Huang: AI needs a billion times more compute

Bg2 Pod brought back Jensen Huang for a third conversation and the Nvidia chief executive opened on what he called an existential issue, where functionality, capability, safety and security sit at exactly the same place and priority number one is not to stop.

Huang said if we stop today, we lose.

The hosts pressed him on his own track record. Two years ago Huang said the market was underestimating cumulative compute buildout and it would be at least two trillion dollars, not one trillion, and that inference compute would rise a billionfold because of reasoning at inference time. Last year he said the industry was entering the age of agents, before Anthropic released Opus 4.6 and Claude Code took off. Huang argued both calls played out and the logic still holds, thinking and problem solving take many more inferences than regurgitating memorized text, and as models get useful enough to be hired for more jobs, usage and compute compound.

That set up the interviewer register of the hour. The hosts, investors Brad Gerstner and Bill Gurley, framed acceleration against the way economies are usually modeled as linear deceleration and pointed to Elon Musk, Sam Altman and Demis Hassabis all saying the field is entering the foothills of the singularity, with models now solving problems like Navier Stokes on relatively small compute, perhaps two to three gigawatts each for OpenAI and Anthropic. Huang did not dispute the direction but reframed the mechanism. He said OpenAI did not solve Navier Stokes from nothing, it built on a large foundation of prior math and science and then used exhaustive exploration of search space, which is why a breakthrough that looks sudden still required massive intelligence and compute.

Huang described the unlock of the past year as the harness, or exoskeleton, around the large language model. The model itself keeps scaling, he said, but the harness now manages working memory, short and long term memory and knowledge graph creation, plus tool use, so agents can be connected to a company's ground truth and to the tools humans already use, from Spice and Verilog and simulators to Photoshop, AutoCAD and Siemens software. Coding was the first on ramp, he said, because customers and tools tolerated friction, and the next phase he pointed to was design and creative work, citing Astra rebuilding a photo into editable CAD as an example of an agent producing a first implementation that humans then tune.

What Huang declined to quantify was the very question in the episode title for 2027, where bottlenecks, revenue and capex collide. He argued the cost of intelligence will fall fairly quickly as more compute comes online and smarter models become more efficient, invoking three decades when computing cost halved every year and saying intelligence costs must fall by a similar factor to fill what he called a massive vacuum of opportunity. He illustrated it with a personal scale, the original GPT might have cost about 20 million dollars to train and now costs perhaps 100 dollars, while ambition has expanded to training 3 trillion, 10 trillion and 30 trillion parameter models and usage is up perhaps a million fold. The hosts pushed that this double exponential, three to five times more compute next year for the frontier labs, should yield major breakthroughs on long running agentic swarms, Huang agreed the precipice is crowded with problems ready for that exploration, but neither side put a number on when revenue catches capex or what breaks first in 2027.

The gap that remains is deployment. Huang said the challenge is how to deploy agents safely and securely into the enterprise, the same safety and security question he tied to functionality at the start, yet the conversation left open how companies verify those harnesses, who governs tool permissions, and what measurable threshold would tell a buyer that 10,000 agents working together for 4 million dollars of compute, the math the hosts cited for the Navier Stokes run, is ready for regulated work.