OpenAI and Anthropic face valuation test on the eve of going public: The rise of "affordable AI", the trillion-dollar myth encounters resistance from "Chinese substitution"

This financial reporting season, the costs of artificial intelligence are beginning to show in financial data. Meta (META.US), Shopify (SHOP.US), Spotify (SPOT.US), and Pinterest (PINS.US) all pointed out that the rising costs of artificial intelligence and inference are dragging down profit margins. Shopify stated that economies of scale are "partially offset by the rising costs of large language models (LLMs)". This is the cost of pricing models that support the expected IPO valuations of OpenAI and Anthropic (expected to exceed $800 billion each). These valuations are based on an assumption that OpenAI and Anthropic will maintain their market share and pricing power - competitors will struggle to catch up, and enterprise customers will continue to pay premiums as there are currently no real alternative solutions. The "constraining advantages" of Chinese models: a startling leap from 1% to 60% However, more and more data is pointing in the opposite direction. Cutting-edge artificial intelligence is becoming richer and cheaper. Chinese labs charge only a fraction of what American labs charge for similar work. Meanwhile, competitors such as NVIDIA Corporation (NVDA.US), Cohere, Reflection, and Mistral are developing cheaper, smaller, and more efficient alternative solutions for enterprises. By the time OpenAI and Anthropic submit their S-1 filings (OpenAI's confidential documents may be submitted as early as this week), the core premise of their valuations may have been challenged. The cost gap is significant and continues to widen. Enterprise AI budgets are skyrocketing. A survey by cloud cost company CloudZero showed that around 45% of surveyed companies stated that by 2025, their monthly spending on artificial intelligence will exceed $100,000, a 20% increase from the previous year. The allocation of these funds is becoming increasingly important. Artificial Intelligence benchmarking company Artificial Analysis conducts the same 10 evaluations on all mainstream models and tracks the total cost. For the strongest models from various labs: Anthropic's Claude costs $4,811; OpenAI's ChatGPT costs $3,357; DeepSeek costs $1,071; Kimi costs $948; KNOWLEDGE ATLAS's GLM costs $544. Claude's cost is nearly nine times that of the cheapest Chinese alternative under equivalent workloads. Even Alphabet Inc. Class C (GOOGL.US) is supporting innovative clouds. At this week's I/O developer conference, CEO Sundar Pichai stated, "Many companies have already exhausted their annual token budgets, and it's only May", and highly recommended Alphabet Inc. Class C's more cost-effective Flash model as a solution. Pichai pointed out that if Alphabet Inc. Class C's largest cloud customers were to migrate 80% of their workloads from edge models to Gemini 3.5 Flash, they would save over $1 billion annually. Alphabet Inc. Class C acknowledges the need for more economical options for enterprises. Furthermore, these low-cost alternative solutions are no longer trailing behind. DeepSeek, the Chinese AI lab that sparked a sell-off in US tech stocks last year, released a preview of its next-generation model last month, which is on par with OpenAI, Anthropic, and Alphabet Inc. Class C's latest models in terms of encoding, intelligence, and knowledge benchmarking tests, sometimes even surpassing them. Over the past four months, other Chinese labs including Beijing Dark Side of the Moon, Xiaomi, and KNOWLEDGE ATLAS have also released models with similar performance levels. Ali Ghodsi, CEO of Databricks, has real-time observations on this shift. The company's AI gateway connects thousands of enterprise customers with the models they are using, and Ghodsi stated that the revenue from this product is growing rapidly. He pointed out that the technology being deployed by enterprises is called "consultant models". By default, most of the work is done by a cheap open-source model. When faced with tasks that cannot be solved, it will call on the advanced models of OpenAI or Anthropic for help. "This can effectively control costs," Ghodsi said. This shift is remarkable in its speed. On OpenRouter, developers can access hundreds of AI models through a single interface, and the usage rate of Chinese models has skyrocketed from about 1% in 2024 to over 60% in May. Suppliers are also starting to sell cost reductions as a product. Dylan Field, CEO of Figma, stated that enterprises are experiencing three stages in AI application: in the first stage, almost no one uses it; in the second stage, everyone has to use it, and some are "competing in token consumption"; in the third stage, companies realize that "everyone is over-consuming" and must cut expenses. He stated that many companies are currently in the third stage. Figma is selling features that can reduce customers' token consumption by 20% to 30%. The "constraining advantages" of Chinese models and the responses of American labs The cost gap reflects the architectural differences between the two sides. American cutting-edge labs invest billions of dollars in capital expenditures, using NVIDIA Corporation's most expensive chips to train