OpenAI posts job listings for engineers to optimize AI models for TPUs; Anthropic also taps TPUs through Broadcom deal

Major AI-based apps including ChatGPT and Google's Gemini [AFP]
Major AI-based apps including ChatGPT and Google's Gemini [AFP]

Google's tensor processing units are gaining ground as an alternative in the AI chip market long dominated by Nvidia, with OpenAI — one of Nvidia's most important customers — now moving to purchase Google TPUs, following similar moves by Anthropic and Oracle.

The Information reported Tuesday (local time), citing sources, that OpenAI plans to buy Google TPUs.

The specific purchase amount and deal terms were not confirmed. The outlet said the deal, if completed, would be a significant boost for Google's TPU business as it challenges Nvidia's GPUs.

OpenAI recently posted job listings seeking engineers to optimize its cutting-edge AI models for TPUs. The postings noted that experience in either training or inference is welcome.

The Information interpreted the listings as a sign that OpenAI's TPU optimization efforts are still in their early stages.

Bloomberg also reported the same development, describing the purchase plan as part of OpenAI's broader push to rapidly expand its AI computing capacity.

Google began developing TPUs more than a decade ago, initially to improve the speed and efficiency of its search engine. The chips have since proven useful for AI workloads and have emerged in recent years as an alternative to Nvidia GPUs.

OpenAI struck a partnership with Google on Monday, allowing its AI models to run on Google Cloud. With that deal, OpenAI's GPT models became the second, after Anthropic's Claude, to be available across all three major cloud platforms — Microsoft Azure, Amazon Web Services and Google Cloud.

OpenAI also plans to begin using "Jalapeño," a proprietary AI chip it co-developed with Broadcom, later this year. The chip's development is led by Richard Ho, who previously worked on TPU development at Google.

Earlier, Anthropic disclosed in an IPO filing with the SEC on Tuesday that it had signed a five-year agreement with Broadcom to lease Google TPU-based equipment worth $161 billion.

Anthropic projected it would spend an additional $150 billion on TPUs from 2028 onward.

Also on Tuesday, Broadcom announced it would arrange $63 billion in financing from financial institutions to support Anthropic's Google TPU purchases.

The transaction will be structured through a special purpose vehicle. The SPV will raise funds from financial institutions to purchase Broadcom's AI accelerators and networking equipment, then lease them to Anthropic. The AI accelerators are TPUs co-developed by Google and Broadcom.

Oracle also announced Tuesday that it will offer Google TPUs on its cloud platform starting next year — the first time Google has supplied TPUs to a major cloud provider outside its own platform.

Despite OpenAI's plans to purchase Google TPUs, The Information assessed that OpenAI's relationship with Nvidia remains deep.

In August, Nvidia agreed to provide up to $105 billion in credit support for a data center project in Ohio that OpenAI plans to lease.

Nvidia CEO Jensen Huang said in a blog post at the time that OpenAI had committed to deploying a substantial amount of Nvidia infrastructure through 2030, representing roughly $600 billion in computing business opportunities for Nvidia.

TPUs incorporate HBM chips, and Samsung Electronics and SK hynix are understood to supply HBM for Google's TPUs.


mokiya@heraldcorp.com