Rasyn's new platform reflects an industry caught between cloud convenience and data sovereignty demands, as regulators tighten rules and chemical companies rush to reformulate legacy products.
When Rasyn unveiled Marigold Science this past August, the Y Combinator-backed startup made a promise that might have sounded mundane five years ago: your data never leaves your building. For pharmaceutical formulators and materials scientists working under the European Union's freshly enacted AI transparency rules, that assurance carries weight. The platform bundles 54 computational chemistry tools—RDKit, Chemprop, OpenMM, and others familiar to bench chemists—into a self-hosted workspace where natural-language prompts trigger calculations that once required scripting expertise.
It's a pitch tailored to an industry under unusual strain. Chemical manufacturers are contending with PFAS reformulation deadlines after 3M shuttered its PFAS production lines at the close of last year, according to the company's regulatory disclosures. Data center operators, meanwhile, are hunting for coolants that work in warm-water direct-to-chip systems without relying on the very fluorochemicals being phased out. And enterprise R&D groups face compliance headaches now that the EU AI Act's general-purpose AI transparency obligations took effect in early August.
Grand View Research projected the addressable market at $1.5 billion in 2025, climbing to $11 billion by 2033. Whether Rasyn can carve out a defensible position remains an open question, but the company arrives as foundation models trained on molecular data move from academic curiosities to production tools.
A Fragmented Software Landscape
AI adoption in chemical R&D has accelerated past the pilot stage. Deloitte's outlook for the chemical industry, drawing on 2025 data, found that 51 percent of U.S. manufacturers already deploy AI in operations; four-fifths expect it to become essential by decade's end. Yet the software ecosystem splits awkwardly between cloud-only platforms and legacy desktop applications that lack modern machine-learning capabilities.
Materials informatics software alone is projected to reach $211.58 million in 2026 and $583.82 million by 2032, according to a report from 360iResearch released last week. The broader lab automation category stands at $10 billion this year, climbing to $18.4 billion by 2033 in Grand View Research's estimates.
Foundation models specific to chemistry gained traction through 2025 and into this year. ChemFM, detailed in a January paper in Nature Communications Chemistry, demonstrated one-to-three-billion-parameter models pre-trained on 178 million molecules. A Nature Reviews Chemistry article published in July identified "structural asymmetries" that limit shared infrastructure in self-driving labs, while a Communications Materials piece the same month called multi-agent AI and agentic workflows "keys to scaling" autonomous experimentation.
The technical pieces are falling into place, in other words. What's less certain is whether commercial platforms can deliver on the promise of meaningfully faster discovery cycles.
Converging Pressures
Three forces are reshaping formulation workflows. Foundation models are consolidating around base architectures with domain-tuned adapters for reactions, mixtures, and spectral analysis—a trajectory that mirrors the evolution of large language models. Hyperscalers moving to warm-water direct-to-chip cooling create fresh demand for optimized, PFAS-free coolants with tighter specification windows around conductivity, viscosity, and corrosion inhibitor packages. Microsoft announced in June that new AI data centers would achieve near-zero operational water use, the company said on its blog, intensifying the hunt for thermal fluids that perform under those constraints.
Data residency requirements, the third pressure point, have intensified quickly. The EU AI Act's transparency rules for general-purpose AI began enforcement on August 2, and NIST's AI Risk Management Framework has shaped U.S. enterprise procurement controls over the past two years. "Your data stays home," Rasyn states on the Marigold Science product page, framing self-hosting as the antidote to intellectual property concerns. The platform's privacy policy, dated February 18, is explicit: "We do not use Your Scientific Data to train our models unless you provide explicit, written opt-in consent."
Regulatory action on PFAS has accelerated reformulation timelines across the industry. The European Chemicals Agency issued opinions on broad PFAS restriction proposals this past March, according to the agency's hot topics page, and the EPA finalized the first-ever PFAS National Primary Drinking Water Regulation in April 2024, with updated guidance following in April of this year. A November 2025 report from EY cited a German specialty chemicals firm using AI to identify non-PFAS alternatives, though that account is now more than a year old and specific outcomes remain unpublished.
Early Adopters and Validation Efforts
L'Oréal partnered with IBM in January of last year to build a "Formulation Foundation Model" for sustainable cosmetics, according to IBM's newsroom. Shiseido launched VOYAGER this past January, an AI platform the company says generates "tens of thousands of formulation candidates" per project; the first AI-developed suncare product is slated to ship this summer, the company announced.
