Nvidia bets the farm on open-source agent openclaw and a trillion-dollar chip pipeline
Jensen Huang wants every company on Earth to clone itself—digitally. On the GTC 2026 stage he barked a new marching order: OpenClaw or bust. The Nvidia CEO claims the scrappy open-source agent, once a Valley side project, is now ‘the new PC’, the pivot on which the next decade of AI productivity will swing.
From weekend repo to enterprise mandate
Two years ago OpenClaw was a GitHub curiosity called Clawdbot, then Moltbot, maintained by Austrian coder Peter Steinberger. Steinberger was quietly hired by OpenAI last winter, yet the repo kept forking. Huang’s team counted 43 000 enterprise pull requests in 90 days. That traction convinced Nvidia to stop observing and start architecting. The result: a full-stack pitch that turns commodity servers into agent swarms.
Huang’s logic is brutal and simple. CPUs drove the information age; GPUs drive the inference age; but agents, he argues, will drive the automation age. Miss this cycle and your data center becomes a museum. He framed OpenClaw as the Linux of this transition—foundational, unruly, unavoidable.
Security nightmare wrapped in a trillion-dollar bow
Agents that read mail, query ledgers, and write code also open a highway for exfiltration. Huang didn’t flinch. He unveiled NemoClaw, a hardened fork laced with Nvidia-built microvisors, privacy routers and traffic-filters that keep corporate secrets from strolling out the front door. Early adopters—Goldman Sachs, Foxconn, Mayo Clinic—get the toolkit under NDA this quarter. Wider release lands in September, paired with a road-show where developers can assemble a bespoke agent in 45 minutes on a DGX Spark box.
The sales target is equally muscular. Huang told analysts to pencil in one trillion dollars in cumulative demand for Blackwell and Rubin accelerators through 2027, a figure that eclipses the GDP of Switzerland. Wall Street scribes scrambled to check their models; the supply chain, already sweating CoWoS capacity, just received a fresh injection of adrenaline.
Tokens are the new stock options
Then came the HR grenade. Huang proposes paying engineers partly in compute tokens—call it salary 2.0. A senior GPU architect earning 400 k USD might receive an annual allowance of 200 k tokens, each token worth roughly a cent of H100 time. Burn them on pet experiments, rent them to other teams, or hoard them like crypto; productivity, Nvidia claims, can jump 10× when experimentation is frictionless. Recruiters at Google, Meta and Tesla are already listing ‘token packages’ in offer letters, confirming that compute is becoming the fourth component of compensation after cash, bonus and equity.
Loopholes abound. Token pricing floats with cluster load; a midnight training binge could halve their value by dawn. Labour lawyers warn of taxable events each time tokens change hands. Yet the talent market doesn’t care. In a hallway poll at GTC, 62 % of attending engineers preferred 1 M tokens over 50 k USD of RSUs, gambling that scarcity will drive token appreciation faster than equity dilution.
The subtext is clear: whoever controls compute currency controls the talent pipeline. Nvidia, already the gatekeeper of GPUs, now eyes the payroll layer. Competitors can either match the game or watch their best model builders migrate to Santa Clara.
Huang left the stage after a final jab: ‘Windows gave you a desktop, OpenClaw gives you a workforce.’ The lights dimmed to Blade-Runner synths, but the message was stripped of Hollywood gloss. Adopt agents this year or spend the next decade licensing your own obsolescence from someone who did. The countdown started Tuesday; the bill arrives in 2027.