--- title: "AI can design a drug in 18 months. Why aren't there any on pharmacy shelves yet?" date: 2026-10-06 category: AI in Action site: NeuroAI canonical: https://neuroai.site/a/na-app-ai-drug-discovery language: en --- # AI can design a drug in 18 months. Why aren't there any on pharmacy shelves yet? > China's AI drug-discovery leaders Insilico Medicine and XtalPi have gone public and built real pipelines — but the industry's hit-rate is still unproven. A drug candidate normally spends four to six years just reaching the clinic. In a lab in Shanghai, a team watched one go from a blank sheet to a testable compound in eighteen months — then sent it into real patients. The question hanging over the celebration was simple: was that speed, or a head start that biology will still erase? China's AI drug-discovery sector has moved from demo videos to balance sheets. Two names carry the argument: Insilico Medicine (英矽智能) and XtalPi (晶泰科技). Both are now public companies with real pipelines, real revenue partners, and real skeptics. ## Two bets, one question Insilico Medicine, founded in 2014 with roots in the United States and a headquarters in Hong Kong, builds generative-AI platforms that nominate disease targets and design molecules. XtalPi, founded in 2015 out of work at MIT, pairs quantum-physics simulation with robotic wet labs to predict molecular behavior and run experiments at scale. Neither sells a finished drug. Both sell a claim: that machine intelligence compresses the costliest, slowest steps of pharmaceutical R&D. ## Insilico's pipeline and its record IPO Insilico went public on the Hong Kong Stock Exchange on 30 December 2025 under code 03696.HK. The listing raised gross proceeds of HK$2.28 billion (≈ US$293 million), netting about HK$2.03 billion (≈ US$260 million) after expenses — the largest Hong Kong biotech IPO of that year, according to the exchange's own allotment document. Cornerstone investors included Eli Lilly, Tencent, Temasek and Schroders. The company says its Pharma.AI platform has produced more than 30 programs, of which 10 candidates have received clinical-trial (IND) clearance. Its lead asset, Rentosertib (ISM001-055) for idiopathic pulmonary fibrosis, became the first AI-discovered drug to post a peer-reviewed clinical proof-of-concept in a Phase IIa trial, published in *Nature Medicine* in 2025, and is now advancing toward Phase III. Insilico also points to partnerships as validation: out-licensing deals with Exelixis and Menarini worth up to US$2.1 billion in potential value, plus co-development work with Sanofi, Lilly and Fosun Pharma. ## XtalPi's robots-in-the-loop model XtalPi listed on the Hong Kong exchange on 13 June 2024 under code 2228, the first company to debut under the exchange's Chapter 18C specialist-technology regime. Priced at HK$5.28, it raised net proceeds of about HK$900 million (≈ US$115 million). Where Insilico bets on end-to-end discovery, XtalPi sells a hybrid: physics-based computation plus a fleet of automated lab robots that test thousands of compounds. The company states it serves more than 300 biopharma and research clients worldwide, including 16 of the top 20 by 2022 revenue — names such as Pfizer and Johnson & Johnson appear in its disclosures. ## The honest number no one celebrates Both companies cite the same headline efficiency: traditional early discovery averages about 4.5 years, while Insilico reports compressing target-to-candidate work to 12–18 months. That is a real, documented acceleration of the *discovery* phase. But discovery is the cheap half of the problem. The industry's open secret is the hit-rate. As of 2025–2026, very few AI-originated candidates have reached late-stage trials, and none has become a broadly approved, marketed therapy. Most AI-designed molecules are still in Phase I or preclinical. Attrition — the normal death of drug candidates in larger, stricter trials — has not been abolished; it has been deferred to the most expensive stage. According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, the bottleneck for AI drug discovery is no longer finding a candidate but proving it survives large, rigorous trials — where the old rules of biology still apply. ## Why the bottleneck moved, not vanished What AI changed is the front end. Generating a plausible molecule is now largely a computation problem. Proving that molecule is safe, effective and manufacturable in humans remains a biology problem with decade-long timelines and billion-dollar price tags. Investors are betting the faster front end produces more shots on goal; patients will only benefit if some of those shots actually land. The fair claim is narrow but real: AI has demonstrably shortened how long it takes to get a credible drug candidate into the clinic. It has not yet demonstrably shortened the road to an approved medicine. ## Honest limitations - The 12–18 month discovery timeline and the 4.5-year traditional baseline are company-stated figures, not independently audited benchmarks; they describe Insilico's own programs, not the whole field. - "More than 30 programs" and "10 IND clearances" are drawn from company disclosures and Chinese financial press; we did not verify each program's stage against a clinical registry. - The HK$2.28 billion Insilico IPO and HK$900 million XtalPi net proceeds are taken from HKEX listing documents and exchange filings; exchange rates are approximate (≈ US$1 = HK$7.8). - Claims that "no AI-designed drug is yet marketed" reflect the public record through 2025–2026 but could shift as late-stage trials report. - Rentosertib's specific trial numbers are covered in a separate NeuroAI article and are intentionally not restated here. ## What readers can do now - **Track Phase III, not the IPO** — a 52-week, several-hundred-patient readout on Insilico's lead asset is the number that will actually validate or bury the AI-discovery thesis. - **Separate "discovered by AI" from "approved by regulators"** — when a company says AI designed a molecule, ask what was de-risked (target + structure) versus what remains (safety at scale, efficacy, reimbursement). - **Watch the robot labs, not just the models** — XtalPi's automated wet-lab throughput, not its algorithms, is the closer analog to a real industrial moat; monitor client count and repeat contracts as the truer signal. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-app-ai-drug-discovery Free to quote with attribution and a link to the original.