Anthropic's slowdown and AI's financing chain, reverse engineer anything, and a spy in OpenAI's boardroom

On Wednesday Anthropic told investors gathered in San Francisco that it had taken $16.8 billion in revenue in the third quarter, up 45% on the quarter before, and that on an adjusted basis it was now profitable and generating cash. By Thursday Reuters was reporting that some of those same investors were discussing a stock market valuation nearer $1.5 trillion than the $2 trillion that had been circulating. CoreWeave's shares and bonds were falling, and the general reading had settled into a single word: slowdown. Growth of more than 100% in the second quarter had become 45% in the third, and that was enough to unsettle a market that had priced the company, and much of the infrastructure around it, on the earlier unprecedented pace.

The slowdown reading is fair, but other key figures received far less attention. Anthropic reportedly generated $1.70 of revenue for every dollar it spent on compute, which puts compute at roughly 59% of sales. Bloomberg's Ed Ludlow reported an 85% gross margin, meaning the cost of actually serving customers was around 15% of revenue. The gap between those two numbers, about 44% of everything Anthropic earned, is compute consumed by research and training, and it amounts to roughly three quarters of all the computing the company bought. Josh You at Epoch AI reached a similar estimate, putting 70 to 75% of Anthropic's compute into research. Anthropic's profit, in other words, is a decision about how much of its capacity to point at the next model rather than at paying customers, and its growth rate is partly a consequence of that decision. A company devoting three quarters of its compute to research, in a year when rents for Nvidia's Blackwell chips have risen by more than a quarter, was not necessarily short of demand. It may have been short of chips to train models and serve customers at the same time. (Anthropic also told investors that distillation, rival labs paying to train their own models on Claude's answers, had made up 5 to 7% of its revenue, and that it cut this back materially in the third quarter.)

Growth may still be held back by capacity and by the race to train the next model, and 45% a quarter would count as strong in almost any industry. The concern lies in how that growth is financed. A financial stability report from the IMF, published the day before the investor meeting, examined 14 disclosed AI deals since 2023 and found a common three-part structure. The large cloud companies put equity into financially weaker AI firms, around $300 billion in total, and because that equity is revalued every time the AI firm raises money at a higher price, the gains flow straight into the investor's reported profit. Lenders then supply debt, around $650 billion, some of it to the AI firms directly and some guaranteed or arranged by the investors, to be repaid from the AI firm's future revenue. In return the AI firms commit to spend on the investors' clouds and chips, about $1.4 trillion of it, which the cloud companies book as contracted future revenue and which the lenders behind the data centres treat as security.

How the IMF draws the loop: $300 billion of equity in (blue), $1.4 trillion of revenue commitments back (green), $650 billion of debt alongside (red), across 14 firms. Anthropic's $518 billion of cloud commitments is the largest single example of the green arcs. Source: IMF Global Financial Stability Report, October 2026.

Anthropic is the clearest case. Amazon and Google own stakes in it; Amazon's second quarter net income included $53.4 billion of non-operating gains, mostly from that holding, and Alphabet's included $99 billion from equity securities, partly from the same holding. Anthropic has in turn promised to spend at least $518 billion over a decade, much of it on their clouds. All three parts rest on one assumption, that Anthropic's revenue keeps growing fast enough to justify the valuation, service the debt and honour the spending. The IMF calls this wrong-way risk, because if the AI firm stumbles the equity falls, the debt sours and the promised revenue fails to arrive, all at once. This week showed a milder version in which the three parts moved separately. The valuation expectation dropped by $500 billion, the paper gains at Amazon and Alphabet shrank with it, and the $518 billion of commitments stayed exactly where they were.

Further down the chain the exposure spreads. Morgan Stanley figures cited by the Bank of England put AI-related debt issued by early September at around $450 billion, more than double the whole of last year, and the Bank found cloud companies accounting for 47% of sterling corporate bond issuance in 2026. Below the bonds sit neoclouds such as CoreWeave and Nebius and power suppliers such as Bloom Energy, which borrow against contracts tied to AI demand, and below them sit the funds. Value Aligned Research Advisors, a small firm in Princeton, New Jersey, had grown to $26.4 billion by June with Alphabet, Amazon, Nvidia and Bloom Energy among its largest holdings, then saw its flagship fund lose 44% in July when crowded AI positions unwound. On Thursday it was denying an unconfirmed rumour that it was in trouble. The same expectation of Anthropic's growth is being financed three times over, by bondholders, by hyperscalers marking up their stakes, and by leveraged investors, and strong results from the lab reassure all three and encourage them to extend further. A profitable Anthropic concentrates the risk rather than relieving it.

