1 in 5 dollars of engineer comp is now non-transferable compute

In February 2026, venture capitalist Tomasz Tunguz of Theory Ventures published a deceptively simple calculation: take a top-quartile software engineer salary of $375,000, add $100,000 in AI tokens, and you're at $475,000 fully loaded. One dollar in five is now compute. The post crystallized a shift that had been building across Silicon Valley for months — and that most engineers still don't fully understand.

Those tokens don't vest on a schedule. They don't appreciate in value. In most cases, you can't sell, transfer, or convert them to cash. If you leave — voluntarily or otherwise — they vanish. The headline number looks like compensation. The fine print says otherwise.

In 2026, the engineers winning aren't the ones with the biggest token grants. They're the ones who read the terms.

Why This Matters Now

AI token compensation has crossed a threshold. What began as a founder-level experiment — a way to give early employees access to expensive compute without handing out more equity — is now a standard retention and performance tool across Silicon Valley. The shift accelerated in March 2026 when Nvidia CEO Jensen Huang stood on the GTC stage and told the industry that a $500,000-per-year engineer should spend at least $250,000 annually in AI tokens. He said he'd be "deeply alarmed" if they spent only $5,000, comparing low token use to a chip designer working with paper and pencil.

The message landed. Companies from OpenAI to Databricks to startups with fewer than 10 employees now embed token budgets into compensation packages, TechCrunch reported. But most engineers still treat these allocations like bonuses or perks — line items that inflate a headline number without changing the underlying economics.

That misunderstanding has real consequences. Token structures carry vesting cliffs, usage restrictions, and departure terms that most job-seekers have never encountered in traditional equity negotiations. Engineers who don't parse these terms leave money and career leverage on the table — sometimes without realizing it until they try to leave.

This isn't a perks story. It's about how AI is redefining labor value, productivity metrics, and job security. And it's happening faster than most people's offer letters can keep up.

The Token Illusion — When "Compensation" Isn't Really Compensation

Jamaal Glenn, a Stanford MBA and former venture capitalist who now serves as a CFO in financial services, has argued bluntly on his Substack that token budgets aren't compensation. They're non-transferable, non-appreciating compute credits with no residual value. Companies can inflate apparent comp packages without increasing cash or equity — and engineers who negotiate only on headline numbers fall for accounting theater.

The math backs him up. Tunguz's breakdown — $375,000 base plus $100,000 in tokens equals $475,000 fully loaded — shows that roughly one dollar in five is now compute. But that $100,000 has zero liquidity and zero portability. It can't be sold on a secondary market. It doesn't vest into anything. It can't follow you to your next role.

Compare that to traditional equity. Restricted stock units vest over four years. If the company IPOs or gets acquired, those shares convert to real money. Even in a flat exit, employees often retain some value. Tokens offer none of that. They're access credits — useful only as long as you're inside the building.

The practical effect is a bait-and-switch that's technically honest. The offer letter says $475,000. The token line item is real in the sense that the company will provision $100,000 worth of compute. But calling it compensation is like calling a company car part of your salary. It's a tool, not a paycheck.

From Perk to Policy — How Token Spending Became a Performance Metric

The token story gets stranger when you look at how companies actually use these budgets. At Meta and OpenAI, engineers compete on internal leaderboards that track token consumption, the New York Times reported March 20. High spenders get visibility. Low spenders get questions.

At Databricks — valued at $134 billion — CEO Ali Ghodsi publicly celebrated an engineer who burned through more than $7,000 in AI tokens over two weeks in January 2026 using the company's internal coding tool "Isaac," Forbes reported. The celebration wasn't subtle. It was a signal: this is what winning looks like now.

At Sendbird, a startup serving businesses like DoorDash and Redfin, CEO John Kim says about 5–10% of engineering staff have reached what the company calls "AI God" tier — defined as using at least 100 million tokens per day. The company runs a leaderboard and awards prizes including coffee gift cards and company swag, with plans to add extra vacation days. At Writer, valued at $1.9 billion, non-engineering business users have hit the "billion-token-a-month" club; the top user on the internal leaderboard racked up almost 5.9 billion tokens, senior partnerships lead Matt Sobel wrote on LinkedIn.

