NVIDIA Frames AI Agents as the Workload Driving Its Compute Stack

NVIDIA framed the central claim in direct terms: “Useful AI has arrived,” with agents “working by your side.” The agent was presented as the common workload behind the company’s hardware, enterprise software, PC, cloud, and robotics announcements. The point was not only that agents are becoming more capable; it was that agent work creates demand for more inference, cheaper tokens, faster CPUs, memory support, interconnects, enterprise guardrails, and local PC execution.

The agent story began with a shift in status. “Agents used to be misunderstood,” the narration said, before turning them from a Hollywood joke into “teams making dreams come true.” The most concrete use case was small-company creation: people “building companies from living rooms.” That line immediately led to the infrastructure constraint: “But they need so much compute, we hear ya.”

Vera Rubin was the answer NVIDIA placed closest to that compute problem. The recap described Rubin as delivering “the cheapest tokens” and “ten times faster inference,” with the speed claim reinforced visually by a carnival multiplier display running up to 10X. NVIDIA also tied BlueField to agent memory, saying it “keeps agents’ memory true,” and connected the CPU claim back to agents rather than presenting it as an isolated specification.

The CPU message was simple: “Fifty percent faster.” A chalkboard labeled “CPU” and a billboard reading “50% Faster!” accompanied the line. NVIDIA then clarified the design center in the narration: “Not for Vera, it’s built for agents.” In the compressed logic of the announcement, faster inference, cheaper tokens, memory support, and CPU gains all served the same thesis: agents are moving from novelty to work, and that work requires more compute beneath them.

NVIDIA’s Vera Rubin claims combined economics, performance, and availability. After the references to cheaper tokens and faster inference, a cleanroom-style visual showed two robots pressing a large button. The on-screen label changed from “Start Production” to “In full production,” and the narration stated the point plainly: “Vera Rubin’s in full production.”

That production claim followed the argument that agent workloads need substantial compute. NVIDIA was not only presenting Rubin as a performance direction; it was saying Vera Rubin is “in full production” in the same sequence that positioned agents as a growing compute load.

NVLink Fusion sat beside Rubin in the compute sequence. The narration said “NVLink Fusion blends ASICs smartly,” while a neon sign read: “NVLink Fusion presents All Silicon Welcome.” NVIDIA’s phrasing was expansive — “Everyone’s welcome to the NVLink party” — and presented NVLink Fusion as blending ASICs into NVIDIA’s silicon ecosystem.

The result was a cluster of agent-related compute claims: Vera Rubin for inference speed, token cost, and production status; BlueField for agent memory; the CPU for a 50% speed improvement “built for agents”; and NVLink Fusion for a more open silicon mix around NVLink.

NVIDIA’s enterprise AI segment bundled several product names into a short workflow claim. “Nemotron Ultra leads the run,” the narration said, followed by “Five X faster work gets done.” The same sequence attached NemoClaw to guardrails and OpenShell to sandboxing: “NemoClaw keeps the guardrails right. OpenShell keeps the sandbox tight.”

The product names appeared as glowing signs above arcade claw machines: “NemoClaw,” “Enterprise AI,” “A PLUSH,” and “OpenShell.” The playful setting did not change the business claim. NVIDIA positioned enterprise AI around agents doing work faster while staying inside guardrails and sandboxes.

The code-use case was explicit. “Your code migrated and reviewed,” the narration said, “all before this song is through.” The line is comic in form, but it identifies the kind of enterprise work NVIDIA wanted to associate with these tools: moving code, reviewing it, and doing so within stated constraints.

That makes the enterprise pitch narrower and more practical than a generic claim about AI assistants. NVIDIA described agentic software activity in terms of speed, safety boundaries, and code operations rather than only model capability.

NVIDIA reduced its broader AI infrastructure pitch to a metaphor: “AI is a five-layer cake.” A large cake labeled “AI Five Layer Cake” appeared on screen while the narration described “compute to revenue.” The layers were not named one by one, but the associated ingredients were clear: global AI clouds, gigawatts of power, DGX systems, and watt-level optimization.

The scale claim was expressed through energy and efficiency. NVIDIA referred to “Global AI clouds with lots of gigawatts,” said “DGX keeps power lean,” and added that “every watt” is optimized. AI infrastructure was framed as a production path from compute to revenue, with power efficiency treated as part of the system.

The PC claim extended the same agent story to local machines. NVIDIA said “RTX 40’s finally here” and called it the “biggest PC moment in forty years.” A glowing laptop sat beneath a neon sign reading “New PC Era,” followed by the line: “Agents powering all workflows.”

The Windows point was reach: agents would be “running anywhere Windows go.” NVIDIA then split execution in plain terms: “Harnesses run on CPU. Models fly on GPU.” The PC becomes another place where agents run, with CPU and GPU each assigned a role in the workflow.

The final technology cluster was “physical AI.” NVIDIA named Cosmos, Omniverse, Groot, Unitree robots, and Thor in a sequence that moved from simulation and synthetic data to movement and humanoid robots.

Cosmos was described as building “worlds that robots need” and turning “compute into synthetic feeds.” Omniverse was given the role of perception and reasoning: it “sees and reasons through” and “understands worlds like people do.” Groot was described as how robots “learn to move,” acquiring skills and “finding grooves.” The sequence ended with “Unitrees powered by Thor” and the claim that “the future’s humanoid to the core.”

The robotics segment mirrored the rest of NVIDIA’s framing. Compute produces synthetic worlds; those worlds support understanding and movement; movement becomes embodied in robots. Physical AI was therefore presented as another expression of the same demand pattern: agents and robots need generated environments, reasoning systems, learned skills, and hardware capable of running them.

Across the claims, NVIDIA’s highest-order message was consistent. Agents were not presented as a single product category. They were the workload connecting Vera Rubin, BlueField, NVLink Fusion, Nemotron Ultra, NemoClaw, OpenShell, DGX, RTX PCs, Cosmos, Omniverse, Groot, Thor, and humanoid robots. Useful AI, in this framing, means agents at work, and agents at work require infrastructure from cloud-scale compute down to PCs and physical machines.