The high cost of tokens in the artificial intelligence sector has sparked widespread debate, even leading Silicon Valley giants to restrict employee usage. However, a seemingly contradictory phenomenon exists where the current token consumption within AI subscription services is actually supported by massive hidden subsidies, with some token subsidies reaching as high as 70 times the subscription fee. According to market analysis, as OpenAI and Anthropic enter their IPO sprint phases, concerns are mounting that once publicly listed, these companies may emulate the subsidy wars of the internet era, adjusting token prices by increasing average revenue per user. An analysis reveals that SemiAnalysis conducted a deep dissection of the subscription models of OpenAI and Anthropic, using AI to simulate tasks and calculate API consumption. The results show that the more expensive the subscription plan, the higher the subsidy ratio. This "reverse pricing" strategy aims to lock in enterprise decision-makers through massive losses, thereby drawing entire teams and business lines onto the platform. This logic of trading losses for scale was successfully validated in the mobile internet era by companies like Uber Technologies Inc (NYSE: UBER) and Meituan, based on the core assumption that users are "locked in" during the subsidy period and face high switching costs. However, the underlying logic in the AI field is entirely different. As interfaces between different platforms become increasingly standardized and development frameworks widely support multi-model switching, developers can migrate API calls from Claude to GPT or Gemini within a day. For ordinary users, switching platforms is even easier. Tokens are essentially a standardized resource; regardless of the issuer, their value attributes converge. This means that once subsidies cease, users will quickly flow to lower-priced platforms. Furthermore, complex tasks such as code writing and document analysis consume 5 to 30 times more tokens per conversation than ordinary dialogues. Under a $100 Claude Max subscription plan, a single programming task can nearly exhaust the entire $100 token allowance. Uber Technologies Inc's (NYSE: UBER) CTO even revealed the company had exhausted its entire 2026 AI budget within four months. In this subsidy war, the funding sources of the combatants have structural differences. Alphabet Inc (NASDAQ: GOOGL) generates over $300 billion in annual advertising revenue, a cash flow that occurs automatically without needing to be explained to investors. Alphabet using ad revenue to subsidize AI tokens is akin to someone owning an oil well starting a price war at a gas station, while OpenAI and Anthropic rely on financing to purchase "oil." OpenAI has raised over $180 billion with a valuation exceeding $850 billion; Anthropic has raised over $130 billion. These funds come from investors expecting substantial returns post-IPO. Analysis notes that Bill Maris, founder of Google Ventures, stated bluntly in a podcast that if Alphabet were to lower token prices by 80%, the business models of OpenAI and Anthropic would face collapse. As a former Alphabet executive who incubated Waymo and Google X, Maris pointed out that if the Gemini API price dropped to a quarter with comparable quality, enterprise clients would inevitably switch, and competitors lacking a "money printer" would be unable to match the price cuts. The technological barriers between large models are rapidly shrinking; today's leading edge could be caught up within three months, which is completely different from the technological generation gap between the iPhone and Nokia. Competition among AI models is more like a sandbar, easily washed away by the tide. Meta Platforms Inc (NASDAQ: META) can open-source free models, China has DeepSeek and Byte, and Amazon.com Inc (NASDAQ: AMZN) is also advancing its own models. When one company cuts prices, competitors follow, ultimately triggering a price war. Historical experience shows this competitive model often ends with a few winners monopolizing the market and controlling pricing power, as seen with Amazon.com Inc's (NASDAQ: AMZN) rise in e-commerce and cloud computing. However, another possible outcome is AI becoming standardized infrastructure, like electricity, bandwidth, and cloud storage. Due to minimal product differentiation and low user switching costs, competition would drive prices down near the cost line, with profit margins approaching zero. OpenAI and Anthropic currently have almost no effective countermeasures. They cannot cut prices like Alphabet can, nor can they rely on technological differences to maintain high prices. If AI ultimately moves towards "utility-like" infrastructure competition, this war will have no final victor but rather an endless race. Meituan founder Wang Xing once noted that the goal of some competitions is not to defeat the opponent but to ensure one stays in the game. OpenAI's high valuation is not due to the massive funds needed to train models but to continue participating in this price war. Alphabet's preparation for an 80% price cut is not to eliminate rivals but to ensure its own core position in the AI era, much like ensuring it didn't fall behind in the mobile era by providing the Android system for free. Anthropic doubled the API price for its flagship product Fable 5 to $10 per million tokens for input and $50 per million for output. This is essentially proactively filtering for enterprise clients willing to pay for premium features, as subsidies on the consumer end cannot defeat Alphabet. Every price war has expanded the scale of AI applications, bringing more data, application scenarios, and developers, making all participants' models more powerful. This is not a zero-sum game but a process where competition makes everyone stronger, yet no one can reap exorbitant profits. This mirrors the history 140 years ago when Edison and Westinghouse vied for electrical standards, ultimately making the electricity industry an infrastructure no one could monopolize. Analysis suggests Bill Maris's remarks may foreshadow a larger trend: in the AI field, tokens will ultimately not belong to any single company. For OpenAI and Anthropic, even with technological advantages and massive funding, their goal of "making big money through AI" may be destined to fail. What they strive to build will likely become the "water, electricity, and roads" for the next generation of society. For users, as long as the token subsidy war continues, people can keep enjoying the benefit of obtaining $400 worth of computing power for a $20 cost.
Comments