AMD Lands Deal For AI Provider Cohere To Expand Use Of Its Instinct GPUs

While AMD is pushing to make deeper ties with top AI software vendors like Cohere to boost adoption of its Instinct data center GPUs, Nvidia continues to expand relationships with many of the same companies that have already embraced the rival’s full-stack computing platform.

AMD said Wednesday that enterprise AI startup Cohere will expand use of the chip designer’s Instinct GPUs as part of a new agreement.

Cohere—a Canadian startup that raised a $500 million funding round from AMD, Nvidia and other investors in August—plans to make its full suite of offerings available on infrastructure running on Instinct GPUs, according to the chip designer.

These products include Cohere’s North platform, which is designed to help businesses build and run private AI agents in a secure fashion, as well as the startup’s Command family of generative AI models. As part of the deal, AMD said it plans to use Cohere’s North platform as an “integral part” of internal AI workloads, including for engineering functions.

AMD positioned Cohere’s expanded use of Instinct-based infrastructure as a way for enterprise and government customers to “meet performance and TCO [total cost of ownership] goals that make their AI plans real.”

“With the AMD AI computing platform, customers benefit from excellent total cost of ownership and energy efficiency—critical advantages as enterprise and governments accelerate their AI transformation,” said Vamsi Boppana, senior vice president of AI at AMD, in a statement.

The deal marks an expansion of AMD’s existing relationship with Cohere, which said at the chip designer’s Advancing AI event in June that its Command models have been deployed on the Instinct MI300X GPUs that launched in late 2023.

“[Giving] public and private sector customers the ability to deploy our full suite of technology across our foundational models and security-optimized enterprise AI products gives them significantly greater flexibility in how they choose to deploy Cohere’s AI,” Nick Frosst, co-founder of Cohere, said in a statement. “AMD has a compelling TCO proposition with its AI infrastructure and is a great option for sovereign AI initiatives in Canada and globally."

The deal was announced as part of AMD’s move to challenge Nvidia’s dominance of the AI computing space with its Instinct GPUs and related offerings, including the ROCm software stack that boosted AMD GPU performance in its seventh major update last week.

However, AMD faces a steep climb in the market, having made only $5 billion in revenue from Instinct GPUs last year in contrast to the $102.2 billion Nvidia made from data center compute products during roughly the same period. The company has since stated that it expects Instinct revenue to reach tens of billions of dollars in the “coming years.”

While AMD is pushing to make deeper ties with top AI software vendors like Cohere to boost Instinct adoption, Nvidia continues to expand relationships with many of the same companies that have already embraced the rival’s full-stack computing platform.

Case in point: More than three months after AMD announced OpenAI as an early design partner for its upcoming Instinct MI450 GPU platform, Nvidia said on Monday that it plans to invest $100 billion in OpenAI and serve as the AI software giant’s preferred strategic compute and networking partner for AI factory growth plans.

AMD’s focus on “high-touch” engagements with its biggest and most influential customers—which also include Meta, Microsoft and xAI—is one of the main reasons the company isn’t yet making Instinct products widely accessible for channel partners to sell, AMD sales executive Kevin Lensing told CRN in an interview published last month.

The chip designer is also refining the software stack, including the recently launched ROCm Enterprise AI, to ensure channel partners can make repeatable sales and integration motions with Instinct-based systems, Lensing added.

“The challenge with doing a channel enablement on Instinct is we can’t enable a model where we have to go one-to-many if we can’t touch them all and deliver a great experience,” he said.