At GTC, Huang Says Agentic AI Is Here and Demand Through Rubin Is Massive

The Real AI Bottleneck May Be Power, Not Chips

At GTC in San Jose this week, Jensen Huang delivered a demand forecast that should concern anyone betting on AI scale: NVIDIA now sees $1 trillion in Blackwell and Rubin chip orders through 2027. Two months ago, that figure was $500 billion. That is not a revised estimate. It is a supply shock dressed up as a keynote talking point.

The $1 trillion headline comes with footnotes that matter. It excludes China H200 sales — chips NVIDIA halted production on last year due to U.S. export restrictions, and is now restarting after receiving export licenses. It excludes central processors, networking chips, the Groq-based LPUs NVIDIA is shipping from its $20 billion acquisition, and the Rubin Ultra variant planned for 2027. The actual pipeline is almost certainly larger than the headline number.

Vera Rubin, the successor to Blackwell now in full production and shipping to customers later this year, is the more instructive part of the story. The system comprises 1.3 million individual components and, by NVIDIA's own figures, delivers ten times the performance per watt of its predecessor. Ten times. In a single generation. That is the number data center operators and utility planners should be staring at, because it means the efficiency gains from new silicon are finally starting to outrun the raw power appetite of the workloads being asked to run on them.

Or they have to. The shift Huang is betting the company on — from chatbots that respond once to agents that loop continuously — is fundamentally a power story. Persistent agents are operational software. They run longer, call more services, and generate downstream compute traffic that compounds utilization across GPUs, memory, storage, and facility infrastructure. If agentic workflows become the dominant deployment pattern, the industry is not building toward a plateau in compute demand. It is building toward a cliff in power delivery requirements.

The counterargument is real and worth holding: enterprises are not smoothly deploying what they are piloting. Governance gaps, latency constraints, cost opacity, and integration debt are stalling production deployments even where model quality is sufficient. The NVIDIA keynote is a supply-side signal, not a demand-side validation. The $1 trillion in orders is a statement of intent from hyperscalers and sovereign AI programs — not a confirmed book of business.

What is less ambiguous is the competitive environment at the system level. Rubin ships with 35 times the tokens-per-watt performance of a standalone Groq LPU rack configured beside it. The Groq 3 LPU itself — built from the December acquisition of the startup founded by former Google TPU designers — ships in the third quarter. NVIDIA is not just selling chips; it is locking in a reference architecture that makes its systems the default path for anyone building at scale.

For founders and infrastructure teams, the practical takeaway is blunt: treat NVIDIA's GTC as a supply constraint announcement more than a capability announcement. The chips are real. The question is whether you can get power to run them at the density the new workloads demand. The grid interconnection queues, the transformer lead times, the data center power density limits — those are the variables that will determine whether the $1 trillion figure is a self-fulfilling prophecy or an optimistic ceiling.

Notebook: The hidden scarcity in the Blackwell-to-Rubin era may be operational power capacity, not model ideas or chip supply.