Anthropic burns sixty-five billion dollars to keep you captured. You are not witnessing market-driven success. You are watching a sophisticated operation in value-laundering.
It is true that the large language models work, in the narrow sense in which working is defined by benchmark scores on held-out test sets, when the models are running on machines that the company—or, more accurately, the Amazon Web Services data centers that host the inference stack at the end of this transaction—pays for at a rate that dwarfs the revenue generated. The press treats the forty-seven billion dollars in reported annualized revenue as a signal of commercial maturity. The trouble is that the spending on computing infrastructure and cloud transit far exceeds the revenue the company reports pulling in from subscription fees and application-programming-interface calls. The architecture is legible enough to read from the public filings: Anthropic is running a capital-burn rate whose magnitude exceeds a small country’s yearly gross domestic product in order to maintain a chokepoint between individual users and the general-purpose computing layer they already own.
The discipline of cryptographic verification, which demands that every claimed guarantee be traceable back to an unbroken chain of evidence rather than resting on a press release, has a useful concept here: you cannot verify a system whose specification you cannot write down. The alignment narrative is not a specification. It is a customer-acquisition document. As Doctorow has documented at book length, enshittification operates as a mechanical consequence of removing competition and regulation from a platform’s operating envelope; once the four constraints that historically kept platforms from abusing their users melt away, as they did across this sector during roughly forty years of lax antitrust enforcement, the platform’s only structural imperative is to claw value back from both end users and dependent business customers until the system collapses.
The architecture choice is visible in the infrastructure layer. The company’s commitment of one hundred billion dollars to the Amazon cloud platform over a decade is not a technology partnership; it is vertical integration by subscription, locking downstream customers and developers into a managed service where pricing variables are adjusted invisibly and machine-speed, the continuous computer-mediated adjustment of prices and visibility that the cloud layer makes possible at scale, ensures that per-user costs are extracted without a transparent marketplace to contest them. The Competition Bureau, when it reviews the structure of a transaction like this, looks for foreclosure of rivals. What is visible here is the foreclosure of the open internet itself, replaced by a walled garden where every token generated, every line of code written, and every piece of text drafted has to pass through a single billing gateway.
What these platforms are actually doing is not selling intelligence; they are selling automated uncertainty at the scale of the entire industrial economy. It is a reverse-centaur operation, as Doctorow has named it, pressing professionals into service as the error-correction layer for probabilistic reasoning engines that cost a fortune to run. The actual cost of this computation—the electricity, the specialized GPU cycles, the cooling—is being socialized, while the risk of the resulting errors is shifted onto the organizations foolish enough to integrate these systems into production. When the company touts its ability to “write code” or “perform tasks,” it is glossing over the fact that the bill for this infrastructure is being deferred, and the gap between the forty-seven billion in revenue and the massive underlying cash burn is the bezzle—that illusory wealth that, as Galbraith taught and Doctorow keeps repeating, feels like prosperity until it is discovered.
The sector is pushing toward blockbuster IPOs as if they were a cure for profitability concerns, but an IPO is merely the point at which the initial investors cash out and pass the bag to public markets. We have seen this movie before; it is the same script run by every leveraged-buyout firm that hollowed out industrial assets in the nineties. Tim Wu has written at length about the monopoly-rent extraction that follows when vertical integration closes off transit routes. The upcoming IPO filings will not change the underlying architecture; they will only expose the dependency to public-market scrutiny, at which point the gap between the alignment promise and the accounting reality will have to be reconciled with shareholder expectations. The future is always hypothetical; the data centre next door is already drinking the aquifer and buying the grid. There is a public consultation open at the Competition Bureau regarding the anti-competitive effects of platform consolidation. The deadline matters because deadlines are the only part of regulatory processes that the regulated actually respect.