The scale is escalating fast. At Alven, a startup with fewer than 10 employees building AI tools for real estate, the company spent $16,000 on tokens in February 2026 and is aiming to spend $60,000 the following month, cofounder Julio-Cezar Scerbina wrote on LinkedIn.

The pattern is clear: high token spend equals high visibility equals promotion path. But it also creates pressure to burn tokens regardless of whether the output justifies the cost. When your performance review is tied to compute consumption, the incentive isn't to be efficient. It's to be loud.

The Vesting Trap — Why Most Token Grants Are Designed to Retain, Not Reward

Traditional equity has a well-understood grammar. Four-year vesting, one-year cliff, acceleration on change of control, early exercise options. Engineers have been negotiating these terms for decades. Token grants don't have an equivalent playbook — and that's by design.

Some companies are experimenting with token grants that vest over time, mirroring traditional equity schedules, according to TechCrunch. At least one startup offers token stipends that roll over quarter to quarter, letting engineers build up reserves for intensive projects. But these are early experiments, and there's no public data on acceleration clauses, early exercise rights, or secondary-market sales.

The contrast with equity is stark. Traditional equity offers liquidity events, IPOs, or acquisitions. Even in a down market, vested shares have some negotiable value. Tokens offer only continued access while employed. Leave the company and the unvested portion disappears — along with, in many cases, the AI-augmented workflows that made you productive in the first place.

That second loss is the one nobody talks about. An engineer who's spent months building agent pipelines, fine-tuning prompt chains, and integrating AI tools into their daily workflow doesn't just lose tokens when they leave. They lose the entire productivity stack those tokens funded. Starting over at a new company means rebuilding from scratch — often with a different toolset, different models, and a different token budget that may not cover what you actually need.

The vesting trap isn't just financial. It's operational. And it's effective.

The Productivity Paradox — More Tokens ≠ More Value

Meta CTO Andrew Bosworth said at a tech summit in San Francisco that his best engineer is spending his salary equivalent in tokens but is "5x to 10x more productive," Forbes reported. His verdict: "This is easy money. No limit."

It's a compelling claim. It's also nearly impossible to verify.

Andreas Welsch, founder of Intelligence Briefing, has noted that roughly 80–85% of AI projects have failed since 2018, CNBC reported. Goldman Sachs estimates AI could automate tasks accounting for 25% of all work hours in the U.S., potentially displacing 6–7% of jobs over the adoption period, while delivering a 15% productivity boost. That's a real gain — but it's a long way from 5x to 10x, and it comes with a jobs cost that Bosworth's framing doesn't address.

Then there's the Ericsson engineer in Stockholm who told the New York Times he probably spends more on Anthropic's Claude than he earns in salary, with his employer covering the cost. That's not a productivity story. That's a cost-center story. There's no indication his output justifies the spend — only that the spend is happening.

The pricing landscape makes this worse. Anthropic's Claude Opus 4.6 charges $25 per million output tokens. Premium models can exceed $100 per million tokens. Cheaper models cost a few cents. The range is enormous, and most engineers aren't optimizing for cost-per-output. They're optimizing for speed, quality, or — increasingly — leaderboard position.

Token budgets can inflate activity metrics without improving outcomes. An engineer burning 100 million tokens a day might be building something transformative. They might also be running redundant experiments, over-engineering prompts, or generating code that gets rewritten by a human anyway. The tokens don't distinguish. The leaderboard doesn't care.

The Talent Squeeze — Why Companies Use Tokens to Hide a Deeper Problem

There's a structural reason token compensation is spreading, and it has nothing to do with productivity. It's a talent strategy.

Career consultancy Mercer Asia reports that 54% of C-suite executives cite talent scarcity as their top macro challenge. At the same time, 98% expect AI to lead to headcount reductions over the next two years. Around 65% expect 11–30% of their workforce to be redeployed or reskilled due to AI by 2026, CNBC reported.

