Our Take
Huang's $200B TAM claim rests on a single year of sales ($20B) and the unproven assumption that billions of agents will require dedicated CPUs—a market that doesn't yet exist and that AWS, custom silicon, and AMD are all pursuing.
Why it matters
Nvidia faces genuine pressure from hyperscalers building their own AI chips, and CPU market share is genuinely contested ground. The Vera bet signals Nvidia is moving beyond GPU dominance, but the numbers are Nvidia-reported and the competitive landscape is crowded.
Do this week
Infrastructure leaders: audit whether your Vera evaluations include benchmarks from AWS Trainium, AMD MI300X, and custom chips built in-house before committing to multi-year CPU contracts.
Huang pitches Vera as the opening of a new CPU market
On Nvidia's earnings call Wednesday, CEO Jensen Huang positioned the company's new Vera CPU as the entry point to a "brand new $200 billion TAM for Nvidia, a market we have never addressed before." Vera, introduced in March, is designed specifically for AI agents and is sold standalone or bundled with Nvidia's Rubin GPU.
Huang claimed that $20 billion worth of standalone Vera CPUs have already been sold this year (company-reported). He framed the addressable market as emerging from a shift in compute workloads: while the "thinking" phase of AI models runs on GPUs, agents spend most cycles on CPU-driven execution tasks. Vera is optimized for token processing speed rather than the multi-core parallelism of traditional cloud CPUs.
Huang stated that "every major hyperscaler and system maker is partnering with us to deploy it," positioning Nvidia at the center of what he calls a rebuild of computing for agentic and robotic physical AI.
The CPU market is Nvidia's real vulnerability
Wall Street's anxiety over Nvidia's future has recently centered on CPU exposure. The company is historically a GPU powerhouse, and CPUs have been the domain of Intel and AMD. Nvidia has made CPUs before, but they are not core to the business.
That vulnerability is material. Last month, Amazon Web Services announced a major contract with Meta to supply millions of custom AI CPUs. AWS CEO Andy Jassy has been explicit that AWS believes it can build AI chips, both GPU and CPU, at parity with or better than Nvidia.
Huang's $200 billion TAM claim is simultaneously an admission and an offensive move: Nvidia is acknowledging that CPU is a distinct and defensible market segment, and it is arguing that Vera's early traction (the $20 billion figure) proves it can win there. However, the $20 billion is company-reported, no independent benchmark exists, and the TAM itself is aspirational. The market for dedicated agent CPUs does not yet exist at scale; it is a hypothesis about how compute will be distributed as AI systems move from training to inference and orchestration.
Evaluate Vera against real alternatives before locking in
The competitive landscape for agent CPUs is active. AWS Trainium chips, AMD MI300X, Intel Xeon processors, and custom silicon built by hyperscalers are all pursuing the same workloads. Vera's $20 billion claimed sales figure is high, but it says nothing about performance, cost per inference, power efficiency, or total cost of ownership relative to alternatives.
Before committing to Vera in production deployments, benchmark it against AWS and AMD offerings on your actual agent workloads. Vendor-reported sales figures are not a substitute for independent performance data. The agent CPU market is nascent enough that switching costs and lock-in are still avoidable. Do that comparison now.