OpenAI and Anthropic are heading toward public listings that would value them near or above $800 billion each. OpenAI raised its last private round at $852 billion. Anthropic hit $30 billion in annual revenue by April 2026, tripling its revenue in roughly four months, and is fielding investment offers at valuations as high as $900 billion. On paper, these are extraordinary businesses.
But the foundation of those valuations is starting to crack. Both companies justify high prices by being the best. The argument to public investors will be: enterprises need the most capable AI for important work, we make the most capable AI, so they will keep paying. That argument gets harder to make every month.
The price gap between American and Chinese AI is not a rounding error. Running a standard benchmark on Anthropic's flagship model costs about $4,800. The same test on DeepSeek costs roughly $1,000. DeepSeek's latest V4 model, released in April, charges about $3.48 per million words of output. OpenAI and Anthropic charge $30 and $25 respectively for the same amount of work. That is roughly an 8 to 9 times difference for work that benchmarks suggest is now only a few months behind in quality.
Anthopic itself acknowledged this in a policy paper in May, saying U.S. models are only "several months ahead" of Chinese ones and that Beijing is winning on global adoption through cost.
The market is already shifting. Open-source and cheaper AI models captured 38% of enterprise usage by early 2026, up from 11% just a year before. The average cost enterprises pay for AI dropped 67% in a single year. Companies are not abandoning expensive models entirely. They are using a mix: cheap models handle the bulk of the work, and the expensive ones are called in only for tasks that demand them.
This matters directly for the IPO math. Both companies have raised money at valuations that assume sustained pricing power and growing market share. OpenAI reportedly missed its 2025 ChatGPT revenue target. Enterprise pricing for OpenAI rose 120% year over year at the contract level, but finance teams are pushing back, scrutinizing whether the cost is justified. At the same time, Anthropic overtook OpenAI in enterprise spending share in April 2026, rising from 9% to 34% of tracked businesses in twelve months, which shows it can compete but also that no premium player has a locked-in position.
The companies are responding in different ways. OpenAI launched a "Guaranteed Capacity" product in May, offering enterprises one, two, or three-year contracts at discounted rates. This locks in revenue and helps the IPO story, but it is also a sign that demand is not quite the vertical wall the company describes internally. Anthropic has moved away from flat-rate pricing toward billing per actual usage, which will give it cleaner revenue data for investors but also exposes just how much the previous pricing obscured real demand.
The one strong defense American labs have is trust. Banks, defense agencies, insurance companies, and other regulated industries will not send data to Chinese servers regardless of cost. Cohere, which sells specifically into those segments, grew sixfold last year on exactly that pitch. But regulated industries are a minority of the total enterprise market. Outside of them, the case for paying a nine-times premium gets thinner with every new cheap model release.
For any business currently spending serious money on AI, the practical implication is already here: the tools to run most of your AI workloads at a fraction of the current cost exist today. The constraint is not availability, it is the internal effort required to change suppliers or run a mixed setup. That is a project worth evaluating, especially as AI costs become a visible line item on income statements rather than a buried IT expense.