Periodic Labs posts 21 AI‑science roles with $275k median pay

The Architecture of a New Institution

Periodic Labs has 21 open roles on its board as of late July, Startup Jobs' data shows, six posted in the past 30 days, JobsRadar's data shows, concentrated in a Menlo Park facility built for on-site work. The company, founded in 2025 by former OpenAI post-training VP Liam Fedus on a $300 million seed round, Fast AI Jobs reported, is not staffing a typical software team. It is staffing a new kind of institution — one where the bottleneck lives in experimental throughput, not model architecture.

Fourteen of the 21 openings sit in Science: research scientists in condensed matter theory, thin films, materials characterization, and data; research associates in thin films; a computational scientist in structural and thermal analysis. Engineering holds six roles: software engineers, an ML systems engineer, a product engineer, two forward-deployed engineers (quantum simulations and physics & simulation), and a research engineer focused on lab automation. Business & Operations covers three: a procurement and finance operations manager, a founding HR leader, and an environmental health and safety specialist. Contract roles account for four positions, including a laboratory technician on a contract-to-hire path and a nanofabrication intern for summer 2026. One remote role exists; the other 18 require presence in Menlo Park.

First-party board data from Zero G Talent confirms the velocity. An ML Systems Engineer role at $250,000–$350,000 (Zero G Talent's figures put the top of the band at $350,000) appeared in the past week, alongside a Founding HR Leader and a Product Engineer at the same band, a Research Scientist in Condensed Matter Theory at $225,000–$325,000, and a Forward Deployed Engineer in Quantum Simulations at $200,000–$275,000. The board's median salaried band sits at $275,000 Zero G Talent found across 13 roles according to Zero G Talent with disclosed pay. JobsRadar's July 30 snapshot shows similar bands clustering between $225,000 and $350,000 for senior technical roles, with contract and intern tiers ranging from $100,000 to $180,000.

Fedus has described the split as "bits and atoms" — mid-training and pre-training roles on the AI side, control engineering and system engineering on the physical side, with product engineering spanning the gap. The org chart bears that out: the Science department alone outnumbers Engineering and Operations combined. That ratio is unusual for a company with Fedus's pedigree; it signals that the bottleneck Periodic Labs is attacking remains throughput, not architecture. The lab automation role, the process development engineer, the powder process engineer — these are not support functions. They are the interface where the AI scientist meets the fume hood.

Six roles opened in the last month. The first-party board shows fresh postings for founding-level HR and product engineering, suggesting the company is moving from a research collective toward a scalable operation. The presence of a Procurement & Finance Operations Manager and an EHS specialist indicates regulatory and supply-chain maturity that most seed-stage AI labs skip. Periodic Labs is not skipping them. The hiring plan reads like a facility commissioning schedule: technicians first, then process engineers, then the scientists who will run the closed loop Fedus has described — simulation, experiment, data, next experiment.

That loop is the hiring thesis. Every role on the board serves it. The question is whether the talent market can supply people who have already lived inside that loop.

What the Pay Bands Reveal

Periodic Labs' public salary bands tell a clear story: the company pays like a well-capitalized growth-stage startup, not a frontier lab. First-party board data shows six roles posted in the past week with base-plus-equity ranges clustering between $200K and $350K. The ML Systems Engineer, Founding HR Leader, and Product Engineer roles each list $250K–$350K. The condensed matter theory role runs $225K–$325K. The quantum simulations role sits at $200K–$275K. The recruiter role spans $175K–$275K. Across 13 salaried roles on the board, the median lands at $275K with a band of $113K–$350K. A separate scrape of 17 public postings from 2025–2026 puts the median at $215K ($190K–$225K). Both figures point to the same neighborhood: solid, competitive, but a tier below the $900K median that OpenAI, Anthropic, and DeepMind now post for senior individual contributors.

Three variables drive the spread: valuation, equity liquidity, and retention pressure. Frontier labs command premium valuations that justify massive equity grants, but those grants stay on paper until a tender offer or IPO. OpenAI's profit participation units (PPUs) vest like stock yet settle on a different schedule; recent hires have not seen a full tender cycle. Anthropic equity remains illiquid until acquisition or public listing. Nvidia sits in its own category. Tenured engineers from 2020–2022 have watched RSU value grow four to five times, making 78 percent of them paper millionaires. Meta's SI Labs has written the largest individual packages on record, with Alexandr Wang's deal reported above $100M.

Periodic Labs operates without that liquidity lever. Its equity is a bet on a future exit, not a tradable asset. That makes the cash component matter more. The board data shows the company leaning into base-heavy offers with meaningful equity upside, a rational structure for a private company still building its valuation narrative. Candidates comparing a $350K Periodic Labs offer against a $1.2M Anthropic package face a real choice: the Anthropic number assumes a liquidity event that may not arrive on schedule, while the Periodic Labs number is spendable today. Risk-averse engineers increasingly treat a $700K Google offer as more valuable than a $1.2M frontier lab offer for exactly this reason.

