Cheap AI could derail OpenAI and Anthropic's IPOs

This earnings season, the financial impact of AI costs has become increasingly evident. Major companies like Meta, Shopify, Spotify, and Pinterest have all reported that escalating AI and inference expenses are negatively impacting their profit margins. Shopify noted that while economies of scale have helped, they have been partially offset by rising costs associated with large language models (LLMs). These developments raise concerns for OpenAI and Anthropic, both of which are eyeing IPO valuations exceeding $800 billion. Such figures rely on the assumption that these companies will maintain their market share and pricing power, with the expectation that competitors won’t easily catch up and that enterprise clients will continue to pay a premium due to a lack of alternatives. However, emerging data suggests that the landscape is changing. Advanced AI technology is becoming more accessible and affordable. Chinese laboratories are offering services at significantly lower prices compared to their American counterparts, while a surge of Western competitors, including Nvidia, Cohere, Reflection, and Mistral, are developing cost-effective solutions tailored for enterprises that prefer to avoid Chinese models. As OpenAI and Anthropic prepare to file their prospectuses, with OpenAI’s confidential filing anticipated soon, the very foundation of their lofty valuations may be at risk. The disparity in costs is widening. A recent survey by CloudZero found that 45% of companies reported spending over $100,000 per month on AI in 2025, a steep increase from just 20% the previous year. The distribution of that spending is becoming crucial. AI benchmarking firm Artificial Analysis evaluated major models, revealing stark differences in total cost. For instance, Anthropic's Claude model costs $4,811, while OpenAI's ChatGPT is at $3,357. In contrast, cheaper alternatives like DeepSeek, Kimi, and Zhipu's GLM cost significantly less, with Zhipu’s model priced at just $544—making it nearly nine times less expensive than Claude for similar workloads. Even Google is acknowledging the need for more affordable options. During its I/O developer conference, CEO Sundar Pichai highlighted that many companies are exceeding their annual token budgets early in the year and promoted Google’s economical Flash model as a solution. If major Google Cloud clients transitioned 80% of their workloads to this model, they could realize savings exceeding $1 billion annually. The shift towards budget-friendly alternatives is accelerating. On OpenRouter, a marketplace for AI models, usage of Chinese models surged from approximately 1% in 2024 to over 60% by May 2025. Vendors are now marketing cost reduction as a product feature, with companies like Figma reporting that clients are moving through three phases of AI adoption—initial reluctance, widespread usage, and ultimately, a realization of overspending. The cost differential reflects the operational approaches of American and Chinese labs. U.S. labs are investing heavily in capital expenditure, training larger models on the most advanced chips, while Chinese labs are optimizing their operations under export restrictions, allowing them to train competitive models more efficiently. The American response is gaining traction. Nvidia, a key player in the AI boom, is now promoting its own systems that can be downloaded and run on company servers, offering a free alternative to both Chinese models and the locked-down options from OpenAI and Anthropic. Similarly, Reflection AI aims to create American open-source models catering to enterprises seeking domestic alternatives. Despite concerns over national security, the push for cost-effective solutions is gaining momentum. The U.S. government's AI Safety Institute has noted a significant rise in downloads of Chinese models, indicating a shift in the market. Meanwhile, Anthropic has admitted that U.S. models are only slightly ahead of their Chinese counterparts and warned that Beijing is gaining ground in terms of global adoption based on cost. OpenAI, on the other hand, maintains a different perspective, stating that each new frontier model release has spurred increased demand for their API and products. They argue that while open-source models play a role in low-stakes tasks, they are not significantly threatening OpenAI’s core business. Yet, an anonymous enterprise AI CEO suggests that the growth potential for frontier models could be even greater without the rising competition from cheaper alternatives. As OpenAI and Anthropic prepare to approach public investors, they will need to demonstrate solid enterprise revenue growth to justify their high valuations. However, as competition intensifies and the premium they once enjoyed diminishes, the path ahead may be more challenging than anticipated.