Where the value lands
This library covers AI and the life sciences end to end: challenge competitions where AI proposes the designs and wet-lab experiments act as the referee, organoids, CROs, the lab-supply upstream, and the capital behind innovative drugs. One question runs through all of it: when AI rewrites biology, which layer captures the value?
The five judgements below come from reading all of that research together. Each pulls out a pattern you cannot see in any single article but that runs through the whole body of material, and each comes with one chart. They were written in Oct 2026.
The source flags every figure as either consistent with the original document, or estimated, second-hand or company-reported. Here the second kind is marked in the text with "(est.)" or "(company-reported)", and with an asterisk (*) in the charts.
01 · AI made the cheapest step cheaper
AI has cut the cost and the number of attempts needed to produce a drug candidate by an order of magnitude. It has not changed whether a molecule works once it is inside a human body, and that is where the money and time in the chain actually sit. The most immediate effect so far is not that downstream work has gone quiet. Downstream is busier.
AI drug discovery means squeezing a 10-year R&D cycle into a few years.
The step being compressed is the first one, which is also the cheapest and shortest. Phase II and the length of clinical trials, which set the cost and the timeline, have not moved. And because there are more candidates, orders for validation and development have risen across the board over the past two or three years.
The money goes on failure: ~8% overall success, 6–9 years in the clinic
Here is the public breakdown for a drug's path from target (the protein or gene the drug acts on) to approval. Phase I, a safety study in healthy volunteers, succeeds 52% of the time. Phase II, which tests efficacy in patients, succeeds 29% of the time. Phase III, large-scale confirmation, succeeds 58% of the time, and regulatory review 91%. Multiplied together, that is an overall success rate of about 8%. Most of the money is spent on failures, and most of the time is spent in the clinic, which takes 6–9 years. Another measure tells the same story: the average R&D cost per approved drug has risen to $2.67B, and the combined success rate from Phase I to approval is about 7%. Keep this yardstick in mind. What makes a drug expensive and slow is the right-hand half of the ladder.
Discovery is compressed by an order of magnitude: 2,500–5,000 compounds → ~250
At the discovery end, AI's impact is real and thorough. One AI drug company says it now synthesizes about 250 compounds per programme, against an industry norm of 2,500–5,000 (company-reported). Another reached a development candidate (a molecule ready to enter preclinical development) with just 78 molecules. Mechanically, computation today acts as an "enrichment filter". Traditional high-throughput screening, where automated equipment tests hundreds of thousands of molecules in one run, usually has a hit rate below 1% (self-reported in the material; est.). Docking against ultra-large compound libraries (computing how a molecule fits its target) has shown measured hit rates of 11–24%, and AI-designed protein binders 9–88%. Raising the hit rate by one or two orders of magnitude cuts the number of molecules to synthesize and test by the same amount. One line in the source is worth copying out: computation changes the economics of experiments; it does not replace them.
But that step is only ~2.7% of cost, and AI molecules pass Phase II ~40% of the time against a historical 37%
So how much of the chain's money sits in the compressed step? From target to hit (the first molecules that show activity), each marketed drug carries about 24 programmes, about $24M in total. That is only about 2.7% of the out-of-pocket R&D cost per marketed drug ($873M). What AI saves here is people and months. What decides success sits downstream. One big pharma's review of 142 programmes found that 65% of Phase II failures were due to lack of efficacy: the molecule was made, but most failed because the target was poorly understood.
More important still, trust in computation decays step by step up the ladder. Structure prediction is now part of the training recipe. At the binding-pose level, no deep-learning docking method beats the classic tools. At the affinity level (how strongly a molecule binds), docking scores are good only for enrichment: the top-scoring bin has a hit rate of about 25%, the bottom bin 0%, and the absolute values are unreliable. At the cell-function level, neither foundation models nor deep models beat simple "no effect" or "additive" baselines. In other words, AI is most confident where it is furthest from the patient.