Geography sharpens the contrast. Frontier labs now pay location-agnostic compensation benchmarked to San Francisco for senior hires, meaning a senior engineer at Anthropic earning $850K in San Francisco often earns the same in Seattle or Austin. Big tech still applies geographic adjustments of 10–25 percent for moves away from SF, Seattle, or New York. International offices at Anthropic London, DeepMind London, and OpenAI Dublin pay 25–40 percent less on a USD basis. Periodic Labs' roles list Menlo Park or "United States" without a stated geo differential, suggesting the company has not yet formalized a remote-comp policy. That will matter as the 21 open roles attract applicants outside the Bay Area.

Microsoft has been the most aggressive big-tech responder, handing level 67–69 AI hires multi-year retention packages that narrow the gap with frontier labs. Apple, after losing senior AI talent to OpenAI and Anthropic through 2025, is rapidly improving its cash and equity offers. The trade-off remains: Microsoft equity vests in liquid MSFT shares; OpenAI PPUs require a tender offer. Periodic Labs offers neither. Its equity is a long-dated call option on a company still proving its commercial model. For the specialized talent the screening process targets (production LLM, distributed training, inference optimization), the market continues to climb. Commodity AI work is getting cheaper. Periodic Labs is not hiring for commodity work.

The Screen: Where AI Meets the Fume Hood

The role breakdown Periodic Labs published — 18 positions across six domains — functions as a de facto syllabus for what the screen evaluates. Five thin-film and novel synthesis roles ask for candidates who can translate AI-predicted structures into physical samples, including "previously unrealized compounds." Four autonomous lab infrastructure openings demand fluency with robotic synthesis, characterization instrumentation, and the control software that stitches them together. Three frontier LLM training roles target researchers who can adapt large-model architectures to scientific reasoning, not chat. Two lab software and data systems positions require building evaluation frameworks and training datasets from experimental output. Three supercompute and infrastructure slots call for scaling distributed training across GPU clusters while managing the I/O burst of instruments generating gigabytes per experiment. One product and scientist tooling role bridges internal researchers and external customers already deploying custom agents on semiconductor heat-dissipation data.

The team assembled before launch signals the bar. Costa Huang, creator of CleanRL, and Vincent Moens, creator of TorchRL, both joined to build reinforcement-learning infrastructure where the reward signal comes from physical experiments — not simulated environments. Rishabh Agarwal, a NeurIPS best-paper winner for offline RL, sits alongside them. Dzmitry Bahdanau, co-inventor of the neural attention mechanism that underpins every transformer, lists his focus as "move bits to construct new things from atoms." Screening for the LLM training roles therefore weighs published contributions to RL algorithm design, offline policy evaluation, and transformer architecture modifications against the specific constraint: the environment is a wet lab with latency, noise, and safety interlocks.

Materials science depth is non-negotiable. The condensed matter theory role explicitly lists superconductivity and magnetism as bonus qualifications. Matthew Horton brought Microsoft Research's MatterGen and the Berkeley Lab Materials Project; Muratahan Aykol arrived from DeepMind, Toyota Research Institute, and Rivian with cathode and battery materials expertise; Eric Toberer holds a professorship at Colorado School of Mines. Candidates for theory and thin-film roles face scrutiny on density-functional theory workflows, high-throughput computational screening, and the ability to judge whether an AI-proposed structure is synthesizable — not just stable on paper. The research scientist data role description makes this concrete: translating scientific workflows into benchmarks and RL environments, building datasets that improve scientific reasoning, creating feedback loops connecting applications, evaluation, and training.

Lab automation experience carries equal weight. Naveen Menon ran cathode manufacturing pilot lines at Tesla. Sam Cross served as senior principal engineer for lab automation at Lila Sciences and Samsung. Nicholas Bergantz brings 20 years in robotics. Jun Feng directed PECVD semiconductor process lines at Applied Materials. The autonomous lab infrastructure screen tests whether a candidate has commissioned, debugged, and maintained closed-loop hardware — not written simulation scripts. The company's own timeline underscores the gap: as of October 2025, TechCrunch reported the lab was set up and processing experimental data, but the robots were not yet running. Five months later, the central diligence question remains whether the system has moved from human-assisted experiments to genuine autonomous synthesis, characterization, and model feedback. Candidates who have only worked in L1–L2 autonomy labs (the majority, per a 2024 community survey) face a steeper climb.

