OpenAI builds own AI chip 'Jalapeño' with Broadcom to cut Nvidia dependency
Key Questions
What steps is OpenAI taking to address compute shortages?
OpenAI is testing its first homegrown "Jalapeño" AI chips and partnering with Broadcom to build custom hardware as part of a broader strategy to reduce reliance on external suppliers. This comes amid reported shortages that prompted spending cuts and a new $122B funding round, alongside a 'Guaranteed Capacity' enterprise offering with multi-year commitments. The company is also pursuing deals with Cerebras and others to secure additional resources.
How is OpenAI tackling the energy demands of its AI infrastructure?
OpenAI-backed Helion has raised $465M and obtained the world's first regulatory licenses to operate a fusion power plant, targeting grid delivery by 2028 to support AI workloads. This aligns with large-scale projects like the Stargate Michigan data center and Altman's comments on soaring token usage costs becoming a major concern. High idle GPU rates and corporate spending caps underscore the need for sustainable energy solutions.
What risks are associated with OpenAI's data center and infrastructure plans?
OpenAI faces regulatory scrutiny, including a California AG complaint regarding Altman's ties to Cerebras, alongside community opposition to the Stargate Michigan project involving Erin Brockovich. Despite progress on groundbreaking and local job creation, Altman has acknowledged criticisms around AI spending ROI and token cost overruns. These factors add pressure amid competitive shifts and massive capex plans.
OpenAI shipped its first inference chip in just 9 months, aiming for 50% cost reduction vs. Nvidia. This is a major infrastructure play to lower API costs and reduce reliance on a single supplier, directly addressing the token cost crisis Altman flagged.