低コスト AI が OpenAI と Anthropic の IPO を阻む可能性（7 分読）

watch now

This earnings season, the cost of AI started showing up in the numbers. Meta, Shopify, Spotify, and Pinterest all flagged rising AI and inference costs as a drag on margins. Shopify said economies of scale were “partially offset by increased LLM costs.”

This is the bill coming due for the pricing model that underpins OpenAI’s and Anthropic’s expected IPO valuations, both projected north of $800 billion. Those numbers assume OpenAI and Anthropic will hold their market share and pricing power — that competitors can’t easily catch up, and that enterprise customers will keep paying a premium because there’s no real alternative.

But increasingly the data is pointing the other way. Cutting-edge AI is becoming abundant and cheap. Chinese labs are charging a fraction of what American labs do for comparable work, while a wave of Western challengers — Nvidia, Cohere, Reflection, Mistral — are building cheaper, smaller, more efficient alternatives for enterprises that won’t touch a Chinese model. By the time OpenAI and Anthropic file their prospectuses, with OpenAI’s confidential filing coming as soon as this week, the central premise of their valuations may already be gone.

The cost gap is wide and getting wider. Enterprise AI budgets have surged. Some 45% of companies surveyed by cloud cost firm CloudZero said they spent more than $100,000 a month on AI in 2025, up from 20% the year before. Where that money goes increasingly matters. AI benchmarking firm Artificial Analysis runs every major model through the same 10 evaluations and tracks the total cost. For each lab’s most capable model: Anthropic’s Claude came in at $4,811. OpenAI’s ChatGPT: $3,357. DeepSeek: $1,071. Kimi: $948. Zhipu’s GLM: $544. Claude is nearly nine times more expensive than the cheapest Chinese alternative for the same workload.

watch now

Even Google is making the case. At its I/O developer conference this week, CEO Sundar Pichai said “many companies are already blowing through their annual token budgets, and it’s only May,” and pitched the company’s cheaper Flash model as the answer. If the largest Google Cloud customers shifted 80% of their workloads from frontier models to Gemini 3.5 Flash, Pichai said, they would save more than $1 billion a year. The company is acknowledging that enterprises need cheaper options.

And the cheap alternatives are no longer a step behind. DeepSeek, the Chinese AI lab whose model triggered a U.S. tech selloff last year, released a preview of its next-generation model last month that matches or nearly matches the latest from OpenAI, Anthropic, and Google on coding, agentic, and knowledge benchmarks. Models from other Chinese labs, including Moonshot, Xiaomi, and Zhipu, have shipped at similar capability levels in the past four months.

Databricks CEO Ali Ghodsi has a real-time view of the shift. The company’s AI gateway sits between thousands of enterprise customers and the models they’re using, and Ghodsi said revenue from that product is climbing sharply.

The technique enterprises are deploying, he said, is called an “advisor model.” A cheap open-source model handles the bulk of the work as the default. When it hits a task it can’t solve, it’s given a tool that lets it call out to a frontier model from OpenAI or Anthropic for help.

“You can curb costs really well this way,” Ghodsi said.

The speed of the shift is striking. On OpenRouter, a marketplace that lets developers access hundreds of AI models through a single interface, Chinese models went from about 1% of usage in 2024 to more than 60% in May.

And vendors are starting to sell cost reduction as a product. Figma CEO Dylan Field said companies are moving through three phases of AI adoption: first, nobody uses it; second, everyone has to, with some “literally holding competitions of who can spend the most with tokens.” And third is the realization that “everyone’s spending too much” and has to cut back. Many enterprises, he said, are now entering that third phase. Figma is selling features that cut customers’ token consumption by 20 to 30%.

U.S. vs. China

The cost gap reflects how the two sides are built. American frontier labs are running on hundreds of billions of dollars in capex, training ever-larger models on the most expensive chips Nvidia sells, inside a U.S. power grid that can’t add capacity fast enough. Those costs get passed through to customers. For Chinese labs, constraint has become the strategy. Working under chip export restrictions, they’ve been forced to optimize aggressively — training competitive models with less compute and running them more efficiently.

The American labs’ best defense is trust. Cohere CEO Aidan Gomez, whose company sells AI models specifically to banks, defense agencies, and other regulated industries, says those buyers won’t touch Chinese models regardless of price. Cohere’s revenue grew sixfold last year selling into exactly that segment. But it’s a relatively narrow slice of the broader enterprise market. Outside of regulated industries, where security and compliance rules are looser, the case for paying a premium gets harder to make.

The American response is taking shape. Nvidia, the company that has profited most from the AI boom, is now publicly pushing a different model, releasing its own AI systems that any company can download and run on its own servers, free of charge, as an alternative to both Chinese options and the locked-down models from OpenAI and Anthropic. Reflection AI raised at a multibillion-dollar valuation specifically to build American open-source models for enterprises that want a domestic alternative. Both are well-capitalized and explicitly targeting the same gap — capable models, cheaper than the frontier, deployed on infrastructure U.S. enterprises already trust.

watch now

The case against this shift has rested on national security. But the objection is dissolving in practice. Even the U.S. government’s AI Safety Institute, which flagged DeepSeek models as lagging American ones on security and performance, documented that downloads have risen nearly 1,000% since the R1 release in January 2025.

And Anthropic itself acknowledges the pressure. In a policy paper released in May, the company said U.S. models are only “several months ahead” of Chinese ones, and warned that Beijing is “winning in global adoption on cost.”

OpenAI sees it differently. A person familiar with the company’s thinking said every release of a new frontier model, including GPT-5.5 last month, has driven a surge in API and product usage, with enterprise demand growing in what they described as a “vertical wall.” Open source has a role in low-stakes tasks, this person said, but isn’t eating into the company’s core business. Pricing pressure isn’t on the company’s top ten list of concerns.

But an enterprise AI CEO, who asked not to be named to protect customer relationships, offered a different read. The growth is real — “but it would expand even faster for frontier if this technique wasn’t used.”

This is the market OpenAI and Anthropic are expected to ask public investors to value. At nearly trillion-dollar valuations each, the S-1 has to show enterprise revenue growth and concentration that justifies the multiple. But the premium that justifies the valuation is eroding fastest in exactly the segments the labs need to dominate.

WATCH: OpenAI preparing for confidential IPO filing

watch now