Jensen Huang says engineers should burn $250,000 a year in AI tokens — and a Microsoft candidate refused to sign without a daily compute floor

At a GeekWire event in Seattle on March 24, 2026, Charles Lamanna, Microsoft's executive vice president for agents and business apps, described a job candidate who refused to join unless Microsoft guaranteed a daily allocation of AI tokens — "upwards of hundreds of dollars per day." Not a signing bonus. Not extra equity. A compute floor, written into the offer as a condition of acceptance.

Lamanna did the math out loud. If a fully loaded engineer costs $500,000 a year and $100,000 in tokens makes them three times as productive, "that's a great deal for everyone involved."

That exchange is no longer an anecdote. It's a market signal.

AI tokens and compute credits are emerging as a fourth pillar of tech compensation, sitting alongside salary, bonus, and equity in the offers that frontier labs and compute-intensive startups put on the table. The shift is being driven by the explosion of agentic AI tools that engineers now use daily (tools that cost real money to run at scale) and by a labor market where top AI/ML engineers command median equity grants up 31% and salaries up 9.1% over the past two years, according to Carta's H2 2025 State of Startup Compensation report published in May 2026. Employers who treat compute access as a perk rather than a core compensation lever risk losing candidates to competitors who already understand that in 2026, a signing bonus might be denominated in GPU hours.

Nvidia's CEO Said It First — and the Market Listened

At Nvidia's GTC conference in March 2026, CEO Jensen Huang made a proposal that landed somewhere between thought experiment and executive directive: engineers should receive roughly half their base salary again in AI tokens. For top engineers, that could mean burning through $250,000 a year in compute, TechCrunch reported.

Huang called tokens a "recruiting tool" and predicted the practice would become standard across Silicon Valley, per CNBC's coverage of the event. Only the CEO of the company whose chips power the entire AI stack could make that statement with a straight face, and have the market treat it as a planning assumption rather than a provocation.

Huang's framing went beyond compensation. He described a workforce of 42,000 biological employees complemented by "hundreds of thousands of digital employees." The logic is structural: if a company's output is increasingly generated by AI agents running on its engineers' prompts, then compute access isn't a fringe benefit. It's infrastructure. Denying it to an engineer is like giving a carpenter a workshop with no electricity.

That message resonated because it formalized something the market had already started doing on its own.

The Market Was Already Moving Before the CEOs Spoke

By the time Huang took the stage in March, the compensation architecture had already shifted. Tomasz Tunguz, a venture capitalist at Theory Ventures, wrote in mid-February 2026 that tech startups had begun treating inference costs as a functional fourth component of engineering pay. Using data from Levels.fyi, Tunguz put a top-quartile software engineer's base salary at $375,000. Add $100,000 in tokens and the fully loaded cost hits $475,000 — meaning roughly one dollar in five is now compute, he wrote.

The accelerant, by most accounts, was OpenClaw. The open-source AI assistant, designed to run continuously, shipped in late January 2026 and immediately made the compute question unavoidable, TechCrunch reported. Any engineer expected to use AI agents daily now has a visible, recurring cost attached to doing their job well. Employers who don't budget for it are effectively asking staff to pay out of pocket, which, as it turns out, many already are. Heavy AI users can easily spend hundreds of dollars a month in API credits, TechCrunch found.

Candidates Are Negotiating for Compute — and Saying No Without It

The power dynamic in tech recruiting has flipped. Candidates no longer passively receive offers; they stipulate compute access as a condition of acceptance.

Lamanna's Microsoft anecdote is the cleanest example: the candidate who set a daily token floor before agreeing to join. But it's not isolated. Thibault Sottiaux, engineering lead at OpenAI's Codex team, wrote on X that he is "increasingly asked during candidate interviews how much dedicated inference compute they will have to build with Codex," Business Insider reported. OpenAI president Greg Brockman framed it plainly: "The inference compute available to you is increasingly going to drive overall software productivity."

Hakeem Shibly, a data specialist at Levels.fyi, spotted a compensation submission from a software engineer that listed a "Copilot subscription" as part of pay and benefits, Business Insider reported. The boundary between tool access and compensation had already blurred in practice before anyone gave it a name.

