Biotech Is Losing the AI Talent War. Money Isn’t Why.

_{Senior-level base salary ranges, July 2026, extended to senior-level median total compensation where reported. Sources and leveling notes in methodology.}

The numbers hold. These are senior-level individual-contributor bands, with multi-level postings cut down to their senior band. The pharma postings ask for a PhD plus two to five years. The labs mostly ask for four or five years of experience and treat the degree as optional. Anthropic will pay $300,000 to $320,000 in base salary for a research scientist on its life sciences team, right in line with what its senior engineers report company-wide. OpenAI’s senior engineers report a base of $280,000 to $340,000. Pfizer’s posting for a senior machine learning research scientist, at the same PhD-plus-two leveling, tops out at $166,500. Takeda’s tops out at $215,270. One pharma group has closed the base gap: Genentech’s AI-for-drug-discovery unit posts its senior band to $297,300, squarely inside both labs’ ranges. The rest of pharma is not close. And base is not where the game is decided.

If base were the whole package, the best of pharma would be competitive. It is not the whole package. On Levels.fyi, the median senior-level package runs $563,000 at Anthropic and $681,000 at OpenAI, and $841,000 to $937,000 one level up, with equity making up roughly half to two-thirds of the total. Pharma grants stock too, but at the individual-contributor level it runs zero to about twelve percent of a package a third the size. The pharma grants are at least liquid public stock while the labs sell private paper. No engineer weighs $25,000 of Roche against $250,000 of anything. A biotech that heroically matches base salary is still offering half the package. The bidding war is unwinnable on cash, and it is even less winnable on equity.

So the interesting question is not why biotech loses on salary. It is why biotech loses even when the salary gets close. Look at what the engineer is actually choosing between, because it is not two numbers.

At a lab, the engineer owns a system end to end, sits among hundreds of peers who speak their language, reports to someone who can judge their work, and watches their model ship in weeks. At most biotechs, the same person is a service function: tickets in, dashboards out, one of three ML people in a building of biologists, reporting to a scientist who cannot evaluate what they built, waiting a quarter for the wet lab to return an answer. Salary is the visible half of the offer. The job is the other half, and biotech routinely loses both.

I lived through this once already, the last time biology tried to hire its way into a new technology. Ten years ago, in the deep learning era, I watched it from the inside at Exscientia, Insitro, Freenome, Tempus, and many others. There weren’t enough people trained in both biology and machine learning, so the early teams recruited their friends, ran out of friends, then went hunting at Google, Facebook, Amazon, and Uber for raw ML talent. Some of the best engineers I ever saw came out of those places. Most of the companies that hired them never made it work. The biology was foreign, the problems too abstract, and the two sides spoke different languages. Putting them in the same building did not make them speak the same one.

The companies that got it right did three unglamorous things. They are as actionable today as they were then.

First, they put a dual-trained person in charge of the group. Not a biologist with an ML advisor, not an engineer with a scientific liaison, but one person who could hear a modeling result and a biology result and judge both. Hire that person before you hire the engineers. They are rarer than the engineers, and they are also the magnet: the best people follow a leader who can judge their work, and once the magnet is in place, everyone else gets easier to hire and easier to keep.

Second, they designed roles an engineer could own end to end without a PhD’s worth of biology. Write the role before you open the req: name the system the engineer will own, and name the interface where biology hands them ground truth. If you cannot name either, you are not hiring a scientist. You are staffing a service desk, and the engineer will know it by the second week. Then keep the owner close to the top: every layer between the engineer and the person who sets the science subtracts from the ownership you just promised.

Third, they built a culture where biologists and engineers learned from each other and were treated as equals: shared authorship, shared credit when a model kills a bad program, one seminar series instead of two. It broke the moment one side won. When biology dominated, the engineers left. When ML dominated, the science got sloppy.

Here is the part I did not expect when we pulled this year’s data. You can read job design straight off the postings, and it tracks pay almost perfectly.

_{Role-design markers read from the language of live postings, July 6, 2026. “—” means not stated in the posting.}

Genentech’s AI-for-drug-discovery group posts machine-learning scientist roles that read like an AI lab wrote them. No biology degree required: the posting invites ​“anything from a BS+7 to PhD+2” and gates on a first-author ML publication instead. Ownership language throughout, seated inside research, with a principal band that tops out at $344,600. GSK runs its group under its own banner, gsk​.ai, and posts senior AI engineers to $266,750 on the same logic. Anthropic went one step further: its life-sciences posting requires drug-discovery experience outright. The dual-fluent role, designed on purpose, at a dual-fluent price.

Takeda’s postings ask for a PhD in computational biology or bioinformatics and describe the mandate as deploying systems that ​“augment scientist productivity.” The AI hire builds tools for the scientists. They are not one of the scientists. The band tops out at $215,270. Eli Lilly’s flagship AI hiring page sits inside IT. Pfizer’s senior ML research scientist band starts at $99,900.

I don’t think that correlation is an accident, and I don’t think it’s simply that richer companies write better postings. The design of the role determines what the organization believes the role is worth. A company that frames AI as internal support benchmarks it against IT, pays accordingly, and posts accordingly, in public, where every candidate can read it. A posting that loses on money and on scope loses twice.

