Memorable moments from Nvidia GTC 2026

Here’s what I’m paying attention to:

1. The constraint is power. Not compute. Not models. Power.

NVIDIA’s Space-1 concept makes the direction clear: when electrons are scarce and permitting is slow, compute moves to where energy is abundant. We are entering the era where AI infrastructure follows power supply curves.

1. Saas -> AGaaS This is really a shift in where energy is monetized. Inference is the meter.

The stack is reorganizing around continuous consumption of compute (and therefore power).

1. "Open" as a distribution strategy. Open models are not just about access – they are about demand creation. Pair that with vertical stacks (finance, healthcare, robotics), and you get: more usage → more inference → more energy demand.
2. "Vertically integrated, horizontally open" Across AWS, Azure, Google, Oracle, CoreWeave, and more, NVIDIA sits in the position every grid operator understands: the system benefits from throughput.
3. Unstructured data is the context layer. Libraries like cuVS are about turning unstructured and dormant data into active inference load – which ultimately means more compute cycles, more power consumption, more demand density.
4. The inference business model is explicit.

Free → Ultra is a pricing strategy that tied directly to: – Tokens per watt (energy efficiency) – Latency (responsiveness) The business model is expressed in energy economics.

Bottom line: Efficiency is the revenue driver. In a power-constrained world, the winners are the ones who can convert electrons into tokens most efficiently.

Curious how others are thinking about where margin accrues in the '5-layer-cake'.

Watch NVIDIA GTC Keynote 2026 | https://www.youtube.com/live/jw_o0xr8MWU?si=a3zgSRUdA5HQz-jL

-------- Clean Power Whisperer™ Perspectives Educational insights on risk, insurance, and capital for AI • Energy • Climate

← All Perspectives · Next: The New Data Center Scorecard →