01 Trigger Event
a16z published an article in 2026 discussing the possibility of Big Pharma and Big Software achieving valuation parity. Its central figures were that Eli Lilly's market capitalization had exceeded $1T; more than 300 drugs worldwide were expected to generate annual sales above $1B; and fewer than 10% of drugs entering human clinical trials ultimately receive approval.
More important for AI builders, a16z has invested in 75+ life-sciences startups since 2015, including insitro, Formation Bio, and BigHat Bio. Its thesis is that AI models for science are improving drug-development success rates while shortening development cycles and reducing costs.
But this is not a model launch; it is a bet. AI is moving beyond a software tool and into high-value, long-cycle regulated workflows.
02 What This Really Means
The question is not whether AI can help pharma run fewer experiments, but whether it can change the return structure of drug-development assets.
Pharma has traditionally been a hits business: R&D spending is heavily front-loaded, failure rates are extremely high, and a small number of successful drugs generate most of the revenue. After approval, it becomes an annuity business, with patents, clinical data, physician prescriptions, and manufacturing capabilities collectively forming a moat.
AI's strategic significance is that it attacks both ends:
- At the front end, it can improve the odds of target validation, molecular design, clinical-trial design, and candidate selection.
- At the back end, it can improve commercialization through better companion diagnostics, drug discovery, and patient matching.
- Across the timeline, it can compress a ten-year development chain, increasing capital turnover.
This is what a16z means by “Big Pharma Can Rival Big Software.” What is being repriced is not a lab SaaS contract, but how many approved drugs emerge per dollar of R&D and how long each approved drug generates cash flow.
The concentration figures cited in the article are revealing: the top 25% of drugs generate roughly 85% of revenue; among the top 10 pharma companies, the top three assets account for about 50% of revenue; and 63% of Eli Lilly's 2025 revenue came from its GLP portfolio.
If AI genuinely improves hit rates, it is not raising an average. It is increasing the probability that a small number of critical clinical programs clear their gates. For large pharma companies with vast data and capital, this creates leverage. For small AI-native drug companies, it may make developing their own drugs viable for the first time.
That said, I have not run these science models internally and cannot conclude that the article's optimistic claims have already been realized. It currently offers an industry logic and investment thesis, not prospective-trial success-rate data.
03 Historical Analogy / Structural Comparison
This resembles how AWS transformed software companies from 2004 to 2009, rather than the terminal replacement associated with the iPhone in 2007.
The iPhone changed the user interface and distribution; AWS changed software companies' fixed-cost structures and capital allocation. It allowed more companies to shift from building heavy assets first to renting computing capacity and validating demand first.
AI for Pharma could have a similar effect. Traditional pharmaceutical companies build laboratories, clinical networks, and manufacturing systems first. AI-native companies may instead assemble a lighter R&D network from models, automated wet labs, and external infrastructure.
Pharma, however, is harder than software to “rent” infrastructure for. A drug cannot win through faster deployment or benchmarks alone. Clinical endpoints, regulatory approval, patent protection, and manufacturing scale are all indispensable.
The historical analogy also has a counterexample. Database compression was once expected to eliminate database companies; instead, it simply enabled more applications to be built on top of them. AI is unlikely to eliminate Lilly. It is more likely to give it more data, stronger models, and faster iteration.
The real structural competition is therefore not AI drug companies replacing Big Pharma. It is two organizational forms competing for the same high-value targets:
- Big Pharma has real-world data, clinical execution, distribution, and a strong balance sheet.
- AI-native biotech has model speed, experimental feedback loops, and lower organizational friction.
The former has a deep moat but slow decisions; the latter iterates quickly but has weak validation capabilities. Ultimately, excess returns will go to whoever can marry AI's speed with pharma's execution system.
04 What It Means for AI Builders
The immediate adjustment is not to register a drug company, but to reassess the productization boundaries of AI for Science.