Clinical evidence agrees. Of 117 AI-derived assets that entered humans, 60 completed Phase I and only 8 completed Phase II. Phase I success is 80–90% (historically 52%). Phase II success is about 40%, against a 37% historical baseline in the same study; the 29% quoted above comes from a different sample with a different basis. At the gate that actually decides success, there is no difference. No AI-discovered drug has been approved yet.
Downstream is busier over the past 2–3 years: CDMO +72.7%, tox new orders +98%
The counterintuitive side effect is that, over the past two or three years, downstream has become busier, not quieter. With more candidates, orders for validation, toxicology, CMC (process and manufacturing) and production have all risen. One leading CDMO (contract development and manufacturing organisation) grew its development and manufacturing business 72.7% in H1 2026. A leading drug-safety-assessment company's new orders rose 98%. A leading US preclinical company's book-to-bill ratio hit 1.19, a four-year high.
To be honest about it, part of this is the biotech funding recovery and the metabolic-drug cycle, and the source material already strips out the share that cannot be attributed to candidate volume (inference; est.). As for the story that "AI compresses 10 years into a few", that is the field's own narrative, not evidence. What the evidence currently supports is only the first, cheapest, shortest step.
The sample is tiny (only 8 AI assets have completed Phase II), so the conclusion could flip at any time. The strongest evidence for better Phase II success today comes from human genetics, not generative models. "Downstream is busier" describes only the past 2–3 years.
What to watch
- Whether the cumulative number of AI-derived assets completing Phase II rises well above 8, and whether their success rate moves away from the ~40% historical average.
- A leading AI drug-platform company's Phase II readout on Nov 2, 2026: a single-point test of whether the "build your own pipeline" route can pay off in one go.
Sources (library notes): Frontier model companies building their own drug R&D: the six-step ladder and Phase II evidence · The target-to-binder simulation toolchain and sim2real · Simulation and synthetic data: the fidelity ladder · LLM opportunities in pharma R&D: substitution and creation · Accelerating drug validation: breaking down success rates · AI penetration and demand shifts by stage · Virtual cell reviews compared · Recursion analysis
02 · Biology has data. It lacks outcomes
Public, free biological data with hundreds of millions of entries is so abundant that nobody treats it as a moat. What is scarce is paired records labelled with outcomes, the "what happens if we do this" data. Each record needs someone to pay for an experiment, and these records reach models through only three narrow gates: evaluation, post-training alignment and federated fine-tuning.
Bio AI is a data-buying race: whoever stockpiles the most, or can afford the most, wins.
By record count, the mix is public databases as the base, model-generated data multiplying that by one or two orders of magnitude, and a tiny amount of ground truth to finish. In pre-training, ground truth is a tiny share of records but carries the most weight. The whole field has only two publicly verifiable precedents of paying cash for data, and both were existing clinical data, not experiments run for training.
The free layer records shape: 241M predicted structures vs 261K solved ones
Biology really does have a lot of data. There are 149 million auto-annotated protein sequences, but only 0.4% have been reviewed by hand. There are 241 million model-predicted protein structures, but only 261,000 experimentally solved ones, a gap of three orders of magnitude. Public single-cell data (gene expression measured cell by cell) has reached about 218 million cells, also free. These are the pre-training base (learning general patterns from massive data first) for every foundation model, and the library's verdict is blunt: "free, and not a moat".
Every outcome costs money: from $149 per protein, and only ~2,154 human answers
Sequences, structures and expression profiles describe what a molecule looks like. What models really need to learn is what happens if you do something: does this molecule bind, what did this perturbation (deliberately changing one gene) change, is this compound toxic in an animal. Every paired record labelled with an outcome costs someone a payment.
Public prices exist only in a narrow band. Binding assays priced per protein start at $149, including two replicates, with a standard three-week turnaround. A five-assay ADME panel (absorption, distribution, metabolism, excretion) costs $199 (company-reported). Further down, in living animals, one cynomolgus monkey for research is estimated at $10,600–28,200 based on case-by-case purchase prices (est.). Drug labels tied to human clinical outcomes number only about 2,154 at the most public liver-toxicity prediction company (one company, one endpoint). By comparison, one model-generated record works out to a few cents to a few dollars (est.).
