OpenAI CEO Sam Altman has acknowledged that the company’s ability to release new products is hampered by limited computing capacity.
In a recent Ask Me Anything (AMA) session, Altman explained that the increasing complexity of artificial intelligence (AI) models has introduced hurdles to rolling out new features, with computing limitations being a primary factor.
“All of these models have gotten quite complex,” Altman said in response to a question about delays in OpenAI’s next-generation AI models.
“We also face a lot of limitations and hard decisions about how we allocate our compute toward many great ideas.”
Reports have indicated that OpenAI has struggled to secure enough infrastructure to train and operate its generative AI systems at scale.
According to Reuters, the company has been collaborating with Broadcom for months to develop a proprietary AI chip, aiming to address compute shortages.
The chip could be available as soon as 2026, potentially easing some of OpenAI’s current constraints.
Compute limitations have directly impacted the development of OpenAI’s Advanced Voice Mode for ChatGPT.
Initially, the feature was teased in April as part of an upcoming update that would integrate vision capabilities, allowing the chatbot to respond to visual cues from a smartphone camera.
However, those vision features are now on hold. An investigation by Fortune revealed that the April demo of the new ChatGPT functionality was rushed to draw attention away from Google’s I/O conference.
Many OpenAI insiders were reportedly wary of the demo, believing the feature was not fully ready for release. The voice-only version of Advanced Voice Mode ultimately faced months of delays.
OpenAI’s anticipated image generator DALL-E update is also facing a prolonged development period, with no launch date currently set.
Meanwhile, Sora, OpenAI’s experimental video-generating tool, is grappling with technical setbacks.
During the AMA, OpenAI’s Chief Product Officer Kevin Weil noted that Sora’s development team is focused on enhancing the model’s safety, scaling computing, and tackling issues around impersonation and other safety risks.
The model’s initial version, unveiled earlier this year, reportedly took over 10 minutes to render a one-minute video, falling short of competitors like Luma and Runway.
Tim Brooks, a co-lead on the Sora project, recently departed for Google, adding to the team’s challenges.
In addition, Altman mentioned that OpenAI is exploring the option of allowing “NSFW” content in ChatGPT in the future, stating, “We totally believe in treating adult users like adults.”
For now, OpenAI’s main focus remains on improving its o1 series of “reasoning” models, including some advanced features presented at the company’s DevDay conference in London.
New capabilities include enhanced image understanding, but Altman clarified that OpenAI is not preparing to release a “GPT-5” model this year.
“We have some very good releases coming later this year,” Altman assured during the AMA, underlining that although the next wave of features is imminent, a major model upgrade remains out of reach as compute challenges continue to impact development.