1. Do Not Sell Model Access Alone
The model access offered by opc.club lets external teams invoke general-purpose models, but pharma's core value does not lie in which model can be called. Customers will pay a premium for better decisions after models form a closed loop with experimental data, instruments, lab workflows, and clinical evidence.
A pure model wrapper can easily be flattened by the platform capabilities of OpenAI, Anthropic, and Google. A more defensible position is to own the data interface for a specific assay, an experimental feedback loop, or a regulatory workflow.
2. Prioritize High-Value, Low-Latency Decisions
Drug development is not suited to an entry point such as “AI writes a research summary.” More viable entry points include candidate ranking, response prediction, trial-site selection, and patient stratification.
The criteria can be simple:
- How much experimental money does the current decision waste when wrong?
- Is the decision cycle measured in weeks or years?
- Can outcome data be obtained to validate the model?
- Is the error reversible?
A mistaken selection that wastes $1M, takes only weeks to decide, and produces results that can feed back into training is better suited to productization than generating an unvalidated report.
3. Data Protocols May Be Closer to the Moat Than Models
Switching costs in Science AI come not only from fine-tuning, but from instrument calibration, sample metadata, experiment records, failed samples, and reproducibility across laboratories.
Developer-tooling competition will therefore enter a narrower but deeper domain: whoever controls the lab data schema may control the model-training and validation flywheel. Models can be replaced; accumulated data loops are much harder to replicate.
4. Be Conservative About Model-API Consumption
In the near term, drug discovery will still use a mix of foundation models, specialized models, simulations, and experimental results. AI-for-Science narratives are not a reason to migrate all production traffic immediately.
I would treat AI-for-Science traffic as evaluation workload: compare models' actual costs in structured reasoning, tool use, long-context retrieval, and code execution while preserving multi-model routing to avoid dependence on a single provider.
The value of specialized models may be underestimated here. If customers will pay a substantial premium for a particular assay or molecule class, a general-purpose API is not the only procurement model; proprietary deployment and a data flywheel may be more valuable.
05 Counterargument / Risks
I may be overestimating AI's impact on drug-development economics, at least in the near term.
First, the article's most important claim—“AI models for science are getting better, fast”—lacks rigorous evidence. Better public benchmarks do not imply better randomized clinical outcomes, much higher approval rates, or fewer clinical failures. Between molecule generation and an approved drug lie biological failure, clinical failure, regulatory risk, and manufacturing execution.
Second, cheaper and faster candidate molecules may expand the pipeline without increasing successful drugs. Grove's survival logic applies: if development costs fall by 50% while candidate volume rises by 100%, the total number of failed projects increases. AI may make pharma better at producing hypotheses without making it better at proving them.
Third, synthetic data creates a self-reinforcement problem. Models train on existing literature and experimental data, generate more similar candidates, and then absorb the resulting experimental data. If the data distribution does not genuinely expand, AI may scale low-quality hypotheses rather than create new biology.
Fourth, AI-native drug companies face harsher economics than software companies. Software can retry through subscriptions many times; biotech has longer capital cycles and little runway left after failure. Better model capabilities may not compensate for financing, patent, and clinical-execution risks.
Fifth, a16z is both an investor and an industry advocate, so its case selection naturally favors the idea that AI is improving pharma. The article discusses a portfolio and market structure rather than prospective statistics tracking specific projects from AI prediction to clinical success.
I therefore would not treat Eli Lilly's $1T+ valuation or GLP-1 growth as evidence that AI has rewritten pharma. These trends are consistent with the industry's existing hits-business logic. The real inflection signal would be a reproducible, control-relative improvement in clinical success rates for AI-assisted projects in prospective studies, with development-cycle gains not offset by lower-quality candidates.
Until such data appears, AI for Science is more like a long-term call option. Models can be sold, tools can be built, and workflows can be captured, but treating AI as validated pharma operating leverage is premature.