“All of these models have gotten quite complex,” he wrote in response to a question about why OpenAI’s next AI models were taking so long. “We also face a lot of limitations and hard decisions about [how] we allocated our compute towards many great ideas.”

Many reports suggest that OpenAI has struggled to secure enough compute infrastructure to run and train its generative models. Just this week, Reuters, citing sources, said that OpenAI has for months been working with Broadcom to create an AI chip for running models, which could arrive as soon as 2026.