AI research and product company Zyphra has signed a multi-year agreement with IBM for access to a cloud-based GPU cluster.

The cluster - accessed via IBM Cloud - is described as large and comprised of AMD Instinct MI300X GPUs. The number of GPUs in the cluster has not been shared.

Zyphra will use the compute to train "frontier" multimodal foundation models, including language, vision, and audio modalities, for its Maia agent.

According to IBM, it is the largest such deal for an AMD stack on IBM Cloud.

The MI300X GPU, launched in December 2023, is built on CDNA 3 architecture and has 1.5× more memory capacity (192GB) and 1.7× more peak theoretical memory bandwidth (5.3TBps) than the previous M1250X version, delivering nearly 40 percent more compute units. AMD claims its MI300X GPUs exceed the speed of Nvidia's H100 chips, offering 1.3 petaflops of FP16 and 2.6 petaflops of FP8 performance.

In addition to the MI300X GPUs, the IBM cluster features AMD Pensando Pollara 400 AI NICs and AMD Pensando Ortano DPUs. An initial deployment was made available to Zyphra in September 2025, and is planned to be expanded in 2026.

"This collaboration marks the first time AMD's full-stack training platform—spanning compute through networking—has been successfully integrated and scaled on IBM Cloud, and Zyphra is honored to lead the way in developing frontier models with AMD silicon on IBM Cloud," said Krithik Puthalath, CEO and chairman of Zyphra. "We're excited to partner with IBM and AMD to power the next era of open-source, enterprise superintelligence."

"Scaling AI workloads faster and more efficiently is a key differentiator in achieving ROI for established enterprises and emerging companies alike," added Alan Peacock, GM of IBM Cloud. "We are delighted to support Zyphra's strategic roadmap as we collaborate with AMD to deliver scalable, economical AI infrastructure that can accelerate Zyphra's model training." 

IBM announced it would be offering the AMD MI300X GPUs via its IBM Cloud in November 2024, with plans to deploy in 2025. More recently, the two companies revealed they would be partnering on the development of quantum-centric supercomputing, or next-generation computing architectures based on a combination of quantum computers and high-performance computing (HPC).

Zyphra is an open-source AI research and product company based in San Francisco, California, with a focus on innovations in novel neural network architectures, long-term memory, and continual learning. The company closed its Series A financing round with a $1 billion valuation.