A neural network designed a drug and Lilly paid $2.75 billion for it
Last Saturday, Eli Lilly wrote a $115 million check to a company most people have never heard of. The full deal is worth up to $2.75 billion. What did they buy? Drugs designed by a neural network. Not assisted by AI. Not "AI-enhanced." Designed end-to-end by generative models that identified the disease target and then invented the molecule to hit it.
The company is Insilico Medicine, based in Hong Kong. They've been at this since 2014. For most of that decade, the pitch was theoretical. Not anymore.
The first AI drug that actually worked in humans
In June 2025, Insilico published results in Nature Medicine that made people pay attention. Their drug rentosertib, designed entirely by AI (both target and molecule), hit its primary endpoint in a Phase IIa trial for idiopathic pulmonary fibrosis. Patients on the highest dose saw lung function improve by 98.4 mL. The placebo group declined by 62.3 mL.
That gap matters. Pulmonary fibrosis kills. Treatment options are thin. And the drug that showed real efficacy in actual human lungs was drawn not by a medicinal chemist but by a generative model trained on molecular data.
From target identification to Phase I took 30 months. Traditional drug discovery averages 10 to 15 years and north of $2 billion before a single patient is dosed.
Why Lilly is betting billions on this
Lilly isn't doing this out of curiosity. They're sitting on the most valuable drug franchise in the world right now. Tirzepatide (Mounjaro and Zepbound) is projected to be the #1 selling drug globally in 2026, with revenue around $45 billion. The GLP-1 market overall is approaching $100 billion.
That kind of money attracts competition. Novo Nordisk's semaglutide is right behind them. Dozens of biotechs are racing to develop oral alternatives — because right now, most GLP-1 drugs require injections, and the company that cracks a better oral formulation wins a massive chunk of that market.
The Lilly-Insilico deal specifically targets oral therapeutics. Lilly gets access to Insilico's Pharma.AI platform and an exclusive worldwide license for certain preclinical oral drug candidates. They're using AI to find the next generation of drugs before their competitors' chemists finish drawing the first molecular structure.
The part that should make you skeptical
No AI-discovered drug has received FDA approval. Not one. The pharmaceutical industry's roughly 90% clinical failure rate hasn't budged. AI compresses the early stages, but biology doesn't care about your algorithm's confidence score.
Recursion, another AI drug discovery company, discontinued its lead candidate REC-994 after long-term data didn't support continued development. DSP-1181, another AI-designed molecule, was shelved after Phase I despite a clean safety profile. Speed to the clinic doesn't mean speed to the pharmacy.
There are currently 173 AI-discovered drug programs in clinical development. Maybe 15 to 20 will enter pivotal trials this year. The forecast from researchers who actually work in this space: validation and disappointment in roughly equal measure.
Where this actually gets interesting
The headline is "$2.75 billion AI drug deal" and that's dramatic enough. But I think the real story isn't one deal. It's the cost curve.
Traditional drug development costs over $2 billion per approved drug. AI-assisted preclinical work is showing 30-70% cost reductions through virtual compound screening and predictive modeling. McKinsey puts the total pharma value at $60 to $110 billion annually.
That math reshapes who can develop drugs. Right now, only mega-cap pharma companies can absorb the cost of a $2 billion R&D failure. If AI brings that number down to $500 million — still enormous, but manageable for mid-size biotechs — then the number of entities capable of discovering drugs multiplies. More shots on goal means more drugs that actually work. Eventually.
This is the Jevons paradox again, just like we saw with AI compute last week. Make drug discovery cheaper, and you don't get less of it. You get more. A lot more.
What to watch
- Insilico's Phase IIb trial for rentosertib is the single most important data point in AI drug discovery right now. If it holds up in a larger trial, the field is validated. If it doesn't, expect a cold winter.
- Oral GLP-1 competition is getting crowded. Novo Nordisk's oral semaglutide hit 26,000 prescriptions within two weeks of its January launch. Lilly is using AI to find something better. The oral GLP-1 race is where AI drug design meets the biggest pharmaceutical market on earth.
- The 90% failure rate is the number to watch. If AI-discovered drugs start clearing Phase II at higher rates than the historical average, that's the signal. Everything else is marketing.
- 2026 is the proving year. After a decade of promises, the first wave of AI-designed drugs is hitting late-stage trials. By this time next year, we'll know whether AI drug discovery is real or whether it's just a faster way to fail.
A neural network drew a molecular structure. That structure improved lung function in dying patients. And the company that makes the world's best-selling drug just paid $2.75 billion for the system that drew it. The question isn't whether AI belongs in drug discovery anymore. The question is what happens when it gets good at it.
References:
1. Eli Lilly reaches $2.75 billion deal with Insilico to bring AI-developed drugs to the global market
3. AI Drug Discovery 2026: 173 Programs, FDA Framework & Market
4. AI Drugs Reach the Clinic: 2026 Is the Year of the First Large-Scale Test