increasingly large models, while the growth rate of the American grid's capacity cannot keep up. These costs are ultimately passed on to customers. For Chinese labs, constraints have become a strategy. Due to chip export restrictions, businesses are forced to optimize actively - training competitive models with fewer computing resources and improving operational efficiency. American labs that are less affected are mainly those that focus on models for sensitive data. Aidan Gomez, CEO of Cohere, stated that the company specifically sells artificial intelligence models to banks, defense agencies, and other regulated industries, where buyers avoid choosing Chinese models regardless of price. Last year, Cohere's revenue grew sixfold, mainly due to sales in this niche market. However, this is only a relatively small part of the broader enterprise market. Beyond industries with lenient security and compliance requirements, reasons for paying premiums are becoming increasingly difficult to justify. American responses are beginning to take shape. NVIDIA Corporation, which has benefited the most from the AI boom, is now promoting a different model by releasing its self-developed AI system for free download for any company to run on their servers, as an alternative to Chinese models and closed models for companies like OpenAI and Anthropic. Reflection AI, valued at tens of billions of dollars, specifically targets the market gap by building American open-source models for companies looking for domestic alternatives. Both companies are well-funded and explicitly target the same market gap - powerful features, priced lower than cutting-edge technology, and deployed on infrastructure that American enterprises already trust. Arguments against this shift primarily focus on national security. However, these opposing views are beginning to crumble in practice. Even the US government's AI Safety Institute once pointed out that the DeepSeek model lags behind American models in terms of security and performance, but data from the institute shows that since the release of the R1 version in January 2025, downloads of the DeepSeek model have increased by nearly 1000%. Anthropic acknowledges the pressure it faces. In a policy document released in May, the company stated that American models are only "a few months ahead" of Chinese models, and warned that China is "winning globally in terms of costs". OpenAI has a different perspective. A source familiar with the company's thoughts stated that the release of each new cutting-edge model, including GPT-5.5 released last month, has driven a significant increase in API and product usage, with strong growth in enterprise demand, which they describe as "vertical growth". This source stated that open-source plays a role in low-risk tasks but has not eroded the company's core business. Pricing pressure is not one of the company's top ten concerns. However, an unnamed CEO of an enterprise AI company expressed a different view. He believes that growth is real - "but if this technology is not used, the pace of advancement in cutting-edge technology would be even faster." The collision of valuations and reality: What will the S-1 filings reveal? OpenAI and Anthropic are expected to seek public investor valuations. Both companies' valuations are close to a trillion dollars, so the S-1 filings must demonstrate the revenue growth that can support such high valuations. However, the premiums that support such high valuations, precisely where these two companies need to dominate the niche markets, are fading at the fastest pace. First is the risk of revenue concentration. Counterpoint Research data shows that in Q1 2026, Anthropic led the global large language models market with a 31.4% revenue share, slightly higher than OpenAI's 29%-1. However, around 80% of Anthropic's revenue comes from enterprise customers, while OpenAI's is around 40% - meaning Anthropic is more reliant on a few large customers, and any decision by a large customer to reduce costs and increase efficiency could significantly impact its revenue growth. Second is the difference in profit paths. Market information shows that Anthropic is expected to break even in 2028, two years ahead of OpenAI's target for 2030-1. This difference is not just a financial detail but the most important variable in "a capacity cost-heavy, profitability-uncertain industry". In 2026, Anthropic's annualized revenue soared from $9 billion at the end of 2025 to $45 billion in May, a 500% increase in five months-2. But even with such rapid growth, whether they can realize a trillion-dollar valuation in the public market depends on whether investors believe this growth can be sustained - and in a downtrend, maintaining growth is precisely the most difficult part. Regarding OpenAI, a source familiar with the situation stated that the release of each new cutting-edge model, including GPT-5.5, has driven a significant increase in API and product usage, with strong growth in enterprise demand, stating that "open-source plays a role in low-risk tasks but has not eroded the company's core business." This assessment may be correct, but it avoids a more fundamental question: how much of the "core business" remains when "low-risk tasks" cover the majority of AI calls in enterprises?