The Case for Drug Abundance
Why the cost of making a medicine will soon stop deciding who gets cured
When I founded Valo Health in 2018, the core question we focused on was how do you know, as early and as cheaply as possible, whether a program is going to work? Most of the cost of a medicine is spent finding out the answer late. We had the view that human data, models trained on that data, and a continuous and integrated rather than campaign-based process could move the answer earlier. Those concepts have been percolating through the industry.
The cost of turning biology into candidates is falling, driven by a set of technologies that are already in use. As they mature together, we can arrive somewhere the industry has never been: a world of drug abundance, where the limit on what gets made is biology rather than budget; and imagination rather than capital.
Drug abundance is a simple idea with large consequences. Today a medicine is scarce the way a skyscraper is scarce: each one is a bespoke, decade-long, two-billion-dollar project, so the industry builds tens a year and primarily where the market can repay them. Abundance is the state in which a new medicine costs so little to attempt, and fails so early when it is going to fail, that the number of programs is set by the number of diseases worth treating rather than by the number of bets an industry can afford. Getting there requires a design cost that approaches zero, a laboratory that learns continuously, and a development process in which every step predicts the one after it.
The shape of scarcity
The cost to develop a drug is about $2.6 billion, according to the Tufts Center for the Study of Drug Development, which includes the cost of failures and the cost of capital over a development cycle that runs more than a decade1. Deloitte’s annual analysis of the largest pharmaceutical companies put the average cost per asset at $2.23 billion in 2024, with cycle times lengthening2. Across all therapeutic areas, a drug entering Phase I has roughly a 14 percent chance of reaching approval3. The single largest reason a program fails in Phase II or III is that it does not work well enough in people despite success in preclinical development, which accounts for about half of late-stage failures4.
Those numbers describe a specific kind of scarcity. There are more than 7,000 rare diseases, affecting an estimated 30 million Americans, and fewer than five percent have an approved treatment5. Talent and effort are not the constraint. Arithmetic is. When a shot on goal costs on the order of two billion dollars and succeeds less than one time in seven, an industry can only afford a few hundred shots a year, and it has to aim them at diseases where they can repay the cost.
The deeper issue sits underneath the headline figure. Most of the money in drug development pays for the time between a wrong hypothesis and the discovery that it was wrong. The industry has become very good at running the process. What it has lacked is the ability to see failure early and at low cost. That is what technology is starting to change.
Designing each step to predict the next
A development program is a sequence of gates: target selection, hit finding, lead optimization, preclinical safety, Phase I, Phase II, Phase III. Each gate breaks down into smaller steps that also tend to run one after another. Historically each gate has been designed to answer its own question. Does the molecule bind? Is it tolerated in a rat? Is it safe in healthy volunteers? Each answer is real, and each is only loosely connected to the question that matters, which is whether the drug helps the patient at the end.
The alternative is to design each step so that its output is predictive of the next step and of long-term success. The target is chosen because human genetics and human tissue data say the mechanism matters in this patient population, which is why genetically defined diseases tend to have a higher probability of success. The assay is chosen because it has been calibrated against clinical outcomes, so that a result in the dish carries information about a result in the clinic. The Phase I study measures a biomarker that reports whether the mechanism is engaged, so that a Phase II outcome becomes more predictable before Phase II begins. Every stage becomes a forecast of the stages after it, and every forecast can be checked and improved.
When each step predicts the next, failure surfaces early, where it is cheap, and capital flows to the programs most likely to reach patients. With only what we know today, and the tools and data that already exist, I believe we could triple the probability of success in drug development.
Six breakthroughs already in motion
Molecular design is progressively approaching zero marginal cost. Ten years ago, producing a set of candidates with usable potency, selectivity, and drug-like properties against a well-characterized target was a multi-year medicinal chemistry effort. Today a small team runs it in the background. Structure prediction models now cover proteins, nucleic acids, and their complexes with small molecules6. Generative design tools produce binders for targets that were long considered undruggable, design antibodies against epitopes chosen in software, and optimize a dozen properties simultaneously. The most striking results are in de novo protein design, where entirely new proteins are built to a functional specification and confirmed in the lab, work recognized with the 2024 Nobel Prize in Chemistry7. The number of molecules you can design and test increases by orders of magnitude at minimal cost. A program’s timeline is then set by how fast you can test, with design becoming a query you re-run every time you learn something new.
