Vultr to offer AMD's MI455X chips for AI cloud push

Vultr will offer AMD's Instinct MI455X graphics processors and support AMD's Helios rack-scale system, making it among the first cloud providers to make the new hardware available.

The move expands Vultr's artificial intelligence infrastructure as cloud providers compete to attract companies training models and running inference workloads. It is taking pre-orders for reserved capacity tied to the new systems, with deployments planned across 2027 and 2028.

AMD's latest products target customers that need large clusters of accelerators for AI and high-performance computing. Vultr said the systems will support AI training, inference, fine-tuning, agentic AI and scientific computing. Helios is designed as a rack-scale configuration that can support 72 GPUs.

The announcement also highlights the growing importance of memory and cooling in AI infrastructure. Vultr will incorporate direct liquid cooling options for denser deployments in response to the heat and power demands of large-scale AI systems.

Cloud providers have been racing to secure access to the newest chips from major semiconductor groups as demand for AI computing rises. For Vultr, which positions itself as a global cloud infrastructure provider, early access to AMD's newest accelerator offers another way to compete in a market shaped by demand for specialised hardware and broad geographic reach.

Vultr linked the launch to a shift in customer demand from experimentation to production use. That change has pushed buyers towards systems that can handle repeated inference tasks, larger models, and workloads spread across multiple servers and racks.

J.J. Kardwell, Chief Executive Officer of Vultr, said demand is changing. "Customers are rapidly moving from AI experimentation into production, and increasingly that means agentic and inference-heavy workloads that require infrastructure built for scale," Kardwell said.

"The AMD Instinct MI455X and Helios rackscale architecture give our customers the flexibility, control, and price-to-performance they need to scale their AI initiatives. Making this architecture available across Vultr's global infrastructure enables teams to build and deploy without constraints," he added.

AI demand

AMD said the Instinct MI455X is built on its CDNA architecture and includes HBM4 memory, higher memory bandwidth, and support for lower-precision AI data types. Those features matter because developers are trying to fit larger models into memory while reducing the cost of serving AI applications once they are in use.

Helios adds another layer by combining GPUs with AMD Pensando networking technology in a rack-scale design. The approach reflects a broader industry shift away from buying individual accelerators in isolation and towards purchasing systems that package compute, networking, and cooling together.

The market is also becoming more focused on agentic AI, where software systems perform sequences of tasks and call models repeatedly rather than answering a single prompt. That pattern can sharply raise demand for inference, memory, and networking compared with simpler chatbot-style applications.

IDC expects a tenfold increase in the number and complexity of AI agents used by enterprises over the next five years, according to figures cited by Vultr. The research firm said those workloads will require infrastructure that can handle data-intensive processes and constant inference loops.

Dave McCarthy, Group Vice President, Cloud and Datacenter Infrastructure at IDC, outlined that view in a statement carried by Vultr. "As enterprises transition into the agentic era, the massive influx of complex, data-intensive workflows requires a fundamental shift in infrastructure design," McCarthy said.

"Vultr's deployment of the AMD Instinct MI455X GPU provides the massive memory capacity and rack-scale efficiency necessary to support the continuous inference loops that define next-generation AI workloads," he added.

AMD push

For AMD, the deal is part of a broader effort to win more share in AI data centres, where Nvidia remains the dominant supplier. The chipmaker has been expanding its software and hardware offerings to persuade cloud providers and enterprise buyers they can build large AI systems on an alternative stack.

That effort includes ROCm, AMD's software platform for AI training and inference, as well as a focus on open standards. Software support has become central to competition between chipmakers because customers need tools that work across model frameworks and allow them to manage training and deployment at scale.

Andrew Dieckmann, Corporate Vice President and General Manager, Data Centre GPU Business Unit at AMD, said the rack-scale system is meant to meet that demand. "AMD Helios rackscale solution brings the best of AMD compute and networking technology together into a single rack-scale building block, delivering leadership performance, performance per watt, and cost per token for the next wave of AI," Dieckmann said.

"As demand accelerates for large-scale inference and frontier-model training, AMD Helios gives customers a clear path to scale from rack to cluster with the efficiency, economics, and openness needed to power the most demanding AI workloads," he added.

Vultr is taking pre-orders for reserved capacity for deployments based on the AMD Instinct MI455X and AMD Helios across 2027 and 2028.