Ground truth enters models through only three gates: evaluation, post-training alignment, federated fine-tuning
In the library's words: "a tiny share of records, but fine-tuning and evaluation carry the most weight". Public databases do hold existing experimental results. The affinity data (binding strength) used by publicly documented models comes to about 1.2 million records, all curated from the literature into public databases, all noisy, and none of it commercial. New ground truth, made specifically to probe a model's weaknesses, has only three ways into a model:
- Evaluation. One public closed loop for binders (designed proteins that attach to a target protein) produced 1,320 designs and 354 true binders, 26.8%, assayed independently by two outside organisations.
- Post-training alignment (correcting a trained model's direction with a small amount of high-value data). One protein model reported that "larger models respond more strongly to alignment on lab data" (company-reported).
- Federated fine-tuning (data stays with each company; only model parameters are exchanged). Five pharma companies each contributed thousands of experimental structures, and the work was done in under 10 weeks.
All three gates come after pre-training, and all are counted in thousands. Stockpiling data does not buy a higher model ceiling.
Not buying data but trading it: only two public cash-purchase precedents
There are only two publicly verifiable cases of paying cash. One pharma company bought data from a clinical-data company for $42M over three years. The same data company charged another pharma company $200M in data licensing plus model-development fees. Both involved existing clinical and multimodal data, not experiments run for training.
What replaces buying is trading. One big pharma opened its model to partners for free; the entry ticket is at least 1,000 experimental endpoints contributed per institution per year. How the work is funded decides whether the data ends up a public good or a private asset. Similar perturbation data can be company-funded and then released as CC0 (a public-domain waiver), funded by government and industry and then published, or circulated only inside a members-only consortium and kept out of every open model's training set.
The free layer has access barriers and systematic bias (about 79.7% of experimental structures come from X-ray crystallography). "Nobody buys data" only means nothing shows up in public material. The 2,154 figure is one company and one endpoint, and cannot be extrapolated.
What to watch
- Whether a third verifiable precedent of paying cash for biological training data appears, and whether it covers experiments run for training rather than existing clinical data.
- Whether perturbation atlases and in-vitro toxicology, two kinds of outcome data, get a public unit price, meaning the supply of top-layer ground truth starts to have a market price. Section 04 looks at the same thing from the seller's side.
Sources (library notes): Public databases and the literature · AI4S data hub overview · Prices, costs and mixes · Model training data recipes · Data vendors and policy · Producing wet-lab ground truth · Validation loops and feedback from self-driving labs · Axiom Bio research
03 · The higher the score, the more it matters who set the exam
To judge whether a scientific AI scorecard is credible, ignore the score and check two things. Was the ground truth produced by an experiment run after the prediction? And are the people who set the questions the same people being graded, or selling the service?
Public benchmarks and public competitions give an objective yardstick, so scores can be compared side by side.
The most mature, highest-scoring public benchmark was audited and found to leak widely. And the most impressive "prospective" competitions in new fields are often set, graded, run in the lab and sold as a service by the same company. The sign of a healthy leaderboard is the opposite of what you might expect: whether human experts are still winning.
Ask only two things: was the answer made after the prediction, and who is grading?
When you see a scientific AI scorecard, skip the score and ask two questions. First, was the exam's "right answer" (the ground truth, the experimental result a model is compared against) produced by an experiment run after the prediction? If the answer predates the prediction, the model may be reciting rather than reasoning. Second, are setting the questions, grading, running the experiments and selling the service all in the same hands? These two questions are enough to place any leaderboard. Laid out this way, the cell for "experiment after prediction, grader separate from graded" holds little more than a few long-running blind tests.
