OpenAI execs can't stop talking about not having enough GPUs

Executives at OpenAI are openly discussing the company’s relentless quest for more computing power, particularly GPUs, which are essential for advancing their AI technologies. In a recent conversation on the "Moonshot" podcast with XPrize founder Peter Diamandis, OpenAI's Chief Product Officer, Kevin Weil, emphasized that any GPUs added to their resources are utilized almost immediately. This sentiment echoes comments made by OpenAI's CEO, Sam Altman, who indicated last month that the company plans to acquire over 1 million GPUs by year-end. In contrast, Elon Musk’s xAI recently revealed their use of a supercluster of over 200,000 GPUs, named Colossus, to enhance their AI training capabilities. Altman humorously remarked on social media that while he is proud of their progress, the team needs to expand their GPU capacity significantly. Musk, who has shifted from being an ally to a competitor of Altman, has set ambitious goals for xAI, aiming for 50 million units of Nvidia's H100 chip equivalents over the next five years. The competitive landscape in AI necessitates these advancements, as Jonathan Cohen, VP of Applied Research, described GPUs as akin to “currency” for AI researchers. Priscilla Chan, co-founder of the Chan Zuckerberg Initiative, noted that GPUs also serve as a recruitment tool within their philanthropic organization. Weil explained the reasoning behind the intense demand: "The more GPUs we acquire, the more AI applications we can develop and utilize." He likened the impact of increased GPU availability to the way enhanced internet bandwidth enabled the widespread use of video content. To tackle this pressing need for computational resources, OpenAI announced the Stargate initiative, a $500 billion collaborative project with Oracle and SoftBank. During its introduction at the White House, Altman stated that Stargate aims to propel the United States towards achieving artificial general intelligence (AGI). OpenAI’s CFO, Sarah Friar, highlighted the urgency of their GPU requirements, stating, "We are constantly under-computed. That’s why Stargate was launched and why we are focusing on larger builds." Weil elaborated on the various applications for additional GPUs, whether it’s enhancing product performance, reducing latency, or expanding access to new features across different user tiers. However, he also acknowledged the challenge of meeting the insatiable demand for GPUs among researchers within the organization, prompting ongoing efforts to enhance their computational capacity.