# AI Research Equity Grants Pass $4 Million While Base Pay Stays Close

> Pave's new dataset puts senior AI research equity above $4 million while base pay stays near engineering bands. The split changes job design, dilution, retention, and pay transparency.

- Published: 2026-08-30
- Author: Gene Dai
- Canonical: [https://digidai.github.io/2026/08/30/ai-research-equity-grants-base-pay/](https://digidai.github.io/2026/08/30/ai-research-equity-grants-base-pay/)
- Topics: Artificial Intelligence, AI Talent, Compensation, Equity, Workforce Planning, Deep Investigation

---

At 10 a.m. Eastern on August 24, Pave and the Nua Group released a compensation report with a number large enough to reset an offer meeting. In their dataset, the median annualized intended new-hire equity grant for an AI Research Scientist reached $4.09 million at the P6 individual-contributor level and $4.72 million at the M6 manager level. Median base pay for those roles was $311,000 and $340,000.

Machine-learning engineers at the same levels earned higher median base salaries: $321,000 at P6 and $347,000 at M6. Their median equity grants were far lower, at $1.60 million and $2.02 million. The expensive part of a senior research offer was not hiding in the salary band. It sat in the ownership line.

Later that day, [Axios reported that Luke Metz had left OpenAI for Meta](https://www.axios.com/2026/08/24/meta-hires-openai-luke-metz). He joined Meta Superintelligence Labs and would report to Alexandr Wang. Metz had already moved from OpenAI to Thinking Machines Lab and back before making this turn. Earlier in August, Lilian Weng returned to OpenAI after a period at Thinking Machines. Those moves did not disclose anyone's pay. They showed how quickly a scarce research labor market could move while an annual salary survey was still being assembled.

Read the report and the personnel news as different evidence. Compensation systems produced the benchmark. Reporters traced named people changing institutions. Neither proves that equity caused a move. Compute access, colleagues, research freedom, model strategy, and personal mission can matter more than another dollar. Yet a compensation committee has to fund an offer before it can test those motives.

A company-wide "AI premium" would hide the decision. The [State of AI Talent report](https://explore.pave.com/The-State-of-AI-Talent.html) maps more than 15,000 employees into three job families across Pave's dataset of more than 9,000 companies. AI Research Scientist, ML Engineering, and AI Engineering do not share one pay curve. Nor do they produce the same work.

Pave sells compensation data, and Nua advises employers on compensation. Their access to HR systems, applicant systems, and equity records gives the report unusual detail. Their commercial position also requires a clear boundary. Role mapping, level matching, private-company valuations, and the conversion of a multiyear grant into an annualized intended value are methodological choices. A $4.09 million figure is not cash received this year. It is not a guaranteed liquid value. It is not necessarily the face value of the full grant.

Even with that boundary, the split matters. It changes which role a startup can afford and what a public salary range reveals. It also changes how a recent hire compares with an incumbent. Most of all, it shows whether a board is funding research or merely buying the label.

## August 24 moved the premium out of salary

Put the three senior job families on one page and base pay looks close. At P6, the distance from the lowest to the highest reported median is $27,000. At M6, it is $43,000. Those are meaningful sums, but they are small beside the gap between research and engineering equity.

| Pave job family | P6 median base | P6 median new-hire equity | M6 median base | M6 median new-hire equity |
| --- | ---: | ---: | ---: | ---: |
| AI Research Scientist | $311,000 | $4.09 million | $340,000 | $4.72 million |
| AI Engineering | $294,000 | $1.48 million | $304,000 | $1.12 million |
| ML Engineering | $321,000 | $1.60 million | $347,000 | $2.02 million |

Pave labels the equity measure as gross intended annualized value for a new hire. Each word matters. "Intended" points to the value used when the grant was made, not the amount ultimately realized. "Annualized" spreads a grant across its intended life. "Gross" precedes taxes and any cost of exercise. A private-company candidate still needs the share count, fully diluted ownership, strike price, and preferred price. Vesting, the exercise window, liquidation preferences, and a candid account of liquidity complete the picture.

Compensation owners need those details too. Two grants with the same annualized intended value can impose different dilution and retention effects. A four-year option grant at a young startup is not the same instrument as restricted stock in a public company. A marked private valuation can rise while common-share liquidity remains unavailable. A down round can change the value without changing the employee's work.