The laboratory is becoming a closed loop. The next generation of systems is the self-driving lab, where software proposes an experiment, automated hardware executes it, instruments read it, the model updates, and the next experiment is chosen, all without a person in the cycle. Autonomous chemistry platforms have synthesized and tested compounds end to end, and cloud laboratories let a company with no wet-lab footprint run thousands of experiments from a laptop. The design-make-test-learn loops can enable automated and accelerated learning that can even be extended across programs. These systems still work best on narrow, well-defined assays and remain expensive to build. But when the cost per experiment falls toward the cost of reagents, which is where the trend points, the number of hypotheses a program can afford to test grows again by orders of magnitude, and the closed loop becomes the default way a program learns.
Models of human biology are learning to predict humans. Animal models and immortalized cell lines are imperfect proxies for people, and the difference does not show up until Phase II, years and hundreds of millions of dollars later. The tools for measuring people directly are the ones that matter now. Patient-derived organoids and engineered tissues recapitulate disease biology well enough to screen against. Large perturbation atlases map how thousands of genetic and chemical interventions ripple across thousands of cell states. Single-cell and spatial profiling of patient tissue is accumulating at an immense scale. On top of that data, foundation models of the cell are beginning to predict the effect of a perturbation before anyone runs it. The regulator is moving in the same direction: in April 2025 the FDA published a roadmap to phase out animal testing requirements in favor of human-relevant methods, and a year later reported draft guidance for antibodies, a searchable database of accepted alternatives, and the first AI-based development tool qualified for regulatory use8,9. If human-relevant models keep improving, the gap between what we see preclinically and what we see in patients narrows, and the largest single cause of late failure starts to shrink.
The clinical trial itself is becoming a data problem. Trials are slow for reasons that have little to do with the drug. Eligible patients are found by clinic and by chance. Participation is a burden. Endpoints are coarse: a survival curve or a symptom score at twelve weeks. A mechanism that is working may be invisible to the endpoint, or may take years to declare itself. Each piece is being rebuilt. Health-system records and molecular diagnostics identify eligible patients by signature before they ever reach a site. Decentralized and hybrid designs bring the study to the patient’s home. Biomarker and digital endpoints report within days whether the drug is engaging its target. Software drafts protocols, forecasts enrollment by site, watches safety signals in real time, and simulates a study under alternative designs before the first patient is dosed. External control arms built from real-world data are replacing placebo groups in rare disease, and platform trials let one master protocol test many drugs against one disease on shared infrastructure. The revised international good-clinical-practice guideline, finalized in January 2025, explicitly accommodates decentralized elements and risk-based monitoring, and the FDA has issued draft guidance on the use of AI in regulatory decision-making10,11. When these pieces are assembled, a trial that enrolls in weeks and reads out on a mechanistic endpoint becomes ordinary, and the clinic becomes one more stage in the learning loop rather than the place the loop stops.
Manufacturing is following the same curve. Continuous manufacturing, modular biologics facilities, and cell-free protein synthesis are shrinking the fixed cost of making a drug. Programmable platforms, mRNA being the clearest example, let one production process serve many products, so a new medicine inherits a validated process instead of building one. If this holds, producing a drug for a few thousand patients becomes economically routine, which is the precondition for serving the long tail of disease at all.
The data is becoming a shared asset. Every one of the technologies above depends on high-quality, longitudinal human data linked to outcomes. That data is being generated faster than ever, through national biobanks, health-system consortia, patient registries, and the trials themselves. Federated learning and privacy-preserving computation let models train across institutions without moving patient records. When the human data layer becomes broad, deep, and connected, every model in the field gets better at once, and the predictive design of each step described above becomes possible at industrial scale.
