On August 19, Merck and Moderna announced something medicine had never seen before: a positive Phase 3 result for an individualized mRNA cancer therapy.
The trial enrolled 1,137 people whose high-risk melanoma had been surgically removed. For every patient receiving Moderna’s intismeran autogene, the process began with that person’s tumor. DNA and RNA sequencing identified its mutations. A set of algorithms then selected as many as 34 neoantigens, the abnormal protein fragments most likely to provoke an immune response. Moderna encoded those targets into a custom strand of mRNA, manufactured a therapy for one person, and combined it with Merck’s Keytruda.
The study met its primary endpoint of recurrence-free survival and a key secondary endpoint of distant-metastasis-free survival. Detailed data have not yet been released, overall survival is still being followed, and the treatment remains investigational.
The signal is the system behind it.
A repeatable system connected tumor sequencing, algorithmic target selection, programmable medicine, manufacturing and clinical feedback. The algorithm was one component inside a tightly integrated discovery-and-production loop.
That loop is becoming the new institution of science.
For most of the postwar era, we pictured discovery as a relay race. Government funded basic science in universities and national laboratories. Industry received the baton later, turning established knowledge into products. Vannevar Bush gave this arrangement its canonical argument in his 1945 report, Science: The Endless Frontier.
Eighty-one years later, that division is breaking down. The frontier is moving toward organizations that can generate a hypothesis, run a simulation, perform an experiment, absorb the result and decide what to try next without waiting for the annual grant cycle, the next paper or a handoff between institutions.
Many of those organizations are companies.
The old map no longer describes the territory
The White House’s July 2026 report, Science: A New Golden Age, is the clearest official recognition of the change. It says the old “linear model” has become inadequate because discovery now works as “an iterative loop between fundamental and applied work.” Industry and engineering, the report argues, increasingly spur basic research rather than merely commercialize it.
The funding data point the same way.
According to the National Science Foundation, businesses funded 75% of all U.S. R&D in 2023, while the federal government funded 18%. Even within basic research, the business share rose from 21% in 2012 to an estimated 35% in 2023; the federal share fell from 52% to 41%, despite a real increase in federal dollars. U.S. businesses performed $722 billion of R&D in 2023, including $43 billion classified as basic research.
In fields where progress depends on large-scale computation, proprietary data, automated instruments and rapid engineering, private labs can now pursue fundamental questions with an intensity universities often cannot match.
The White House report puts it bluntly: Ph.D. students and professors can raise hundreds of millions of dollars to form companies that chase fundamental breakthroughs. Small startups increasingly conduct basic research themselves. These firms “blur the distinction between basic and applied science.”
The examples are already arriving. Look at what has appeared in the last eighteen months.
From AI assistant to AI scientist
Discovery Loop launched this month with a founding team that reads like a compressed history of modern AI: Jeff Dean, Sanjay Ghemawat, Quoc Le and Oriol Vinyals. Their work helped produce Google’s distributed computing infrastructure, TensorFlow, TPUs, sequence-to-sequence learning, AlphaFold and Gemini.
Their new company is not selling a chatbot for scientists. Its stated aim is to automate the experimental loop itself: propose an experiment, implement it, run it, evaluate the result and iterate. Thousands of times in parallel. It will begin with machine-learning research, where experiments can be executed entirely in software, before moving toward medicine, energy, materials and other engineering problems.
A software experiment can be run overnight, but a drug still has to be synthesized, tested in cells and animals, manufactured, reviewed by regulators and eventually tried in people.
Still, Discovery Loop shows where the ambition has moved. The goal is no longer to help a scientist write a paper 20% faster. It is to raise the clock speed of discovery.
Lila Sciences is attacking the physical half of the problem. The private company has raised $550 million to build what it calls AI Science Factories: automated facilities where models generate hypotheses, design experiments, operate instruments, read results and update the next experiment. Lila says its first facility has run hundreds of thousands of AI-directed experiments across life science, chemistry and materials. Its laboratories put flow cytometers, plate readers, X-ray diffractometers and electron microscopes under a shared software layer.
Periodic Labs, backed by a $300 million founding round, is building AI scientists alongside autonomous materials laboratories. Its opening target is the discovery of new materials, including superconductors. The company’s argument is simple: scientific models trained only on published literature are trapped inside the record of what humans chose to report. An autonomous lab can produce its own data, including failed experiments that journals rarely publish.
