Cheaper AI Threatens OpenAI & Anthropic IPOs

This earnings season, the escalating costs associated with artificial intelligence are beginning to make a noticeable impact on corporate bottom lines. Companies like Meta, Shopify, Spotify, and Pinterest have all cited rising AI and inference expenses as a drag on their profit margins. Shopify, for instance, noted that while economies of scale are a benefit, this advantage is being “partially offset by increased LLM costs.”

This financial reality is directly challenging the pricing models that underpin the projected IPO valuations for OpenAI and Anthropic, both of which are anticipated to exceed $800 billion. These ambitious valuations are predicated on the assumption that OpenAI and Anthropic will maintain their market dominance and pricing power. This hinges on competitors being unable to rapidly match their capabilities and on enterprise clients continuing to pay a premium for what is perceived as a lack of viable alternatives.

However, an increasing body of evidence suggests a diverging trend. The landscape of cutting-edge AI is rapidly evolving towards abundance and affordability. Chinese AI laboratories are offering comparable services at a fraction of the cost of their American counterparts. Simultaneously, a new wave of Western competitors, including Nvidia, Cohere, Reflection, and Mistral, are developing more cost-effective, compact, and efficient AI solutions tailored for enterprises that are hesitant to adopt Chinese models. By the time OpenAI and Anthropic prepare for their IPO filings, with OpenAI reportedly making its confidential filing as soon as this week, the fundamental assumptions driving their valuations could be significantly undermined.

The cost disparity is substantial and is widening. Enterprise AI budgets have seen a significant surge. A survey by cloud cost management firm CloudZero revealed that approximately 45% of companies reported spending over $100,000 per month on AI in 2025, a sharp increase from 20% the previous year. The allocation of these significant investments is becoming increasingly critical. AI benchmarking firm Artificial Analysis conducts rigorous evaluations of major AI models across standardized assessments to track total costs. For their most capable models, Anthropic’s Claude incurred a cost of $4,811, while OpenAI’s ChatGPT was priced at $3,357. In stark contrast, DeepSeek’s model cost $1,071, Kimi’s cost $948, and Zhipu’s GLM was a mere $544. This illustrates that Claude is nearly nine times more expensive than the most affordable Chinese alternative for performing the same computational tasks.

Even industry giants like Google are acknowledging this market shift. At its recent I/O developer conference, CEO Sundar Pichai highlighted the burgeoning costs, stating that “many companies are already blowing through their annual token budgets, and it’s only May.” He then introduced Google’s more cost-effective Flash model as a solution, estimating that a 80% migration of workloads from frontier models to Gemini 3.5 Flash by large Google Cloud customers could yield annual savings exceeding $1 billion. This signals Google’s recognition of the growing enterprise demand for more economical AI options.

Crucially, these more affordable alternatives are no longer lagging in performance. DeepSeek, a Chinese AI firm that previously triggered a U.S. tech selloff, recently previewed its next-generation model. This new iteration demonstrates capabilities that either match or closely rival the latest offerings from OpenAI, Anthropic, and Google across key benchmarks in coding, agentic tasks, and general knowledge. Furthermore, models from other Chinese laboratories, including Moonshot, Xiaomi, and Zhipu, have been released with comparable performance levels within the last four months.

Databricks CEO Ali Ghodsi offers a real-time perspective on this evolving market dynamic. The company’s AI gateway acts as an intermediary between thousands of enterprise clients and the AI models they utilize. Ghodsi reports a sharp acceleration in revenue growth for this product. He explains that enterprises are increasingly adopting an “advisor model” strategy. In this approach, a cost-effective open-source model handles the majority of tasks by default. When it encounters a problem beyond its capabilities, it can access a specialized tool to call upon a more advanced frontier model from providers like OpenAI or Anthropic for assistance.

“You can curb costs really well this way,” Ghodsi commented, underscoring the economic efficiency of this hybrid strategy.

The speed at which this adoption is occurring is remarkable. On OpenRouter, a platform that provides developers with unified access to hundreds of AI models, the usage of Chinese models has surged from approximately 1% in 2024 to over 60% as of May. This dramatic shift highlights the accelerating preference for more affordable AI solutions.