A future of abundance
Picture these six innovations converging. A program begins with a mechanism identified in human genetics and confirmed in patient tissue. Design produces candidates overnight, having screened in software more molecules than the industry has synthesized in its entire history to this point. An autonomous lab tests the selected candidates against organoids that have been calibrated against clinical outcomes, and the platform’s models forecast, with stated confidence, how each candidate will perform in people, then optimized rapidly through a closed loop process. The best candidate enters a first-in-human study whose endpoint is the biomarker the model predicted, and the result confirms or corrects the forecast within weeks, against success criteria defined before the first patient was dosed, with predictivity prebuilt to enable higher-confidence design of the next human study. A Phase II study populated from health-system data enrolls in a month. Manufacturing inherits a validated platform process. The whole arc runs in a fraction of the time and cost of today’s, and, more importantly, the program’s chance of success was known and rising at every step.
Now multiply that across an industry. The number of programs in development is measured in hundreds of thousands rather than thousands. Most are small: a rare mutation, a molecular subtype of a common disease, a mechanism that matters in one tissue at one stage of illness. Each is viable because it costs a small fraction of what a program costs today, so prevalence stops being the gate on what gets made. The 7,000 rare diseases become a pipeline rather than a list. A diagnosis is followed by a match against a deep catalog of options, many tuned to the molecular signature of the patient in front of you. The words “no approved therapy” become the exception.
The industry grows in this world. What changes is its unit of production, from a handful of very large bets to a very large number of small, well-instrumented ones. Success rates rise because each step predicts the next. Capital that once paid for late failure pays for early learning. The economics start to resemble software more than oil exploration, and the people who work in drug development spend their careers on medicines that reach patients.
The technology doors opened by abundance
Abundance of candidates is not abundance of cures. Some diseases are hard because the biology is unforgiving; neurodegeneration, many solid tumors, and most of what we call aging will not yield merely because molecules are cheap. At least not at the stage where we currently catch them, once they are chronic and entrenched. Abundance offers many more attempts, at much lower cost, with far higher resolution on why each one fell short. What follows is to treat those diseases as long-running, portfolio-scale learning programs rather than single bets, where every failed candidate returns a measurement that improves the model for the next. That turns a decade of isolated failures into a decade of accumulated understanding.
Diagnostics and data science can also catch these diseases earlier, before they are entrenched, and a disease caught early is a different target from the same disease caught late. That is where the next generation of targets will come from.
Safety has to scale with throughput. A platform that produces a billion candidates a week can produce a billion toxic ones. Abundance without safety is only a faster way to hurt people. Safety science has historically advanced more slowly than efficacy science, in part because toxicity is rarer and harder to model. The same tools apply: human tissue models of liver, heart, and kidney, calibrated against clinical adverse events; toxicity prediction trained on the growing record of what has actually harmed people; and safety biomarkers built into early studies so harm is caught at the first patient rather than the thousandth. A mature abundance world is one where safety has been industrialized as thoroughly as discovery.
Access is a separate problem from supply. A hundred times more medicines does not mean patients receive them. Pricing, reimbursement, and distribution each have their own bottlenecks, and abundance stresses them harder. A payer cannot evaluate a thousand new therapies a year the way it evaluates fifty today. The solution is to move the evidence infrastructure to the payer: outcomes-based contracts that pay when the drug works, continuous real-world evidence that keeps updating the value of a therapy after approval, and diagnostic-first coverage that routes each patient to the therapy their biology predicts. Done well, this turns the payer from a gatekeeper into a participant in the same learning loop, and the medicine that works for a given patient becomes the medicine that is covered.
It is easy to imagine that with abundance comes extreme longevity. Maybe. Metformin is a longevity drug if you have diabetes, and chemotherapy is a longevity drug if you have cancer. Every cure is a longevity drug for someone.
We will progressively treat and ideally solve the diseases that are the demons of today. Do not be surprised when new ones take their place. The difference is that we will meet them with a better toolkit.
Where I see some of the investment opportunities today
The human data layer. Whoever assembles the highest-resolution longitudinal measurements of human disease, especially from the time of apparent health through the development of chronic disease, linked to outcomes and made computable, owns the ground truth every model is trained against. The quality, depth, and relevance of that data matters as much as its size.
The closed-loop platform. Discovery companies that run as learning systems, measured on cycles per week, cost per cycle, and improvement in predictive accuracy over time. These should be valued on the learning rate of the platform, and the lead asset is evidence of the platform rather than the whole thesis.
