NVIDIA Recasts the Data Center as Infrastructure for Agentic AI

At GTC Taipei, NVIDIA presented its infrastructure agenda around a shift in AI workloads: from generative systems, to reasoning models, to agents that can plan, use tools, and act continuously. The practical implication, as NVIDIA describes it, is that data centers are no longer being designed primarily for human-paced application use. They are being built for autonomous software loops running across models, tools, memory, context, and governance layers.

The stages are separated plainly. Generative AI produced “words, image, code, answers.” Reasoning models “think through problems step by step.” Agents, in NVIDIA’s account, go further: they understand intent, plan, use tools, take action, and complete work without requiring a person to open applications and click through menus. The user explains the desired outcome; the agent decides how to pursue it.

NVIDIA says this is no longer speculative. It points to GitHub activity as evidence that “useful AI has arrived,” with pull requests merged at 90 million, commits at 1.4 billion, and new repositories per month at 20 million across a chart spanning 2023 to 2026. Software commits, NVIDIA says, tripled in early 2026 after years of flat growth. The company also attributes to AI agents “$9 trillion worth of economic output” from work previously described as $3 trillion in human engineering work.

NVIDIA uses the GitHub chart and the $9 trillion figure to support a strategic conclusion: every company will run agents, and therefore every company will need infrastructure built for them.

That conclusion drives the rest of the positioning. Agents, NVIDIA says, do not operate on “human time scales.” They run in nanoseconds, orchestrating models, calling tools, processing outputs, and looping back “millions of times a second across thousands of simultaneous workflows.” The NVIDIA Agent Toolkit for Enterprise AI places this loop inside a broader stack: prompt, orchestration, memory, context, observe, reason, act, tools and skills, security and governance, CUDA-X libraries, and Nemotron open models. The load-bearing point is not just that agents need GPUs for inference. NVIDIA is arguing that agentic workloads require a different end-to-end computing pattern.

Against that agentic-workload framing, NVIDIA presents Vera Rubin as a production system “built from the ground up for agentic AI.” The claim is specific: not merely faster inference, but agent execution at scale. Vera Rubin is said to deliver 10 times the agent throughput of Blackwell and to be deployed at pod scale, optimized from silicon to the data center floor. NVIDIA’s formulation is blunt: “This is the supercomputer that agents live in.”

That sentence captures the architecture claim NVIDIA is making. The company is not describing Vera Rubin as a general product refresh. It is tying the system to the operational model of agents: persistent, tool-using, looping workflows running at high concurrency.

The companion announcement is the Vera CPU, described through Jensen’s framing as “the beginning of a market that never existed before.” NVIDIA’s claim is that previous CPUs were designed for humans, while Vera is “built for agents.” In practical terms, Vera is presented as 1.8 times faster on agentic workflows than x86 CPUs. The comparison breaks that into compilation performance at 1.7 times x86 and Python performance at 1.9 times x86.

NVIDIA also gives Vera a concrete systems profile: an NVIDIA-custom Olympus core, 88 cores and 176 threads with spatial multithreading, 10-wide instruction fetch/decode, 2MB L2 per core, 164MB L3, a 250W to 450W TDP range, 1.2 TB/s LPDDR5X ECC, 40% lower loaded latency versus x86, 3.4 TB/s core-to-core bandwidth, 1.4 TB/s PCIe Gen6, and 1.8 TB/s NVLink C2C.

NVIDIA names Anthropic, OpenAI, Oracle, and CoreWeave as adopters. Its broader Vera market slide lists early adopters as OpenAI, Anthropic, and SpaceX; cloud companies as Nebius, CoreWeave, Oracle, Firmus, Lambda, and together.ai; and ecosystem partners as Dell Technologies, HPE, Supermicro, ASUS, Foxconn, GIGABYTE, Inventec, Pegatron, QCT, Wistron, and Wiwynn.

The market-size argument is expressed in one blunt line: “There will be more agents than there are people on the planet.” NVIDIA uses that forecast to explain why a CPU “for agents” is presented as a new market category rather than a conventional server component.

For data center operators, NVIDIA shifts from compute architecture to power economics. The premise is that the world is “absolutely power-constrained.” In that context, a 1-gigawatt allocation is fixed: “1 gigawatt means 1 gigawatt. You don’t get more.” The variable, NVIDIA argues, is how much revenue-producing compute can be safely placed inside that envelope.

DSX, described as NVIDIA’s Data Center Optimization platform, is presented as the mechanism for changing that equation. Current AI factories, NVIDIA says, over-provision power by up to 40%. DSX is said to let operators safely deploy more GPUs within the same power budget, thereby unlocking additional annual revenue from infrastructure that has already been built.

NVIDIA’s DSX MaxLPS example uses a 96 MW power budget. Without DSX MaxLPS, the displayed configuration has 26,322 total GPUs and $19 billion in annual revenue. With DSX MaxLPS, it has 38,304 total GPUs and $26 billion in annual revenue. The same material also displays a larger case of 99,720 total GPUs and $67 billion in annual revenue.

The economic framing is simple: “Throughput per watt is your revenue.” NVIDIA’s claim is that DSX matters because it improves utilization of an already scarce input. A tenant-level data center view reinforces the same operating problem: one tenant is shown at 95% GPU utilization, another at 25%, and another at 75%, with GPU memory and CPU usage varying across the same floor.

NVIDIA closes the argument by tying Vera Rubin, Vera CPU, and DSX to the partner ecosystem building the physical systems. The company names Foxconn, Quanta, Wistron, ASUS, GIGABYTE, Pegatron, Wiwynn, and other Taiwan-based system partners as essential to making Vera Rubin real.

The scale claim is substantial: 150 partners, millions of square feet, and hundreds of sites, “all moving in extraordinary precision together.” Jensen’s term for it, according to NVIDIA, is “extreme co-design.” NVIDIA’s own description is “the most remarkable manufacturing partnership this industry has ever seen.”

The partnership language is more than ceremonial. It completes the infrastructure thesis. If agents require new compute patterns, and Vera Rubin and Vera CPU are built for those patterns, then the ability to manufacture, integrate, and deploy those systems at scale becomes part of the product argument. NVIDIA presents the Taiwan ecosystem not as a supplier backdrop, but as a coordinated industrial base for what it calls AI factories.