscription therefore purchases recurring exposure to the product as well as revenue.

• The catch is R&D: training spending is larger than serving costs and increased more during the first half even as usage exploded. Better competing models force further investment. “The speed with which they need to spend money on that is out of their control.” Sharp emphasizes that this is an ongoing expense, rather than a problem solved once.

2. Cloud distribution improves access and complicates margins

• Amazon’s early investment forced Anthropic onto Trainium, something Thompson doubts it initially wanted. That constraint proved useful: starting with Nvidia, then supporting Trainium, TPUs and an AMD deal required hardware flexibility. Anthropic built “a layer of abstraction” across chips. Amazon benefited too, because a demanding frontier customer improved Trainium’s capabilities and tooling for other customers.

• Availability across AWS, Google’s cloud and Microsoft also helped Anthropic win enterprise customers while OpenAI was tied to Microsoft. Thompson’s concrete example is straightforward: a company whose data already lives on AWS would rather consume AI there. About 50% of Anthropic’s revenue comes through those three providers.

• Thompson challenges how that revenue is recorded. If a customer pays Amazon $100, Anthropic recognizes $100, then books the platform’s roughly 30% share—about 37% in estimates cited—as sales and marketing. He argues for recognizing net revenue or treating the payment as cost of revenue. Either would reduce the headline margins under discussion; he still calls the business’s margins excellent.

3. The $80 billion loss exposes the mechanics of dilution

• Thompson explains the reported $80 billion first-half 2026 loss through Amazon’s financing. Amazon supplied money as a loan that Anthropic could use for compute; Amazon could receive repayment or take shares. As fundraising rounds increased Anthropic’s valuation, its obligation was repriced and the increase booked as a loss. Thompson believes Amazon will receive shares at the IPO: “It’s not money out the door.”

• The corresponding investment appreciation appears in Amazon’s income. That creates a broader earnings-reading problem: hyperscalers’ reported profits can swing with investments unrelated to their operating performance. Thompson contrasts Amazon’s Anthropic gains with a decline in Google’s SpaceX holding and advises focusing on operating income.

• The economic bill survives the accounting adjustment through dilution. Thompson identifies outstanding share obligations involving Broadcom, Google and roughly two-and-a-half additional Amazon arrangements. He also describes a provision permitting issuance of 2% of outstanding shares after roughly 50% stock appreciation. “The people who are paying are existing shareholders”: their ownership funds the compute purchased with those arrangements.

• Equity financing shares upside with new investors, which Thompson considers healthier than relying entirely on debt. Sharp prefers public access and disclosure to permanent private ownership. Still, Thompson questions the opportunity available to buyers at illustrative valuations of $1.5 trillion or $2 trillion. His caricature of the pitch: “We’re going to dilute you endlessly, but we’re going to control the world.”

4. The IPO snapshot may conceal a changing growth trajectory

• Sharp suspects, without claiming certainty, that insufficient compute made first-half financials look better. Thompson explains that Anthropic did not have enough capacity: because compute is a major cost, the shortage kept costs lower even as customers experienced outages. Sharp jokes that those customers “did your part for Anthropic’s IPO” by helping make the margins look better.

• The S-1’s numbers stop at June 30. Thompson asks whether spending indicators such as Ramp’s index showed growth slowing, given OpenAI’s push and Astra’s arrival, or whether it reaccelerated after Opus 5.5. Sharp argues that Opus 5.5 could restart growth. If model improvement is what restarts demand, however, that reinforces the need to keep funding the training race. The growth slowdown remains uncertain, and Thompson notes that extraordinary growth cannot simply repeat indefinitely.

• Thompson says the financial information becomes stale on November 12, while reporting points to a mid-November listing. Whether Anthropic lists before that threshold or supplies an update could materially change the picture.

5. Owning the interface offers an escape from replaceable models

• Thompson describes a second battle beyond OpenAI versus Anthropic: frontier labs versus the surrounding software industry. Microsoft and Palantir tell customers to build around models that remain interchangeable. Satya Nadella and Alex Karp emphasize flexibility, better prices and protection against handing a future competitor the details of one’s business.

• That proposition is uncomfortable for labs whose largest expense rises with frontier competition. If customers can switch by changing an API, pricing power stays limited. Thompson argues that, “in the fullness of time,” labs need to move up the stack and own the customer experience. Sharp sees potentially enormous software revenue; Thompson emphasizes the resulting attachment.

• Their own usage shows why interfaces matter independently of model rankings. Thompson compares Codex to people who organize email into folders: projects and threads suit structured work. Sharp, a former lawyer, calls Codex “the associate” and Claude “the partner.” Thompson says Claude handles varied incoming tasks, while Sharp has done most of the concrete building on his GeckoBot project in Codex, with Claude powering it.

