AMD has talked of its great success last year selling its Instinct MI300X GPUs, and partners, customers, and industry analysts are hotly anticipating the arrival of the MI350X series in mid-2025 and the MI400 to follow, but in between all that another member of the Instinct family is due for its moment in the sun.
Vultr announced it is deploying AMD’s MI325X GPUs in its Chicago data center region, making it the first cloud provider to tout its use of the 325X. Vultr’s need for the MI325X, after its initial deployment of the MI300X last year, may speak to how just strong demand continues to be for GPUs in the data center as AI inference activity ramps up, and how urgently some providers are working to fill that demand now instead of holding out for the next-generation GPUs waiting in the wings. Also, while the MI325X may have a brief time window in which to make an impression before the new GPUs arrive, it still presents an opportunity for AMD to sell more chips into customers who either can’t afford Nvidia GPUs, or might be on a long waiting list for them.
“There is currently a serious shortage of AI compute at data centers worldwide,” said Jack Gold, president and principal analyst at J. Gold Associates, in an email to Fierce Electronics. “Nvidia can’t provide enough chips, and what they do provide is expensive. AMD chips do an admirable job on AI processes that may not require tens of thousands of [Nvidia] Blackewells, and that includes the expansion of inference processes. AMD chips are filling in a gap that Nvidia can’t right now.”
The MI325X is the last generation of AMD GPU using the company’s CDNA3 architecture, and the MI350X series, including the MI355X, will be the first to leverage the CDNA4 architecture, which promises more powerful performance and memory, and greater power efficiency. While AMD has positioned the 325X as a chip that favorably competes with Nvidia’s H100 [which Vultr also uses], the 350X family will take a shot at Nvidia’s Blackwell family.
During AMD’s recent fourth quarter 2024 earnings call, AMD President and CEO Lisa Su told analysts that customer demand drove AMD to move its timeline for release of the MI350X series from the second half of 2025 into “the middle of the year,” and that she expects the second half of 2025 to be stronger in terms of revenue than the first half due to that acceleration in the release schedule.
“The newer AMD chips are coming soon, but if you have customers waiting to use your service and pay you, you put in place whatever you can to generate the revenues,” Gold said. “And the MI325s won’t all of sudden become obsolete once the 350s come out. They will still be useful in providing compute for many AI workloads.”
The Vultr news comes not long after AMD Ventures participated in a $333 million financing round for the cloud infrastructure firm that valued Vultr at $3.5 billion. That does not exactly put Vultr in the same weight class the cloud giants ike AWS, Google Cloud, and Microsoft Azure, but the Florida-based company is actively growing its data center footprint around the world. For Vultr, the MI325X GPUs are hosted on Supermicro AS -8126GS-TNMR 8-U servers.
“The AMD Instinct MI325X sets new AI industry standards, delivering incredible performance and efficiency for inference,” said J.J. Kardwell, CEO of Vultr, in a statement. “The Instinct MI325X gives our customers priority access to state-of-the-art GPU hardware, empowering them to scale AI deployments in production. Together, we are unlocking new possibilities for AI innovation and shaping the future of cloud GPU infrastructure on a global scale.”