Retrospective benchmarks leak widely: of 22 endpoints, only 3 models pass every check
Leakage in retrospective benchmarks (where every question comes from published historical data) is not a guess. It has been audited. A review of 51 benchmark configurations, including a 22-endpoint drug-property library and a 9-property molecular set, found conflicting labels and train–test leakage in most of them. Several leading co-folding models (which predict the protein and the bound ligand structure together) were retested on about 2,600 protein–ligand systems released after their training cut-off; performance fell as similarity to the training set fell. After re-splitting the data by protein domain, one class of deep-learning docking methods dropped to a 7.1% success rate. Ablation experiments (removing inputs one at a time to see whether the model can still predict) went further: with the protein or ligand input removed entirely, some models performed about as well. That is memorisation, not learning interactions.
The sharpest example is in the most mature direction. On the public drug-property (absorption, metabolism, toxicity) leaderboard, inhibition-prediction AUC (a discrimination metric where 1 is perfect) has reached 0.90–0.92. Yet a 2026 critical review found that, across 22 endpoints, only 3 models passed every check for reproducibility, leakage resistance and overfitting.
In new fields the problem is structural: setter, grader and lab are often the same company
In new fields the problem is not cheating; it is structure. In protein design and virtual cells, the two newest arenas, the benchmark targets, the standard readout, the hit definition and the lab work are often held by one company. A 7-target binder-design benchmark, a measured protein library with 1,000+ proteins at launch (company-reported), and all the wet-lab ground truth for the competition come from the same lab, which is also the validation vendor, the competition host and the benchmark provider. In the virtual cell challenge, the evaluation set was produced by the organiser, and partway through the scoring rule was found to be gameable (predicting the mean was enough to push the error down), so a special prize was added.
That does not make these untrustworthy. Publishing everything including negative results, running a fully open track as a control group, and deliberately picking "fresh targets" to prevent memorisation are real checks. But it does mean what you should look at is the governance, not the score. A word on the yardstick itself: in one set of competition data, the same round's hit rate is 14% using expressed proteins as the denominator, and only about 3% once designs that failed to express are counted too. The same report gives an overall 26.8% alongside group-level figures of 26.7% and 35.1%. So this page makes no side-by-side comparison of hit rates.
The healthy signal is counterintuitive: human experts still win (novel RNA targets 0/42)
Only a few setups are structurally clean: a long-running blind test whose targets have no published experimental structure yet, a weekly blind assessment of structures about to be released, and a third-party multi-stage competition that publishes all results, negatives included, after a three-month embargo. The most reliable health signal is counterintuitive. In RNA 3D structure, across all previously unseen natural targets, not a single prediction reached a TM-score of 0.8 (a structural-similarity score where above 0.8 usually counts as a correct fold). Human experts still clearly beat automated methods. The material treats this as a selling point rather than a weakness, because it shows the questions are real and have not been gamed yet.
To be clear, several hit rates and public prices cited elsewhere on this page come from the same "vendor who also sets the exam" structure. This page uses them only for what they show about structure (public prices exist, and the bottleneck in turnaround is the wet lab), not to argue how strong any model is.
Vendor-set exams have also produced rare public releases of negative data. Leak verdicts depend on the similarity threshold chosen. Prospective blind tests have few questions and long cycles; "clean" and "useful" are a trade-off.
What to watch
- Whether the first protein-design or perturbation-prediction benchmark appears that is third-party run, preregistered and cross-validated across multiple labs. This kind of independent check is clearly missing today.
- Whether the science-agent benchmark reported as near saturation (92% on the Verified subset, vendor self-reported) is overturned by an independent review or replaced with a new question set.
Sources (library notes): Benchmarks, evaluation and data formats · Validation loops and feedback from self-driving labs · Protein binder challenges and wet-lab platforms · RNA structure prediction challenges and evaluation
04 · Only what can be ordered per run gets paid
AI multiplies demand for experiments. Whether that demand turns into revenue depends on whether the experiment has been standardised into "fixed input, fixed output, public price, orderable by software". Thirty years and five generations show one pattern: selling results wins, selling machine time loses, and the ceiling is a legally required signature.