Look closely at the headline tail. Pave puts the P6 and M6 90th percentile at roughly $45 million to $60 million. The filter covers private San Francisco Bay Area companies that have raised $1 billion to $5 billion or more. That is an extreme percentile inside a heavily financed regional set, not a general market price for a senior researcher. A Series A founder who imports it into a board deck has already lost the denominator.

Geography narrows the lens further. California accounts for 49.3% of US AI and ML professionals in the Pave dataset, while the United States accounts for 57.3% of the global total. Those are dataset shares, not population estimates. A company hiring in Toronto, Paris, Bengaluru, or a lower-cost US city needs a local comparison, plus an explicit policy for remote pay. Bay Area grants can affect expectations elsewhere without describing the local market.

Still, the tail affects negotiations below it. A research candidate does not need a $45 million offer for its existence to change the conversation. Recruiters cite known packages. Investors ask whether the option pool can support one exceptional hire. Incumbents compare grants issued under an older valuation with the ownership offered to a new colleague. Managers decide whether another researcher would be more useful than five engineers, more compute, or six additional months of runway.

Matt Schulman, Pave's chief executive, framed the market as one where recent hiring data can age in six months. That is also a vendor argument for continuously updated benchmarks. It should not become an excuse to let a benchmark make the decision. Faster data can show that offers moved. It cannot determine whether the company should follow.

Benchmarks can also push the market they measure. In a thin specialty, several boards may see the same upper percentile and approve exceptions to avoid losing a candidate. Those offers enter the next dataset. The feedback does not make the observations false, but it can turn a descriptive price into a coordination point. Finance needs an internal ceiling tied to expected value, not only an external percentile tied to other companies' fear.

Before asking for an AI premium, identify the work that has become scarce enough to make ownership the unit of competition.

## Three job families carry different pay curves

Research begins where a predictable feature schedule ends. Scientists may improve model capability, efficiency, safety, or understanding. The output can be an experiment, training method, evaluation, paper, data insight, or model behavior that does not ship next week. Research judgment matters, often alongside expensive compute, proprietary data, and peers who can challenge the result.

Production changes the equation for ML engineers. Their work can include data pipelines, distributed training, inference, evaluation systems, model monitoring, and the difficult path from a promising experiment to a repeatable system. High median base pay makes sense when the job carries continuous operating responsibility.

Closer to the product, an AI engineer applies models to a customer experience or business process. The person may design retrieval, tool use, agent workflows, model routing, evaluation, guardrails, and the surrounding application. This role can deliver revenue or labor savings without inventing a foundation model. It is also the family most likely to collide with ordinary software engineering when a title is written badly.

Ryland Bauer, a Total Rewards Advisor at Nua Group, described employers asking for "AI Engineering talent" before deciding which of these jobs they actually needed. The report's useful move is to force that decision upstream. It sorts companies into Innovators, Integrators, and Implementers. An Innovator builds new models or core methods. An Integrator embeds models into a differentiated product. An Implementer uses available tools inside an existing business.

Those categories are a starting point, not permanent identities. A healthcare software company may implement a general assistant in finance, integrate a model into clinical administration, and run a small research program on a proprietary dataset. One company can occupy all three lanes. One job cannot do all three at once.

Consider a founder hiring the first "Head of AI." The job description asks for publications, production infrastructure, enterprise integration, sales calls, hiring, and policy ownership. Finance finds one executive band. Candidates arrive from research labs, cloud infrastructure teams, and application startups. Their past pay is not comparable because the proposed work is not coherent.

Splitting the role changes the budget. The founder may discover that a senior AI engineer and a part-time research adviser cover the next twelve months. A model company may reach the opposite conclusion: one research scientist with a rare specialty could unlock a training result that ten application engineers cannot produce. Job architecture is not administrative cleanup. It is capital allocation.

Candidates benefit from the same discipline. A software engineer should not accept a research title that carries no research environment, publication policy, compute allocation, or scientific manager. A research scientist should not be judged on feature velocity when the actual mandate is uncertain exploration. An AI engineer should not inherit production risk without an on-call model, observability budget, and authority to reject an unreliable release.

