Cheap AI tokens are fueling an explosion in artificial intelligence applications, but their price does not necessarily reflect the actual cost. Behind the seemingly friendly API tariffs, model providers have to pay for data centers, chips, energy, and technology development on a scale that is difficult to match with regular software companies. OpenAI once projected cumulative cash burns of about US$115 billion by 2029 due to the ever-growing computing needs.

The problem doesn't stop with model makers. AI startups that only pack third-party models are in the most vulnerable position. They pay tokens for every user request, don't master the main engine, and often sell services at a margin that depends on the current price of the API. When providers raise rates, change usage limits, or deliver similar functionality, products that previously seemed promising can lose their edge in just one update.

The risk arises through vendor dependency and the practice of Sherlocking. Data, prompt, integration, and customer workflow are becoming increasingly attached to one provider. At the same time, that provider can incorporate popular functionality directly into its core product. The industry no longer has much patience for that, said Darren Mowry, leader of the global startup organization Google, when discussing businesses that rely solely on other companies' models.

Ironically, the financial pressures also opened up new revenue streams. ChatGPT's ad trial in the United States exceeded annual revenue of $100 million in six weeks and involved more than 600 advertisers. The figure shows that AI businesses are not out of the way, but it does emphasize that cheap tokens are not permanent gifts. For startups, the term rug pull is more properly read as cost alarm, not fraud charges.