AI-driven automated labs are something new, and the robots are what matter: whoever has the most robotic arms, the lab that looks most like science fiction, catches the demand AI creates.
The most sci-fi generation, the cloud lab, has failed most often, while chemical e-commerce ("order, receive a parcel"), which worked back in the 1990s, has been steadily profitable. The fastest growth today is in interfaces that charge per experimental result. What decides success is how standardised the work is, not how advanced the technology is. Profit is not in unit price either, which keeps deflating. It is in being hard to replace, and in one step's output automatically becoming the next step's order.
Results win, machines lose: cloud labs failed most, pay-per-result validation grows fastest
Lay "experiment outsourcing" out over time and five generations appear.
| Generation | Model | Evidence |
|---|---|---|
| 1. Chemical e-commerce | "Order, receive a parcel", working since the 1990s | One on-demand synthesis supplier lists 9.45 billion orderable structures as catalogue products, 3–4 week delivery, >80% success. Steady double-digit growth and profitable since the 1990s. |
| 2. Service marketplaces | Brokering between people | Annual transaction volume above $1B, +66% year on year. But it still matches people with people; for an automated caller it is "a window you can email, not an interface you can call". |
| 3. Cloud labs | Renting a whole automated lab by the hour | The most sci-fi and the most failures: one ended public subscriptions in 2023 and moved to private deployments; another has 80% of capacity tied to one university. |
| 4. Pay-per-result validation APIs | Upload a design; machine-readable measurements come back per run within days | One provider's throughput grew more than 5× in a year and revenue about 10× in 18 months (company-reported). List price from $149 per protein. |
| 5. Aggregator layer | Labs become plugins inside software | A lab that is not on the plugin list is invisible to the caller. |
The ceiling is legal sign-off: GLP tox and similar steps can never be bought per run
The dividing line across the five generations is not technical sophistication but standardisation. The material gives six criteria: the input can be fully described in a file; the output is a number; orders are small and frequent; the process is fixed; turnaround is short; and there is no legal signature or regulatory liability. Testing stage by stage, in-vitro safety panels (hERG, CYP and other toxicity pre-screens) "meet every criterion on paper, but nobody has built a window yet". GLP toxicology (formal toxicology studies for regulatory filings), process development and clinical execution will never become windows, because the study director is the legal signatory, and China is not a member of OECD Mutual Acceptance of Data (under which member countries accept each other's GLP study data). That is the root reason they "can't be bought per run", and it is the horizontal line in the chart.
Unit prices deflate, profit comes from lock-in: price-compared consumables 39% GM vs process-embedded 79%
The first thing orderability brings is a price war. Western CROs (contract research organisations) quote $2,000–5,000 for an ADME panel (in-vitro tests of absorption, metabolism and other pharmacokinetic properties); Ginkgo's package is $199 per panel (company-reported). Per-base gene synthesis prices fell about 22% in each of two consecutive years. That leaves profit in only two places.
- Chain expansion. GenScript's integrated "gene to protein" workflow now makes up nearly 65% of its life-science services revenue (company-reported), and gene-synthesis orders flowing into that platform rose 160%. The company attributes this segment's growth to structural demand driven by AI. The pattern: make one step's output automatically become the next step's order.
- Being irreplaceable. Generic plastic consumables, which buyers can compare unit by unit, have seen gross margin squeezed to about 39% and net profit fall 55% year on year. Research antibodies and recombinant proteins, with SKUs so fragmented that nobody compares prices, earn 75–90% gross margin (a range because accounting standards differ). Best of all is being written into a manufacturing process or tied to an instrument: chromatography resins (packing for purifying biologics) earn 79% gross margin, with commercial-stage projects at 25% of that category's revenue, and the single-cell sequencing leader gets more than 90% of revenue from consumables at 70–74% gross margin.
