At CES 2026, NVIDIA introduced the Vera Rubin platform, its next-generation AI infrastructure architecture. While NVIDIA presented a long list of component-level announcements — new GPUs, CPUs, networking, and memory — the more important signal was architectural. Vera Rubin reflects NVIDIA’s view that AI infrastructure is no longer defined by individual accelerators, but by how entire systems behave under sustained inference-heavy workloads.

For enterprise IT leaders, this matters because the assumptions that governed AI infrastructure over the last several years — training-centric, accelerator-first, and benchmark-driven — are increasingly misaligned with how AI is actually being deployed inside organizations. Inference, not training, is becoming the dominant workload. And inference stresses systems in different ways.

Vera Rubin as a Platform, Not a Product

NVIDIA was deliberate in positioning Vera Rubin as a new platform, not a new chip generation. At its core is the Rubin GPU, which succeeds Blackwell and introduces NVIDIA’s third-generation Transformer Engine with support for NVFP4 precision. NVIDIA cited up to 50 PFLOPS of inference performance and 35 PFLOPS for training, paired with HBM4 memory delivering 22 TB/s of bandwidth.

The specifications are notable, but the architectural intent is more important. Rubin is designed for workloads such as mixture-of-experts models, long-context inference, and agentic pipelines where memory bandwidth, interconnect latency, and synchronization overhead often constrain performance more than raw compute does.

To address those constraints, NVIDIA introduced NVLink 6, increasing inter-GPU bandwidth to 3.6 TB/s per GPU and enabling dense scale-up configurations. This interconnect is foundational to the Vera Rubin NVL72, a rack-scale system integrating 72 GPUs via NVLink switching. NVIDIA also emphasized modular design, cable-free GPU trays, and serviceability features intended to support upgrades and maintenance without interrupting running workloads.

This focus on operational characteristics — serviceability, upgrade paths, and sustained utilization — suggests NVIDIA is optimizing for long-lived enterprise deployments, not just hyperscale showcase systems.

Data Movement and the Inference Bottleneck

A recurring theme throughout the Rubin announcements was data movement. NVIDIA has expanded its networking portfolio with ConnectX-9 SuperNICs and BlueField-4 DPUs, offering up to 1.6 Tb/s per GPU and PCIe Gen6 switching. These components offload networking, storage, and security functions from CPUs and GPUs, allowing more cycles to be devoted to application-level work.

NVIDIA also extended its Ethernet strategy with Spectrum-X switches supporting 102.4 Tb/s co-packaged optics, reinforcing the company’s belief that Ethernet will play an increasing role in AI fabrics alongside InfiniBand.

One of the more practically relevant announcements for enterprise AI was NVIDIA’s Inference Context Memory Storage Platform. As inference workloads become more stateful — particularly in agentic systems — the cost and inefficiency of maintaining large KV caches has become a growing concern. NVIDIA’s approach introduces an additional memory tier between GPU memory and traditional storage, tightly coupled with networking and DPUs. NVIDIA cited up to 5x improvements in throughput and power efficiency for context-heavy inference workloads. Whether those gains translate cleanly into real-world enterprise environments will depend on workload characteristics, but the problem NVIDIA is addressing is real.

Why the CPU Re-Enters the Conversation

Midway through the Rubin story, the role of the CPU becomes clearer. The platform is anchored by the Vera CPU, which NVIDIA positions explicitly as a control-plane processor rather than a general-purpose server CPU. Vera integrates 88 cores built on Arm’s Olympus microarchitecture, supporting 176 threads via spatial multithreading. It is paired with SOCAMM memory delivering 1.2 TB/s of bandwidth and 1.8 TB/s of chip-to-chip NVLink connectivity to GPUs.

The choice of Olympus is telling. Olympus is not a replacement for Arm’s Neoverse cores, nor is it a consumer or mobile design. It reflects a recognition that AI systems place new demands on CPUs: orchestration, scheduling, memory management, and coordination of accelerators are often latency-sensitive and do not parallelize cleanly. In those roles, single-thread performance and predictable behavior matter more than core density or power efficiency alone.

For enterprise IT, this underscores an important point: CPUs are not becoming less relevant in AI systems — they are becoming more specialized.

Operational Features and Enterprise Readiness

NVIDIA also highlighted features aimed squarely at enterprise deployment. These include rack-scale confidential computing, second-generation RAS capabilities designed for non-disruptive diagnostics, and deeper integration across compute, networking, and security.

These features are unlikely to generate headlines, but they matter in practical terms to organizations running AI workloads as part of production IT environments rather than experimental labs. As those organizations are acutely aware, issues like downtime, upgrade complexity, and security boundaries increasingly shape total cost of ownership. So NVIDIA is directly addressing enterprise pain points here.

What This Means for Enterprise AI

The most immediate implication of Vera Rubin is for inference economics. Reduced-precision formats, improved context memory handling, and high-bandwidth interconnects are aligned with enterprise workloads where cost-per-token, latency consistency, and power efficiency matter more than peak training throughput.

At the same time, adopting Vera Rubin implies deeper alignment with NVIDIA’s networking, software, and operational ecosystem. Organizations pursuing multi-vendor or accelerator-agnostic strategies will need to consider those trade-offs carefully.

What Vera Rubin Says About NVIDIA’s Product Strategy

Vera Rubin reflects NVIDIA’s view that the next phase of AI infrastructure is defined by systems, not components. GPUs remain central, but CPUs, memory hierarchies, networking, and operational design are now equally critical to delivering predictable, cost-effective AI at scale.

For enterprise IT leaders, the significance of Vera Rubin lies less in any individual specification and more in the architectural direction it represents. As inference becomes the dominant AI workload, infrastructure decisions will increasingly hinge on orchestration efficiency, data movement, and operational maturity. In this context, NVIDIA and AMD are offering distinct paths forward, and the real challenge for each enterprise will be aligning its organizational priorities with one of those paths.