Level matching matters as much as title matching. P6 in one company may mean a staff engineer who leads work across teams. Elsewhere it may map closer to principal. Manager scope varies with team size, technical depth, and hiring authority. A clean benchmark can become false precision when a recruiter maps the internal role to the most convenient external level.

Before pricing, write three sentences:

1. Name the business or research output expected in the next twelve months.
2. Name the system, team, budget, and decision rights the person will control.
3. Name the evidence that would show the work requires research, model-lifecycle engineering, or application integration.

If the sentences describe three jobs, redesign the job. The compensation range will be easier to defend once the work is real.

## Small startups spend their flexible currency

Cash is visible every month. Equity is where a startup can make a bet it cannot fund from revenue.

[Carta's H2 2025 startup compensation report](https://carta.com/data/startup-compensation-h2-2025/) shows that the initial equity offered to AI and ML engineers rose 31% from January 2024 through February 2026, while salary rose 9.1%. Among companies valued from $1 million to $10 million, initial AI and ML engineering equity increased 64% over two years. The source is Carta's private-market customer data, and its board-share records can lag. It nevertheless points in the same direction as Pave: young companies are using ownership more aggressively than cash.

Sometimes the strategy is rational. A founder preserves runway, shares upside, and recruits someone whose work may materially change the company's value. The candidate accepts more variance in exchange for participation in that outcome. Both sides can prefer the trade.

Bad offers can hide inside the same trade. Percentage ownership without a fully diluted denominator is incomplete. A dollar value based on the latest preferred round can overstate what common shares would receive in a sale. A short exercise window can turn vested options into a tax and cash decision after departure. Future fundraising can dilute the stake. A tender offer may never arrive.

Cash and equity should appear on separate lines in the offer model. So should ownership percentage and estimated dollar value. Finance can then test at least three scenarios: the latest valuation holds, the next round dilutes existing holders, and common equity finishes below the preferred price. A candidate can make a personal risk decision without treating a private-company mark as a bank balance.

Personal risk changes the answer. A candidate paying a mortgage, supporting relatives, managing an immigration deadline, or leaving already vested stock may prefer more cash and less theoretical upside. Another person may be able to hold illiquid options for a decade. Treating the second candidate as more committed rewards financial capacity rather than job fit. A defensible offer process permits a bounded cash-equity choice and records its exchange rate.

New-hire grants create an internal problem before they create external wealth. Pave reports that recent research hires can receive median base salaries up to 10% above incumbents at the same level. A person who built the research environment may watch a new colleague enter with more salary and a grant priced for the current race. If the company treats that difference as confidential market reality, the incumbent learns it from the next recruiter.

Turnover makes the gap expensive. In Pave's customer data, 12-month turnover was 21.8% for AI and ML individual contributors and 22.6% for managers. Software-engineering turnover at the same companies was 16.8% for individual contributors and 18.9% for managers. These totals include voluntary and involuntary departures, so they do not measure poaching alone. They do show that a new-hire budget without an incumbent review is unfinished.

Reviewing incumbents does not mean matching every new grant. It means answering why a gap exists. Scarcity may have changed. The new person may bring a different specialty or level. Earlier employees may already own more shares because they joined at a lower valuation. Performance and scope may differ. Once those facts are recorded, the company can decide among a base adjustment, refresh grant, promotion, retention grant, expanded scope, or no change.

Silence is also a decision. It transfers the explanation to rumor.

For a 20-person startup, the trade-off is sharp. Ten million dollars of option-pool value can fund a frontier researcher, several ML engineers, a product team, or no one while the company buys model access. A large laboratory may absorb an outlier grant across a broad capital base. The smaller company may reshape everyone else's ownership to make one hire.

Run the dilution in shares, not only dollars. Show the board the option-pool increase, the stake after the next financing, the cash runway under each hiring plan, and the milestone each plan is expected to reach. If a researcher is supposed to create a proprietary model advantage, state the technical evidence and deadline. If an AI engineer can reach the same customer result with an external model, price that route too.

Equity is flexible currency. It is not free currency.

## Prompting earns a two-percent premium

Set the $4.09 million research grant aside for a moment. A German employer experiment offers a useful counterweight.