• Cloud workspaces deepen that relationship. A convenient virtual machine that remembers work also means, as Thompson puts it, “you’re now stuck in their cloud.” Applications and familiar workflows make switching progressively harder. That is the commercial purpose of climbing the stack: “Lock in is an important part of making money. Like, what’s your moat?”

6. Anthropic’s moral positioning raises a governance question

• A listener compares Dario and Daniela Amodei’s criticism of OpenAI during their Walter Isaacson interview to “someone badmouthing their ex during their wedding vows.” The complaints span advertising, sycophancy, government dealings and honesty. Thompson’s objection is strategic as well as personal: defining decisions through opposition lets OpenAI set the agenda. “Who’s actually setting the agenda here?”

• Thompson says he has never taken Anthropic’s claimed moral authority seriously. Sharp, by contrast, says he believes they are trying to act as moral arbiters. His concern is concentration: “There is no small collection of people that get to decide it for everyone else.” Thompson agrees that teaching Claude morals is not the right approach.

• Thompson connects the governance concern to the S-1’s unusual structure: roughly seven share classes and special voting arrangements he describes as a “magic share.” He prefers conventional corporations with understandable profit incentives. Those incentives have problems, but give counterparties a predictable framework for decisions and investment. Public buyers receive disclosure; they also knowingly accept dilution and governance that is not simply organized around maximizing profit.

7. Final demand must eventually finance the entire chain

• A listener brings back Craig Moffett’s question about 2027 and 2028: are customer payments entering Anthropic and ChatGPT sufficient to cover the whole system? Transactions among Nvidia, TSMC, clouds and labs can distract from that test. Moffett’s formulation, quoted by Sharp: “All that really matters is enough money coming into the system to support the system.”

• Thompson says current Anthropic revenue comes nowhere close to aggregate capex, but that spending also builds capacity for two, three, four or five years ahead. The uncertainty is timing. Strong revenue growth makes investors tolerate dilution and unusual governance; slowing growth makes those same features suddenly salient.

• Google, Amazon and Microsoft have the greatest capacity to supply free cash flow, borrow against diversified businesses and issue shares. They are the “girders of the bridge” across the funding gap. If labs require repeated rescues, the hyperscalers could acquire substantial ownership and control. Sharp imagines Amazon offering a distressed Anthropic terms even less favorable than its early financing.

• Nvidia’s support for neoclouds reflects the risk that funding runs out before revenue catches up. Thompson’s railroad analogy preserves the distinction: useful infrastructure can survive a financial crash. But a successful bridge financed by hyperscalers creates another problem for Nvidia. Those customers gain negotiating power, develop competitive chips and push pricing; the entities sustaining demand can also weaken its supplier’s position.

8. Multiplying people creates demand on a slower clock

• Drawing on his exchange with Palantir’s Shyam Sankar, Thompson distinguishes improving the bottom line by replacing people from improving the top line by “multiplying humans.” The latter creates workflows, opportunities and businesses that did not previously exist. It also usually takes longer and often comes from new companies. AI-native challengers may displace incumbents that struggle to bolt AI onto unsuitable software structures.

• Thompson could scarcely be more bullish on AI. Sharp says he is in the middle ground and does not think AI will replace everyone. Thompson argues that the economy is not purely utilitarian: a huge portion of it involves status, pleasure and things people want but do not need. Infrastructure spending happens now; the businesses and cash flows required to repay it may arrive much later.

• His defense of advertising follows the same optimism about new demand. Meta can show people things they did not know they wanted: “It’s manufacturing human desire.” Economic activity includes expensive World Series tickets, not merely necessary tasks AI can perform efficiently. Thompson calls Anthropic’s criticism of advertising evidence of misunderstanding humanity. His answer to claims that people cannot invent valuable new activities: “I bet we can.”

• Thompson’s timing example is the internet: 1993 preceded search advertising around 1999–2000 and the feed in 2006. Discovering a powerful technology’s interface, businesses and monetization takes time. Sharp says AI will be good but does not think it will replace everyone; Thompson says he is an unequivocal believer in AI while emphasizing that monetization and adoption take time.

9. Muse makes Meta’s training expenditure a strategic commitment

• Sharp cites Meta shares falling about 10% over a week after increased AI spending plans and a new nonvoting share class. The report discussed potential stock offerings of up to $95 billion. Sharp says dilution is a reason the shares should be down; Thompson notes that companies like Meta, Google and Microsoft may be able to get away with it.

• Anthropic’s disclosure clarified the debate for Thompson by exposing training’s share of costs. Muse could become the “aggregator of aggregators,” owning the interface between users and other services. But Meta must control a sufficiently good underlying model; outsourcing that dependency to a lab seeking the same interface would be dangerous. “If you believe in Muse, you should be happy they’re spending money.”

