2026-10-09商业判断Youcheng Li · 利友诚

AI 药物许可卖什么？

从英矽与 Lilly 的交易和 Isomorphic 的战略，拆解软件、研发交付与资产权利，解释证据交接、里程碑、收入确认和项目组合怎样决定商业持续性。

AI 药物研发公司的模型变得更强之后，客户究竟为什么付钱？一个分子生成得快，是技术能力；一套结果能被另一支团队接手并继续开发，是交付能力；某些权利能被明确许可，是交易对象。从技术能力到药物资产，中间隔着实验依据、权利边界与接手后的开发风险。这几个接口，决定了平台能力能否变成可持续的商业价值。

本文以 Isomorphic Labs 2026 年 9 月 29 日的战略文章为近期切入点，对照英矽智能与 Lilly 在 3 月 29 日披露的合作，以及英矽 8 月 26 日发布、覆盖截至 6 月 30 日六个月的中报。旧公告用于核对交易结构，不作为今日新闻。公司战略和技术案例属于公司自述；合同披露与财务数字也必须按各自口径阅读。近期战略文章；港交所合作公告；正式中报。

一、三种买卖对象，不能共用一套增长逻辑

先把“AI 药企签了合作”拆成具体合同对象。软件授权让客户在约定范围内使用工具；研发服务承诺一组研究活动或成果；药物资产许可则让对方在特定范围内使用相关资产权利并继续开发。三者可以并存，也可以组合在同一协议里，不能画成必然从低级到高级的升级路线。

合同对象	客户买到的主要内容	规模化需要检验什么
软件使用权	按范围、期限和条款访问软件或模型能力	真实使用、续约、支持成本与客户自行完成验证的能力
约定的研究交付	针对明确问题的研究工作、数据或候选结果	交付是否可验收，新增项目需要多少科学与实验人力
药物资产许可	按领域、地域与义务界定的资产使用权，以及约定交接内容	证据能否接手，权利是否可许可，后续开发与付款条件是否推进

英矽与 Lilly 的公告正好说明了组合合同：一部分是针对某些适应症、处于临床前开发阶段的新型口服疗法，授予全球独占的开发、制造和商业化许可；另一部分是围绕 Lilly 选择的靶点进行多项研发合作。两部分不能全部说成“卖软件”，也不能全部归为某个现成药物资产的出售。许可授予的是约定使用权，不必然转移所有知识产权的所有权、全部训练数据或平台模型。公开公告没有给出完整合同，不能自行补上这些条款。来源：2026-03-29 公告。

这种区分也解释了客户关系的不同。软件可以在约定范围内被多位客户使用，但同一资产的同一独占权利不能无限重复许可。研究合同可能在不断接新项目中增长，却同时增加交付负担。资产合同可以把此前投入积累的成果变成一次交易，但持续增长需要新的可交接项目或既有项目继续达到条件；模型本身可复用，并不自动让每个资产的收入都可复用。

二、模型分数怎样变成可接手的证据包

技术背景不是一句“生成式 AI 更快”。药物设计同时处理许多约束：对目标的活性、对其他靶点的选择性、体内暴露、毒性风险、合成与制剂的可行性。模型可以提出或排序候选；若排名依据与实验体系不匹配，更高的计算分数也可能只把错误方向推进得更快。买方最终关心的是哪部分不确定性已经被数据缩小，哪部分仍要自己承担。

一个便于检查的交接流程有四步。首先，把目标、适应症和需要达到的性质写清楚，否则不同任务的“成功”无法比较。其次，让预测对应到有方法、对照、重复与原始记录的实验，区分计算值、细胞结果和动物结果。再次，检查这些结果在接收团队的条件下是否能重现，并把失败结果与适用范围一并保留。最后，将数据、材料、方法、相关知识产权及使用权限映射到合同承诺。每一步解决一个接口，不能用一张最佳曲线代替整套交付。

这里的“证据包”是本文的分析框架，不是说上述公司的公开合同逐项采用同一清单。药物研究阶段和交易内容不同，尽调范围也不同；化合物身份与合成路线、测定条件、暴露与安全边界、权利来源等，要按项目具体判断。拿到了专利或独占许可也不能直接推定自由实施不受其他权利限制。证据可复核、材料可追踪、许可权限清楚，才使另一支团队能够评估继续投入。

FDA 对临床前与临床研究的说明有助于划清科学阶段：临床前研究回答包括毒性在内的基础安全问题，不能替代人体研究。本文不把这些美国监管说明当作某项全球合同的完整合规标准，而用它们提醒读者：发现、临床前、临床和上市验收各回答不同问题。FDA：临床前研究；FDA：临床研究。