Calibrated translation. Companies whose product is the link between a preclinical measurement and a clinical outcome: assays and models validated against what actually happened in trials, sold as a way to know a Phase II result before running it. This is the practical form of designing each step to predict the next.
The integrated trial stack. Patient identification, decentralized operations, biomarker endpoints, external controls, protocol generation, and trial simulation are mostly sold today as separate services. The larger prize is integrating them so a study is designed, populated, and read out as one software-driven process.
Industrialized safety. Human-relevant toxicity prediction, organ-on-chip safety screening, and early safety biomarkers, offered as a platform that every discovery engine will need to buy. The value sits in making human tissue, human cells, and human patients measurable at industrial scale, and in the models that turn those measurements into predictions that hold up in the clinic. In the traditional world, the further a measurement sits from the whole human, the more error it lets into the system, but as we can enable better predictivity, we may be able to create highly relevant surrogates.
Long-tail portfolio structures. Once cost per program falls far enough, thousands of indications become viable businesses, and many fit better in a portfolio than in a standalone company: shared platform, shared regulatory and manufacturing infrastructure, and a stream of small programs financed as a pool, structured by design rather than by the accident of which science happened to work. Building that structure, and the financial products around it, is a new category.
Evidence-linked access. Infrastructure that lets payers and health systems evaluate, contract for, and monitor many therapies at once, with reimbursement tied to measured outcomes.
Why I am optimistic
We can now start to imagine a different world. A child is born with a mutation that today would mean a lifetime with no treatment and a family learning to live with it. In this world, the mutation is matched, potentially within days, to a mechanism and a program, and if no program exists, the cost of starting one is low enough that someone does. A person diagnosed with a common disease learns which molecular subtype they have and receives the therapy built for that subtype rather than the one that works on average. A patient with a disease we still cannot cure enrolls in a study that reads out in weeks, and whatever the result, the next attempt is better. A physician in a small clinic anywhere in the world opens a catalog rather than a short list.
That is what abundance means in practice. It is hope distributed to everyone rather than to the fortunate few whose disease happened to be common enough, or profitable enough, to be worth a two-billion-dollar bet. The technologies to build it exist. The work now is to connect them, and I cannot think of a better way to spend the next decade than building the tools that do it.
References
[1] DiMasi JA, Grabowski HG, Hansen RW. Innovation in the pharmaceutical industry: New estimates of R&D costs. Journal of Health Economics, 2016. https://pubmed.ncbi.nlm.nih.gov/26928437/ (see also https://www.nature.com/articles/nrd4507)
[2] Deloitte, Measuring the return from pharmaceutical innovation, 2024 edition. https://www.deloitte.com/us/en/industries/life-sciences-health-care/research/measuring-the-return-from-pharmaceutical-innovation.html
[3] Wong CH, Siah KW, Lo AW. Estimation of clinical trial success rates and related parameters. Biostatistics, 2019. https://academic.oup.com/biostatistics/article/20/2/273/4817524
[4] Harrison RK. Phase II and phase III failures: 2013–2015. Nature Reviews Drug Discovery, 2016. https://www.nature.com/articles/nrd.2016.184
[5] National Center for Advancing Translational Sciences, NIH. Our impact on rare diseases. https://ncats.nih.gov/research/our-impact/our-impact-rare-diseases
[6] Abramson J et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature, 2024. https://www.nature.com/articles/s41586-024–07487‑w
[7] The Nobel Prize in Chemistry 2024. https://www.nobelprize.org/prizes/chemistry/2024/summary/
[8] FDA. Roadmap to Reducing Animal Testing in Preclinical Safety Studies, April 2025. https://www.fda.gov/media/186092/download
[9] FDA. FDA Achieves Year 1 Goals in Reducing Animal Testing in Drug Development, April 2026. https://www.fda.gov/news-events/press-announcements/fda-achieves-year-1-goals-reducing-animal-testing-drug-development
[10] ICH E6(R3) Good Clinical Practice, adopted January 2025. https://www.ich.org/page/efficacy-guidelines
[11] FDA. Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products, draft guidance, January 2025. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/considerations-use-artificial-intelligence-support-regulatory-decision-making-drug-and-biological
Recent Insights
