Anthropic’s Alchemy

While the race to regulate dominates headlines, armies of analysts are poring over the improbable and fantastical financial projections emanating from AI giants. This isn’t new. The last four years have seen a steady stream of pitch decks and keynotes where AI executives say "a trillion dollars" as often as possible. In many respects, they have little choice. The scale of AI’s theoretical reach is so vast that any global projection naturally strays into trillion-dollar territory. There is also an obvious incentive to frame today's eye-watering capital expenditure against a massive speculative horizon—without that narrative, the cash burn would be indefensible.

But there’s a limit. Market sizing and financial modelling are usually sober affairs when preparing for an IPO. Traditionally, a public offering is designed to raise capital while establishing a viable baseline valuation so public markets don't rout the stock at the first sign of headwinds. The playbook is different for frontier AI companies, it seems.

Anthropic is a prime example, though far from an exception. In recent investor disclosures ahead of its planned public offering, the company charted a $30 trillion Total Addressable Market (TAM)—simply by tallying up the total global compensation of all human labour exposed to automation. Capturing even a fraction of a percent of that pool, they argue, justifies a multi-trillion-dollar valuation.

Historically, tech companies modelled market opportunity by evaluating established spend; SaaS providers, for instance, size their opportunity against existing corporate software budgets (around $2.4 trillion globally across all S&P 1500 tech firms).That’s not enough for AI giants. Consequently, Anthropic’s model abandons traditional software entirely, operating on a far grander premise: anything a human knowledge worker can do, AI can do also.

Yet even if public markets accept this wild departure from traditional market sizing, operational viability remains in short supply. Serving inference at scale is expensive—and in many cases, more expensive than the legacy salaries Anthropic aims to replace.

To bridge this gap, Anthropic claims to be enjoying positive "adjusted operating income"—a financial euphemism for profitability achieved by carving its largest cost centres out of the equation. Specifically, the company touts gross margins above 80%. However, this headline margin explicitly strips out two existential costs: partner revenue sharing (the substantial top-line cuts surrendered to cloud distribution partners like Amazon Bedrock and Google Cloud) and, more pressingly, model training costs (the multi-billion-dollar compute spend required to train frontier models and prevent product obsolescence).

Doing so is cheeky, to say the least. Development of frontier AI is not a one-off capital expenditure like building a factory; model training is a continuous, recurring operating cost needed to stay competitive against rivals like OpenAI.

Re-incorporating training compute, distribution splits, and stock-based compensation back into the model sees Anthropic continue to burn through immense amounts of cash. Bulls argue that algorithmic efficiencies, cheaper compute, and hardware advances will eventually fix unit economics. However, there is little evidence that these cost declines will arrive fast enough to offset the escalating spend required to train the next generation of models.

This is where the alchemy falls apart. Scaling up to arbitrage human labour does not automatically solve the underlying economics of the platform. In practice, it could make them worse.

Human labour is not a static baseline; it is an active, competitive counterparty. As AI moves in as a direct substitute for traditional workforce capacity, it exerts immediate downward pressure on wages—a mechanism familiar to labour economists, and one that mirrors contemporary political arguments around immigration-fuelled labour surges destabilising local economies. But at a global scale, established human labour markets will flex to defend their market share. Indeed, we are already seeing a flurry of enterprise reports detailing businesses rehiring human workers after discovering the real bills for their AI replacements shot well over budget.

This creates a fatal paradox for frontier AI models. Sizing a market against total global human compensation assumes those wages remain fixed while AI captures them. In reality, as AI depresses labour rates and global workforce markets adjust downward to remain competitive, the very benchmark Anthropic is attempting to underbid moves lower with it. Basing a $30 trillion market size on a dynamic, self-correcting global labour pool is a form of accounting alchemy that does little to reassure us of the long-term potential of AI giants. Rather, it fuels the conspiracy theory that their go-to-market is simply to save companies the burden of having to pay pesky humans.