Your Next Raise Will Be Measured in Tokens, Not Dollars. AI Compute Is the Fourth Component of Tech Compensation.

Something shifted in how Silicon Valley thinks about paying engineers, and it happened quietly enough that most people outside the industry missed it.

OpenAI's Codex engineering lead, Thibault Sottiaux, posted on X this week that job candidates are now asking during interviews how much dedicated inference compute they will have access to. Not salary. Not equity. Compute. How many tokens per month can I burn through Codex to build software?

OpenAI President Greg Brockman put the implication plainly: "The inference compute available to you is increasingly going to drive overall software productivity."

Tomasz Tunguz of Theory Ventures connected the dots in a way that should make every founder and hiring manager pay attention. He argues that AI inference is now the fourth component of engineering compensation: salary, bonus, equity, and tokens. With Levels.fyi pegging the 75th percentile software engineer salary at $375,000, adding $100,000 in annual inference costs brings the fully loaded cost to $475,000. Just over 20% of the total compensation cost could come from AI usage.

This is not a prediction about 2028. Tunguz says it is starting to happen right now.

Why this matters beyond HR

The shift from "AI as a tool" to "AI as compensation" reveals something deeper about how software gets built in 2026. An engineer with unlimited Codex access produces fundamentally more software than one without it. The output gap is not marginal. Tunguz estimates it could be 8x.

That gap creates a two-tier labor market. Engineers who negotiate AI compute budgets alongside their salary will outproduce colleagues who do not. Companies that provide generous inference budgets will attract better talent than those that limit access. And the CFO's office will need entirely new frameworks for measuring whether the AI spend is generating returns.

Peter Gostev of Arena suggested that OpenAI and Anthropic should create recruitment sites where companies can advertise roles listing the token budget alongside the salary range. That has not happened yet. But the fact that the idea is circulating tells you the market is moving in this direction.

The CFO problem

For finance leaders, AI inference is a new category of employee cost that behaves differently from anything they have managed before.

Salary is fixed. Bonus is predictable. Equity is modeled. Inference is variable, usage-dependent, and can spike unpredictably when an engineer tackles a complex project. Tunguz himself automates 31 tasks a day at a cost of about $12,000 per year in inference. A power user running large code generation workloads through Codex could easily hit $100,000 annually.

CFOs need to track this cost, attribute it to individuals or teams, and correlate it with output. The question Tunguz poses is sharp: what is the productive work per dollar of inference? It is the new unit economics of engineering talent.

Almost no company has good answers to this yet. That is the gap.

Five ways to build a business around this shift

AI compute benchmarking platform

The most direct opportunity. Build a SaaS that helps companies measure the return on their AI inference spend.

The product connects to OpenAI, Anthropic, and other API providers. It tracks token usage per engineer, per team, per project. It correlates inference spend with output metrics: pull requests merged, features shipped, bugs resolved, code review velocity.

The dashboard shows a CFO exactly what they need: which teams are getting the most value from their AI compute budget, and which are burning tokens without proportional output.

Price it at $99 per month for small teams, $499 for mid-market, enterprise pricing for large organizations. The comparison to cloud infrastructure monitoring is direct: companies already pay for Datadog to track cloud spend efficiency. The AI compute equivalent does not exist yet.

At 500 companies paying an average of $249 per month, that is $1.5 million in annual recurring revenue. And the market is growing as fast as AI adoption itself.

Inference cost optimization consulting

Most engineering teams are wasting 30% to 50% of their AI inference spend. They send unnecessarily large context windows. They retry failed requests without backoff. They use expensive models for tasks that cheaper models handle equally well. They do not cache responses that could be reused.

An optimization engagement looks like this: audit the team's API usage for two weeks, identify the top waste patterns, implement caching, model routing, and prompt optimization, and measure the savings.

If a 50-person engineering team spends $500,000 per year on inference and you cut it by 35%, you saved them $175,000. Charge $15,000 for the engagement. The ROI is obvious and immediate.

You need to understand the pricing models of OpenAI, Anthropic, Google, and the major inference providers. You need to know the difference between when GPT-5.4 is necessary and when a smaller model handles the task. And you need to be able to instrument API usage to identify waste patterns.

This is a service business you can start this week. The demand will grow every quarter as inference budgets expand.

Levels.fyi for AI compute

Levels.fyi became a $100 million company by letting engineers compare compensation packages across companies. One data field, "Copilot subscription," already appeared in a compensation submission. But there is no structured way to compare AI compute benefits across employers.

Build the platform. Engineers self-report their compensation packages including: base salary, equity, bonus, and AI compute budget (monthly token allocation, which models they have access to, whether usage is capped or unlimited).

Companies would voluntarily participate because the platform becomes a recruiting tool. Listing a generous AI compute budget signals that the company takes engineering productivity seriously.

Monetize through job board listings, employer branding pages, and recruiting tools. The same model Levels.fyi uses, adapted for the AI era. Being first to own this data set creates a durable competitive advantage.

Per-engineer inference dashboards

Closer to the team level. Build a tool that engineering managers use to understand their team's AI compute usage and correlate it with shipping velocity.

The dashboard shows each engineer's daily token consumption, the models they use, the types of tasks they run, and how their usage patterns correlate with their output. Not as surveillance. As a coaching tool. Helping engineers optimize their AI workflows the way a running coach helps optimize training load.

Price it at $29 per month per seat for small teams, $99 for larger organizations with advanced analytics. The buyer is the engineering manager who has a $500,000 annual inference budget and no visibility into whether it is being spent effectively.

At 2,000 seats paying $49, that is $1.2 million in ARR from a focused product.

AI productivity coaching

Tunguz says the engineer burning $100,000 in inference had better be 8x more productive. But most engineers have never been taught how to maximize their AI compute budget. They use Codex or Claude Code the same way they used Stack Overflow: reactively, when they get stuck.

The engineers who are 8x more productive use AI differently. They decompose problems before prompting. They provide structured context. They use cheaper models for routine tasks and reserve expensive models for complex reasoning. They build personal automation workflows that compound over time.

A cohort-based course teaching these techniques could price at $200 to $500 per seat. The audience is every software engineer who has access to AI tools but is not getting the full value from them.

Run cohorts of 25 to 30 engineers. At $350 per seat, each cohort generates $8,750 to $10,500. Two cohorts per month is $175,000 to $210,000 annually. And the content updates itself as new models and tools launch.

The market timing

This trend is at the inflection point where awareness is growing but infrastructure has not caught up. OpenAI's president is talking about it publicly. A VC is writing about it on LinkedIn. Business Insider is covering it. But the tools, services, and platforms that companies need to manage this new cost category barely exist.

The comparison to cloud cost management is instructive. When AWS bills started growing in the early 2010s, companies scrambled for tools to understand and optimize their cloud spend. CloudHealth, Cloudability, and Spot.io built billion-dollar businesses by solving that problem. The AI inference cost management market is at the same stage.

The difference is that inference costs are growing faster than cloud costs ever did. Usage per user at OpenAI is growing faster than user growth, meaning existing engineers are consuming exponentially more tokens. The measurement and optimization market will grow proportionally.

What to take from this

Salary, bonus, equity, tokens. The fourth component of engineering compensation is arriving, and every business that helps companies manage, measure, and optimize it has a clear market opportunity. The tools do not exist yet. The demand does. Build now.