Context

a16z general partner Jorge Conde interviews Moderna CEO Stéphane Bancel after Moderna and Merck reported positive Phase 3 results for a personalized cancer vaccine in melanoma, the culmination of ten years of work. The therapy sequences a patient's own tumor, compares it to their healthy cells, picks the mutations most worth attacking, and manufactures a one-of-a-kind mRNA medicine that teaches the immune system what to hunt. Most of the conversation is not about the biology. It is about the operating problem behind it: how to build, ship, and get regulatory approval for a product where nearly every unit is different, made one customer at a time. That makes it unusually useful for a PM, because the hard parts, personalization as architecture, the make-it-work-before-you-make-it-efficient sequencing, the unit-economics levers, validating a system whose output varies every time, are the same problems software teams hit when they build anything bespoke or generative at scale.

The Big Idea

When personalization is fundamental rather than cosmetic, you cannot bolt it on. The entire system, from data intake to manufacturing to regulatory approval, has to be built for a batch size of one.

The data forced this: Moderna found that about 90% of the target mutations differ from one patient to the next, so a single shared product could never have worked. That one fact is why every part of the pipeline, sequencing, algorithm, factory, and even the regulatory filing, is designed around producing a unique output for each individual.

Key Insights

1. Personalization is the architecture

The field spent 20-plus years and over a thousand failed trials trying cancer vaccines built on shared antigens, the same molecular targets used for every patient. Moderna's bet was the opposite: build for one patient at a time. When they started they had no idea how much of each patient's cancer would be unique; they guessed maybe 2 to 5% overlap difference. The real answer was that about 90% of the useful targets are different from person to person. That single finding means personalization is not a premium tier or an edge case here, it is the only thing that can work. The PM lesson: before you build, find out whether personalization is decorative or load-bearing, because if it is load-bearing, a generic core with a customization layer on top will quietly fail.

2. Ship information, not atoms

The reason this scales at all is that the product is really a data file. Moderna does not need the patient's physical cells; it needs the sequence of their healthy and tumor DNA, an information molecule. From that file it synthesizes the mRNA chemically, in water, using enzymes, in tiny reactors. Bancel contrasts this directly with CAR-T therapy, another personalized cancer treatment, which must physically extract a patient's immune cells, re-engineer them in a large reactor, and ship the living cells back. Because CAR-T handles fragile cells, everything about it is big and slow. Because Moderna's input is information, the physical footprint shrinks dramatically. The takeaway for any product: designing the value chain around moving information instead of moving physical goods changes what is possible to scale.

3. Make it work before you make it efficient

Moderna deliberately built its first personalized-manufacturing machine to be reliable, not efficient. It was the size of a large fridge, "big and clunky," and they told the team explicitly not to chase efficiency yet. The reasoning is sharp: if you spend five years perfecting a beautiful, optimized robot and the underlying science turns out not to work, you have wasted the five years. Worse, a half-working machine could produce a false negative in the trial, making a therapy that actually works look like a failure, which Bancel calls a potential "disaster for humanity." So they optimized only for quality until Phase 2 proved the science, and only then poured engineering into speed and cost. The lesson: do not pre-optimize a pipeline before the core hypothesis is validated, and be especially wary that a sloppy system can kill a good idea by generating a false negative.

4. Obsess over the two real cost levers

Once the science was proven, Bancel says he became obsessed with exactly two numbers. The first is cycle time: currently 42 days from biopsy to the finished dose ready in the hospital ("needle to needle"). Shorter cycle time means each machine can turn over more times per year. The second is square inches: the physical footprint of each machine. In a fixed clean-room envelope, fitting two or ten times more machines multiplies throughput and spreads fixed costs, so he pushes to move even the compute out of the clean room to save floor space. The lesson: for an operations-heavy product, unit economics usually come down to a small number of dominant levers. Name them explicitly, then relentlessly work those, rather than optimizing everything equally.

