There’s a memorable scene in “The Dark Knight” where the Joker sets a huge pile of money on fire. Launching your own frontier AI model today might resemble that act of waste — just ask OpenAI’s chairman, Bret Taylor.
Taylor, who has experience with major companies like Google, Facebook, and OpenAI in training large language models (LLMs), described developing new AI models as a “great way to burn through millions of dollars.”
On a recent episode of the Minus One podcast, Taylor advised AI entrepreneurs to focus on creating services and use cases instead of attempting to build new models from scratch. “Unless you work at OpenAI, Anthropic, Google, or Meta, you’re likely not able to create one of those,” he stated. The financial requirements are so significant that it leads to industry consolidation.
This high cost barrier has prevented the emergence of a viable “indie data center market,” noted Taylor. Instead, he recommended collaborating with major AI players. OpenAI, for instance, profits by selling “tokens” for its API, which developers can integrate into their applications.
While the American LLM market is largely dominated by a few companies, Taylor’s theory has been put to the test by international players. In January, DeepSeek launched its R1 reasoning model, accompanied by a chatbot, using simpler and less powerful chips to minimize expenses.
DeepSeek became the top app on the App Store, surpassing ChatGPT, and sparked discussions among tech enthusiasts and investors about whether major tech companies are overspending on AI model development.
During the podcast, Taylor outlined alternative routes for entrepreneurs in the AI space, abstracting away from training new models. He pointed to the “AI tools market,” likening it to providing pickaxes during a gold rush. However, he cautioned that many tools today could soon be replaced by foundational model providers.
He also suggested the potential for “applied AI companies,” predicting that what we now recognize as SaaS applications will evolve into agent-based companies by 2030.
Building an entirely new model, as Taylor warned, is a surefire way to “burn your capital.” He described handmade models as “fast-depreciating assets” that come with hefty price tags. He advised against this approach, suggesting exploring alternatives like leasing or open-source models instead.