• Sharp says Muse’s model is not good enough; Thompson replies that he does not really experience that limitation. Sharp instead highlights 1Password integration as a major improvement. He explains how an agent-specific vault can supply API keys and passwords without awkward terminal files or secrets pasted into chat. Thompson sees 1Password’s unpopular transition toward cross-platform enterprise software as prescient: businesses pay, Apple pressures consumer password tools, and agents need credentials.

• The training-versus-serving distinction also changes the discussion of Anthropic’s enterprise push. Sharp says the S-1 gives him some sympathy for it: enormous training bills create pressure to sell the resulting models somewhere. Thompson notes that serving enterprise customers consumes relatively little compute compared with training, easing the concern that those customers necessarily take capacity away from other products.

10. Meta’s cash engine funds ambition and pressures Netflix

• Thompson says Sharp wrote the most compelling investor objection: “Meta doesn’t actually need Muse.” Its advertising business remains formidable. Investors seeking straightforward optimization would prefer the behavior of professionally managed Google or Amazon. Meta instead has a founder, Mark Zuckerberg, pursuing another platform opportunity. Sharp considers this direction more promising than some of Meta’s previous detours.

• Thompson views Amazon’s infrastructure and inference focus as rational for a company already built around lower margins and logistics. It can serve Anthropic and OpenAI without spending billions to lead frontier development itself. The hosts defer their discussion of Google to the next episode.

• Thompson’s Netflix thesis is indirect. AI expenses pressure Meta to produce cash, encouraging more ads and more engagement. Earlier restraint over feed ad loads and video consumption gives way to “all the dials turned to 11.” Someone spending an hour watching Reels has less time for Netflix. Sharp initially doubts familiar AI-video displacement arguments, then accepts this cost-pressure mechanism as more persuasive.

11. State power shapes AI access and military infrastructure

• Responding to a listener who understood him as demanding Anthropic serve the government, Thompson clarifies his argument. If AI becomes civilization-altering, governments will assert authority over its deployment: “They have the guns.” He is describing power and inevitability. The separate normative debate over surveillance or weapons targeting does not resolve whether a private supplier can dictate terms to a state possessing overwhelming coercive power.

• Thompson distinguishes commercial Starlink from military Starshield. “As I understand it,” the US military controls Starshield’s use; he acknowledges uncertainty about the precise boundaries of Musk’s authority. On Ukraine, he notes that the country was using commercial Starlink rather than Starshield and raises the possibility that Washington asked Musk to enforce the service’s terms.

• Both hosts stress how little they know about the reported Ukraine slowdown. Thompson guesses Musk was in close alignment with the Trump administration; Sharp guesses the US government asked him to do it. These remain hypotheses. Their shared point is that Ukraine has less power than Washington or the infrastructure supplier it depends on.

• Thompson’s industrial argument is that consumer demand supplies military scale. Defense primes serving one government customer tend toward expensive, low-volume systems; drones and missiles increasingly require volume. His analogy is car factories repurposed for wartime production. His proposed route to domestic drone capacity is categorical: “Ban all foreign drones and completely legalize drone delivery across the US.” Sharp embraces the commercial-volume argument.

12. Persuasive filmmaking can carry a disputed premise

• Thompson calls You Can See Everything “the Zootopia of documentaries”: excellent craft carrying a premise he finds untenable. He rejects a sympathetic reading of Elizabeth Holmes as someone manipulated by men and interprets her relationships differently. His concrete objection is the theater’s alarm at Holmes granting Billy Evans power of attorney, which Thompson regards as normal when someone enters prison and a partner cares for the children.

• Sharp’s defense rests on the film as commentary about media and truth, rather than as an Elizabeth Holmes documentary. He sees multiple layers of construction and interpretation and thinks it should win best picture. Thompson grants the filmmaking’s persuasive power while remaining frustrated by what that persuasion accomplishes.

13. The Brewers’ title rewards an outlook built around the season

• Thompson had made peace with the Brewers’ small-market limitations and learned to enjoy summer evenings without expecting a World Series. Basketball with Giannis felt different: a real chance creates relentless pressure. The Brewers’ victory initially produced surprisingly little feeling because he had “cauterized the postseason in my heart.” He then becomes emotional explaining the unexpected reward of enjoying both the season and its championship ending.

• The connection back to AI is uncertainty and control. Thompson objects to people depressing themselves about an unknowable future while missing the life they can live now. “Be nice to people around you. Learn new things. Explore new things.” His sports lesson is that outcomes lie beyond a fan’s control; his broader advice is, “Figure out the future when it gets there.”