Company evidence: Schrödinger lives on subscriptions; Ginkgo's bespoke foundry fell from $478M to ~$80M annualised
Company-level evidence agrees: selling results wins, selling machines and bespoke work loses. Of four AI platform companies, Schrödinger is the only one that has made it work. 78% of its revenue comes from software subscriptions (annual contract value about $208M, 96% gross retention, meaning the share of last year's customers still paying this year). Its exposure to drug outcomes is carried by equity and milestones in partners: three partners were acquired, cashing in more than $240M in total, more than its drug-discovery revenue in any single year. From 2021 to 2025 it spent about $500M building three clinical programmes of its own; one was terminated and two went looking for partners. At the other end, revenue from Ginkgo's bespoke foundry model fell from $478M to about $80M annualised.
Turnaround is a lock too. Design has been squeezed to 24–48 hours, but wet-lab validation still takes weeks (about three weeks as standard), so the wet lab sets the tempo of each closed loop. Cloud labs fail precisely because they sell the slowest step as "machine time", while buyers want results, not machines.
Organ-on-chip, the arena with the strongest policy tailwind, has actually seen VC fall 78%; orderability is necessary but not sufficient. Pay-per-result validation is still a small base and prices are rising. Experiments that need judgement will never be bought per run.
What to watch
- Whether a third per-run-priced, software-orderable validation provider appears (in-vitro safety panels are the most likely candidate), and whether the two existing providers keep raising unit prices or start cutting them.
- Whether the first organ-on-chip qualification for a defined context of use comes through. If it does, the whole arena's valuation anchor will shift from "tool company" back to "regulatory platform", and the gap between hot policy and cold capital will close.
Sources (library notes): Machine-orderable experiment windows: five generations over thirty years · DMPK and safety assessment: market and entry paths · Data hub: validation loops · Four AI platform companies compared · Schrödinger analysis · Ginkgo analysis · The "salt" of drug R&D: ranked summary · Nucleic-acid inputs · Protein and enzyme reagents · Plastics and automation consumables · 10x Genomics · Three protein-binder validation providers compared · Five organ-on-chip companies · Organoids, organ-on-chip and AI computational toxicology · Organoids: US and Chinese labs and startups · Internal research note (history of Chinese CROs: clinical CRO gross margins)
05 · China makes the molecules, the US sets the price
On the same industrial chain, China has won the middle: engineering, capacity and speed. The US holds both ends: basic science, high-end instruments, the paying market and capital. So for the same molecule, China keeps very little cash. And capital markets price the same China in two opposite ways: a premium on the tools layer, a discount on the asset layer.
Seeing "China out-licensed well over $100B, and eight of the world's top ten licensing deals", people conclude China already leads the value chain. Or, conversely, that the "China discount" is a single, uniform phenomenon.
What is booming is the sale of half-finished goods: upfront payments are a single-digit percentage of total deal value. And the discount holds only at the drug-asset layer; at the tools and A-share layer it is a premium. Together they run exactly opposite to where value actually flows.
Output is catching up: $135.7B out-licensed, ~49% of the global total
Start with output. In 2025 China approved 76 Class 1 innovative drugs (new molecules never marketed anywhere in the world), up from 48 the year before. About 4,751 innovative drugs are in development, roughly a third of the global total (est., by pipeline count). There were 157 out-licensing deals worth $135.7B in total, about 49% of the global figure. Chinese assets' share of global deal sourcing rose from 17% in 2020 to 31% in 2025. On these numbers alone, China looks like it has taken the head of the table.
The moat is in the middle: discovery costs 20–30% of US levels
The moat is real, but it is in the middle. By BCG's 2026 estimate, molecule discovery in China costs 20–30% of the US level, and preclinical development less than 50%. Outsourcing quotes are 30–50% of the leading Western providers', with timelines at 60–70%. Taking one bispecific antibody (an antibody that binds two targets at once) to a development candidate costs about $7–10M and 24–36 months in China; in the US it costs 3–4× more and takes 12–24 months longer. The mechanism is physical: labour-rate gaps, capacity right next door, cheap space. One leading CRO's chemistry business alone has about 18,000 people, roughly 8–10× a big pharma's internal discovery team (the latter estimated).
