Detailed Analysis
Anthropic, the AI safety company behind the Claude family of large language models, faces a pointed structural critique in a provocative analysis that draws a systematic parallel between the company's current trajectory and WeWork's catastrophic 2019 collapse. The argument centers on four interlocking mechanisms: theological branding that inflates perceived value far beyond underlying business activity, circular capital flows between strategic investors and the company itself, accelerating commoditization of the core product, and accounting conventions that obscure the true ongoing cost of remaining competitive. The author contends that Anthropic's fundamental commercial activity — renting GPU capacity from Amazon and Google, running inference on transformer-based models trained via widely published techniques, and reselling tokens to developers — constitutes compute arbitrage analogous to WeWork's real estate arbitrage, and that the AI safety mission narrative performs the same function Adam Neumann's "consciousness elevation" rhetoric did: converting a margin business into an asset deserving platform-company multiples in the eyes of investors.
The circular capital critique is among the most structurally specific claims in the analysis. Anthropic has received approximately $8 billion from Amazon and additional billions from Google, with deal terms requiring the company to spend the preponderance of that capital back on AWS and GCP compute. The author argues this creates a situation in which reported revenue figures, including a cited $30 billion annualized run rate, are partially an artifact of investor money recycling through Anthropic's profit-and-loss statement and back onto the top lines of the same entities that provided the capital. The parallel to SoftBank's $18.5 billion investment in WeWork — where capital flowed back through leases and cross-portfolio arrangements to create the appearance of robust organic revenue — is presented not as metaphor but as isomorphic financial structure. Whether or not Anthropic's organic enterprise demand is genuinely robust, the argument establishes a legitimate analytical challenge: it is unusually difficult to isolate demand that would exist absent the strategic investor relationships that both fund the company and constitute a portion of its customer base.
The commoditization argument adds a time-sensitive dimension to the valuation critique. The author observes that Claude, GPT, and Gemini converge on most practical benchmarks, that open-weight models from Meta, DeepSeek, and Alibaba continue closing the capability gap on coding and reasoning tasks, and that API switching costs are trivially low — a single line of configuration change. This framing connects to a well-documented dynamic in technology markets: when product differentiation collapses and switching costs approach zero, pricing power gravitates toward whoever controls the lowest-cost input, which in this case is the compute infrastructure providers, not Anthropic. The WeWork analogy holds here as well — WeWork's supposed differentiation through design and community proved instantly replicable by any landlord willing to invest in aesthetics, just as frontier model performance is now replicable by well-resourced open-source developers.
The accounting argument is perhaps the most technically substantive. The author challenges the classification of model training as research and development expenditure rather than cost of revenue, arguing that training is not a one-time capital investment producing a durable asset but a recurring, unavoidable cost of maintaining competitive product standing in a field where the frontier advances every six to twelve months. This framing — that training compute is functionally "rent" in the WeWork sense — suggests that reported gross margins on inference are materially misleading because the largest recurring cost of the business is being held off to the side as though discretionary. This mirrors the critique of WeWork's "community-adjusted EBITDA," which famously excluded rent — the company's dominant and inescapable expense — from profitability calculations. If Anthropic's training costs were amortized into cost of revenue on a rolling cycle basis, the economics of the business would look structurally different from what current investor communications imply.
The broader significance of this critique extends well beyond Anthropic specifically. The analysis articulates a failure mode that may apply across the frontier AI sector: companies that combine genuine technical capability with mission-driven branding, strategic investor entanglement, and accounting conventions that obscure cyclical capital intensity may be systematically overvalued on metrics that do not survive contact with commoditization pressure. The governance observation — that Anthropic's Long-Term Benefit Trust structure, like WeWork's dual-class shares, concentrates decision-making authority in a way that delays public accountability — adds a structural layer to the concern. WeWork's 2019 IPO attempt forced transparency that the private market had not demanded; Anthropic remains private, and the secondary market valuations that imply a trillion-dollar company operate without the disclosure discipline that a public offering would require. The argument is not that Anthropic will fail, but that the mechanisms by which it could fail are more precisely analogous to a documented historical case than most participants in the current AI investment cycle appear willing to acknowledge.
Read original article →