Moderna was only company 18. The real product is the company factory

Sarah Guo opens No Priors with the arc: Noubar Afeyan arrived as a teenager after his family fled war-torn Beirut, earned MIT's first PhD in biochemical engineering, and spent three decades building Flagship Pioneering — the firm behind Moderna, more than a hundred other biotech companies, and more than $100 billion of value. Afeyan says he will work hard to fit that description. The motivation for Flagship, he explains, started with what he was already doing.

He started a company in 1987, when 24-year-old immigrants did not start companies in the United States. The people entrusted with what then counted as massive venture capital — $2–3 million a round — were former Merck or IBM senior executives. He got the chance right out of graduate school, raised a fair amount of money, and went down an entrepreneurial path. Along the way he was put off by the assumption that the process was supposed to be random, improvisational, idiosyncratic, almost emotional and game-like. In the early 1990s he went around asking why entrepreneurship was not a profession, and if it were, what that profession would look like.

One answer: professionals work in parallel. Venture capital is a parallel investing activity; if investors could only do one company at a time they would be serial venture capitalists. Lawyers operate in parallel too. Entrepreneurs, somehow, were supposed to be serial. Learning cycles of doing several things at once are completely different, because you are forced to think about the essence of what you are doing, what is reproducible, and what has to be different in each case. In the late 1990s, while still running his first company, he co-founded several others to try this as a solo player, then realized it was too hard alone.

So he created what he calls the first company-creation company. It was named Nucogen, for new company generation, and they operated under that name for three years until people told him it sounded like a disease. He changed it to Flagship. The deeper motive, he told MIT entrepreneurship students for 16 years and Harvard Business School innovation students for three, is that starting companies is the most value-creating activity in the current human endeavor. Students guessed the internet and satellites. He told them the invention is the startup: Google, Tesla, Facebook, Genentech. Why outsource that to a random, gamey activity of winners, losers, and keeping score? In healthcare, climate, agriculture, and food security, he says, you cannot treat the work as shots on goal. You have to get better at it.

Flagship is famous for allowing breakthroughs to emerge rather than being made by people. After 23 years of practice, Afeyan and Gary Pisano laid the idea out in Harvard Business Review as emergent innovation. Human design is goal-based, he says, but you cannot define the goal of creating Nvidia. Jensen Huang's team thought they could make valuable products that would first affect people in gaming, and then the AI opportunity emerged. It was not in the business plan. Where else does novelty happen in that unpredictable way? Nature.

In nature, variation, selection, and iteration create unbelievable novelty and esoteric complexity. Life is the result; we still have many living things and not many living silicon things. Run those three forces on anything and you get emergence: on human thought you get revolutions, memes, political ideas, religions; on running shoes you get Air Jordans. Participants then describe the result as though they came up with it, converting reality into a story in which the human agent presided. "Whoever wins the war gets to write history," he says. After 38 years he does not believe anything he has been involved in innovating is actually the product of his work. It has been emergent. The organization's job is to create an environment in which emergence can happen, then have the humility not to describe it as an act of genius or superintelligence they wielded.

Generative AI, he adds, is one hell of a technology for emergence, whether that means new proteins or new consumer products. Guo notes that the best entrepreneurs are often guilty of rewriting the story as if they had predictive power from day one. Afeyan says successful people who have done unexpected things also secretly worship at the altar of chance, because their explanations do not suffice for what happened. Nobody writes the book that says they did this, then that, and then something out of the blue they cannot explain is what let them succeed. The honest version is: a lot of things happened, they got fortunate, and they were ready because they had done everything else as best they could. Selection pressure, as in nature, comes from the sum of what already exists; consumer preference often reacts to what is put in front of it. You have to build an environment where variation and selection can interact.

Flagship is careful about leaving medical biotech, but it will venture out where it has intellectual property or a daring that comes from not knowing enough. The first thing they do in a space informs the next five; if all five fail, they conclude they cannot get paid for innovation there. The method is to bring to life today what might otherwise exist in five years, and to make sure it will also be valuable — which is not true of every sector. Renewable energy taught them that the hard way.

Between 2008 and 2012 they engineered photosynthetic bacteria that usually grow in the ocean depths, a cool organism that sees only a little light, to consume CO2 and secrete diesel. People thought it was impossible. They did it, and they built dirt-cheap reactor systems. The company was called Joule. Then the market spoke. Diesel is diesel, renewable or not. Carbon started at $50 a ton and ended at $5. The United States went from energy-dependent to gushing oil and gas. No matter what they did as innovators, they could not get a premium. People still ask him to go back. He still does not know if anyone will pay for it. They have also started one-off companies in supercomputing and networking, and recently in semiconducting and carbon-capturing materials. Not every domain, he says, needs or rewards scientific leaps.

