AI Can Design a Drug. It Still Can’t Prove It Works.

On October 9, 2026, AI drug discovery stocks on China’s A-share market jumped in the afternoon. Berry Genomics hit the daily limit, its third limit-up in six trading days. Tainuomaibo rose more than 10%. Jinshi Pharmaceutical, Tellgen, InnoCare Pharma, and Sinocelltech climbed sharply.

That same day, Anthropic was reported to have built a wet lab in the San Francisco Bay Area. It uses Claude to drive robots through real biological experiments.

A few days earlier, Anew Labs, spun out of ByteDance, closed a $290 million first round at a $1.5 billion valuation.

Capital is rushing in. Big tech is entering. Policy is backing it. All of it circles back to one date.

I. The cracks in the double-ten law

Traditional pharma runs on a brutal rule: ten years, a billion dollars, and a clinical success rate below 10%.

Where does the money go?

Drug discovery comes first. Identify a target. Validate its link to the disease. Then find the one molecule among hundreds of millions that works. Researchers synthesize and test in the lab, over and over, like looking for a needle in a haystack. In the traditional model, this takes months or years.

Insilico Medicine’s Rentosertib went from target to preclinical candidate in 18 months. The traditional average is 4.5 years. That is a time cut of more than 60%.

Guangzhou Pharmaceutical Holdings, working with Stonewise on small-molecule drug discovery, compressed early-stage work from one or two years into three to six months. R&D costs dropped by 70%.

A Southwest Securities report gives a broader number. With AI, drug discovery is compressed by 70% to 90%. Preclinical and clinical stages each save more than half their time. Total R&D costs fall by about 50%.

$54 billion.

That is the money AI saves at the front end. Instead of human screening, AI “draws” molecules. It uses massive biomedical datasets to generate molecular structures or protein sequences on a computer. Work that took months now takes seconds to hours.

But a molecule generated on a computer is just a blueprint.

To become real, the blueprint goes to a wet lab. There it is synthesized, expressed, and tested. Can it bind? Does it work? Is it safe? Designing a blueprint may take seconds. Making it and running a test can take weeks or months.

The data then has to flow back. Whether an experiment succeeds or fails, the real-world results are fed into the AI model. Each cycle adds data and improves predictions. R&D shifts from one-way trial and error to a loop that keeps iterating.

That loop is why AI pharma has completed only the first half of the journey.

II. Three types of players, three ways to survive

The AI pharma supply chain has three kinds of companies.

The first are upstream AI technology companies. They draw the blueprints.

Insilico Medicine is the fastest mover. Its Pharma.AI platform covers target discovery through clinical prediction. It has software licensing deals with 13 of the world’s top 20 pharmaceutical companies, including Servier, Eli Lilly, SK Biopharmaceuticals, and Takeda. Since the start of 2026, publicly disclosed partnership value has approached $10 billion.

Last fiscal year, its revenue was $56.24 million.

Its core pipeline, Rentosertib, has an AI-discovered target. Phase IIa data showed patients in the 60 mg dose group had an average lung capacity improvement of 98.4 mL. The control group declined by 20.3 mL. It was the industry’s first clinical proof of concept for AI drug discovery. In July 2026, Rentosertib entered a Phase III trial in China. It plans to enroll 320 participants.

Anew Labs is taking a different path. It positions itself as end-to-end drug discovery, with self-developed small-molecule candidates. Its $290 million first round valued it at $1.5 billion, only months after it was spun out of ByteDance.

The second are midstream CRO and CDMO service providers. They build from the blueprints.

They do not bet on whether a pipeline succeeds. They care how many molecules need validation. The more candidate molecules upstream AI companies design, the more orders they receive.

They sell shovels to prospectors.

Twist Bioscience’s AI pharma orders have doubled for three straight years. About $25 million in 2025, about $50 million in 2026, and a target of $100 million in 2027. Swiss startup Adaptyv Bio grew revenue roughly tenfold in 18 months. After GenScript partnered with Eli Lilly’s TuneLab, AI-driven protein orders are expected to keep growing at triple-digit rates.

Novo Nordisk teamed up with Anthropic. Protein candidate sequences generated by the Claude Science model go to external providers for wet-lab validation. Eli Lilly’s TuneLab platform joined GenScript, Twist, and other providers to build a service system covering protein validation, data testing, and antibody characterization.

The third are downstream traditional pharmaceutical companies. They handle industrialization and clinical implementation.

