Nvidia-obsessives will pore over 20-odd releases from GTC, and there is lots to chew over; some of it is interesting: the AI-RAN stuff with T-Mobile and Nokia; the idea of a new category of “agent computers” (a likelier hit than AI glasses, surely); all the focus on physical AI, animated by agents running inference models at the edge. But today’s industry trends are yesterday’s roadmap items for Nvidia; the biggest talking point at GTC – as ever with Nvidia – is in the framing, which more than MWC or whatever other sub-sector trade show, sets the global tech focus.

As such, Jensen Huang’s talk during his GTC keynote about the firm’s new Grace Blackwell architecture with NVLink 72 (“a huge technological gamble, not easy for any of our partners”) and NVFP4 tensor cores – delivering 35–50 times throughput improvements over Hopper, apparently – featured the most interesting stuff, and worked as a segue to something else: about tokens as a commodity, driving AI infrastructure into a new economic phase. Grace Blackwell / NVLink 72 proposition also cuts costs per token to the lowest anywhere, he said.

And then Huang stood in front of a graph showing the throughput (number of tokens per second at a fixed power level) on the vertical axis and token speed (response rate per inference step) on the horizontal axis. Data centers are no longer just compute hubs, but AI ‘factories’ – per the new terminology – that produce tokens at scale, and measure efficiency in terms of throughput, latency, and ultimately revenue per watt. The faster the speed, the larger the usable model, the longer the context, the more intelligent the AI – is the point. 

This is what every CEO in every enterprise will be looking at, he said – token efficiency in their factories, as a core operating metric. The token is a new commodity, which will be tiered in pricing once it matures: from gratis (high throughput, low speed), as we consume AI today, through intermediate, advanced, high-speed, and ultra high-speed plans (going to $150 per million tokens). Nvidia’s NVL72 is presented as the “inference king”, delivering 50-times higher performance per watt at 35-times lower cost versus its old NVL8 line, and a jump on its competitors.

So the takeaway from GTC is not that Nvidia has new hardware, software, blueprints, frameworks, initiatives, coalitions. It is that all of these things, together with ecosystem developments on top, redefine how hardware is understood and monetised. Whoever controls the most efficient “token factories” will shape the next phase of the AI economy. And, as with previous AI cycles, Nvidia is not just supplying the tools, but setting the terms. Be interesting to see if and how operators might tokenize their networks, which they are pumping so much investment into.