The AI Labs Are Cutting Prices and Filing for $3 Trillion IPOs at the Same Time

Anthropic is about to ask public markets to value it at two to three trillion dollars ^[1]. It is also, this month, negotiating to spend $6 billion on a company it is buying to lower the computing and training costs of its Claude models ^[2]. A company that needs outside help to make its own product cheaper to run is not a company whose prices are about to rise. The price cuts are everywhere at once. OpenAI cut GPT-5.6 Sol pricing by more than 20 percent — its third reduction in two months ^[3]. Google cut Gemini 3.7 Flash token costs in half ^[4]. Across OpenAI, Anthropic, xAI, and Meta, the cuts track the same pressure point: Chinese open-weight models undercutting the proprietary API price ^[5]. None of this is what a company does when it believes it can charge more later. The reason the cuts won't reverse is that the switching costs that would let prices rise have been dismantled, by three separate forces. Enterprises are routing around the labs: model routers cut inference costs by up to 30 percent, and 62 percent of organizations say token costs have already changed their business decisions ^[6]. Stripe paid up to $10 billion for OpenRouter, a cost-containment layer built on top of the labs' own pricing ^[6].

Our durable belief is that AI tokens “will be a massive line item for every business’ operating expense," — OpenRouter, Inc.

Chinese developers keep shipping open-weight models that remove the cost of switching away from a US lab entirely ^[7]. And the labs' own chip suppliers are funding the disruption: Nvidia and AMD both invested in River AI, which raised $1.1 billion to build open-weight tools at two to four times lower cost than closed models ^[8]. Even the bank underwriting the story concedes the model layer doesn't stand alone. Morgan Stanley's own analysis says model providers can earn 20 to 60 percent returns on Nvidia infrastructure only if cheap model access works as a loss leader for more profitable cloud spending on storage, security, and managed APIs ^[9]. The model is the free sample; the cloud is the product. The field evidence runs the same direction. Morgan Stanley's data-center math: a fully optimized AI data center costs $25 billion a year to rent and generates $23 billion in output ^[10]. Uber burned its entire 2026 AI budget in four months with no measurable improvement in customer experience ^[11]. Canva cut its revenue forecast by a third after cutting cost-per-task by 90 percent — users made three times as many designs and the total bill went up, not down ^[12].

This validated the demand, but also showed us we needed to reduce the cost of completing an AI task to support a broad rollout. — Melanie Perkins

ServiceNow's index shows enterprise AI spending up 110 percent while operational maturity rose only 16 points — spending running roughly seven times ahead of readiness ^[13]. The labs are behaving like companies that have read the same numbers. Both have pivoted to enterprise implementation services — OpenAI with a $4 billion joint venture, Anthropic with a $1.5 billion one — because model-layer revenue alone can't carry the valuation ^[14]. Anthropic is filing for a $100 billion-plus IPO while simultaneously seeking a $10 billion revolving credit line and $15 billion in debt financing for data centers ^[15][16]. The IPO proceeds, on their own, won't cover the compute buildout. There is real adoption underneath all this — Anthropic leads US business AI adoption at 43.5 percent, and the top 1 percent of spenders pay $7,400 per employee each month ^[17]. But adoption is not the same as margin, and the pricing structure itself gives the game away. Pylon's annual Anthropic bill was projected to jump from $400,000 to $1.4 million once its introductory pricing expired at a 150-seat threshold ^[18]. The headline cuts are promotional rates with a shelf life. The real price is three and a half times what enterprises budgeted — and the IPO is landing while the promotional rate is still on the shelf.