Data quality awareness separates applicants who have wrestled with real instruments from those who have trained on public corpora. Periodic's leadership has stated that published literature carries noise floors too high for effective training; npj Computational Materials reviews cite noisy datasets, missing metadata, poor reproducibility, and absent standardized formats. The research scientist data role centers on this problem: developing evaluation strategies, identifying capability gaps, building datasets that improve model performance. Screening probes whether a candidate has designed data pipelines that capture provenance, instrument state, and negative results, the last of which are seldom published but critical for RL reward shaping.

Supercompute and infrastructure roles add a distinct constraint: the training loop must ingest streaming instrument data while checkpointing models that control the next experiment. The board's first-party data shows that role at $250k–$350k, reflecting the premium for engineers who can co-design distributed training and real-time lab orchestration. The quantum simulations position at $200k–$275k signals demand for translating these capabilities into customer environments. The semiconductor manufacturer using custom agents for heat-dissipation iteration is a live deployment, not a pilot.

The screen does not reward breadth alone. A candidate with transformer expertise but no materials synthesis judgment will not clear the thin-film panel. A robotics engineer who has never designed an evaluation benchmark for scientific reasoning will not clear the data systems panel. The company's advisory board — Carolyn Bertozzi, Mercouri Kanatzidis, Steve Kivelson, Zhi-Xun Shen, Chris Wolverton — represents the external validation layer: their domains define the technical depth the screen measures. The filter is simple: Periodic is hiring for the intersection that barely exists in the market — researchers who have closed the loop in a physical lab, published in RL or transformer architecture, and can defend a materials synthesis pathway under cross-examination from a Nobel laureate.

Culture as Filter: Bits, Atoms, and the Verification Gap

Periodic Labs organizes its hiring around a "bits and atoms" taxonomy that functions as both an organizational chart and a cultural filter. The "bits" side seeks mid-training and pre-training AI researchers plus infrastructure engineers. The "atoms" side recruits control engineers, systems engineers, and project engineers. This split is not administrative convenience — it reflects the company's founding premise that intelligence alone is insufficient; new knowledge only emerges when ideas confront physical reality. Candidates who cannot articulate how their work crosses that boundary rarely advance.

The evaluation starts with the role definition itself. That role is not hired to publish papers; they are hired to generate the priors and verification loops that let an AI scientist propose and test hypotheses in a closed autonomous lab. A Forward Deployed Engineer for Quantum Simulations is not a traditional solutions engineer; they sit inside a semiconductor manufacturer's workflow, training custom agents that ingest experimental data and accelerate iteration. Every open role maps to a specific node in the bits-to-atoms pipeline: data generation, model orchestration, simulation fidelity, or physical actuation. Interviewers probe whether the candidate has operated at that node before, or at least understands its failure modes.

Fedus, co-creator of ChatGPT and former VP of post-training at OpenAI, frames the cultural benchmark explicitly: the company does not build foundational language or coding models from scratch. It uses existing frontier models, citing Claude Code by name, as an orchestration layer, then concentrates all internal ML effort on domains where open-source capabilities fall short. That strategic choice shapes the type of researcher they want. They screen for people comfortable outsourcing the "commodity" layers of the stack and obsessive about the narrow, high-leverage problems where Periodic's proprietary data and closed-loop architecture create defensibility. A candidate who wants to pre-train a 70B-parameter LLM from scratch signals misalignment; a candidate who wants to design the active learning loop that directs the next synthesis experiment signals fit.

The "atoms" evaluation is equally specific. Fedus has noted that physical lab setup involves long lead times and carefully calibrated systems, a different bottleneck class than digital infrastructure. Control engineers and systems engineers are assessed on whether they have shipped hardware that runs reliably for months without human intervention. Project engineers are tested on their ability to coordinate across the simulation-to-synthesis gap: translating a neural network's proposed experiment into a robotically executable protocol, then feeding the resulting gigabytes of spectral data back into the model. The company currently uses commoditized, off-the-shelf robotics supplemented by human operators; candidates who over-index on advanced robotics R&D and under-index on making existing automation reliable tend to miss the near-term priority.

Multidisciplinary collaboration is not a buzzword here. It is the daily operating rhythm. Fedus identifies the close collaboration between top-tier physicists, chemists, AI researchers, and engineers as the most intellectually exciting aspect of the company. The screening process looks for evidence that a candidate has worked in genuinely mixed teams, not just adjacent ones. A condensed matter theorist who has co-authored with ML engineers, or an infrastructure engineer who has debugged a failed synthesis run alongside a chemist, carries more weight than a longer publication list in a single discipline. The company's commercial model, selling an intelligence layer that serves as a system of record and control plane for enterprise materials engineering, reinforces this. Every technical hire must be able to explain their work to a customer's process engineer without jargon.