In coatings, a June 2025 study in npj Materials Degradation showed machine learning suggesting novel corrosion-protective formulations from sparse historic data. AkzoNobel provided the dataset, and subsequent testing validated the model's recommendations. A photolithography case study published in AZoM this past January detailed AI-guided optimization that extended well beyond the initial 100 experimental datapoints. An arXiv platform paper in July reported two-to-nine-times reductions in experimental effort compared to random search across published cases, though peer review has not yet confirmed those figures.
Rasyn describes itself as both research lab and product company, a dual identity reflected in its site tagline: "We design new formulations and materials through the fusion of computation and experimentation." The Marigold Science workspace, announced via Y Combinator's LinkedIn in August, allows users to "give it any chemistry task in plain language," the post said. The system then "pulls the literature, picks the tools, runs the calculations, and hands back the numbers with the reasoning."
The platform bundles those 54 tools and added self-hosting this year with single-command installation for macOS and Linux, the company's site indicates. Pricing runs $89 monthly for individual researchers with 3,000 compute units, $249 per seat monthly for lab teams with a three-seat minimum, and from $2,000 monthly for institutional deployments that include bring-your-own-model endpoints and single sign-on.
Rasyn lists proprietary models alongside the open-source toolkit: Synthon for retrosynthesis, which the company claims achieves 66.2 percent top-one accuracy on the USPTO-50K benchmark; Marcus for joint catalyst-solvent-reagent prediction; Larmor for per-atom NMR shift; and ChromPeakNet for chromatography peak detection. A February paper page on Rasyn's site reports an F1 score of 0.889 on real data versus MZmine's 0.476, though independent validation of these numbers remains pending.
The company was founded by Ansh Tiwari, Daood Hashmi, and Ayush Chauhan, according to Y Combinator's August post. CB Insights lists a $0.5 million convertible note dated August 1, though the recipient remains unverified as Rasyn and lacks company confirmation.
A Crowded Field
Rasyn enters a competitive arena. Citrine Informatics sells enterprise SaaS for materials and formulations optimization across chemicals, coatings, and semiconductors. Schrödinger introduced RetroSynth, combining physics and AI for synthesis and materials discovery, in a February webinar. Orbital Materials raised a $50 million Series B in May, Fortune reported, and partners with AWS on decarbonization and cooling materials. Atinary opened a self-driving labs facility in Boston in February, the company announced.
Academic infrastructure is scaling in parallel. The Acceleration Consortium at the University of Toronto and UBC expanded self-driving lab capacity this year and funded translational research programs, according to the consortium's news page. A July feature in npj Robotics highlighted multi-robot orchestration in self-driving labs, and a Royal Society of Chemistry benchmarking study in 2026 proposed standardized acceleration metrics after surveying median factors across the literature.
Greg Mulholland, CEO of Citrine Informatics, told the Association for Materials Protection and Performance in April of last year: "Every problem in materials and chemistry is an exercise in balancing priorities." That observation captures the challenge facing Rasyn and its competitors. The Open Compute Project and ASHRAE formed an alliance in October 2025 to align liquid-cooling standards and best practices, the OCP blog noted, as hyperscalers coalesced around warm-water direct-to-chip loops using propylene glycol and water blends. OCP's Deschutes specification referenced PG25 blends, deionized water with additive packages, and elastomer compatibility testing for EPDM hoses and sub-100-micron filtration channels.
Self-driving labs and agentic orchestration will face demands for validated acceleration metrics and auditability as reviewers push for standardized benchmarking, a Digital Discovery article argued this year. Foundation models for chemistry appear headed toward consolidation into base models with domain-tuned adapters for reactions, mixtures, and spectra—early signs emerged from ChemFM and retrosynthesis agent papers published on arXiv through 2025 and into this year.
Rasyn's self-hosting pitch addresses intellectual property concerns in pharma and chemicals, and aligns with the EU's transparency timeline for general-purpose AI that began in August. The FDA announced plans in April of last year to phase out some animal testing requirements for monoclonal antibodies and other drugs, a shift that could shorten preclinical timelines for formulation programs down the line.
The next milestone for Rasyn and its competitors comes down to proving acceleration claims with transparent benchmarks and moving pilot deployments to production scale. The market opportunity is real—perhaps $2 billion near-term by some estimates—but so is the scrutiny from enterprise buyers who've watched AI hype cycles come and go. Whether Marigold Science can deliver the kind of validated, repeatable gains that turn curiosity into contracts remains the central question.