In the August 2026 IZA discussion paper [The Value of Prompting Skills](https://www.iza.org/index.php/en/publications/dp/18871/the-value-of-prompting-skills), Lara Fleck and five co-authors embedded a preregistered choice experiment in the BIBB Training Panel. The working sample included 992 firms and 15,660 repeated candidate-choice observations. Employers compared hypothetical applicants with completed vocational training, relevant experience, and five years of work. Candidate profiles varied prompting skill, occupational skill, social skill, gender, and salary expectation.

High prompting skill raised the probability of being selected by about four percentage points. Employers' estimated willingness to pay was about 2% above the salary of an average skilled worker. Among large firms and firms already using or planning to use AI, the selection increase was about nine points.

Two percent is not $4.09 million. It is not meant to be.

Each study measures a different object. The IZA paper isolates one attribute in a broad skilled-worker hiring decision. Pave observes compensation attached to job families, levels, companies, and grants. PwC's [2026 AI Jobs Barometer](https://www.pwc.com/gx/en/news-room/press-releases/2026/pwc-2026-ai-jobs-barometer.html) reports a 62% wage premium for jobs requiring AI skills, using a much broader job-ad classification.

| Reported premium | Unit being compared | Evidence type | Claim it cannot support |
| --- | --- | --- | --- |
| About 2% | High prompting skill within a hypothetical skilled-worker profile | German employer choice experiment | Market pay for an AI engineer or researcher |
| 62% | Jobs requiring AI skills versus jobs without that classification | Job-ad analysis across occupations | Causal value of adding one AI skill to one worker |
| $4.09M at P6 | Annualized intended new-hire research equity in the Pave dataset | Compensation-system benchmark | Liquid cash value or company-wide AI premium |

Prompting skill also failed to erase an occupation-specific skill gap. Strong social skill was worth more to employers in the experiment. An experienced machinist, nurse, accountant, or sales operator who learns to prompt has a different value proposition from a candidate who can prompt but cannot perform the occupation.

Training and pay design should reflect the difference. A company rolling out generative AI to 2,000 employees does not need to create 2,000 premium AI titles. It can define proficiency inside existing jobs: safe tool use, task decomposition, verification, domain judgment, and escalation. A modest skill differential, completion payment, or progression step may fit better than a separate salary band.

Specialist roles need a different test. An AI engineer accountable for evaluation infrastructure, model routing, and a production incident owns more than prompt technique. An ML engineer responsible for training or inference performance owns a technical system. A research scientist owns an uncertain program whose result may change model economics. Paying all three for "AI fluency" flattens responsibility while pretending to recognize it.

Employees see the flattening quickly. A customer-support lead may redesign a workflow and train peers while a new "AI specialist" receives a premium for using the same tool. A software engineer may become the unofficial reviewer for model failures without a title or on-call allocation. The company celebrates broad adoption, but the burden falls on people whose job records never changed.

Build ordinary AI proficiency into the career ladder. Pay specialist accountability through the relevant engineering, product, risk, or research family. Reserve exceptional equity for a documented exceptional dependency. That is less dramatic than declaring every AI-skilled worker scarce. It is also more durable.

## Salary ranges leave equity off the page

A $300,000 salary range can be precise and incomplete. The largest part of the offer may sit elsewhere.

California's [Equal Pay Act guidance](https://www.dir.ca.gov/dlse/California_Equal_Pay_Act.htm) says employers with 15 or more employees must include the salary or hourly wage range they reasonably expect to pay in a job posting. The state's FAQ also says bonuses, tips, and other benefits do not have to be included, although an employer may add them. This is a statement of the rule, not legal advice about a specific posting.

For senior AI research, the required pay scale can reveal the compressed line while leaving the differentiating line outside the frame. A posting may show $280,000 to $360,000 in salary. One candidate could receive a routine grant and another a grant whose annualized intended value exceeds $4 million. Both offers can fit the visible range.

Voluntary disclosure helps only when the units are intelligible. "Competitive equity" tells a candidate nothing. A dollar estimate without the share count, ownership basis, vesting, exercise terms, and valuation basis can look precise while remaining impossible to evaluate. A percentage without expected dilution or the company's capital structure is better, but incomplete.

