Nvidia’s huang bets every company will run on open-source agents, pays staff in compute tokens
Jensen Huang took the stage in San José and told 4,000 engineers that the next computer is not a box on a desk but a swarm of autonomous agents. Then he offered to pay them with the same currency those agents burn: tokens.
Huang declares openclaw the new windows
The Nvidia CEO rebranded a scrappy open-source project—once nicknamed Clawdbot and Moltbot—into OpenClaw, a runtime he claims will do for personal AI what Windows 95 did for the PC. Every company needs an OpenClaw playbook,
he said, pacing in front of a neon claw icon. Agents are the new apps.
Silicon Valley has already downloaded the prototype 120,000 times since Peter Steinberger open-sourced it last fall. Steinberger now works at OpenAI, but the repo keeps forking; Huang’s engineers committed 47 patches last night alone.
The catch is enterprise paranoia. An agent that can read every spreadsheet at midnight is a compliance officer’s nightmare. Nvidia’s answer is NemoClaw, a hardened fork that routes every query through privacy filters and on-device sandboxes. Early adopters include Shell and Kaiser Permanente, both piloting supply-chain agents that never leave their firewalls.
Tokens become the fourth salary component
Huang saved the mic-drop for the tail of the keynote. Starting next quarter, new hires will receive up to 50 % of base pay in inference tokens—GPU cycles measured in real time. Recruiters in the hall swapped LinkedIn messages before he finished the sentence. Engineers will ask ‘How many tokens is the package?’ the same way they ask about stock refresh,
Huang predicted. A senior CUDA architect burning one million tokens a day could, in theory, 10× productivity and still have surplus compute to sell on internal marketplaces.
The math is already rippling through salary bands. A mid-level ML engineer earning $280 k at a rival firm could jump to $320 k plus 400 k monthly tokens—worth $40 k retail on Nvidia’s cloud. Recruiters call it compensation arbitrage
; accountants call it a clever way to keep idle A100s humming.
Wall Street liked the stunt. Nvidia stock ticked up 4 % after hours, adding $110 billion to its market cap—enough to buy three Intels. Meanwhile, Groq’s inference chips, powering the demo rigs on the show floor, saw inquiries triple. The startup’s CEO quietly told investors that $1 trillion in demand for Blackwell and Rubin accelerators is no longer fantasy
if companies really do spin up one agent per employee.
Huang exited to a standing ovation, leather jacket still reflecting the stage lights. Outside the convention center, a food-truck vendor had already spray-painted Pay us in tokens
on his van. The line for $18 ramen stretched past it, every engineer calculating how many agent cycles one lunch costs.