AMD and Nvidia launch competing rack-scale AI systems

The shift to rack-scale systems

The competition between AMD and Nvidia has moved beyond comparing individual GPU speeds. As of late 2026, the primary focus for AI buyers is the 'rack-scale' system, which bundles GPUs, CPUs, networking, and software into a single unit. This approach allows large-scale operators to deploy massive amounts of compute power more efficiently than by buying individual components.

AMD's Helios racks combine Instinct MI450 Series GPUs with 6th Gen EPYC processors and Pensando networking. These systems are already being used by major companies like Anthropic, Meta, Microsoft, and OpenAI. By offering a complete rack, AMD aims to simplify the deployment of large-scale AI infrastructure for these high-demand customers.

Nvidia is following a similar strategy with its Vera Rubin NVL72. This system bundles Rubin GPUs with Vera CPUs and NVLink switching. While AMD's Helios is built on a design from Meta's Open Compute Project, Nvidia's rack is a proprietary extension of its own MGX line. Both companies are now marketing massive multi-year plans to provide gigawatts of total capacity to their clients.

Performance and power claims

Both vendors are using specific metrics to frame their hardware's advantages, though these figures are often compared against their own previous generations. Nvidia reports that its Vera Rubin NVL72 provides a 30x increase in throughput per megawatt when measured against the GB300 NVL72. This metric is significant because power availability is often the primary constraint for building new data centers.

AMD's growth is evidenced by financial records rather than a direct head-to-head benchmark. The company reported a record $8.5 billion in Data Center revenue for Q3 2026, a 97% increase from the previous year. This surge is attributed to high interest in EPYC processors and the debut of the Instinct MI450 Series.

Independent benchmarks comparing the MI450 against the Rubin architecture do not yet exist. Consequently, buyers must grade claims based on who made them and what they were measured against. For example, Nvidia's 30x throughput claim is attributed to a combination of five different hardware and software levers working together.

Memory and architecture trade-offs

Memory capacity remains a key differentiator for different types of AI workloads. AMD describes its MI455X GPU as having 432 GB of HBM4 memory using 3D chip-stack packaging. Higher memory capacity can be beneficial for training large models or serving long-context agents because it reduces the need for frequent cross-GPU communication.

Nvidia takes a different approach by focusing on system-level memory management. The Vera CPU can connect up to 1.5 TB of LPDDR5X memory to the Rubin GPU. Nvidia argues this design is superior for agentic workloads that pause often, as it allows the system to park the cache in CPU memory instead of using scarce high-bandwidth memory.

The choice between these two designs depends on the specific use case. If a user is running dense batch inference with no pauses, raw HBM capacity might be the priority. However, for agent traffic with long idle gaps, Nvidia's method of staging the cache in CPU memory may offer a distinct advantage in managing resources.

Software and availability

Software maturity is a critical factor in how quickly a new accelerator can be put to work. Nvidia's CUDA ecosystem is the current industry standard, powering 81% of the TOP500 supercomputing list. However, AMD's ROCm software has reached version 10.1, with new model support notes being released as recently as November 4, 2026.

Availability also varies between the two companies. Nvidia's Vera Rubin NVL72 is generally available on Oracle Cloud Infrastructure as of October 27. In contrast, while AMD's Helios racks are already running workloads for several major firms, the company is still scaling its production and expansion plans for 2026 and 2027.

For smaller buyers who cannot commit to full rack-scale builds, both vendors offer more accessible entry points. AMD offers the air-cooled MI350P for existing data centers that cannot support liquid cooling. Nvidia provides the HGX Rubin NVL8, an eight-GPU configuration that fits into standard x86-based server designs. These options allow smaller teams to test new hardware without redesigning their entire facility.