OpenAI explores alternatives to Nvidia chips, sources say

OpenAI is exploring alternatives to Nvidia’s AI inference chips amid ongoing negotiations and product needs.

The company has engaged with chipmakers including AMD, Cerebras, and Groq since last year.

It seeks hardware with faster inference performance for applications like ChatGPT.

OpenAI relies heavily on Nvidia’s hardware but is dissatisfied with its speed for tasks like coding and software communication.

Nvidia CEO Jensen Huang dismissed reports of tension, affirming Nvidia’s commitment to OpenAI.

OpenAI recently partnered with Cerebras but did not license technology from Groq. Nvidia licensed Groq’s chip designs and hired their engineers.

OpenAI aims to secure hardware to handle about 10% of its future inference workload.

🔗 Source: Reuters

🧠 Food for thought

Implications, context, and why it matters.

OpenAI is hunting for inference hardware, a shift away from training AI models as a capital expense toward inference as an operating cost tied to revenue 1.
This “earning phase” already brings in huge money. Inference makes up about 40% of Nvidia’s data center revenue, based on a LinkedIn post by David Pantera 1.
Serving AI to millions of users needs different infrastructure than training. It leans on reliability at scale and quicker answers for complex queries rather than the raw throughput GPUs deliver in training runs 1.

OpenAI’s search for options also tracks a wider race where rivals challenge Nvidia on more than chip speed, including system-level integration or price 2.
Amazon Web Services (AWS) is promoting its Trainium chips as a cheaper choice with more memory per dollar, according to theCUBE Research’s analysis 2.
A broader mix of suppliers may lower systemic risk. One LinkedIn post describes “circular dependencies” where top AI labs, cloud providers, and Nvidia are deeply tied to each other’s success 1.