FutureHouse offers a different organizational experiment. The nonprofit lab built scientific agents for literature search, synthesis, chemistry and data analysis, then spun out a commercial company, Edison Scientific, to deploy the tools at scale. FutureHouse’s Robin system generated hypotheses, designed experiments and identified the existing glaucoma drug ripasudil as a candidate for dry age-related macular degeneration. The full path from initial question to paper took about two and a half months. Its successor, Kosmos, can read roughly 1,500 papers and run tens of thousands of lines of analysis code in a single research run. FutureHouse says about 80% of Kosmos’s findings were judged accurate in its evaluations.
And then there is Google DeepMind. AlphaFold has predicted more than 200 million protein structures and is used by over three million researchers. GNoME predicted 2.2 million candidate crystal structures, including 380,000 judged stable; a robotic laboratory at Lawrence Berkeley National Laboratory synthesized 41 materials with help from the broader workflow. Alphabet’s Isomorphic Labs is now turning the AlphaFold lineage into a commercial drug-design engine, with partnerships spanning Novartis, Eli Lilly and Johnson & Johnson and $2.1 billion of fresh Series B capital.
These organizations differ, but they share an architecture. The defensible asset is not the model alone. It is the loop.
Why this is happening now
Four changes arrived at once.
1. The cost of proposing an experiment is collapsing
Models can search literature, write code, simulate structures, generate candidate molecules and rank possible explanations at a scale no research group can match manually. Microsoft’s MatterGen, for example, generates novel inorganic materials from desired properties rather than screening a fixed catalogue. MatterSim can then estimate which candidates are physically plausible before researchers attempt synthesis.
This changes the scarce resource. When hypotheses were expensive, the scientist’s problem was finding a promising idea. When hypotheses become abundant, the problem becomes deciding which ones deserve contact with reality.
The White House report captures the inversion in one sentence: “While the cost of generation has decreased exponentially, the cost of verification has not.”
2. Science is becoming an infrastructure business
A leading AI-science lab needs compute, clean data, models, robotics, instruments, software engineers, domain scientists and enough capital to run the system continuously. Those pieces work better when one organization owns the interfaces between them.
A university lab can invent a brilliant method. It is harder for it to operate a fleet of robotic laboratories, maintain frontier-scale compute and integrate both into a product-quality system. Grant funding is usually project-based and temporary. Infrastructure requires patient capital and full-time engineering.
This helps explain why NVIDIA is moving beyond selling chips. BioNeMo packages protein folding, molecular docking, generative chemistry, genomics and biomarker tools for scientific agents. NVIDIA says 18 of the world’s 20 largest pharmaceutical companies use BioNeMo. It is also working with Eli Lilly on an AI co-innovation lab and with Thermo Fisher on autonomous laboratory infrastructure.
Compute is only the entry point. The larger opportunity is to become the operating layer through which experiments are proposed and executed.
3. Basic and applied research are feeding each other
Moderna’s cancer program is a clean example. Clinical outcomes can improve the algorithm that selects neoantigens. The improved algorithm changes the next therapy. Manufacturing constraints influence design. Biology, software, production and medicine sit inside one feedback system.
Biogen’s work with Mila follows the same logic at an earlier stage. Their collaboration applies machine learning to neuroscience data, medical imaging, disease prognosis and clinical-trial enrichment. Biogen has also used Envisagenics’ SpliceCore platform to search RNA-sequencing data for disease-associated splicing events. The aim is to ask better biological questions and improve target or patient selection.
The old categories, with basic research and product development at different places, become difficult to separate once every downstream result creates upstream scientific information.
4. Capital can now follow the scientist
The postwar research system attached talent to institutions. Today, a small group of recognized researchers can raise a nine-figure round before it has a product.
That can give a scientist something a prestigious university cannot: enough compute, engineering and laboratory capacity to test an ambitious idea at industrial scale.
Follow the people
Capital flows are one signal. Talent flows are harder to fake.
This summer, Fields Medalist Jacob Tsimerman said he would leave the University of Toronto for OpenAI. Nobel laureate John Jumper, who led the AlphaFold team, left Google DeepMind for Anthropic. Jeff Dean, Sanjay Ghemawat, Quoc Le and Oriol Vinyals left Google to form Discovery Loop. Periodic Labs brought together Liam Fedus, formerly an OpenAI research leader, and Ekin Dogus Cubuk, who led materials and chemistry research at Google DeepMind, then recruited scientists from AI, physics and materials research.
The transfer can happen at the level of a team as well as a star. Xaira Therapeutics launched with more than $1 billion committed and hired several researchers behind RFdiffusion and RFantibody from David Baker’s University of Washington Institute for Protein Design. Baker remained at the university while the proprietary data, full-time engineering, automated experiments and drug development moved inside the company.