For a startup, the list is shorter and the odds are better. Start with the equity itself. The labs now pay theirs the way big tech does, because that is what they have become: the grants are enormous in dollar terms, but they are late-stage paper in companies already valued in the hundreds of billions. Taking a lab offer in 2026 is taking a FAANG offer in 2015. Richly paid, and the upside already spent. A startup holds the only equity left with real convexity, and it can pair it with what no lab or pharma can match. It can seat the engineer in the room where the science happens, because at ten people there is no IT department for a role to sink into. It can offer a feedback loop where the model redesigns next week’s experiment instead of waiting on a quarterly committee. And it can make someone employee eight instead of headcount. What a startup cannot afford is the standard mistake of hiring ML talent before there is a dual-fluent leader to report to and a designed role to own. An engineer hired into an undesigned role at a startup doesn’t stay longer than at Pfizer. They just leave faster.

Biotech has always had one advantage the paycheck doesn’t capture. The work matters. Steve Jobs recruited John Sculley by asking whether he wanted to sell sugar water for the rest of his life or come change the world. Curing disease is the same pitch, and it lands even harder. But look at who is making that pitch now. Anthropic’s life-sciences posting promises work on ​“problems that matter for human health.” The Chan Zuckerberg Biohub recruits ML engineers toward ​“the ultimate goal of curing disease.” The labs are selling biotech’s own mission back to the talent, and pairing it with lab-scale pay. Mission is no longer a moat. It only lands when the job is designed so the engineer feels that impact every day, owns something real, and sits at the table as an equal.

One more thing changed this year, and it is the most hopeful data point in this piece. Computer science enrollment at US four-year universities fell 8 percent in 2025–26, the first decline in a decade, with CS proper down 11 percent. Students read the same headlines we do: AI is eating entry-level coding. Health and biology fields grew. The generation entering college right now is choosing biology, and it is the first generation that will treat AI the way mine treated pipettes, as standard equipment.

Which means the dual-fluent person, the one every AI-bio company in the deep learning era ran out of friends trying to find, is finally in the pipeline. In five years they start graduating in volume. They will not interview at companies where the AI role reports through IT, and they will not need a translator, because they are the translator. The scarcity was always temporary. The design problem is the part you control, and the companies that fix it now are the ones standing where that generation lands.

You cannot outbid the labs. You can out-design them.

Methodology & sources

_{Base-salary ranges are base pay only from live postings fetched July 6, 2026 via Greenhouse, Ashby, and company career sites, normalized to senior-level individual-contributor bands: Anthropic (Research Scientist, Life Sciences), OpenAI (Researcher, Health AI; the posting publishes one $295K–$445K band spanning mid-to-senior IC and asks for 4+ years of deep-learning/LLM research experience with a ​“Ph.D. or other degree,” so a doctorate is not a hard requirement; the chart therefore shows OpenAI’s reported senior-level base of roughly $280K–$340K from Levels.fyi rather than the posted band), Altos Labs (Senior Scientist I band), Genentech (Senior ML Scientist band, AI for Drug Discovery; the Principal band reaches $344,600), GSK (Senior Applied AI Engineer, via job-board mirror of a live req), insitro (Senior ML Scientist, Imaging), Takeda (Research Senior Scientist, AI/ML; PhD + 2 yrs), and Pfizer (Senior ML Research Scientist; PhD + 2 yrs; via job-board mirror; carries a 15% bonus target). CZ Biohub was excluded because its single posted range ($150K–$350K+) spans all levels; Eli Lilly, AstraZeneca, Bayer, and Neurocrine were reviewed but no posted US salary range could be verified. Total-compensation figures are medians of self-reported US packages at the stated levels for company-wide engineering ladders (the labs do not break out team-level compensation, so life-sciences and health roles are assumed to sit on the same bands); titles differ by company, and pharma IC packages include modest RSU/LTI grants (typically 0 to 12% of total; Genentech E5A about $23K/​yr, Lilly principal about $3K/​yr, Merck ICs below principal report none). Lab extensions on the base-salary chart end at senior-level median total compensation from Levels.fyi per-level pages retrieved July 8, 2026: OpenAI L4 at $681K ($279K base plus $402K/​yr stock) and Anthropic Senior at $563K ($316K plus $247K/​yr); one level up, OpenAI L5 reports $937K and Anthropic Lead $841K. Pharma extensions add reported annual LTI to the posted top (Genentech E5A about $23K/​yr, Levels.fyi; Pfizer about $30K/​yr, CompBioJobs equity guide). Self-reported senior packages are widely dispersed (p10–p90 spans approach $1M at the top levels), so lab figures are directional. Altos Labs, insitro, GSK, and Takeda offer equity or LTI without disclosed values and are marked accordingly. Lab equity is illiquid private paper. Several figures rest on small samples and are illustrative. Data source: Levels.fyi (https://www.levels.fyi). Job-design markers are read from the language of the same postings; a dash means the posting does not address the question. Enrollment figures are from National Student Clearinghouse data for the 2025–26 academic year as reported by the Washington Post and EdSource (computer and information sciences down 8.1% at four-year institutions; computer science down 11.2%; health fields up roughly 3% at four-year and 10% at two-year institutions).}