You Load Sixteen Tokens, What Do You Get?

The AI economy is reinventing the company town — minus the black lung, plus the keynote presentations.

Look, I know. Three days in a row. But when the CEO of Nvidia spends a single keynote publicly constructing the exact economic trap I've been warning about — and the coverage keeps revealing new layers all week — I'm supposed to just... not write it down?

Monday, Jensen took the GTC stage and announced all of it: the $1 trillion demand projection, the DLSS 5 demo on a $10,000 rig, the token compensation idea, the 7.5 million AI agents. Same keynote. Same leather jacket. The press just kept unpacking it piece by piece.

Wednesday I wrote about the demand and the $10k demo rig. Thursday, Micron's earnings call filled in the automotive angle — 300GB of RAM per vehicle. Today, CNBC published the token compensation piece, and... here we are.

I'm not chasing the story. The story was already there. We're just finding out how deep it goes.

Coal mining was brutal, dangerous work that killed people. Cave-ins. Explosions. Black lung. Nobody's comparing sitting at a laptop in San Jose to that.

But the economic mechanism coal companies used to trap workers — paying them in company-controlled currency they could only spend at company-controlled stores — that playbook is getting dusted off in Silicon Valley and presented as innovation.

The Pitch

On Monday, Jensen Huang took the stage at GTC 2026 and floated a novel compensation model: give engineers an annual token budget worth roughly half their base salary.

"They're going to make a few hundred thousand dollars a year, their base pay," Huang said. "I'm going to give them probably half of that on top of it as tokens so that they could be amplified 10x."

At the allowance levels Huang described, that's $100,000-$150,000 in "compute credits" on top of a $200,000-$300,000 base salary. Sounds generous. Engineers would have access to billions of tokens annually, unlocking massive productivity gains.

Here's the thing though: where do those tokens get spent?

On Nvidia's infrastructure. Running on Nvidia's chips. Generating demand metrics that Nvidia reports to Wall Street.

The Pattern

Let's do a side-by-side.

The tokens run on Nvidia's own Vera Rubin infrastructure — meaning the company is effectively consuming its own chips. They "pay" you in something that costs them almost nothing to produce internally. But it shows up on the books as "$100k-$150k in compensation value."

Oh, and Jensen also mentioned that in ten years, Nvidia expects to have 75,000 human employees working alongside 7.5 million AI agents. A 100-to-1 ratio. "As small as possible, as big as necessary," he said — about the humans.

Meanwhile, every token you spend shows up as "inference demand" in the metrics Jensen uses to justify his $1 trillion projection.

Wait, Didn't We Already Know This?

A month ago, Jason Calacanis revealed on the All-In podcast that he's spending $300 per day to run a single AI agent built on Anthropic's Claude — roughly $100,000 a year — and that agent is only operating at 10-20% capacity.

His question was blunt: "When do tokens outpace the salary of the employee?"

Chamath's response: "We're getting to a place where we have to basically now say, 'What is the token budget that we're willing to give our best devs?' ... That is actively happening inside my business, because otherwise I'll run out of money."

That was February. A month later, Jensen picks it up at GTC and reframes it.

Chamath's problem — "my devs need to be 2x productive or I run out of money because of token costs" — becomes Jensen's perk: "I'm going to give engineers tokens worth half their salary as compensation!"

Same math. Completely different framing.

One guy's existential cost problem is another guy's recruiting tool announcement.

The Money Has To Come From Somewhere

Nothing's free. If companies are adding $100k+ in token budgets to compensation packages, that money comes from somewhere:

1. Base salaries quietly flatten — "but your total comp is higher with the token budget!"
2. Equity grants get cut — "tokens are the new equity, they're how you generate value"
3. Bonuses disappear — "your productivity bonus IS the token budget"
4. Headcount gets reduced — fewer humans, same budget, but now "tokens" are a line item that looks like compensation but is actually just operating cost rebranded

And here's the really cynical read: the tokens cost Nvidia almost nothing to produce internally. It's their compute on their chips. But it gets booked as compensation value. And it gets counted as inference demand.

They turned payroll into a demand generation engine and called it a benefit.

Sixteen Tons

The 1946 song "Sixteen Tons" wasn't subtle about what company scrip meant for workers:

You load sixteen tons, what do you get? Another day older and deeper in debt Saint Peter don't you call me 'cause I can't go I owe my soul to the company store

The song was about economic captivity. Not physical danger — though that existed too — but the structural trap of being paid in a currency you couldn't save, couldn't transfer, and could only spend in one place.

Nobody's asked the hard questions about tokens yet:

• Will the IRS treat a $100,000 token grant as taxable income?
• Can an engineer take unused tokens with them if they leave?
• What happens to your "compensation" if Nvidia changes the token valuation?
• If you don't spend your tokens, did you just forfeit half your pay?

The Pullman Company tried this model. Ford tried it with company towns. It keeps getting reinvented because it keeps working — until it doesn't.

The Chain Is Complete

Three days. Three pieces of the same puzzle:

• Day 1: $1 trillion in demand, features that need $10,000 rigs
• Day 2: 300GB of server-class RAM per vehicle, 20-year growth vector
• Day 3: Pay engineers in tokens they can only spend on your infrastructure

The demand comes from making tokens a compensation line item. Pay people in tokens. They have to spend the tokens. Token consumption goes up. Consumption justifies the $1 trillion projection. Projection justifies more chip production. More chips mean more tokens to "pay" people with.

It's circular all the way down.

The AI bubble isn't just making your GPU more expensive. It isn't just making your car unrepairable. It's reinventing the company store and calling it the future of work.

You load sixteen tokens, what do you get? Another day older and deeper in debt Saint Peter don't you call me, 'cause I can't go I owe my soul to the company AI model

The mine got safer. The model stayed the same.

Find me on Mastodon at @ppb1701@ppb.social if you want to talk about it. Because apparently the thread keeps not running out.