But the cash does not stay: upfronts are ~5%, university assets zero
But the cash does not stay. Upfront payments are only about 5% of total deal value (all-China, 2025; 5.8% in H1 2026, against a median of 6.7% and a value-weighted 8.1% across 30 core deals by big US pharma; the three measures should be read separately). The rest is milestones and royalties, which depend on success and arrive years later. Upfronts for Chinese assets are 30–50% lower than for Western assets at the same stage, and historically were 60–70% lower.
The sharpest contrast is Akeso and Summit Therapeutics. Summit holds only the overseas rights to one bispecific, with a market cap of about $13.5B, roughly equal to Akeso (about $12.6B), which holds the home-market rights, its entire pipeline and royalties. Most of the value of the overseas rights went to the buyer. Further upstream it is emptier still. At the 2025 out-licensing table, worth well over $100B, every seller was a pharma or biotech company; university assets were zero. All of Chinese academia sells an estimated RMB 2–4B of biomedical IP a year, less than a tenth of the upfront licensing payments pharma companies receive in a year (about RMB 50B). Only 11 of the 76 approved drugs have a new target or new mechanism.
Not China vs US but "everyone else vs the US": the US is 53% of global drug spend
This is not just a bilateral China–US issue; it is "everyone else vs the US". From 2019 to 2025, 66 of 67 European biotechs listed outside the EU. Japanese investors sum up the model as "Japanese science, American companies". Singapore's version is "R&D at home, capital and exits abroad". In all four places the common exit is selling assets into the US market or to US buyers. At list prices, the US accounts for 53% of global drug spending (China about 20–23% of the US figure), and 70–73% of global biopharma VC is in the US. Pricing power and the exit door are both at the right end. However cheap the middle is, it just sends more half-finished goods across.
Two opposite prices: drug assets at a 30–50% discount, A-share tools at a premium
The most counterintuitive half is the bottom of the chart: the market prices the same China in two opposite ways. The asset layer is systematically discounted 30–50%, and the median 10-year annualised return of listed Chinese innovative-drug companies is +9.7%, against +15.3% for the US sample. The tools and platform layer, by contrast, trades at a premium. Five A-share life-science upstream companies grew revenue 10–39% over the past year, while their share prices rose 56–141%, mostly from multiple expansion rather than earnings. Of 209 companies listed in both Shanghai/Shenzhen and Hong Kong, 203 (97%) are pricier as A-shares, with a median A/H price ratio of 1.71. Two Chinese AI drug-platform companies are each worth more than three first-generation US software companies combined (Sep 2026 prices, not re-sourced, and mixing Hong Kong and A-share market factors, so not purely a generational or national difference; est.).
To be precise: US peers are also in a "flat earnings, high share price" position, so this is not "China expensive, US not". It is an extra premium on China's tools layer relative to peers, plus the closed-market A/H premium. Put another way, the market pays a premium for the middle's "shovels" and a discount for the "drugs" the middle produces, exactly opposite to where value actually flows.
The discount is narrowing (average upfront up from $102M to $141M). Of 12 in-licensing deals by buyers, only 2 made money; this is a power-law ledger. Geopolitical shocks cut both ways. The middle's cheapness is only labour and space.
What to watch
- Whether the upfront share of China's out-licensing deal value leaves the 5–6% range. This is the most direct reading of whether "selling half-finished goods" is changing.
- Whether a second case appears of a big US pharma acquiring a Chinese biotech outright at a premium. Today there is only one verifiable example (about $1.2B, 2024).
- Final rulings on the relevant lists and bans, and whether a new control list appears.
