Nvidia ceo tells firms: build your own ai army or get left behind
Jensen Huang wants every boardroom to hire a digital employee that never sleeps. On Tuesday he handed them the toolbox and warned that hesitation equals obsolescence.
Speaking to a packed hall at GTC 2026 in San José, the Nvidia chief baptised an open-source agent framework, OpenClaw, as “the new computer” and urged the crowd to treat it with the same urgency once reserved for Windows migrations. The comparison is deliberate: Huang believes personal ai agents will do for productivity in the 2020s what the PC did for spreadsheets in the 1980s.
From viennese side project to valley darling
OpenClaw began life as Clawdbot, a weekend hack by Austrian developer Peter Steinberger. After a brief rebrand to Moltbot, the repo exploded on GitHub, forked by startups eager to dodge ChatGPT lock-in. Steinberger was poached by OpenAI last December, yet the code stayed behind, GPLv3-licensed and already woven into a dozen commercial pilots. Huang’s anointment on the GTC stage gives the project enterprise oxygen—and a semiconductor supercharger.
Nvidia’s engineers have grafted inference kernels from Groq onto OpenClaw, cutting latency for agent-to-agent chatter to sub-200 µs on a single DGX node. Translation: swarms of agents can now negotiate supply-chain reorders before a human finishes a sip of coffee.
Security panic meets hardware answer
Handing the keys to autonomous code is, in Huang’s own word, “terrifying.” His response is NemoClaw, a hardened distro that slips network filters and privacy routers between the agent and corporate data. Early adopters—JPMorgan, Foxconn, Maersk—get sandboxed blueprints this quarter. The price: a subscription tied to Nvidia’s new Blackwell and Rubin accelerators, the same chips Huang expects to drive $1 trillion in cumulative revenue by 2027.
The sales pitch is already in motion. At a bootcamp adjacent to the convention centre, 400 developers spent Monday night wiring custom agents to Slack, SAP and legacy mainframes. Nvidia mentors hovered, clipboards in hand, scoring each demo for optimisation potential on the next-gen Rubin rack. The implicit message: if your agent runs faster on our silicon, your company buys more of it.
Tokenised pay cheques rewrite hiring rules
Huang saved the grenade for the closing keynote: compensation in compute, not just cash. Under a scheme being piloted internally, senior engineers receive tokens—think prepaid API credits—worth roughly half their base salary. The tokens are metered on Nvidia’s cloud and translate into raw GPU hours for personal prototypes. Recruiters whisper that star candidates now ask “How many tokens come with the role?” before discussing equity. Recruiting platform TripleByte confirms four late-stage offers already include a token clause, calling it “the fourth component of pay” after salary, bonus and RSUs.
The logic is brutal and simple: whoever commands the most compute commands the smartest model. Huang’s bet is that engineers granted surplus cycles will prototype products overnight that pay back the investment tenfold. Critics call it a risky reinvention of scrip; valley VCs call it leverage.
Either way, the gauntlet is down. Build agents, feed them Nvidia silicon, pay talent in the same currency the models consume. Firms clinging to quarterly GPU rental budgets may find their smartest competitors already traded that budget for an army that codes while they sleep.