Whether the infrastructure can survive a stumble at the lab was the question put to Paul Kedrosky and Azeem Azhar on Ezra Klein's podcast this week. Kedrosky argues that large buildouts tend to end in overcapacity because cheap capital keeps them going past the point of demand, and that this time the whole structure depends on the private-market valuations of the labs. Azhar replies that the data centres get used whoever wins: if enterprises shift to cheaper open-weight models, Anthropic's margins suffer but the capital expenditure is still monetised.

“Everything is levered off the valuation of these companies that are currently sitting in private markets.”

Takeaways: Anthropic has answered the question of whether customers will pay for frontier AI, and left open the question of whether that payment can grow fast enough to support AI's financing stack. The upcoming IPO prospectus will be the first public audited look at a company critical to the future of AI economics.

Brandon Thomas had spent 15 years writing software and a decade of that in Rust when he decided he had paid Adobe enough. At the start of the month he released seven free open-source applications under his ArtCraft banner, modelled on Creative Cloud, with PhotoCraft standing in for Photoshop, FilmCraft for Premiere Pro and EffectCraft for After Effects. He built them in a couple of weeks with Claude Opus 5.5, announced that "software is over", and watched PhotoCraft alone pass 43,000 GitHub stars in under a fortnight. Reviewers who tried the alpha found gaps in colour handling and professional workflows, but the timing helped. Adobe agreed in March to a $150 million settlement of a federal case over hidden early termination fees and hard-to-cancel subscriptions, and the resentment behind it had been waiting for a tool.

Some have started calling this "spite SaaS". A user gets one too many upsell pop-ups or spends an evening with a chatbot that cannot cancel their plan, describes the 5% of the product they actually use to a frontier model, and has a working replacement two hours later. You can see the same capability applied for sheer fun in the screenshot from modder 0xProbability, where Minecraft's blocky character stands outside a Los Santos petrol station in GTA 5 with a full crafting inventory bar along the bottom of the screen. The modder had Opus 5.5 write and test the mod that brings the two worlds together, showing how closed commercial titles are now being treated as raw material that anyone can pull apart and remix.

Beneath these stories sits a change in tooling and a reflection of how capable the latest models have become at software engineering. The number one trending repository on GitHub this week is REA, short for Reverse Engineer Anything, now past 285,000 stars. See a feature in an app you want in your own product, ask your agent to investigate, and it will explain how it works, show the evidence and help you build your own version. One showcase traces Notion's clipboard bridge through its renderer, preload and main processes.

The utility is considerable. Any organisation sitting on undocumented legacy binaries can extract the business logic and rebuild it on a modern stack, and product teams can study how a rival handles a workflow before designing something better. Vendors face a different calculation, since compiled code only protected them because reading it was slow, and bundles only held because customers could not build the slice they needed. Both assumptions have now failed.

But REA's own README warns that lookalike repositories are not official, and one has already been caught serving malware, and a vibe-coded clone of a trusted tool can look identical whilst missing input validation or quietly mishandling files. An agent could edit local game assets to render opponents through walls, bypassing anti-cheat systems that only watch for injected code. When any client-side binary can be unpacked, altered and rebuilt at home, vendors can no longer assume the software running on a customer's machine is the software they shipped.

Takeaways: Opus 5.5 and GPT-6 Astra, paired with harnesses like REA, have turned every compiled program into readable, editable source material, and ArtCraft is the first widely noticed result rather than the last. Legacy systems become recoverable and competitor features become study material, while pricing power, lock-in and client-side trust all weaken. The sensible response is to treat your own software as if it were already open, and compete on service and data rather than secrecy.

This week's image comes from the FBI complaint against OpenAI engineer Abdulrahman Alfozan. On the left, a wad of toilet paper sits in a USB port behind a television in an executive meeting room. On the right is the device it allegedly hid: a Seeed XIAO ESP32-S3 Sense. It's a $13.99 hobby board with a microphone, microSD storage and a 2.4GHz Wi-Fi antenna.

Prosecutors say he asked a productivity agent where executives met and which meetings were booked. He then allegedly used a ChatGPT account linked to his wife to work through the build. His requests covered continuous recording, live listening, an encrypted memory card, a power-only USB connection and silent start-up. None of these would look unusual on a maker forum, but together they match the hardware in the photo.

We don't yet know his motivations, but whatever ultimately turns out to be behind it, the case shows that frontier labs are now worth spying on for their conversations as well as their code. Most AI security planning concentrates on model weights and source code, the assets that would let a rival copy a model. The room targeted here hosted leadership meetings and briefings with government officials and business partners. That information is easier to steal than a model and can be put to use as soon as it's heard. Labs will need to treat physical security and insider risk with the same care as their networks.