Mission alignment also surfaces in how candidates reason about verification. The central tension Fedus describes is the fundamental asymmetry between digital and physical verification: software self-improvement runs millions of iterations at near-zero cost; physical experimentation is constrained by equipment lead times, synthesis messiness, and the fact that atoms cannot be simulated at zero cost. Candidates who treat simulation as ground truth, or who dismiss negative experimental results as noise, fail the cultural screen. The company's closed-loop system is designed to hunt aberrations and inconsistencies between simulation, literature, and fresh experimental data, and to treat each mismatch as the signal that drives the next experiment. That mindset, more than any specific technique, is what the hiring process selects for.

The compensation structure reinforces the mission. Board data shows a salary band of $113k–$350k with a $275k median across 13 salaried roles, competitive with frontier AI labs but structured to attract researchers who might otherwise stay in academia. Fedus has highlighted the disparity between a Stanford postdoc's pay and a machine learning engineer's, framing it as a structural undervaluation of scientific talent. The hiring pitch explicitly targets scientists re-evaluating how their fields operate under intelligent systems. Candidates who negotiate primarily on cash versus equity, or who optimize for short-term liquidity, tend to self-select out. The ones who stay are betting that the scaling laws that transformed language modeling will rewrite physical sciences, and they want to be the ones writing the code that runs the loop.

Talent Market Pressure and Competitor Moves

Periodic Labs' push to 21 open roles arrives as the AI talent market tightens along lines that were barely visible two years ago. Ravio's dataset shows AI/ML hiring grew 88% year over year in 2025. Lightcast recorded a 109% jump in AI-skill job postings from 2024 to 2025, after a 73% rise the year before. The supply side has not kept pace. One 2026 industry analysis puts the global AI talent gap at 3.2 open roles for every qualified candidate: 1.6 million openings against roughly 518,000 qualified people worldwide. ManpowerGroup's 2026 Talent Shortage Survey found 72% of employers struggling to fill roles, with AI-related capabilities displacing traditional engineering and IT skills at the top of the global shortage list for the first time.

That scarcity reshapes how candidates evaluate offers. A 2026 Talent Board study found 68% of candidates say their hiring experience influences their decision to accept a job, yet only 38% rate their most recent experience as positive. The gap is a competitive disadvantage for companies that treat recruiting as a one-way evaluation. Flexibility and transparency are emerging as key differentiators, per Hiring Lab's 2026 global trends report. Hybrid work remains widely valued, though some countries are tightening in-office requirements. Salary transparency is rising as policies shift and job seekers demand data to compare offers.

Periodic Labs' posted bands match that range and sit inside the premium that AI skills now command. Lightcast found AI-skill postings carry a 28% salary premium, roughly $18,000 more per year than non-AI postings. Motion Recruitment documented a 9% rise in machine learning salaries in 2026. Robert Half's 2026 U.S. guide benchmarks AI/ML engineers at $134,000–$193,250 and AI architects at $142,750–$196,750. Periodic's top of band exceeds those ranges, signaling that frontier labs price for the scarce intersection of AI science and lab automation.

Competitors are responding in divergent ways. OpenAI, Anthropic, and Mistral AI sharply reduced hiring of star-performing researchers in 2025 compared to 2024, pivoting toward go-to-market teams; OpenAI alone lists 408 jobs overall, heavily weighted to commercial execution. Meanwhile, "neolabs" are making the boldest moves for research excellence. Together AI, fresh on the scene, has almost as many open research roles as OpenAI. Goodfire, a Stanford-born startup, secured 30 of the world's best AI safety experts from a pool of just 200. In the AI-for-science niche, Periodic Labs competes with Latent Labs (generative models for biological sciences), Lila Sciences, Berkeley A-Lab, Emerald Cloud Lab, and academic self-driving labs, each racing to convert AI predictions into physical outcomes.

The signal is clear — commercial execution is the epicenter, not just bigger and better models.

Avature's 2026 AI Impact Report shows 57% of organizations saw AI-related skills increase significantly or moderately in job descriptions over the prior 12 months, and 83% expect that increase to continue. Yet Avature also finds organizations struggling less with AI enthusiasm than with operationalizing AI in real workflows. Periodic Labs' screen, demanding both deep AI-science expertise and rapid-prototype lab skills, targets exactly that operationalization gap. Candidates who clear it demonstrate they can move models from notebook to wet lab. Rivals watching the pipeline will likely raise their own bars for hybrid wet-dry competence, accelerating a shift that has already made "AI scientist" a distinct hiring category rather than a flavor of ML engineer.

The fume hood in Menlo Park is still mostly human-tended. But the job board (21 roles, six added in a month, every one mapped to a node in the simulation-synthesis-data loop) is the clearest signal yet that the loop is closing. The next hire won't just run experiments. They'll teach the machine which experiment to run next.

Working in frontier tech? Zero G Talent tracks the openings: see every open Periodic Labs role, browse frontier tech jobs, the companies hiring, and the people building the field.