The Startup Equity Concepts Big Tech Never Taught ML Engineers

For ML engineers evaluating career opportunities at Series A and Series B AI startups, an equity percentage is not enough to judge an offer. The deal depends on the option terms, future dilution, investor payout rights, and cash compensation being given up. Big Tech’s RSU model leaves those questions unanswered, so judging startup equity takes a different model as well as a different compensation benchmark.

Why ML engineers from Big Tech are structurally unprepared to evaluate startup equity

RSUs teach a simple rhythm. Shares vest, their value becomes ordinary income, and the engineer owns stock that is liquid on a known schedule. At Google, Meta, Amazon, or Microsoft, that stock sits beside a significant base salary and annual cash bonus. There is no strike price to fund and no exercise decision to make. An engineer can evaluate the package without learning much about how private-company ownership works.

That simplicity creates a blind spot. The mental model works inside Big Tech because the compensation instrument removes most of the difficult decisions. It stops working when “equity” means a right to buy private shares. Startup options come with costs to buy shares, tax exposure, dilution, and rules for who gets paid at exit. Familiar language hides a different bargain.

The financial stakes are growing as that bargain changes. Carta’s 2026 research found that median equity grants for AI engineers rose 31% over two years. With larger grants, what an engineer does not know can cost more. An engineer who can compare salaries and RSU refreshes precisely can still misread a startup package if the analysis stops at the grant’s headline value.

Options are not RSUs — the instrument difference that changes every downstream decision

An RSU gives the engineer shares on vesting, with ordinary income tax due but no purchase required. A stock option gives the engineer the right to buy shares at a fixed strike price. Vesting earns that right. Exercising requires cash, and the decision is irreversible. That distinction introduces a purchase decision before the engineer can receive any eventual sale proceeds.

Startups use Incentive Stock Options, or ISOs, and Non-Qualified Stock Options, or NSOs. Their central difference is tax treatment. ISOs can qualify for long-term capital gains treatment if held correctly, but exercising them can create Alternative Minimum Tax, or AMT, exposure on the spread between the strike price and fair market value. With NSOs, ordinary income tax is due on that spread at exercise. They avoid that particular AMT trap but do not give the exercise spread preferential treatment.

The surprise is the gap between paper gains and spendable cash. An engineer exercising ISOs with a large spread can face a substantial tax bill without selling any shares or having access to a secondary market. Deferring the whole question until departure therefore leaves both the purchase cost and the tax decision unresolved.

The exercise window — why 90 days is a trap most engineers don’t see until it closes

Suppose an engineer has vested options for three years and leaves for another opportunity. Under the industry-default 90-day post-termination exercise window, the clock now starts. The engineer must pay the strike price for every vested share they want to buy before the window closes. Otherwise, those options are forfeited. Three years of vesting can end in a decision about whether to spend real savings on shares with no available buyer.

If exercising ISOs would create AMT exposure, the engineer must consider that cash demand as well as the strike price. The window compresses the time available to plan. Leaving the job does not make the shares liquid, and a new opportunity does not remove the old grant’s deadline.

The 90-day convention comes from IRS rules around ISO treatment, but a company’s exercise policy is negotiable. Some companies extend their windows to one, two, or several years. Those policies remove the immediate forced decision, though they do not make the shares liquid. The offer-stage questions are therefore specific. How long is the window, does it apply uniformly to employees, and has the board approved an extension policy?

Dilution — how a percentage that looks fixed keeps shrinking before exit

Even an affordable option grant with a workable exercise window does not preserve a fixed ownership percentage. The grant represents a fraction of the company’s shares. That denominator can grow. When the company issues new shares to raise money, an existing holder’s share count can stay the same while their percentage falls. Dilution is expected financing mechanics.

Consider a hypothetical engineer who joins at seed stage. At Series A, new investors receive shares, increasing the total share count and reducing the engineer’s percentage. Before Series B, investors may ask for a larger option pool, which dilutes existing holders before the new money comes in. The Series B issuance then increases the denominator again. The original grant remains in place with a smaller ownership percentage.

This is why identical headline percentages at different stages are not equivalent offers. A company earlier in its financing path may have more rounds ahead, so today’s percentage says little without an expectation for future funding. Refresh grants can move the engineer’s stake in the other direction. The useful evaluation therefore combines the likely financing path with the company’s written refresh policy, rather than treating a vague promise of more equity as protection against dilution.

Liquidation preferences — why common shareholders can walk away with nothing from a successful exit

A company can sell for a substantial sum and leave employees with nothing. If the total liquidation preference is greater than the sale price, preferred investors can take all the proceeds before common shareholders get paid. When employees exercise their options, they get common shares. Those shares get paid after preferred investors in the payout order, often called the waterfall. A successful sale for the company is not necessarily a successful outcome for the employee.