Isomorphic 9 月 29 日的文章描述了把数据、模型、软件与物理设施放在一起的垂直整合方向。其分子搜索展示之后仍要选择化合物、合成并在实验中验证；文章说临床前数据正在帮助公司准备临床开发。因此，计算搜索的时长不能当作端到端研发周期，更不能直接解释为临床疗效或整个项目组合的成功率。这个案例支持“公司在连接设计与验证”的战略判断，尚不能独立证明 AI 贡献了多少临床成功。来源：公司战略与示例。

原创交易机制示意，非实验或交易效果数据。软件使用权、约定的研究交付、药物资产权利与证据包是三种并列的合同对象，不是必然升级的顺序。图中只展开药物资产许可：证据和知识产权尽调核对证据、权属与许可权限；合同定义许可范围、地域与双方义务；买方按合同分工进行下游开发、制造和商业化。三类付款分别审视：首付款需核对支付条件与实际到账；开发、注册和商业里程碑取决于约定触发条件；销售提成取决于约定的销售口径与条款。标题总额不能直接视为现金到账或当期收入。签约、交接与履约、会计收入确认是不同判断，现金和收入的时点分别核对。技术效率可能增加资产选择，价值兑现仍依赖证据、权利与下游进展；更快发现不意味着临床成功。箭头表示信息与合同分析流程，不是任何公司的实测结果或必然商业结果。 — 原创分析示意，非实验数据或任何公司的已实现交易结果。三类合同对象彼此并列，图中仅展开药物资产许可；许可的范围、地域、义务和下游分工以合同为准。首付款、附条件里程碑与销售提成需分开评估，不能把标题总额直接当作现金或当期收入。签约、交接与履约、收入确认并非同一事件；技术效率与临床成功也不能画等号。

三、许可交易重新分配工作，也重新分配上行收益

为什么大型药企可能愿意许可一个临床前项目，而不只采购软件？这是一个能力配置问题。自己使用工具，仍需组织科学问题、实验、知识产权和项目管理；购买可继续开发的权利与交接内容，则可能把一部分早期探索打包交给外部团队。买方可以将其与内部项目、其他许可机会和外包研发比较，而不是只与另一个模型的预测分数比较。

许可方也面对选择。较早许可，可能更早获得约定付款并由合作方承担部分下游工作；保留自有项目，可能保留更多后续价值，却需要更多资本、临床执行与风险承受能力。这不是“许可必然更划算”的结论：保留权利、双方继续履行的义务、成本分担和付款触发点都会改变结果。药物合同不能默认把全部风险转给买方。

Isomorphic 在 2024 年 1 月 7 日宣布与 Lilly、Novartis 的战略研究合作；到 2026 年 9 月，公司的战略文又强调垂直整合和推进潜在药物进入临床。两种路径能够并行，不能据此说公司已经完全放弃合作，也不能说自有项目已经完成临床验证。2026 年 5 月 12 日公司宣布的 21 亿美元 B 轮融资，是融资资本，不能与客户许可收入或已实现药物销售混算；其公告把资金用途指向引擎与管线扩展。早期合作公告；融资公告。

我的趋势判断是有条件的：如果 AI 使提出候选方案的成本下降得快于可靠验证与交接的成本，竞争会更重视能重复产出可继续开发项目的团队。实验组织、项目筛选、权利配置和下游协同的议价作用可能增加。这里的优势来自接手后仍可使用的成果，而不是单纯保密一个模型名称。买方保有临床执行、制造和商业化能力，也会继续影响价值如何分配。

四、标题总额只是付款条件的目录

3 月 29 日公告称英矽可获 1.15 亿美元首付款，后续开发、注册与商业里程碑可使总交易金额达到约 27.5 亿美元，另有未来销售的分级提成。这个披露没有证明这些金额已经全部到账、全部确认为收入，也没有给出每项资产的独立价格或完整触发条件。把交易标题直接叫作“AI 赚到了 27.5 亿美元”，跳过了最关键的不确定性。来源：付款结构。

可以用一个原创的现金账本说明要补哪些信息。按离散期间 $t=0,\ldots,H$ 记录，令 $U_t$ 为该期实际收到的首付款类现金，$M_j$ 为第 $j$ 项固定金额里程碑付款，$I_{j,t}$ 为约定条件已满足且该笔款项在该期实际收到的指示量；条件可以在此前满足，每个单次付款最多出现一次。$R_t$ 为按具体条款计算并收到的销售提成。则合同相关现金为：

\mathrm{Cash}_t=U_t+\sum_{j=1}^{J}M_j I_{j,t}+R_t

若还关心许可方从当前信息出发的未来经济贡献，需要减去其在该期仍承担的现金支出 $C_t^{\mathrm{ret}}$，并考虑时点。令 $D_t$ 为确定的折现因子，$\mathcal I_0$ 为当前已知信息。金额采用同一币种与期间口径，相关期望存在，可写为：