Guo asks about categories regulators do not know what to do with — microbiome therapies, gene editors — on top of cyclical markets. Afeyan draws a circle: everything inside is known today; the ring just outside is the adjacency where most innovation happens and where people can estimate risk and reward. That is what due diligence is, aggregating key opinion leaders. Push farther and people can no longer estimate the likelihood of success. He would not call that risk anymore. He would call it uncertainty. Fusion, even with claims of a shorter-than-35-year path, is uncertainty because nobody can tell you the probability. Wall Street wants everything on a risk matrix. He does not believe that.

Adjacencies have their own problem in science-driven technology: everyone is already there — every startup, academic lab, and large company — so the overlooked risk is commoditization. Facing uncertainty, you do the experiment that can first bring the thing to life. mRNA, Flagship's 18th company, which became Moderna, is the example. There was no mRNA drug or vaccine; academic labs had worked on it and given up. They needed regulatory acceptance, pricing, and manufacturing they had no idea how to do. The reward, with or without a pandemic — the pandemic distorted the path, he notes, but the company created plenty of value without it — is why they carefully embrace uncertainty, resolve it on the way to understanding risk, and then mitigate. If you will not do that, you work on me-too value pools. People at Flagship know they are no smarter, no harder working, and no better connected than anyone else. What they can do is underwrite uncertainty.

Guo asks what has changed since 1999. Afeyan reminds her that the world was being overtaken by internet e-commerce — sunglasses.com, diapers.com — while the human genome was being sequenced. He was involved in Celera, the private sequencing effort. Money was siphoned into the fad of the day, and it was really hard to fund life science. They focused on the intersection of biology and technology anyway. From day one they believed you could conceive and create companies systematically. What they did not initially bet on was systematically making breakthrough innovations the way they now believe they have learned to do, a second capability that arrived in the late 2000s.

Today Flagship is 550 people, about 200-plus of them scientists, engineers, and MDs. They file 600 to 700 patents a year centrally, and every project is meant to have essentially no connection to what has been done before. Seven years ago they were about 50 people. Only recently have they brought in-house the ability to scale companies, not just conceive them, so many people now build companies in parallel and the learning cycles have accelerated. Generative AI is the new tool for leaping conception and hypothesis generation, but Afeyan rejects the revisionism that this is Flagship's first encounter with machine learning.

In 2001 they started Affinnova, which used machine learning and evolutionary algorithms to evolve consumer products online, six years of dynamic evolutionary tools that were metaphorically doing what nature does — nothing to do with DNA or RNA. Their 100th company is now developing that idea with modern generative AI. They have also moved along the innovation value chain, generating breakthroughs incumbents can use, with large partnerships at Pfizer, Novo Nordisk, and GSK in pharma, and Thermo, Analog Devices, and Samsung in tech, aimed at expanding reach and impact.

If AI is large-data models that begin to look like cognition, Flagship has gotten more ambitious about using it. Moderna used early deep neural nets on mRNA data to inform the next generations of design and manufacturing. About six or seven years ago they opened a separate what-if: could you computationally design a protein of any desired function? They refused to use AlphaFold or quantum folding models. They asked whether enough instances of a desired function paired with the underlying DNA sequence would let a learning algorithm generate new ones. Critics said you also need the protein sequence, the fold, and the rest. Afeyan's retort: when DNA is handed to the next generation there is no manual. DNA has no idea what a protein is or what folding is, and yet the function follows. The knowledge must be encoded in the sequence.

The incremental cost of asking had come down, in AI and in experimental setups. Within a couple of years they could show computational leaps for antibodies and their targets that you could not make experimentally on the same timeline. That became Generate Biomedicines, one of Nvidia's first large partnerships in biology. Generate now has more than 15 computationally designed antibody programs, some in the clinic, some advancing. Asked to show a computationally designed drug, he notes that showing a drug is a lot more than design and early testing. They have since applied the approach to cell models, DNA, RNA, other molecules, and lipid nanoparticle design.

