This earnings season, the costs of AI are starting to show up in the numbers. Meta, Shopify, spotifyand pinterest All cited rising AI and inference costs as a drag on profits. Shopify said economies of scale were “partially offset by increased LLM costs.”
This is a bill coming due regarding the pricing models that will underpin the expected IPO valuations of OpenAI and Anthropic, both of which are expected to exceed $800 billion. These numbers assume that OpenAI and Anthropic continue to hold market share and pricing power, meaning competitors cannot easily catch up and enterprise customers continue to pay a premium due to the lack of substantial alternatives.
But data increasingly points in a different direction. Cutting-edge AI is becoming abundant and cheap. While Chinese labs are charging a fraction of what U.S. labs charge for comparable research, a wave of Western challengers— NvidiaCohere, Reflection, and Mistral are building cheaper, smaller, and more efficient alternatives for companies that can’t afford the Chinese model. By the time OpenAI and Anthropic file their prospectuses, with OpenAI’s confidential filing expected as early as this week, the core assumptions of their valuations may already be gone.
The cost disparity is large and widening. Enterprise AI budgets are skyrocketing. About 45% of companies surveyed by cloud costing firm CloudZero said they will spend more than $100,000 per month on AI in 2025, up from 20% a year ago. Where that money goes becomes increasingly important. AI benchmarking firm Artificial Analysis runs all major models through the same 10 evaluations to track total costs. For each lab’s most capable model, Anthropic’s Claude came in at $4,811. OpenAI’s ChatGPT: $3,357. Deep Seek: $1,071. Kimi: $948. Zhipu’s GLM: $544. Claude is almost 9 times more expensive than the cheapest Chinese alternative for the same workload.
flat google claims. “Many companies have already exhausted their annual token budgets, and we’re only in May,” CEO Sundar Pichai said at the I/O Developers Conference this week, touting the company’s cheaper Flash model as the answer. If Google Cloud’s largest customers migrated 80% of their workloads from the Frontier model to Gemini 3.5 Flash, they could save more than $1 billion annually, Pichai said. The company recognizes that businesses need cheaper options.
And cheaper alternatives are already a step behind. DeepSeek, the Chinese AI research lab whose models sparked a U.S. tech stock crash last year, last month released a preview of its next-generation model that matches or closely matches the latest models from OpenAI, Anthropic, and Google on coding, agent, and knowledge benchmarks. Models from other Chinese labs, including Moonshot, Xiaomi, and Zhipu, have also shipped with similar functionality levels in the past four months.
Databricks CEO Ali Ghodsi is seeing this change in real time. The company’s AI gateway sits between thousands of enterprise customers and the models they use, and revenue from the product is growing rapidly, Ghodsi said.
He said the technology companies are implementing is called the “advisor model.” Cheap open source models handle most of the work by default. If you run into a task that you can’t solve, it provides tools that allow you to call on OpenAI or Anthropic’s frontier models for help.
“We can control costs very well this way,” Godi says.
The speed of shifting is amazing. On OpenRouter, a marketplace where developers can access hundreds of AI models through a single interface, the usage of Chinese models has increased from about 1% in 2024 to more than 60% in May.
And vendors are starting to sell cost savings as a commodity. figma CEO Dylan Field said companies are going through three stages of AI adoption. Second, everyone needs to do so, and some are “holding literal competitions to see who can spend the most tokens.” And the third is the recognition that “everyone is spending too much” and must cut back. He said many companies are now in that third stage. Figma sells features that reduce customers’ token consumption by 20-30%.
US vs. China
The difference in cost reflects how both are constructed. Frontier Labs in the U.S. is pumping hundreds of billions of dollars in capital investment and training even larger models on the most expensive chips sold by NVIDIA in a U.S. power grid that can’t add capacity fast enough. Those costs are passed on to the customer. For Chinese research institutes, constraints have become a strategy. Working under chip export restrictions, they are forced to aggressively optimize, training competitive models on less compute and running more efficiently.
The greatest defense of American laboratories is trust. Cohere CEO Aidan Gomez, whose company sells AI models specifically to banks, defense agencies and other regulated industries, said those buyers would stay away from Chinese-made models regardless of price. Cohere’s revenue increased sixfold last year by selling to that very segment. However, this is a relatively small segment of the broader enterprise market. Outside of regulated industries, lax security and compliance rules make it difficult to argue the need to pay a premium.
America’s response is beginning to take shape. Nvidia, the company most benefiting from the AI boom, is now publicly promoting a different model, releasing its own AI system that any company can download and run for free on its own servers as an alternative to both Chinese options and OpenAI and Anthropic’s lockdown models. Reflection AI was raised at a multibillion-dollar valuation to build a U.S. open source model for companies seeking domestic alternatives. Both are well-capitalized and clearly target the same gaps. That means it’s a capable model that’s cheaper than Frontier and deployed on infrastructure that U.S. companies already trust.
Litigation over this change is a matter of national security. However, in reality, this opposition is beginning to disappear. Even the US government’s AI Safety Institute, which warned that the DeepSeek model was lagging behind American models in terms of security and performance, has documented a nearly 1,000% increase in downloads since the R1 release in January 2025.
And Anthropic itself recognizes the pressure. In a policy paper released in May, the company said the U.S. model was only “several months” ahead of the Chinese model and warned that Beijing was “winning global adoption on cost.”
OpenAI sees it differently. People familiar with the company’s thinking said usage of its APIs and products has surged with each new frontier model released, including last month’s GPT-5.5, and enterprise demand has grown amid what it described as a “vertical wall.” The person said open source plays a role in low-risk tasks, but has not penetrated the company’s core business. Price pressure is not among the company’s top 10 concerns.
But one enterprise AI CEO, who requested anonymity to protect customer relationships, took a different view. This growth is real. “But if this technology had not been used, the expansion of the frontier would have been even faster.”
This is a market where OpenAI and Anthropic are expected to seek valuation from public investors. With valuations approaching $1 trillion each, S-1s must demonstrate corporate earnings growth and concentration to justify their multiples. But the premiums that justify valuations are eroding fastest in precisely the segments that labs need to dominate.
WATCH: OpenAI prepares for secret IPO filing