Sources (library notes): Why innovative-drug value growth differs between China and the US · A+H share price ratios · Innovative drug R&D at Chinese universities and pharma · China's innovative-drug ecosystem and the value of university assets · How startups do wet labs: China vs US · Staffing and workflow in wet-lab stages · Equipment and reagent purchasing and market share in drug R&D · Seven A-share upstream companies in depth · One-year results and share prices of US and Chinese lab instrument and consumables companies · Drug R&D simulation software vendors · Pharmaron vs WuXi AppTec · Organisational boundaries between drug companies and CROs · Global biopharma investors overview · European VC · Japan and Korea · Singapore biomedical ecosystem overview · Returns on US pharma in-licensing of Chinese assets (internal research note) · Prices, costs and mixes (lab-animal purchase prices)
The library's sections
One thread runs through all the research: in bio, the real moats are manufacturing processes and validation loops, not design itself. The library is organised into five main sections plus background reading.
| Section | Notes | What it covers | Starting points |
|---|---|---|---|
| AI drug discovery | 16 | Starts from a master table of US and Chinese drug-discovery companies (AI-native, non-AI-native, "shovel sellers", NewCos). Then looks at how these companies set up their wet labs, followed by deep dives on shovel sellers and platform companies, frontier model companies and computational toolchains. | Master table of US and Chinese drug-discovery companies · Wet-lab view: how drug-discovery companies build labs · How new drug startups run wet labs: China vs US |
| Validation and alternative methods | 40 | AI designs, the wet lab referees: how public challenges and validation providers work and what an in-house validation lab costs; the regulatory turning point for organoids, organ-on-chip and AI computational toxicology (NAMs), with company-by-company checks of 30+ firms in China and abroad; and the commercial reality of brain–computer interfaces and in-vitro neuron "wetware" computing. | Adaptyv protein-binder challenges · RNA structure prediction challenges and automation · Absci & Claude |
| CRO / CDMO | 11 | The global and Chinese CRO/CDMO landscape, the division of labour between drug companies and CROs, which stages see rising demand as AI spreads, and the share-price performance and growth drivers of the leaders. | Pharmaron vs WuXi AppTec · Kangjun Capital and Tonghe Yucheng · CROs in the US, Europe and Southeast Asia |
| Lab tools and upstream | 41 | The "shovel-selling" side: scientific instruments, reagents and consumables, the "salt" inputs of drug R&D, gene synthesis and single-cell, and lab automation. | US and Chinese lab equipment, reagent and instrument makers · Leading general lab equipment and consumables makers · Listed lab hardware, instrument and consumables companies |
| Innovative drug industry and capital | 38 | China's innovative-drug ecosystem (universities, hospitals, pharma, capital), why innovative-drug value growth differs between China and the US, Singapore's biomedical ecosystem, and profiles of about 290 global biopharma investors (VC, public-market, CVC, state and sovereign, large asset managers) by segment, with cross-regional comparison. | China's innovative-drug R&D and lab supply chain · Innovative drug R&D at Chinese universities and pharma · China's innovative-drug ecosystem and the value of university assets |
| Background reading | 1 | A short history of biology and its major breakthroughs. | History of biology and key results |
Recent updates
The twelve most recent additions, all added on Oct 19, 2026 to the Innovative drug industry and capital section, map biopharma investors worldwide.
| Note (all added Oct 19, 2026) |
|---|
| Global biopharma investors: overview and cross-regional comparison |
| China's tier-one generalist and specialist biopharma VCs |
| China's second-tier and niche specialist biopharma VCs |
| Chinese state funds, government guidance funds and pharma corporate venture capital (CVC) |
| China's new-wave VCs and generalist VC life-science teams (AI + bio, synthetic biology, brain–computer interfaces) |
| Chinese public-market (A-share and Hong Kong) biopharma investors and fund managers |
| US company-creation and early-stage biotech VCs |
| US crossover, late-stage and growth life-science investors |
| US tech-background generalist VCs with bio/health teams, plus AI × bio specialists |
| US public-market biopharma specialists (hedge funds, long-only funds, activists) |
| Big-pharma corporate venture capital (US, Europe, Japan) |
| European life-science VCs |
These are personal research notes. Views and figures are sourced and rated for reliability within each note, all judgements stand as of the date written, and nothing here is investment or R&D advice.
Adapted from the Spark Science research notes (originally published in Chinese at jimmyresearch.com). Last updated Oct 24, 2026.