TokenEyez
In March 2026, on the final day of Nvidia's GTC conference, Jensen Huang sat down with the All-In podcast hosts. One of them — Jason Calacanis — asked whether Nvidia was spending around $2 billion a year on AI tokens for its engineering team. Huang's answer was "we're trying to."
Then he delivered the line that did the rounds for a week:
"If that $500,000 engineer did not consume at least $250,000 worth of tokens, I am going to be deeply alarmed." — Jensen Huang, Nvidia CEO · All-In podcast, GTC 2026
It's the kind of quote that gets meme'd as "Nvidia CEO wants engineers to set $250,000 on fire," but that misreads it. Huang wasn't celebrating cost. He was making a productivity argument, and a sharp one. The full context is what matters.
The CAD-versus-paper analogy.
Huang compared an engineer who doesn't use AI tokens to a chip designer who insists on using paper and pencil instead of CAD tools. Technically possible. Catastrophically slower. Nobody would actually do it — and if they did, you'd suspect they didn't understand the job.
Read that again. He's saying: AI tooling is the new CAD. An engineer who avoids it is voluntarily handicapping themselves by an order of magnitude. So the $250,000 figure isn't a budget — it's a tell. Anyone at the high end of engineering compensation who isn't burning tokens is, by Huang's read, leaving most of their potential output on the table.
"How many tokens come with the job?"
The most quietly important thing Huang said in that interview wasn't the $250k line. It was this:
"It is now one of the recruiting tools in Silicon Valley: how many tokens come along with my job?" — Jensen Huang, same interview
That's a tectonic shift. A few years ago, the question on the recruiting call was equity, then it was GPU hours for research roles, then it was Copilot access. Now it's a number — an actual annual token allowance — and engineers are negotiating it like they negotiate compensation. Because functionally, it is compensation: it's the leverage they bring to their day-to-day output.
What Nvidia is reportedly trying to spend annually on AI tokens for its engineering team. When the largest semiconductor company on earth treats token budget as a strategic line item, the rest of the industry is going to notice quickly.
Where this leaves the rest of us.
Most teams aren't Nvidia. Most engineers don't get a $250,000 token stipend; they get whatever's left in the team's monthly OpenAI bill. But the direction of travel is clear, and three things follow from it:
1. Token consumption becomes a productivity signal.
Not in a punitive, surveillance way — that would miss the point and alienate every engineer worth keeping. In an opportunity way. If two engineers on the same team are doing similar work, and one is using 5x the tokens of the other, the high-token engineer is almost certainly compounding faster. They've built habits around AI. They reach for it sooner. They iterate more cheaply.
The flip side: an engineer who's using almost no tokens probably needs help — onboarding to better workflows, exposure to MCP-style tools, pairing with someone who's further along the curve. That's what the data is good for. Not performance management. Coaching.
2. Engineers need to see their own usage.
The Huang quote frames tokens as something companies provide to engineers. But the engineer who can see their own consumption — who knows that Tuesday morning's session burned $14 of Opus on a single function refactor, or that Friday's Cursor day cost $0.40 in total — has wildly more agency than one who never sees the number.
Personal visibility creates self-correction. Personal visibility lets you defend your token budget at performance review. Personal visibility lets you make the case for upgrading from Sonnet to Opus for the cases where it actually matters. Without it, you're just a cost line item somebody else is staring at.
3. Per-engineer attribution becomes table stakes for teams.
Once "tokens that come with the job" is a recruiting term, finance teams need an answer to "where did the $80k engineering token budget go?" A single team-level invoice from Anthropic doesn't cut it. You need per-developer attribution, per-project attribution, per-model attribution. The same granularity finance has had for AWS spend since 2018, just for tokens.
This is what TokenEyez is built for.
We didn't build TokenEyez because we thought tokens were interesting. We built it because the shift Huang is describing is happening now, and there isn't a serious independent layer making token consumption visible the way the public cloud got its cost-explorer layer a decade ago.
Per-prompt visibility, per-model breakdown, per-platform attribution, privacy-preserving by design (token counts only, never prompt content). It's not glamorous. It's just the dashboard that should have existed two years ago and didn't. Now there's one.
Jensen Huang is right about the direction. The implication for the rest of us isn't "spend more on tokens" — it's "start paying attention to what you already spend."
If you're new to the per-prompt-visibility argument, the foundational piece is Why tracking tokens matters — it covers the spend-visibility problem from first principles. Once you're ready to ask plain-language questions about your own spend from inside Claude or Cursor, the TokenEyez MCP server is the install-it-in-90-seconds route.