V_H=\mathbb E\!\left[\sum_{t=0}^{H}D_t\bigl(\mathrm{Cash}_t-C_t^{\mathrm{ret}}\bigr)\,\middle|\,\mathcal I_0\right]

这只是有限期限 $H$ 下的分析账本，未包含期限外价值，也不是对任何一笔交易或公司股权的估值。它假设现金分类互不重复，固定金额里程碑可识别；复杂的退款、分期或可变付款需要另行扩展。没有合同细节、实际成本与可信概率，就不能给这条式子填出一个貌似精确的数字。

单项潜在里程碑的期望贡献为：

\mathbb E[D_tM_jI_{j,t}\mid\mathcal I_0]=D_tM_j\Pr(I_{j,t}=1\mid\mathcal I_0)

所以，标题金额没有提供的关键变量是触发与收款的概率、时间和许可方后续成本。线性期望不要求不同里程碑彼此独立；事实上临床阶段常相互关联。不能给每一步随便填一个行业平均通过率，再假装得到了这笔合同的真实价值。临床条件、买方是否继续投入、合同履约和收款也不是同一种事件。

从买方看，每个阶段都要重新比较剩余研发与付款成本、预期获批后的价值及其他项目机会。分阶段付款在经济上具有随证据推进投入的含义，但合同上是否可以终止、如何赔偿、必须付出何种努力，不能从一条公告推定。付款结构显示风险被分段安排，不证明风险已经消失。

五、可复用技术，不等于可按季度复现收入

英矽 2026 年上半年披露总收入 1.06303 亿美元，其中“药物发现与管线开发”1.03131 亿美元、占 97.0%，软件收入 269.7 万美元、占 2.5%。前一类别混合了不同研发与权利安排，不能全部称为资产许可收入。公司将增长主要归因首付款，并指出新交易谈判与管线交接进度影响该类收入。中报财务报表经德勤审阅，2026 年数字为未经审计数据。来源：中报第 13、25、55 页。

这是一组已经披露的经营结果，但不能据此把半年的项目收入线性外推成稳定年收入。一个团队可以使用同一套模型开展更多项目；模型重用解决的是一部分生产过程，不等于不同项目的科学难度、权利可许可性和交接日历相同。一个大项目的交接提速，可能让某个期间收入上升，却未必意味着接下来每期都能重复同样规模。

现金和会计收入还必须分开。中报注释 3 披露集团上半年随时间确认的收入为 1.05805 亿美元；这是集团合计，不能用来认定某个 Lilly 合同的具体确认方法。IFRS 15 的原则是围绕履约义务及承诺商品或服务的转移确认收入，可能在某个时点，也可能随履约进度确认。签约、收到首付款和确认收入，不是三个可互换的词。中报第 63 页；IFRS 15 官方说明。

判断持续性，应按相同项目批次追踪：何时签约，何时交接或履约，何时达到后续条件，何时收款与确认收入。软件订阅则要另外检查续约、活跃使用与支持负担。两类业务可能互相支持，但只看汇总增长会把交易结构变化、确认时点变化和技术生产率变化混在一起。

六、多项目可以分散风险，也可能共享同一失败原因

资产路线有一个常见推论：平台同时生成很多项目，就能用项目组合平滑收入。但“项目很多”不是独立性的证据。令 $X_i$ 为相同观察期限内第 $i$ 个项目的随机净现金贡献，假设二阶矩存在，组合波动为：

\operatorname{Var}\!\left(\sum_{i=1}^{N}X_i\right)=\sum_{i=1}^{N}\operatorname{Var}(X_i)+2\sum_{i<k}\operatorname{Cov}(X_i,X_k)

即使每个项目的风险相同，只要它们共同依赖一个未验证的生物学机制、同类化学问题或同一合作方的预算，协方差就可能为正。相反，不同适应症、机制、阶段和伙伴可能提供部分分散，但不能只由项目数量推定。上述式子是通用风险分解，并未估计这些公司的真实相关性。

同样需要观察项目存量的质量。签约项目数若不列停研和失败项目，会形成幸存者偏差；短期收入若主要来自少数交接，不能单凭平台的可扩展性证明组合稳定。阶段、权利类型与研发预算不同的项目应分组比较，保留退出项目并记录其原因，才能区分生产更多可用资产与只是启动更多探索。

七、大药企的认可，是交易证据，不是 AI 因果实验

与头部药企签约，是对双方愿意按约交易的证据。它也可能反映特定靶点稀缺、化学与专利质量、对方补充管线的战略、谈判环境或团队既有药物研发能力。无法仅由交易金额剥离出 AI 的增量贡献。一个漂亮的选择性实例，也可能来自项目和展示结果的选择。

