OpenAI plans to procure Google TPUs

As companies seek cheaper chips suitable for running artificial intelligence models, Google's Tensor Processing Units (TPUs) are gaining popularity as an alternative to Nvidia's market-leading AI chips. Google began selling TPUs directly to some customers this year.

This move also reflects OpenAI's strategic intent to pursue multiple approaches and urgently boost its own AI computing power.

The company had previously announced plans to begin using its own custom chips later this year to support the operation of its AI models. This means it has joined the "custom chip wave" set off by multiple companies, jointly challenging Nvidia's dominance in the computing power field.

It is worth noting that Jalapeno, the in-house chip developed by OpenAI in collaboration with Broadcom, was developed with the assistance of senior Google employees. Richard Ho, who heads OpenAI's internal chip team, previously participated in the development of Google TPUs.

Google began developing its TPUs more than a decade ago. Initially, these TPUs were used to accelerate the company's web search engine and improve efficiency. But these chips quickly proved extremely valuable for handling AI workloads and in recent years have become an alternative to Nvidia GPUs.

Nvidia's computing power hegemony faces a new challenge

This procurement plan will deepen the ties between OpenAI and Alphabet's Google—even though the two companies are in direct competition in the fields of artificial intelligence products and search.

For a long time, OpenAI's relationship with Nvidia has been quite close. From the training of the first-generation GPT model to ChatGPT sweeping the globe, almost all of OpenAI's core model evolution has been built on clusters of tens of thousands of Nvidia GPUs. Jensen Huang personally delivered the world's first DGX AI supercomputer to OpenAI, and OpenAI's explosive popularity directly pushed Nvidia to the top of global market capitalization.

In August this year, Nvidia CEO Jensen Huang said that OpenAI has committed to deploying Nvidia AI infrastructure on a large scale by 2030, with existing and planned deployments of about 12 gigawatts and the opportunity to expand to about 16 gigawatts.

However, this has not stopped OpenAI from seeking to diversify its supply chain. Faced with increasingly high inference costs, a massive computing power gap that often runs into millions of chips, and the bargaining disadvantages and supply disruption risks brought by over-reliance on a single supplier, OpenAI is clearly unwilling to stake its entire fate on Nvidia alone.

It is not an isolated case. Oracle also announced on Tuesday that it will, for the first time, offer customers computing power based on Alphabet's Google TPU chips on its cloud platform.

Regarding this arrangement, Mahesh Thiagarajan, president of Oracle Cloud Infrastructure, said in a statement that different AI workloads have different computing power needs, and this choice allows customers to flexibly select the infrastructure best suited in terms of performance, throughput and economics for each AI workload.

Google executive Amin Vahdat said in a statement announcing the deal with Oracle that "the next generation of artificial intelligence will require diverse computing power."