该产业的支出正迅速加速：在今年前 4 个月，Eli Lilly、Sanofi、Novo Nordisk 以及其他大型药厂签署了超过 50 项 AI 授权交易，几乎与 2025 年全年一样多；若里程碑达成，潜在付款总额将超过 30 亿美元。早期证据喜忧参半。FDA 从未核准过 AI 设计的治疗药物，一些候选药在试验中失败，而阿兹海默症等疾病也仍未出现重大突破。不过，AI 正透过分析蛋白质资料库并使用 AlphaFold 等工具，加速早期工作；Insilico Medicine 表示，一个项目从构想到临床前研究只用了大约 9 个月，而通常需要 3 到 5 年。
目前的资料显示，AI 也许在最早期阶段最有帮助，但未必在后期也是如此：Boston Consulting Group 于 2024 年对 100 多家聚焦 AI 的生技公司所做的研究发现，第一阶段试验成功率超过 80%，高于历史上的 40% 到 60%；但第二阶段成功率只有大约 40%，大致与产业平均相当。研究人员警告，疗效仍然难以预测，而生物学的复杂性也限制了 AI 能解决的问题。有些公司已经遭遇挫折，包括 Recursion Pharmaceuticals 砍掉 20% 的员工，且尚未将任何药物推进到后期试验；而 Generate Biomedicines 则停止了一项 Covid-19 计划，但如今在一项重度气喘药物的后期测试中，该药物可能在 2028 年回报结果，并有可能成为首个获 FDA 核准的 AI 设计药物。
Drug development in the US is slow and expensive: it typically takes about 10 years and more than $1 billion to bring a medicine through clinical trials, and roughly 90% of candidates fail before approval. That backdrop helps explain why pharmaceutical companies are turning to artificial intelligence to make discovery faster and more predictable, especially as executives such as Anthropic CEO Dario Amodei argue AI could help tailor treatments to a patient’s DNA and eventually transform cancer care.
The industry’s spending has accelerated quickly: in the first 4 months of this year, Eli Lilly, Sanofi, Novo Nordisk and other large drugmakers signed more than 50 AI licensing deals, about as many as in all of 2025, with potential payments exceeding $30 billion if milestones are met. Early evidence is mixed. The FDA has never cleared an AI-designed treatment, some candidates have failed in trials, and major wins in diseases such as Alzheimer’s remain absent. Still, AI is speeding early-stage work by analyzing protein databases and using tools such as AlphaFold, with Insilico Medicine saying one program moved from idea to preclinical studies in about 9 months instead of the usual 3 to 5 years.
Data so far suggest AI may help most in the earliest phases but not necessarily later ones: a 2024 Boston Consulting Group study of more than 100 AI-focused biotech firms found first-phase trial success above 80%, versus a historical 40% to 60%, but second-phase success was only about 40%, roughly the industry average. Researchers warn efficacy is still hard to predict, and biology’s complexity limits what AI can solve. Some companies have already hit setbacks, including Recursion Pharmaceuticals, which cut 20% of its workforce and has not yet advanced a drug into late-stage trials, while Generate Biomedicines halted a Covid-19 program but now has a severe asthma drug in late-stage testing that could report results in 2028 and potentially become the first FDA-approved AI-designed medicine.