Anthropic Found a $30 Trillion Market. The Money Doesn't Stay There.

The global software market is worth somewhere between $700 billion and a trillion dollars. The IPO targets floating around the two largest AI labs run one to two trillion. No honest multiplication of the first number reaches the second, so the number changed. Anthropic is preparing to tell investors its addressable market exceeds $30 trillion — measured not in software but in the value of the human labor its models could in theory replace, across engineering, accounting, and legal work ^[1]. It is a clean swap: stop selling a product, start pricing the payroll it might one day absorb. The first difficulty is the price of the labor itself. DeepSeek cut its V4-Pro model's price 75% in May, making it 12 to 19 times cheaper than GPT-5.5 or Claude Opus 4.7 for the same work ^[2]. OpenAI answered by cutting fees on its Luna model by 80 percent, and Anthropic rolled out a high-performance model at half the cost of its top system ^[3]. The per-unit price of the thing they claim to replace is racing toward zero. Morgan Stanley analysts call model access a loss leader — a cheap front door meant to pull in more profitable cloud spending on storage and security, not a profit center in itself ^[4]. The second difficulty sits on the buyer's side. MIT's Project NANDA finds 95% of enterprise AI pilots show no measurable profit-and-loss impact despite $30 to $40 billion invested, and Gartner predicts more than 40% of agentic AI projects will be canceled by 2027 ^[5]. ServiceNow's maturity index shows AI spending up 110% while readiness sits at 51 out of 100 ^[6]. Holly Briedis put it plainly.

The adoption curve stalls not because the technology failed but the data underneath it did. — Holly Briedis

The substitution the pitch assumes is not stalling because the models are dumb. It is stalling because the data underneath is not ready to be replaced. Where substitution does land, the savings stay with the buyer. Klarna replaced 853 customer-service employees with an AI agent and saved $60 million ^[7]. McKinsey cut 3,000 to 4,000 staff while adding 20,000 AI agents ^[8]. Mariner Wealth pays $50,000 per AI worker — a sliver of a loaded human salary ^[9]. Every one of those deltas shows up on the enterprise's own income statement, not the lab's, whose per-token price keeps falling. The one piece of evidence that could rescue the thesis is real. Token prices fell 40% since late June, yet AI spending per employee among the heaviest users rose 50% month-over-month ^[10]. Cheaper input, more consumption. Total spend is rising; the question is where it pools. Jensen Huang has already given his answer, describing Nvidia as an AI factory that turns electricity into tokens, the raw unit of AI output ^[11]. Nvidia is worth over $5 trillion ^[12]. CoreWeave, the cloud landlord sitting between the labs and those chips, holds a $104 billion backlog and still posts negative free cash flow of $5.74 billion ^[13]. The spend flows through the labs to the infrastructure beneath them. The $30 trillion figure does real work: it is the only number large enough to bridge a $2 trillion valuation on paper. What the operational record shows is that the money it describes passes through the labs rather than pooling at them. They are selling labor substitution to investors while their own unit economics price them as a commodity layer — the toll booth, not the destination.