LinkedIn's August [AI Talent Divide report](https://delivery-p143253-e1476319.adobeaemcloud.com/adobe/assets/urn:aaid:aem:81d35729-ba25-4ecd-9365-146d22005a2b/original/as/The_AI_Talent_Divide_2026.pdf) puts the typical listed salary for US AI job postings at $177,000, compared with $80,000 for non-AI postings. The calculation uses annualized midpoints for postings with pay information. It describes visible salary, not realized total compensation.

Women represented 26% of new AI hires in the same report, compared with 50% of non-AI hires. LinkedIn derives hiring and demographic measures from member profiles and its occupation taxonomy. The numbers do not show that hidden equity caused the gap. They do raise a distribution question: who gets access to jobs where the most negotiable and consequential value may sit outside the posting?

Access begins before negotiation. Research roles often require graduate credentials, visible publications, elite referrals, or experience inside a small set of laboratories. LinkedIn reports that 91% of AI roles require a bachelor's degree or more, while its most senior research-heavy categories lean further toward graduate degrees. An employer should separate requirements imposed by the work from signals inherited from the existing network. Otherwise a scarce pipeline and an expensive offer can reinforce each other.

Negotiation amplifies information differences. A candidate moving between well-funded labs may have an agent, competing offers, and detailed knowledge of current grants. An internal engineer shifting into AI work may be compared with an old salary band. A first-time startup employee may focus on the headline dollar value and never see the fully diluted denominator. A worker returning after a career break may anchor to the posted range while the employer reserves equity flexibility for candidates who demand it.

Compensation teams can measure this. For every job family and level, compare proposed and accepted base pay, initial equity, refresh equity, sign-on payments, and offer exceptions by candidate source and relevant demographic groups. Separate legitimate factors such as level, location, experience, and competing offer. Review who received an exception, who asked and was denied, who withdrew, and who was never told flexibility existed.

Do the same for incumbents. A pay-equity analysis limited to salary can miss the instrument carrying the market premium. A total-rewards analysis should value grants under a consistent method, retain both share and percentage data, and avoid treating an uncertain private mark as realized income. The purpose is not to force identical grants. It is to make differences visible enough to explain and correct.

Candidates also deserve time. Deliver the cap-table and option explanation before the acceptance deadline. Identify which values are contractual and which are estimates. State the latest preferred price, the common-share or 409A value when applicable, exercise cost, tax uncertainty, and known liquidity limits. Invite the candidate to obtain independent financial or tax advice.

Transparency does not make a risky grant safe. It lets the risk belong to the person asked to take it.

## A compensation file for AI work

Annual surveys set reference points. Offers require a decision file.

Connect the work, benchmark, instrument, internal comparison, and business milestone. Keep the file short enough for a hiring manager to use and specific enough for finance, compensation, and the board to audit later.

| Decision-file field | Record before approval | Failure exposed later |
| --- | --- | --- |
| Company lane | Innovator, Integrator, Implementer, or a named mix | A fashionable title substitutes for strategy |
| Work output | Model result, lifecycle system, product capability, or internal workflow | Candidate and manager optimize for different outcomes |
| Role family and level | Internal family, external benchmark mapping, scope, and alternatives considered | P6 data prices a role that is not actually P6 |
| Operating environment | Compute, data, team, manager, publication rights, on-call duty, and decision authority | Pay tries to compensate for a job the company cannot support |
| Base benchmark | Dataset, date, geography, company stage, percentile, and range | A vendor median becomes a universal market rate |
| Equity instrument | Shares, fully diluted percentage, strike or reference price, vesting, exercise, and liquidity | Headline dollars conceal risk and dilution |
| Annualized estimate | Valuation basis, grant life, gross or net treatment, and downside scenarios | Intended value is mistaken for cash received |
| Internal comparison | Incumbents, recent hires, adjacent roles, and unexplained gaps | The new offer creates a preventable retention problem |
| Access review | Offer exceptions, candidate source, representation, withdrawals, and negotiation outcomes | Flexibility reaches only candidates with better information |
| Business milestone | Result, date, accountable executive, and cheaper route | Scarcity becomes an open-ended reason to overspend |
| Refresh trigger | Performance, retention risk, market movement, promotion, or financing event | Every external offer becomes an emergency grant |
| Review date | 90-day role check, six-month band check, and annual portfolio review | A fast market turns one exception into permanent architecture |

Apply it to three companies and the answer changes.