Scientific prestige is becoming portable. The most ambitious researchers increasingly choose institutions around the problem, compute and experimental machinery available rather than the traditional hierarchy of campus, tenure and journal.
The White House report notes that more scientists are leaving academia for industry “not because they have abandoned curiosity-driven inquiry,” but because the tools and resources for some fundamental questions now sit outside university walls.
The private lab is not only buying talent. It is becoming a credible place to do the kind of work that earns Nobel Prizes.
The White House is endorsing the loop
The report proposes expanding the Genesis Mission across national laboratories, integrating supercomputers, AI models, scientific instruments and datasets. It calls for opening federal facilities to private builders, strengthening public-private talent flows, creating pre-competitive consortia and investing in autonomous experimentation.
AI can produce papers, grant applications and plausible hypotheses faster than humans can verify them. The report therefore calls for reproducibility, data sharing, falsifiability, machine-auditable replication packages and verification infrastructure.
The most investable insight in the document is the migration of scarcity and value toward verification: physical experiments, trusted data, clinical evidence and reproducible workflows.
The strongest businesses will own hard-to-reproduce data, experimental throughput, validated instruments, regulatory capability or a repeatable path into the clinic.
The watchlist
The opportunity spans public platforms, AI-native biotech companies and private research labs. Here is the watchlist I would use to track where the evidence is appearing.
Public companies
Alphabet (NASDAQ: GOOGL, GOOG)
DeepMind remains the strongest proof that a private AI lab can solve fundamental scientific problems. Isomorphic Labs gives Alphabet a direct route from scientific models to proprietary medicines, although the exposure is financially small relative to Alphabet’s advertising and cloud businesses. Watch for the first Isomorphic-designed candidates entering human trials, expansion of pharmaceutical partnerships and evidence that AlphaFold-class models improve actual drug outcomes rather than only structures and rankings.
NVIDIA (NASDAQ: NVDA)
NVIDIA is positioning itself as the picks-and-shovels provider for computational science: GPUs, BioNeMo, simulation, agent tools and robotics. Its advantage is breadth across the full computational stack. Watch BioNeMo adoption, autonomous-lab partnerships and recurring software or cloud economics beyond chip sales.
Microsoft (NASDAQ: MSFT)
MatterGen, MatterSim and Azure Quantum Elements give Microsoft a credible materials-science stack. The investment question is whether world-class research becomes a durable commercial workflow on Azure. Watch for independently synthesized materials, industrial customers and closed-loop integrations with physical laboratories.
Recursion Pharmaceuticals (NASDAQ: RXRX)
Recursion is the clearest public test of the AI-native, wet-lab-in-the-loop thesis. Its platform combines automated biology and chemistry laboratories, more than 50 petabytes of multimodal data and clinical development. In August, Genentech optioned the first neuroscience target discovered through their collaboration into an early discovery program. Recursion also has several clinical-stage assets, which means the thesis can finally be judged by molecules and patients rather than platform claims. Watch REC-4881, REC-1245 and the initiation of the REC-7735 trial; also watch cash burn and whether partner milestones repeat.
Absci (NASDAQ: ABSI)
Absci pairs generative protein design with wet-lab validation and has moved its lead AI-designed antibody, ABS-201, into Phase 1/2 testing. Positive interim safety and pharmacokinetic data are encouraging, but proof of therapeutic effect remains ahead. Watch the androgenetic-alopecia proof-of-concept readout, progress in endometriosis and whether its six-week design-test cycles translate into better clinical candidates.
Schrödinger (NASDAQ: SDGR)
Schrödinger combines physics-based molecular simulation, machine learning, a software business and a drug pipeline. It is less theatrically “AI-native” than newer entrants, which may be an advantage: pharmaceutical customers already use its tools. Watch software growth, partner economics and clinical validation of internally designed programs.
Moderna (NASDAQ: MRNA) and Merck (NYSE: MRK)
The positive Phase 3 intismeran result is the most concrete current evidence that an algorithmically designed, personalized treatment can work at late-stage scale. It is not yet an approval, detailed results are pending, and Keytruda contributes materially to the combination. Watch the full dataset, regulatory filings, manufacturing turnaround times and results in tumors beyond melanoma. Merck offers the stronger incumbent franchise; Moderna offers more concentrated platform exposure.
Biogen (NASDAQ: BIIB)
Biogen is an adopter rather than an AI-native lab. Its neuroscience data, disease expertise and collaborations make it a useful test of whether AI improves trial design and target selection in one of medicine’s hardest fields. Watch for named programs, measurable development gains and a path from analytics to pipeline value. Until then, AI is a supporting thesis, not a reason to own the stock.
