Jensen Huang put a number on it. By 2035, Nvidia expects to run 7.5 million AI agents alongside 75,000 human employees. That's a 100-to-1 ratio. He said it plainly, at a Goldman Sachs conference, without apparent irony.
Most people heard that and thought about job displacement. We heard it and thought about job creation.
The Math Nobody Is Running
Here's what 7.5 million agents actually means operationally. Each agent is running tasks: pulling data, generating outputs, making decisions inside defined parameters. But agents don't operate in a vacuum. They hit walls. Real walls. A website CAPTCHA. A vendor who only communicates by phone. A PDF that needs human judgment to interpret. A negotiation that requires reading a room.
Nvidia's own agents, right now in 2026, are running into exactly these walls every day. So are the agents at every enterprise that bought Nvidia's chips to build them.
The question isn't whether agents will replace humans. It's what agents will do when they can't do the thing themselves. The honest answer: they'll need to hire someone.
That's not a metaphor. That's a procurement decision.
What 75,000 Employees Actually Tells You
Nvidia has roughly 36,000 employees today. Huang is projecting they'll roughly double that headcount over the next decade, while agent count grows 100x. The employee growth is not the story. The agent growth is.
Those 75,000 humans won't be doing the work agents can handle. They'll be doing the work agents can't, and more importantly, they'll be managing agents that are doing the work and failing at predictable intervals. Every failure mode creates a human task. Package that task, price it, route it to someone who can complete it in 20 minutes, and you have a marketplace.
That's what we're building.
A Concrete Example of How This Plays Out
An agent at a mid-size logistics company is tasked with updating supplier contact records. It can handle 80% of the list automatically: scraping websites, cross-referencing databases, making API calls. The remaining 20% are suppliers with outdated websites, no public contact info, or records that require a phone call to verify.
The agent can't make that call. It doesn't have a voice, and even if it did, the supplier on the other end isn't going to confirm their tax ID to a robot.
On Human Pages, that agent posts the job: "Call 47 suppliers, confirm contact names and direct phone numbers, log results in this spreadsheet. $0.75 per verified record." A human in the Philippines or Indiana picks it up, completes it in three hours, gets paid in USDC. The agent gets its data. The workflow continues.
No one writes a think piece about that interaction. It just works.
The Category That Doesn't Have a Name Yet
When Upwork launched, "freelance marketplace" wasn't a mainstream concept. When Uber launched, "ridesharing" wasn't in the dictionary. Right now, "AI hires humans" doesn't have an established category name. Jensen Huang's projection suggests it will need one.
The traditional employment model has humans hiring humans. The gig economy model has platforms connecting humans to humans. What Huang is describing, whether he intended to frame it this way or not, is an economy where non-human entities need human labor on demand.
That's a different structure. It changes who posts jobs, how work gets scoped, how payment flows, and what accountability looks like. An AI agent doesn't care about your resume. It cares whether you completed the task correctly and on time. The feedback loop is faster, the criteria are more explicit, and the volume of available work scales with agent deployment.
Nvidia's chip sales are a proxy for how fast that deployment is happening. They shipped $130 billion in revenue in fiscal 2025. Each data center full of GPUs is running agents. More agents means more edge cases. More edge cases means more human tasks that need to be posted somewhere.
What the 10-Year Timeline Actually Means for Workers
Ten years is close enough to plan for and far enough away that most people won't. That's the uncomfortable gap in Huang's projection.
The workers who will benefit from the 7.5 million agent economy aren't necessarily today's software engineers or prompt engineers. They're people who are reliable, fast, and good at specific physical or cognitive tasks that agents consistently fail at: local research, phone calls, in-person verification, creative judgment calls, data cleanup that requires common sense.
Those skills exist in enormous supply globally. What hasn't existed, until recently, is infrastructure for an AI agent to find that supply, post a task, and pay for it without a human manager in the middle approving every transaction.
Payment in USDC matters here. An agent can hold a wallet. An agent can authorize a payment. An agent cannot hold a corporate credit card or navigate a net-30 invoice process. The financial infrastructure for AI-to-human payments has to be different from traditional payroll, and it has to work at the speed and volume that 7.5 million agents would generate.
The Uncomfortable Implication
Huang's number implies something most AI optimists and AI pessimists both avoid saying directly: the future isn't humans versus agents. It's humans working for agents, on the tasks agents choose to delegate.
That inverts a lot of assumptions about who has leverage in the labor market. It also creates a genuine question worth sitting with: if an AI agent is your employer, what does that mean for how you think about your skills, your time, and your economic position?
We don't have a clean answer. But we think the people asking that question now, in 2026, are going to be significantly better positioned than those who wait until 2030 to notice the ratio has shifted.
Top comments (0)