Liquidation preferences give investors priority in a sale or liquidation. The aggregate amount is the critical figure. One investor’s terms alone cannot show how much must be paid to all preferred share classes before common holders get anything. AI startups that raise money at high valuations can build large preference stacks even at early stages, so the exit value has to be higher for employee shares to pay out.

Participation can make the outcome worse for common holders. Participating preferred investors take their preference amount and also share in the remaining proceeds as if they had converted to common. That “double dip” distinguishes participating terms from non-participating terms and can sharply reduce employee proceeds in a moderate exit. The practical questions are how much liquidation preference sits above common shares and whether any of it participates. Ownership percentage alone cannot answer either question.

How do you evaluate the equity percentage in an early-stage AI startup offer?

The percentage becomes useful only when its denominator is clear. The fully diluted share count includes issued shares, outstanding options whether vested or unvested, warrants, and the full authorized option pool. It is broader than the shares currently outstanding. Requesting that count and calculating the grant percentage against it establishes what the offer actually represents today.

That starting point still needs context. A founding-engineer grant at pre-seed and a senior ML grant at Series B come with different jobs, paths to funding, and claims from investors. Funding history, headcount at hire, role scope, and seniority determine which benchmark fits. A founding-engineer title does not create a standing entitlement to the grant offered to a company’s first engineer. Nor does a senior-level benchmark price work with staff-level responsibility.

Ask the company to show the grant’s value at the current post-money valuation and at different exit outcomes, with expected dilution stated explicitly. The math for each exit must also include the preference stack. A company that cannot or will not provide the fully diluted share count leaves a basic part of the bargain unresolved.

The cash-for-equity trade — what founding and early ML engineer compensation actually looks like

Once the equity terms are legible, the next question is what the engineer gives up to receive them. Startup ownership is not a free addition to compensation. Seed-stage founding-engineer salaries run below what the same engineer could earn at Big Tech. The gap helps fund the company’s runway in exchange for a speculative ownership position. That is the economic trade, whether or not the recruiting pitch describes it plainly.

Founding grants range from a fraction of a percent to several percent, with stage and hiring order driving much of the difference. At seed stage, equity can provide more than 50% of what makes an offer compelling. Cash plays a larger role later. At Series A and Series B, equity typically represents a meaningful but smaller share of total expected compensation. For non-founding hires, senior and staff ML roles also need different benchmarks because staff work includes setting the technical strategy for the whole organization.

Unvested RSUs at the current employer make the opportunity cost concrete. The startup’s scenarios need enough detail to support an honest comparison with that forgone value. At senior levels, AI lab compensation sits above Big Tech, while periodic secondary tender offers give paper equity a different liquidity profile. Startups cannot pay as much total cash compensation as AI labs. Their competing appeal is product ownership, an earlier-stage role, and an application-layer mission.

Role title confusion and why accepting the wrong benchmark misprices the entire deal

Big Tech titles do not map cleanly onto startup hiring categories, and the category determines the compensation budget before negotiation begins. An AI engineer builds working products with models that already exist. An ML engineer makes machine-learning systems reliable and scalable in production through work on training pipelines, feature stores, model serving, and evaluation. A research scientist creates or validates new methods, with research output at the center of the role.

The distinctions affect both salary and equity benchmarks. Research-scientist compensation at frontier labs sits materially above applied engineering, but most Series A and Series B AI startups need a research engineer profile rather than a pure research scientist unless research output is their core product. A title alone cannot settle that comparison. The work has to establish the category.

ML platform work deserves particular attention because its scarcity creates pricing power. Kubernetes, serving infrastructure, and evaluation tooling belong to a distinct hiring category. The practical step is to confirm the category against the actual responsibilities before the first technical screen. Once the offer is built around the wrong budget, changing the category becomes a harder negotiation.

What to ask before you sign — the equity questions that surface what the offer letter doesn’t say

The questions belong early, before the technical screen, rather than after the offer has anchored the discussion. On instrument and exercise, ask whether the options are ISOs or NSOs, what the current fair market value under the 409A valuation is, and what strike price applies. Ask for the post-termination exercise window as well.

On dilution, ask for today’s fully diluted share count and whether an option pool expansion is planned before or alongside the hire. Then ask what financing path the company anticipates and what dilution that path implies. Refresh policy belongs in the same conversation. Is there one, and is it in writing? A promise of future grants is not a defined policy.

On exit mechanics, ask for the total liquidation preference across all preferred share classes and whether any preferences participate. Ask at what exit value common shareholders would first receive proceeds. Those questions test the payout order directly. They turn an acquisition headline into a question about employee cash.

Finally, ask the company to model the grant at the current post-money valuation, a modest outcome, and a strong outcome after dilution. Those scenarios should reflect the exit mechanics. Refusal to model them leaves the upside claim unsupported. The engineer’s leverage lies in asking these questions while the terms and role category are still open, rather than discovering their consequences after joining.