Consequently, vendors are beginning to market cost reduction as a distinct product offering. Figma CEO Dylan Field observed that companies typically progress through three phases of AI adoption. Initially, AI sees minimal adoption. This is followed by a period where adoption becomes mandatory, sometimes leading to extravagant spending on tokens. The third phase, which many enterprises are now entering, is characterized by the realization that “everyone’s spending too much” and the subsequent need for cost optimization. Figma is actively offering features designed to reduce customers’ token consumption by 20% to 30%.

The widening cost differential between U.S. and Chinese AI development reflects fundamental differences in their underlying infrastructure and strategic approaches. American frontier AI labs are backed by hundreds of billions of dollars in capital expenditures, investing heavily in training ever-larger models on the most advanced and expensive Nvidia chips. This is occurring within a U.S. power grid that struggles to expand its capacity sufficiently, leading to these substantial costs being passed on to clients. In contrast, Chinese AI labs have strategically embraced constraints. Operating under stringent chip export restrictions, they have been compelled to pursue aggressive optimization, developing competitive models with significantly less computational power and running them with greater efficiency.

The primary defense strategy for American AI leaders often centers on trust and security. Cohere CEO Aidan Gomez, whose company specializes in providing AI models to financial institutions, defense agencies, and other highly regulated industries, asserts that these clients are inherently hesitant to adopt Chinese models, irrespective of price. Cohere’s substantial revenue growth of sixfold last year within this specific market segment underscores the demand for secure, domestic AI solutions. However, this represents a relatively niche segment of the broader enterprise market. In less regulated sectors, where security and compliance requirements are more flexible, the justification for incurring a premium cost becomes increasingly difficult to sustain.

In response to this evolving landscape, American AI companies are actively formulating their strategies. Nvidia, which has been a major beneficiary of the AI boom, is now publicly championing a different approach. The company is releasing its own open-source AI systems, which any enterprise can download and run on their own infrastructure, free of charge. This initiative aims to provide a viable alternative to both Chinese offerings and the proprietary, locked-down models from OpenAI and Anthropic. Reflection AI, recently valued at billions, is specifically focused on developing American open-source models for enterprises seeking a domestic alternative. Both companies are well-capitalized and are strategically targeting the same market gap: providing capable AI models that are more affordable than frontier options and can be deployed on infrastructure that U.S. enterprises already trust.

The initial arguments against the adoption of Chinese AI models often centered on national security concerns. However, these objections are increasingly dissolving in practical application. Even the U.S. government’s AI Safety Institute, while acknowledging potential security and performance shortcomings in DeepSeek models compared to their American counterparts, has documented a nearly 1,000% increase in downloads since the R1 release in January 2025, underscoring their growing adoption.

Anthropic itself acknowledges the intensifying competitive pressure. In a policy paper released in May, the company conceded that U.S. models are only “several months ahead” of Chinese alternatives and issued a stark warning that Beijing is “winning in global adoption on cost.”

OpenAI, however, presents a different perspective. Sources close to the company indicate that each new frontier model release, including the recent GPT-5.5, has triggered a substantial surge in API and product usage, with enterprise demand described as a “vertical wall.” This view suggests that while open-source models may play a role in lower-stakes applications, they are not currently eroding OpenAI’s core business. Pricing pressure, according to this perspective, is not a top concern for the company.

Yet, an enterprise AI CEO, who requested anonymity to protect client relationships, offered a contrasting viewpoint. While acknowledging the real growth in AI adoption, they noted that “it would expand even faster for frontier if this technique wasn’t used.” This implies that the widespread adoption of cost-saving strategies like the “advisor model” is potentially capping the growth rate of premium AI services.

This dynamic directly impacts the market that OpenAI and Anthropic are poised to present to public investors. With anticipated valuations approaching a trillion dollars each, their S-1 filings must convincingly demonstrate enterprise revenue growth and market concentration that justify such multiples. However, the premium pricing that underpins these valuations is eroding most rapidly in precisely the market segments where these leading AI labs need to establish dominance.