Private labs worth tracking
Discovery Loop: Watch for a technical result proving that its automated loop improves machine-learning research before extrapolating to physical science.
Lila Sciences: Watch for independently validated discoveries, customer programs and evidence that its autonomous labs improve with accumulated experimental data.
Periodic Labs: Watch whether the autonomous materials lab produces synthesized candidates with useful, reproducible properties rather than only novel predictions.
Isomorphic Labs: Watch clinical entry and partner progression. Alphabet provides indirect exposure, but the lab is increasingly financed as an independent company.
Edison Scientific / FutureHouse: Watch external replication of Robin and Kosmos findings, user retention and the boundary between the nonprofit’s public-interest research and Edison’s commercial data advantage.
Anthropic: Claude Science and the recruitment of John Jumper suggest a serious science push. Watch for domain-specific products, laboratory integrations and validated discoveries rather than general-purpose agent demos.
Xaira Therapeutics: Watch whether the RFdiffusion/RFantibody lineage produces named development candidates, investigational-new-drug filings and clinical entries. Its more-than-$1-billion launch makes it one of the clearest attempts to transplant an academic protein-design capability into a vertically integrated company.
insitro: Daphne Koller’s company combines machine learning, high-throughput biological experiments and proprietary multimodal data. Watch for clinical entry, repeatable partner milestones and evidence that the platform improves decisions rather than merely generating more candidates.
The investable signal
The leading companies will shorten the path from a model’s proposal to a verified result, then feed that result back into the next experiment.
That favors businesses with proprietary experimental data, automated laboratories, scientific instruments, manufacturing capability and a route into regulated markets. Models will spread. A functioning discovery loop will be much harder to copy.
The investment thesis is clear:
AI turns science from a sequence of institutional handoffs into a continuous production system. More discovery will accumulate wherever the loop closes fastest, whether in a university, national lab, nonprofit or company. Private labs currently have an unusual advantage because they can combine capital, compute, engineering, proprietary data and physical experimentation under one roof.
The Moderna result, the formation of Discovery Loop, the capital flowing into AI Science Factories and the White House’s support for AI-native scientific institutions all point in the same direction.
The last scientific age was organized around the grant, the department and the paper.
The next one may be organized around the loop.
Sources and further reading
Merck and Moderna, “Phase 3 INTerpath-001 met recurrence-free and distant-metastasis-free survival endpoints,” August 19, 2026.
Moderna, “Advancing the Fight Against Cancer through mRNA & AI,” December 19, 2023.
White House OSTP, “Science: A New Golden Age,” July 2026.
NSF/NCSES, “Trends in U.S. R&D Performance and Funding,” 2025.
NSF/NCSES, “Business R&D Performance in the United States Increases to $722 Billion in 2023,” September 2025.
Discovery Loop, “Automating discovery to accelerate science and engineering,” accessed August 2026.
Lila Sciences, “Welcoming New Partners in Our Mission to Build Scientific Superintelligence,” updated October 2025; and “Series A close.”
Periodic Labs, “Introducing Periodic Labs,” 2025.
FutureHouse, “About FutureHouse,” accessed August 2026; Edison Scientific, “Kosmos: An AI Scientist for Autonomous Discovery,” November 2025.
Google DeepMind, “AlphaFold” and “Millions of new materials discovered with deep learning.”
Isomorphic Labs, “Series B investment round,” May 2026; “Partnerships.”
Microsoft Research, “MatterGen: A new paradigm of materials design with generative AI,” January 2025.
NVIDIA, “BioNeMo Agent Toolkit,” June 2026; “BioNeMo platform adoption,” January 2026.
Mila, “Collaboration with Biogen on AI/ML in neuroscience therapy development,” February 2023; Biogen and Envisagenics, “Collaboration to advance RNA splicing research,” May 2021.
Recursion, “Second Quarter 2026 results and Genentech neuroscience target option,” August 5, 2026.
Absci, “Investor Relations” and “Technology,” accessed August 2026.
University of Toronto talent move: BetaKit report on Jacob Tsimerman joining OpenAI, July 31, 2026. John Jumper move: Reuters report, June 19, 2026.
Xaira Therapeutics, “Launch with more than $1 billion committed,” April 2024.
insitro, “Leadership: Daphne Koller” and “About insitro,” accessed August 2026.
Disclosure: This article is for research and discussion, not investment advice. Clinical candidates are investigational unless otherwise noted.