The contradiction is the point. Companies can't find enough AI talent, but they're simultaneously planning to reduce headcount through automation. Token compensation sits right in the middle of that tension: it's a way to retain engineers now — make the offer look bigger, make the role feel more cutting-edge — without making the long-term cash or equity commitments that traditional comp requires.

Howard Marks, founder of Oaktree Capital Management, framed the stakes in a memo to investors: AI's new ability to "act autonomously" is what separates a $50 billion market from a multi-trillion-dollar one. The implication for human roles is uncomfortable. If autonomous agents can do the work, the engineers managing those agents become a transitional layer — valuable now, but not necessarily permanent.

Token grants are a stopgap. They retain engineers during the shift to agent-driven development. They don't guarantee those engineers will have a role when the shift is complete.

The Negotiation Gap — What Smart Engineers Are Asking For (and Getting)

The engineers who understand this dynamic are starting to negotiate token terms the way they once negotiated equity. And they're finding that the terms are more flexible than the offer letters suggest.

Some are demanding token cliffs aligned with their equity vesting — a one-year cliff for tokens, then quarterly vesting thereafter. Others are pushing for partial cash conversion of unused tokens upon departure, treating the unused portion as a prepaid expense the company should refund. A few are asking for internal marketplace access, letting them trade unused token allocations with other teams or engineers.

The most aggressive negotiators are asking for portability clauses: the ability to transfer a portion of their token budget to a new employer, or to convert tokens into credits on a specific model provider's platform. None of these terms are standard yet. That's what makes them negotiable.

The leverage isn't in more tokens. It's in better terms. An engineer who treats a $100,000 token allocation as a pure perk will accept whatever structure the company offers. An engineer who treats it as a negotiable instrument — with vesting, rollover, and exit provisions — can often extract concessions that make the difference between a package that traps them and one that actually compensates them.

The Future of Work — When Your Co-Worker Is an AI Agent

Jensen Huang's vision for Nvidia is explicit: 42,000 human workers alongside "hundreds of thousands of digital employees" — AI agents running continuously, handling tasks that used to require human judgment and human hours. Bruno Guicardi, president and founder of CI&T, told CNBC that work that used to take months now takes a couple of days with AI agents, and that software engineers can now instruct computers in plain English rather than programming languages.

The market data supports the direction. Anthropic's Claude Code has reached $2.5 billion in annualized revenue, Forbes reported. Cursor's annualized revenue grew to more than $2 billion over the last three months as of March 2026. OpenAI's Codex has 1.6 million weekly active users. These aren't experimental tools. They're production infrastructure.

At Google, employees in some cases will have their adoption of AI tools factored into their performance reviews, Business Insider reported, as cited by Forbes. The message from the top is consistent: use the tools, or fall behind.

Token compensation is the bridge to this future. Today, engineers are allocated tokens to augment their own work. Tomorrow, they'll be allocated tokens to fund the agents that do the work for them. The performance metric won't be how many tokens you burn personally. It will be how much output your fleet of agents produces per dollar of compute.

Engineers who master token economics now — who understand pricing models, negotiate vesting terms, and build efficient agent workflows — will lead the hybrid human-AI teams of the next decade. Those who treat tokens as a perk and ignore the structure underneath will find themselves managing agents they don't understand, on budgets they didn't negotiate, for outcomes they can't measure.

The Fine Print Is the Future

The shift Tunguz identified — one dollar in five of engineering compensation now being compute — is still early. Most companies haven't standardized their token grant structures. Most engineers haven't learned to negotiate them. And most offer letters still present token allocations as straightforward perks rather than complex, negotiable instruments with real retention implications.

That gap between what tokens appear to be and what they actually are represents both a risk and an opportunity. The risk is that engineers accept inflated headline numbers without understanding the strings attached. The opportunity is that the engineers who do understand — who parse the vesting terms, negotiate the exit clauses, and treat token budgets as seriously as equity grants — will build careers with more leverage, more flexibility, and more real compensation than those who don't.

In 2026, the most valuable skill isn't writing code. It's reading the terms of service. Because in the age of AI compensation, the fine print isn't boilerplate. It's the whole deal.

Working in AI? Zero G Talent tracks the openings: browse AI jobs, the companies hiring, and the people building the field.