The cooler frontier is autonomous scientific discovery: generate a hypothesis, specify experiments, run them, collect data, interpret, iterate — "do science the way Waymo drives cars." They are nowhere near having Waymo on the streets, but they can show the elements, tie them together, and in narrow spaces hint at what a million people playing for a thousand years might look like. Multi-agent systems for emergence are next: new brands and products, and early mental-health interventions that let different agent-based models interplay rather than cloning what a doctor would do. He is not excited about the productivity layer — automated document writing, summarization — that everyone else is doing. Those do not require Flagship's extra effort. The other things feel pioneering.

Guo notes that better hypothesis generation is still the top of the funnel. Afeyan works backward. The last step is filing a BLA or NDA with the FDA. Before that you run a phase three trial: the right dose, a large enough population, statistically significant superiority, no toxicity. Those last steps are regulated. Until data-driven models can replace some analog testing, the field waits on large-scale trials that cost hundreds of millions of dollars. Of course that day should come sooner, he says, especially for the disease he is going to die from. The template is Operation Warp Speed: private sector, public sector, and regulator aligned on a solution, not slowing the process just to be on the safe side while people die. When the consequence of inaction spikes, people act. The sad part is that the consequence in cancer and neurological disease plays out every day. It is very large. It is just slower, so it does not feel like an urgent threat.

He would at least run experiments in data-informed interpretation of trials, including adaptive designs. The work Flagship is doing is to understand the state of a particular patient. Stage one cancer and stage two cancer, he says, are a joke with nothing to do with underlying biology — a super-macro approximation the profession revises now and then. The tools now exist to micro-age disease into what he calls a biostage: a trajectory that is not four stages but 75,000 if you want. Then you can see which mechanisms are on or off and pick a homogeneous subset of people so the others do not defeat the trial. A smaller trial, a smaller indication, then expand. That does not suit a large company that wants to give the drug to everybody. For biotechs who would otherwise die waiting, it might be the interesting innovation — if regulators will be open to it.

Patient data could point to the right mechanism in a subset of Parkinson's, for example, but HIPAA and related rules make using that data to train models slow. Pre-pandemic, everyone thought a vaccine took years, so after Covid there was oddly no hurry either. Technology made a shot possible in three or four months, but people had to die, and economies had to shut, before anyone said they had the money and might as well try. He hopes the next time does not require the same gestation of debate, including from people who still think vaccines cannot work. He is confident there is not a disease they cannot find an antidote or vaccine for if they deploy the best technology, coordinated, with the right incentives. Warp Speed's key move was a value outcome: the government would buy X doses at this price, a clear market signal that made uncertainty acceptable. That mechanism is tried and true in a pandemic, and would go a long way outside one.

Guo puts the classic trade-off to him: invest early in a platform, or in one or two clinical assets, while the investor climate moves under you. Because Flagship goes to far-out places looking for undiscovered value, coming back with one asset is, he says, the definition of insanity. If that is the plan, bet on a slightly different version of a well-known technology and hope your lottery ticket gets pulled. If you are going to do RNA, DNA, gene writing, gene editing, or computational proteins for the first time, you need to diversify, because you do not know which programs will get knocked out for reasons that have nothing to do with the underlying technology. Hence every single one of the 110 companies they have been involved with for 25 years is a platform. No exceptions.

Most people do not do it that way, even if they want to, because of capital. One program is already expensive; two or three more so; a platform that could do 15 things is very high. Investors, he argues, do not properly value platforms. They do not assign value to the correlated option — one program de-risked by three others — so you get the cost deduction without the credit. They also fear the strain on management. In the current environment he sees a mass-extinction event for single-asset companies, accentuated by Chinese central-government bets and startups that go after the same space at much lower cost and a different barrier to clinical data. Good for pharma companies accessing those assets, he says, but he does not know how a higher-cost single-asset biotech avoids commoditization. Platforms at least give you a chance at partnerships and staying alive. Some Flagship companies still meet an unfortunate end. Some do not.

Last question: in his 2025 annual letter he describes polyintelligence, the integration of human, nature-derived, and machine intelligence. Where does human intuition remain most relevant? Intuition, he says, is just a model. Each of us generates one and uses it in short form; it is the closest thing humans have to a large language model. The problem is that LLMs have been trained across millions of people's worth of data, and yours is trained on you. He was not criticizing the word. He thinks the real miss is treating the frontier as a line between human and machine. Most of science is between human and nature, and now there is a machine that can intellectually support inquiries of nature. It is a triangle. The three actors will adapt to each other. Humans compute and act differently from computers and from other forces of nature. That three-way axis of emergence, he says, is where the future of life comes from.