Fosun Pharma completed the V2.0 upgrade of its PharmAID system in the first half of 2026. It covers decision management, early research intelligence, clinical research, and integrated data. As of June 30, two new molecules generated or structurally optimized with AI had entered the PCC stage. Fosun formed a strategic partnership with Insilico Medicine in 2022. In 2025, it pushed Henlius to team up with DP Technology on AI R&D for antibody drugs and ADCs.

Hengrui Pharmaceuticals invested 4.605 billion yuan in R&D in the first half of the year, 29.8% of revenue. AI now runs through target discovery, large-molecule and peptide design, and small-molecule generation and optimization.

Almost no company can cover the entire chain alone. Upstream and downstream collaboration is the norm. Overlapping business lines make competition inevitable.

III. The wall at Phase II

Phase I asks if a drug is safe. Phase II asks if it works. Phase III asks if it beats existing treatments.

AI has lifted Phase I success rates from the traditional 40% to 65% range to 80% to 90%. Molecules are cleaner. Toxicity predictions are more accurate. AI has helped on safety.

At Phase II, the success rate is about 40%. The traditional method is about 37%.

There is no statistically significant difference.

The problem is not molecule design. AI can judge whether a molecule is safe and whether it can be absorbed. Whether a drug works depends on whether the target hypothesis holds. Most Phase II failures happen because the biological mechanism has not been proven.

AI cannot answer that yet.

Verge Genomics’ ALS drug VRG50635 was discovered through an AI platform. It was discontinued after Phase I due to insufficient efficacy. BenevolentAI’s BEN-2293, developed with AI assistance, failed to meet its primary endpoint in Phase II.

As of August 2026, no drug fully discovered by AI has received full FDA approval. The expected date for the first AI drug approval has been pushed back three years in a row.

AI moves the bottleneck downstream. It does not remove it. It improves molecule screening and compresses preclinical research. But whether a drug works in the human body still depends on our understanding of disease biology.

AI can design a perfect molecule. If the target is wrong, the perfect molecule is useless.

IV. The 2029 bet

Everyone is watching the same date.

Insilico Medicine’s Rentosertib is expected to read out Phase III top-line data in 2029. Co-CEO Ren Feng offered a judgment. If Phase III succeeds, “this will be the DeepSeek moment for AI pharma.”

This matters for more than one company. The entire industry needs hard evidence that AI can work in new drug R&D.

Policy is also betting on it. China’s “15th Five-Year Plan for Pharmaceutical Industry Development” lists AI pharma as a key task for the first time. It sets a target that first-in-class drugs account for no less than 25% of the global total by 2030. Innovative drug industry revenue is to grow at no less than 20% annually. It is the first time China has set a quantitative target for its global share of innovative drugs.

The “Pharmaceutical Industry Digital and Intelligent Transformation Promotion Center” was jointly initiated by CAICT, demand-side companies including Hengrui, Fosun, CSPC, and China Resources Pharmaceutical, and supply-side companies including Huawei and XtalPi. Its goal is to tackle pharma companies’ problems of “not daring to transform, not knowing how to transform, and being unable to transform.”

A Guoyuan Securities report estimates the global market for AI drug discovery software and services at about $4.85 billion in 2026. It could reach $18.5 billion by 2030, a compound growth rate above 40%. China’s intelligent pharmaceutical market is expected to exceed 500 billion yuan in the same period.

The next three to four years are a procurement window for pharmaceutical companies.

Whether that window pays off depends on the data readout in 2029.

V. The first half of the disruption

Back to the opening question: is AI disrupting the pharmaceutical industry?

It already is. Only at the front end of R&D.

It has turned drug discovery from looking for a needle in a haystack into following a map. It has connected dry labs and wet labs, at least in part. It has cut trial-and-error costs. It has shown capital that faster, cheaper drug R&D is possible. It has forced traditional pharma to rethink its R&D systems.

AI has not proven it can systematically improve Phase II success rates. It has not achieved large-scale, standardized dry-wet loops. No drug fully discovered by AI has completed Phase III and won approval.

The dry-wet loop cannot be scaled by a handful of companies. It is a shared problem for the entire industry.

In 2029, Rentosertib’s Phase III data will be a key milestone. If it succeeds, it will be AI pharma’s “DeepSeek moment.” If it fails, the industry will have to re-examine the true boundaries of AI in drug R&D.

The direction will not reverse.

AI is pushing pharma from experience-driven to data-driven, from one-way trial and error to closed-loop iteration. The first half of the race is about speed. The second half is about staying steady, completing the loop, and feeding real-world data back into models that keep evolving.

2029 is the date that decides which half we are in.