The labor market data backs up the leverage. Zero G Talent's job board shows 41 Stripe roles added in the past seven days and 23 Waymo roles in the same period, including a Staff Software Engineer position at Waymo paying $251,000 to $310,000 a year and ASIC design verification engineers at $175,000 to $215,000. When companies are hiring at that pace for AI-capable engineers, candidates with the credentials to fill those seats have options. Compute allocation is becoming one of the variables they use to choose between them.

"Tokenmaxxing" Inside Frontier Labs

Inside companies like Meta and OpenAI, token consumption has become a visible, competitive metric. The New York Times reported on March 20 that engineers at these companies are competing on internal leaderboards tracking token usage, a trend that has been dubbed "tokenmaxxing."

It's easy to see why. If your output is measured by what you ship and your tools are AI agents that cost real money to run, then token consumption becomes a proxy for ambition — or at least for resourcefulness. The leaderboard turns compute into a scorecard.

The phenomenon isn't confined to Silicon Valley. The Times spoke with an Ericsson engineer in Stockholm who said he probably spends more on Claude than he earns in salary. His employer picks up the tab. That detail matters: it shows the compute bill has become a material line item even at a telecommunications company in Sweden, not a frontier AI lab in San Francisco. This is not a niche expense. It's a structural cost of employing knowledge workers in 2026.

The Economics Favor Employers Who Say Yes — But the Mechanism Is Still Being Invented

For companies willing to budget for it, token allocation is one of the highest-ROI compensation levers available. Lamanna's math is the anchor: $100,000 in tokens making a $500,000 engineer three times as efficient is not a cost increase. It's a productivity arbitrage.

Goldman Sachs has estimated that AI could automate tasks accounting for 25% of all work hours in the U.S., with a 15% productivity boost from AI, per CNBC. If those numbers hold even approximately, the companies that figure out how to give engineers enough compute to capture that productivity gain will outperform those that don't.

Box CEO Aaron Levie has urged business leaders to start discussing how to budget for AI token bills as usage skyrockets across the workforce, Ken Yeung reported in The AI Economy. This is no longer a tech-industry curiosity. It's an emerging CFO-level concern.

But there are real structural limitations. Jamaal Glenn, a former VC turned financial services CFO, has argued that token budgets don't vest, don't appreciate, and don't carry forward into future negotiations the way base salary or equity grants do, TechCrunch reported. A $100,000 token allocation this year is gone next year. It doesn't compound. It doesn't show up in your net worth. Companies that treat it as a straight substitute for equity may find it harder to retain engineers in year two than they expected.

The smartest employers will design around this, perhaps by tying token budgets to tenure milestones or by converting sustained high-usage allocations into equity grants over time.

Startups and Cloud Providers Build the Infrastructure

The ecosystem is formalizing around token-based compensation. Cloud providers have spent the past two years building credit programs that make employer-sponsored compute scalable without requiring every company to build GPU clusters from scratch.

These programs are the supply-side infrastructure that makes it feasible for a Series A startup to offer a candidate meaningful compute allocations without buying a single server, per AI CERTs News.

That infrastructure matters because the demand side is intensifying. Stripe added 41 roles to Zero G Talent's board in a single week. Waymo added 23 in the same stretch. The competition for AI-capable engineers is tightening precisely as the tooling to support them is scaling.

Companies like ASML and firms across the robotics and autonomous systems stack (including Waymo) are operating in compute-intensive domains where the token compensation question will arrive sooner than their finance teams expect.

The candidate at Microsoft's door asked for hundreds of dollars in daily tokens before signing. Lamanna said yes to the math. Most executives in his position are now doing the same calculation, or they're about to.

In an era when a single engineer with enough compute can do the work of three, the signing bonus isn't a check. It's a token allocation. The talent war isn't coming. It's already being fought in GPU hours.

Working in frontier tech? Zero G Talent tracks the openings: browse frontier tech jobs, openings at ASML, Ouihelp and Stripe, and the people building the field.