Summary: On Nvidia’s May 20, 2026 earnings call, Jensen Huang said Vera, a CPU purpose-built for agentic AI, opens a US$200 billion TAM. He also reported US$20 billion in standalone Vera sales this year, distinct from the larger potential-market claim. The announcement moves Nvidia into CPU competition and makes CPU choice, token throughput, and lock-in risk relevant to agent infrastructure planning.
On the Nvidia earnings call held on May 20, 2026, Jensen Huang put a number on the table that nobody had forecast: Vera, a CPU designed specifically for agentic AI, hands the company an entirely new $200 billion market. “Vera opens a brand new $200 billion TAM for Nvidia, a market we have never addressed before,” he said, adding that standalone Vera CPU sales have already reached $20 billion this year. The chip made its debut at GTC only in March. A ramp that steep signals that what is being sold is a shipping product with real bookings behind it, not a slide in a roadmap.
The backdrop: the front line of the AI chip war is shifting ground. A model’s “thinking” happens on GPUs, yet agents carry out their tasks on CPUs — and once the agent population starts counting in the billions, a market that has always belonged to Intel and AMD abruptly becomes AI’s newest battlefield.
The Claims Made on the Call
By Nvidia’s own description, Vera is “the world’s first CPU, purpose-built for agentic AI,” and it can be bought on its own or bundled with the Rubin GPU. The quarter behind the call was enormous in its own right: revenue set another record at $81.6 billion, and guidance for the coming quarter stands at $91 billion. Huang, though, purposely pulled the spotlight off GPUs and onto CPUs: “The world is rebuilding computing for agentic AI and robotic physical AI,” and “we’re going to need a lot more CPUs.”
He then laid out the market math. Roughly a billion people draw on computing resources today, agents will eventually count in the billions, and he expects each agent to get a dedicated, PC-like setup of its own. The gap between the number of human users and the number of agents is the headroom for CPU demand. On how Vera reaches the market, his framing was that “every major hyperscaler and system maker is partnering with us to deploy it.”
Why Agents Call for a Different Kind of CPU
The core of Huang’s case is that the workloads differ. Conventional cloud CPUs are built around core counts: squeeze many application instances onto one machine, then sell on density and utilization. Vera is instead built around the speed of token processing — every tool call an agent makes, every slice of context it reads, every decision it reaches is tokens in motion. What limits performance is not the number of processes running concurrently but the rate at which tokens move.
There lies the real substance of “purpose-built”: once an agent’s unit of work is a token rather than a process, a CPU engineered for general computing stops being the best-fit part. For Nvidia the framing pays off twice strategically. It recasts the company from “GPU vendor” into a full-stack supplier of AI computing that sells CPUs, GPUs, and networking as one package, and the standalone Vera SKU turns buyers who have never purchased an Nvidia GPU into direct Nvidia customers for the first time.
The Opposition on the CPU Front
Intel and AMD have held the CPU market for decades, but the challenges don’t come only from those two. In April 2026, Amazon signed a contract with Meta under which Meta will buy millions of Amazon’s in-house AI CPUs — a deal TechCrunch described as “another wild turn for AI chips.” That same month, Amazon CEO Andy Jassy pointed straight at Nvidia in his annual shareholder letter, implying that AWS can build AI chips on par with, and possibly ahead of, the incumbent. Scenarios in which Nvidia’s dominance starts to erode are now being seriously weighed on Wall Street.
Huang’s counter is to widen the conflict: rather than stand guard over the GPU, carry Vera into the incumbents’ home market and elevate “who supplies the AI CPU?” into a fresh procurement question. TechCrunch’s read on him lands well — he may be the industry’s most gifted hype man, but he “delivers on the hype, quarter after quarter.” A $200 billion TAM is, to be sure, marketing language; $20 billion of standalone sales is a hard number.
What Developers and Infrastructure Teams Should Take Away
Three concrete consequences follow. One: agent cost structures will move. If agents shift from renting general-purpose CPUs to running on purpose-built ones, inference cost per task and its latency both have room to drop, and teams building agent products should write CPU SKUs and their pricing into cost models rather than watching GPU unit prices alone. Two: CPU sourcing enters the cloud-selection equation. With hyperscalers selling their own silicon while also deploying Vera, any cross-cloud or hybrid agent deployment plan owes lock-in risk a fresh look. Three: the supply chain will regroup around compute built specifically for agents. The ecosystem just spent three years rebuilding itself around GPUs, and rack design, procurement cycles, memory bandwidth, and the CPU-to-GPU ratio will now re-sort around agent workloads. For anyone planning hardware, the question has stopped being “do we need GPUs?” and become “what GPU-to-CPU mix do we need?”