5. When every output differs, validate the process

If every dose is physically different, a regulator cannot approve each dose. So the FDA approves the process, not the product (a "process BLA," the same path CAR-T took). What Moderna has to prove is reproducibility: as Bancel frames the regulator's question, if you feed the same tumor and blood sample in at the start, do you reliably get the same medicine out the end of the black box? This required years of ongoing dialogue with the FDA, not a surprise filing at the end. The transferable idea: for any product that generates a unique output each time, including AI and generative systems, you cannot QA the individual output the way you would a fixed SKU. You validate that the system reliably turns the same input into the same trustworthy output.

6. The failures are the richest data

Bancel is most excited about mining the roughly 20% of patients who did not respond after five years. Because Moderna has their blood samples, sequences, and outcomes, the team can study why the treatment missed and feed that into the next version of the selection algorithm. He frames the current algorithm as "version 1.0," a ten-year-old model, and draws the AI analogy: the current version is the worst it will ever be. The principle he states plainly is that you learn more from what does not work than from what does. The PM habit worth copying: treat your failure cohort as the primary input to the next version, not as noise to explain away.

7. Expand a platform from proven ground outward

With melanoma working, Moderna is expanding along three deliberately ordered vectors, from safest bet to hardest:

• Where the current standard already works: add the vaccine on top of the existing checkpoint immunotherapy (Keytruda) in cancers where Keytruda is already approved, like lung and kidney, expecting the two to be synergistic.
• Early in disease, where the current tool is avoided: in stage 1 lung cancer, use the vaccine alone, because checkpoint drugs are skipped early due to serious lifelong side effects, and the vaccine's side effects are mild.
• Where the current tool fails entirely: the highest-risk frontier, like pancreatic and gastric cancer, where checkpoints have not worked, but a different mechanism might.

The lesson: a platform expands most safely from where it is already proven, into adjacent ground, then into the hard frontier, not by jumping straight to the hardest problem.

Mental Models & Frameworks

Orthogonal beats replacement

Moderna did not try to replace Keytruda, Merck's immunotherapy; it built something that works by a completely different mechanism, which is why the two combine well. Bancel's analogy: Keytruda "opens the gates and lets the dogs out," unleashing the immune system to attack somewhat randomly. The vaccine instead "teaches the dogs exactly what to look for." Because the two mechanisms are orthogonal, they are synergistic rather than competitive. The model for product strategy: a new product that is orthogonal to a strong incumbent can partner with it and ride its distribution, where a direct replacement would have to win a head-to-head fight.

Buy from the store, do not carve your own knife

Asked whether patients or hospitals will "go founder mode" and make personalized medicines themselves, Bancel expects almost all to come to Moderna, for the same reason you buy a knife instead of carving one in a cave: once a high-quality, validated industrial supplier exists, DIY is only worth it when there is no store. For anything where safety, quality, and reproducibility dominate, centralized industrial production beats bespoke self-service, and people spend their freed-up time elsewhere. Useful when weighing whether users will really build their own version of your product, or just want the reliable one.

Trade-offs & Nuance

Quality now, efficiency later

The right sequencing depends on where the risk is. Moderna chose quality over efficiency in its first machines because the binding risk was scientific (does the therapy work?), not operational (can we make it cheaply?). Spending early effort on efficiency would have been optimizing the wrong thing, and a rushed machine could have poisoned the trial with a false negative. The nuance: this ordering is right when the core hypothesis is still unproven. Once the science was validated, the calculus flipped entirely and efficiency became the obsession. The trade-off is not "quality vs efficiency" in the abstract, it is which one carries the fatal risk at this stage.

Personalized power versus industrial safety

A therapy made fresh for each patient is powerful but carries risks a mass-produced drug does not: any injectable made one unit at a time raises the odds of contamination or a single-step manufacturing mistake, and injecting even one copy of a contaminant can seriously harm a patient. The counterweight is exactly the industrialization Bancel describes: validated, standardized, good-manufacturing-practice processes are what make bespoke production safe enough to trust at scale. The tension every personalization-heavy product manages: the more custom each unit, the harder consistent quality becomes, and the more the process discipline has to carry.

