NVIDIA GTC 2026: Vera Rubin and the Agentic AI Era

NVIDIA GTC 2026 opens in San Jose tomorrow with Vera Rubin, the GPU platform succeeding Blackwell, as the centerpiece announcement. Jensen Huang takes the stage at SAP Center at 1 p.m. PT Monday. The AI industry is treating the event less like a trade show and more like a state-of-the-union address for compute. An expected 39,000 attendees from 190 countries will be there in person. The headline items: the official Vera Rubin launch, the first architectural look at what comes after it, and a gigawatt-scale partnership that signals where frontier AI compute is headed next.

NVIDIA confirmed last week a multiyear strategic partnership with Thinking Machines Lab, the frontier AI company led by Mira Murati, to deploy at least one gigawatt of Vera Rubin systems for frontier model training. NVIDIA has also taken a significant equity stake in the company. Deployment is targeted for early 2027, but the deal signals where NVIDIA sees demand heading: not general-purpose cloud workloads but purpose-built agentic and frontier model infrastructure. The Thinking Machines partnership is the most concrete evidence yet that Vera Rubin was designed to serve a specific customer class: labs running continuous, large-scale training on reasoning-heavy architectures.

What Vera Rubin Actually Delivers

The Rubin platform is designed specifically for the workloads that have strained Blackwell: agentic reasoning, long-context inference, and massive Mixture-of-Experts models. NVIDIA’s own figures put the cost-per-token reduction at roughly 10x versus Blackwell for agentic AI and advanced reasoning, while MoE training requires only one-quarter the GPU count of the prior generation. That is a meaningful compression of the economics that have made large-scale AI deployment expensive.

At the architectural level, Rubin introduces next-generation NVLink interconnect, an upgraded Transformer Engine, and the proprietary Vera CPU designed for tight CPU-GPU integration in rack-scale systems. The first systems pair Vera CPUs with Rubin GPUs in full-rack configurations. Meta was announced as the first partner to deploy Grace CPUs at scale, and the company is also evaluating Vera CPUs for its data centers.

Beyond Rubin, Huang is expected to preview Feynman, the generation after Rubin, designed specifically to handle the reasoning and long-term memory requirements of persistent AI agents. Feynman is currently slated for 2027-2028. The preview would be unusual: NVIDIA typically does not roadmap two generations ahead publicly. That it is doing so now reflects how much the agentic AI use case has changed the compute planning horizon for both NVIDIA and its customers.

The Agentic AI Theme Running Through GTC

Agentic AI is not a side session at GTC 2026. It is the organizing principle of the conference. Monday’s pre-show features LangChain CEO Harrison Chase alongside other agent-framework builders, and NVIDIA has set up a dedicated build-a-claw area on the show floor running March 16 through 19, where attendees can deploy persistent long-running agents using OpenClaw, the open-source agent platform that has emerged as one of the fastest-growing projects in the space.

The shift from one-shot model queries to persistent, tool-using agents has direct hardware implications. Agents run continuously, require fast context retrieval, and must coordinate across multiple tool calls in parallel. That is a different compute profile than batch training or even real-time chat inference, and Vera Rubin’s architecture reflects those requirements in ways Blackwell did not anticipate. The compute cost of reasoning has become the dominant variable in AI deployment decisions, as covered in our earlier analysis of the inference economy and what drives it.

The Compute Competition Backdrop

GTC arrives against a complicated backdrop in AI infrastructure investment. CNBC reported last week that OpenAI walked away from expanding its partnership with Oracle at the Abilene, Texas Stargate campus, citing a preference for newer GPU generations rather than additional Blackwell capacity. Oracle confirmed the 1,000-acre campus remains operational with GB200 racks running training and inference loads, but the expansion beyond 1.2 GW has been paused.

That episode illustrates a tension running through the entire AI infrastructure buildout: hyperscalers and model companies want the newest silicon badly enough to pause expansions rather than lock in capacity on what they consider yesterday’s hardware. It is precisely the dynamic that gives NVIDIA’s Vera Rubin launch tomorrow its strategic weight. The question Huang will need to answer is whether supply can match the demand he is about to announce. Given that Oracle and xAI have already committed billions to Blackwell-era infrastructure, the transition to Rubin will test whether NVIDIA can deliver at the pace the market now expects.

For the compute-watchers tracking where AI infrastructure spending goes next, the keynote tomorrow is the most important hour of the quarter. The Vera Rubin era starts Monday.

What Comes After Rubin

Huang is expected to offer the first public preview of Feynman, the next GPU generation targeting 2027-2028. Feynman is being designed from the ground up for the reasoning and long-term memory demands of persistent AI agents rather than the batch training and inference workloads that defined Hopper and Blackwell. The fact that NVIDIA is discussing a chip two generations out at a public keynote is itself a signal: the company wants large customers to plan their infrastructure roadmaps around NVIDIA silicon rather than shopping generation-to-generation.

The Kyber rack, previewed at GTC 2025, is also expected to get details: a 600 kW behemoth with 144 GPU sockets, each with four Rubin Ultra GPU dies, in a standard rack form factor. For reference, a full Kyber rack at that density would represent a step change in how much compute can be physically co-located in a single facility. The power and cooling implications alone are rewriting how hyperscalers design their next-generation campuses.

NVIDIA enters GTC 2026 with more momentum than any prior year, but also more pressure. The Blackwell supply crunch was well documented. Customers who waited months for GB200 allocations are now being asked to plan for Vera Rubin, with Feynman on the horizon. Whether NVIDIA can deliver on that roadmap at the scale its partners need is the central question Huang carries into tomorrow’s keynote.