An Innovator might need a research scientist who can improve pretraining efficiency or solve a model-safety problem. Specify compute access, research peers, experimental authority, and the technical milestone. A large equity grant may be defensible if the person can materially change company value and the board understands the dilution. Paying the grant without the environment buys a departure risk.

For an Integrator, an ML engineer might run evaluation, fine-tuning, serving, and monitoring around proprietary data. Reliability and unit economics can be more important than publications. Base pay, operating scope, and a normal senior-engineering equity curve may fit better than a research outlier. If the product advantage depends on a narrow method, add research capacity deliberately instead of hiding it inside the engineering title.

Implementation calls for another mix: strong software engineers, process owners, security review, and broad employee training. The company can buy models rather than invent them. Its scarce capability may be domain redesign or manager adoption, not frontier research. A $4 million annualized grant would be a poor substitute for a clear workflow, data access, and accountable product owner.

One business can move between lanes. Record the lane for the funded work, not as a permanent corporate identity. Revisit it when the model strategy, capital base, or customer promise changes.

No-hire decisions belong in the file as well. If an external model plus two engineers reaches the milestone sooner, write that down. If the company lacks enough proprietary data to support a research program, wait. If the candidate's value depends on a team that has not been recruited, sequence the team. Restraint should be as legible as approval.

Managers need a line in the file because retention happens in the work environment. Luke Metz's move cannot tell an outsider whether pay, compute, colleagues, leadership, or research direction decided the outcome. A company that answers every departure with a larger grant may retain someone for the wrong job, or lose them anyway. The manager should document the scientific or engineering conditions the candidate asked for and which ones the company can sustain.

This is the strongest case against treating compensation as the center of the talent war. A top researcher may rationally accept less equity to work with a particular collaborator, pursue a method without product pressure, publish, or obtain reliable compute. An engineer may choose the manager who can make decisions and protect technical standards. Those preferences are not soft benefits around the offer. For some jobs, they are the production system. The decision file includes them so that finance does not ask ownership to repair an operating environment.

Employees should be able to see the part that concerns them. They do not need another person's private offer. They do need the family, level, range philosophy, equity method, refresh rules, and route to challenge a mismatch. If recent-hire pressure changes the range, review the people already doing the work before the next recruiter calls them.

## Monday's offer outruns the benchmark

On Monday morning, a compensation partner can open the August report beside an offer request. The hiring manager wants a research title. The job description wants production ownership. Finance has modeled salary but not option-pool expansion. The candidate has another process moving and asks for the fully diluted percentage.

Choosing the research median and adding urgency would be fast. A better response takes an hour.

Rename the output the company needs. Map it to research, ML engineering, AI engineering, or an explicit combination. Check whether the environment can support that work. Put salary and equity on separate lines. Convert the grant into shares, ownership, annualized intended value, and downside cases. Compare the offer with incumbents. Record the next business milestone and the date the architecture will be reviewed.

Then show the candidate the instrument rather than only its headline value.

Pave's dataset will change. Carta's private-market records will change. Another prominent researcher will move, and a new package will become the number repeated in hiring meetings. None of that makes the August benchmark useless. It makes the timestamp part of the evidence.

Internal facts move more slowly. The company knows whether it builds models, integrates them, or uses them. It knows which output the next hire owns. It knows the option pool, runway, compute budget, internal grants, and managers capable of supporting the work. Those facts should decide how much of the external market it follows.

By the time the offer is approved, the $4.09 million median should have done two jobs. It should reveal that senior research ownership operates in a different market from base salary. It should also force the company to prove that it is buying research.

If the proof fails, redesign the role. Save a research grant for work the company can actually support.

## Related Reading

- [AI Talent Raids Reach the Retention Budget](/2026/07/09/ai-talent-raids-retention-budget/)
- [AI Skill Premiums Put Pay Bands on Trial](/2026/06/19/ai-skill-premiums-pay-bands/)
- [AI Training Work Splits the Pay Band](/2026/06/24/ai-training-work-pay-band/)
- [Small AI Teams Still Carry a Talent Bill](/2026/06/25/small-ai-teams-talent-bill/)