Practical Application

Test whether personalization is load-bearing

Before committing to a personalized product, measure how much genuinely differs per user versus how much is shared. Moderna's whole architecture is justified by one number: about 90% of targets differ per patient. If your equivalent number is small, a shared core with a thin customization layer may be enough. If it is large, plan for true batch-size-one from the start, because retrofitting personalization onto a generic core tends to fail.

Validate before you optimize the pipeline

When the core of your product is still unproven, build the supporting pipeline to be reliable and correct, not fast or cheap, and resist the urge to polish it. Reserve heavy optimization for after the central hypothesis is validated. Moderna's clunky-fridge-then-shrink sequence is the template: it avoids sinking years into efficiency for something that might not work, and avoids a shaky system producing a misleading failure.

Name your one or two cost levers

For an operations-heavy product, identify the small number of variables that actually drive unit economics and work those obsessively, instead of spreading effort thin. Moderna's are cycle time (42 days needle-to-needle) and physical footprint per machine. Write down your equivalents, then judge each proposed improvement by whether it moves one of them.

QA for reproducibility, not each output

If your product generates a unique result every time (personalized, or AI-generated), you cannot inspect each output against a fixed spec. Instead build your quality process, and your regulatory or trust story, around proving the system reliably turns the same input into the same trustworthy output. That is the "same sample in, same product out" bar Moderna has to clear with the FDA.

Make your failure cohort the next spec

Systematically capture the cases where your product did not work, with enough surrounding data to diagnose them, and treat that cohort as the primary input to your next version. Moderna is mining the 20% of non-responders, with their full samples and outcomes, to build algorithm 2.0. The discipline is retaining and studying failures instead of rounding them off.

Questions to Consider

• For a product you are considering personalizing, what fraction of the experience genuinely differs per user versus is shared, and does that fraction justify building for a batch size of one?
• Are you optimizing a pipeline or process for cost and speed before the core thing it supports has actually been proven to work?
• What are the one or two variables that truly drive your product's unit economics, and are you working those specifically, or improving everything a little?
• If your product produces a different output every time (personalized or AI-generated), how do you prove to a skeptical buyer or regulator that the system is reliable, given you cannot pre-approve each output?
• Do you capture and study the cases where your product failed for a user with the same rigor you study the wins, and does that failure data feed your next version?

Bottom Line

Moderna's personalized cancer vaccine works because personalization was treated as the architecture, not a feature: the product is an information file, the factory was made reliable before it was made efficient, the unit economics ride on a couple of named levers, and a regulator approves the reproducible process rather than each unique dose. Those operating moves, plus mining the failures for the next version, apply to any team building something bespoke or generative at scale.

Case Studies Mentioned

CAR-T therapy as the contrast

CAR-T is another personalized cancer treatment: doctors remove a patient's own immune cells, re-engineer them to recognize the tumor, and infuse the living cells back. Bancel uses it as the foil for Moderna's approach. Because CAR-T moves fragile living cells, its manufacturing is physically large, slow, and expensive. Moderna moves only information (a DNA sequence) and synthesizes the medicine chemically, so its footprint can shrink far more. The lesson is not that one is better, but that whether your unit of value is physical or informational dictates how small and fast the operation can ever get.

Concepts to Explore

Process approval versus product approval

The regulatory model for products where every unit differs. Rather than approving each output, the regulator approves the manufacturing process and its demonstrated reproducibility. Worth understanding for anyone building personalized or generative systems, since it is a real-world answer to "how do you certify something whose output you cannot inspect in advance?"

Orthogonal mechanisms and synergy

The idea that two interventions working through entirely different pathways can combine for more than either alone, and are complements rather than competitors. Relevant to product positioning well beyond medicine: building orthogonally to a strong incumbent can turn a would-be rival into a partner.

Notable Quotes

"The current version of [the algorithm] is the worst version you're gonna see for the rest of medical history." (Stéphane Bancel)

"You always learn more from things that don't work than things that work." (Stéphane Bancel)