Huang flew into Taipei projecting $1 trillion in revenue while Beijing banned Nvidia's gaming chips back home—a perfect snapshot of the geopolitical tug-of-war over who gets to build the agent economy.
The Summary
- Nvidia CEO Jensen Huang arrived in Taiwan for GTC Taipei keynote as Beijing moved to ban Nvidia gaming chips, supporting domestic rivals Huawei and Cambricon
- Huang forecasts AI will become a multitrillion-dollar industry with "very clear" ROI, while Nvidia eyes $1T in total sales
- The new Vera CPU could command a $200B market in the AI agent space alone
- China's chip ban landed during Huang's Asia tour, crystallizing the supply chain and sovereignty battle that will define who controls the infrastructure layer of Web4
The Signal
Nvidia projects $1 trillion in sales while China simultaneously banned its gaming chips, a move designed to prop up Huawei, Cambricon, and other domestic chipmakers trying to close the gap with US rivals. The timing wasn't accidental. Beijing is watching Nvidia's dominance in AI compute and making calculated moves to ensure Chinese companies can build agents without depending on American silicon.
Huang's Taiwan visit matters because Taiwan remains the critical node in semiconductor supply chains that power both AI training and inference. TSMC fabs on the island manufacture the chips that run the agent economy. Every major AI company, every crypto protocol adding on-chain AI features, every startup building autonomous workflows depends on this island's output.
"AI's projected multitrillion-dollar growth highlights its transformative economic impact, driving competition and innovation across industries."
The Vera CPU announcement projects a $200B market specifically for AI agents. That's not a general AI chip market estimate. That's Nvidia sizing the compute layer for autonomous software that acts on your behalf. Vera is designed for inference at scale, the workload that happens when millions of agents are running simultaneously, making decisions, executing trades, managing workflows, and coordinating with other agents.
Here's what that means in practical terms:
- Training a frontier model is expensive but happens once
- Running millions of agent instances is expensive and happens continuously
- Inference compute is the recurring revenue model for the agent economy
- Whoever controls inference chips controls the economic chokepoint
Huang describes AI ROI as "very clear" because companies are already seeing productivity gains that justify chip spending. But the clearest ROI is in agent workflows that replace entire job categories, not just augment them. Customer service agents, data analysts, junior developers, research assistants. The agents don't need to be perfect. They need to be 80% as good and 95% cheaper.
China's gaming chip ban is theater, but it signals intent. Beijing knows it can't ban Nvidia data center GPUs without crippling its own AI development. But it can carve out adjacent markets, protect domestic players, and slowly build parallel infrastructure. The goal isn't to beat Nvidia today. It's to ensure Chinese companies can build agents that run on Chinese chips within five years.
The Implication
Watch where Nvidia allocates Vera production capacity. If most chips flow to US hyperscalers and Taiwanese partners, that tells you who's winning the agent infrastructure race. If meaningful volume goes to Chinese buyers despite the gaming ban, it means the geopolitical split isn't as clean as headlines suggest.
For anyone building in crypto or agents, your compute costs are about to become a strategic question, not just an operational one. Which cloud provider you use, which chip architecture you optimize for, whether you can run inference on decentralized compute networks instead of centralized data centers—these aren't technical details. They're sovereignty decisions that determine whether your product can operate globally or gets caught in the crossfire.