若要检验 AI 是否改变了研发经济性，需要阶段与任务可比较的项目队列，把停研、失败与追加验证成本一并纳入。记录从什么起点走到什么验收点，是否包含合成、实验、返工和外包开支；控制团队变化、靶点难度与验证标准，或用预先规定的对照来减少混杂。发现更快、可交接项目更多、临床表现更好，是三个不同的结论，应分别给出证据。本文没有运行药物设计或临床实验，也没有估计这些公司的成功概率。

相反的路径也成立：如果客户已经能有效使用开放或商业模型，且自己掌握验证、知识产权与临床体系，独立 AI 供应商可能更适合出售工具或研发服务。若可用候选变得充足，而后续验证仍昂贵，大药企也可能增强筛选与议价能力。技术效率提高，未必让技术供应商按同一比例获得收入或利润；优势取决于哪一项稀缺能力仍由它控制。

八、未来 6–24 个月，怎样推翻这个判断

我预期值得追踪的方向是：AI 平台的商业评价会更需要项目批次、交接质量与分阶段兑现证据。它不是关于某家公司必然成功的预测。以下指标是建议观察协议；如果公司没有披露，就应标为未知，不能拿合作名单代替。

窗口	应追踪的公开证据	怎样改变判断
6–12 个月：交接	同批项目签约、履约或交接、收款、收入确认的衔接；包含停研、终止及延期	签约多但接手后缺少可核验推进，削弱“稳定产出可用资产”的判断
6–18 个月：后续验证	伙伴披露的阶段推进与实际里程碑，按阶段、靶点、权利类型比较；区别开启试验与试验结果	候选更多却需要更多返工，或后续结果不改善，削弱技术转化判断
12–24 个月：持续经济性	新老项目贡献、重复合作、交接与后续研发成本、对少数项目和伙伴的依赖	若收入持续靠个别大额交易、相关失败集中，组合规模不等于稳定增长
持续：替代路径	软件续约与使用、客户自行验证能力、可比外部或内部项目的交付和成本	若通用工具已足以支撑同等交付，独立平台或资产供应商的溢价可能缩小

判断的支点，是技术能力与可交易成果之间的接口。能提出分子、能验证和交接、能许可明确权利、能继续推进并兑现付款，分别需要证明。AI 可以改善其中某些环节；企业选择软件、研究合作或资产许可，决定了改善怎样进入客户预算，也决定了谁继续承担风险。

参考资料

Building a new path to make medicines with AI — Isomorphic Labs · 2026-09-29 · 查阅 2026-10-09
Global pipeline licensing and AI drug discovery collaboration with Lilly — Insilico Medicine, HKEX announcement · 2026-03-29 · 查阅 2026-10-09
Interim Report 2026 — Insilico Medicine, six months ended 2026-06-30, printed pp. 13, 25, 55 and 63 · 2026-08-26 · 查阅 2026-10-09
Isomorphic Labs kicks off 2024 with two pharmaceutical collaborations · 2024-01-07 · 查阅 2026-10-09
Isomorphic Labs secures $2.1 Billion funding to scale its AI drug design engine · 2026-05-12 · 查阅 2026-10-09
IFRS 15 Revenue from Contracts with Customers — IFRS Foundation, official overview · 查阅 2026-10-09
Step 2: Preclinical Research — US Food and Drug Administration · 查阅 2026-10-09
Step 3: Clinical Research — US Food and Drug Administration · 查阅 2026-10-09

2026-10-09Business PerspectivesYoucheng Li · 利友诚

What Does an AI Drug License Actually Sell?

Insilico–Lilly and Isomorphic reveal distinct software, research, and asset-rights contracts. Evidence handover, milestones, revenue recognition, and correlated projects shape commercial persistence.

As an AI drug-discovery company improves its models, what does a customer actually pay for? Fast molecule generation is a technical capability. Results that another team can take over and develop are a delivery capability. Defined rights that can be licensed are a transaction object. Experimental evidence, the boundaries of those rights, and development risk separate a model capability from a drug asset. Those interfaces determine whether reusable technology can create durable commercial value.

The recent starting point is Isomorphic Labs’ strategy article of September 29, 2026, considered alongside Insilico Medicine’s March 29 collaboration disclosure with Lilly and its interim report published August 26 for the six months ended June 30. Earlier announcements establish contract structure rather than being presented as news today. Strategy and technical illustrations are company statements; contract disclosures and financial figures must be read within their own scopes. Recent strategy article; HKEX collaboration announcement; Interim report.

1. Three transaction objects require different growth logic

An announcement that an AI drug company signed a partnership needs unpacking. Software licensing grants use of tools within agreed terms. Research services promise specified activities or deliverables. A drug-asset license allows the recipient to use associated rights within a defined scope and continue development. These can coexist or be combined in one agreement. They are alternatives, not an inevitable progression from a lower to a higher business model.

Contract object	What the customer principally obtains	What scaling requires testing
Software access	Use of software or model capabilities within scope, duration, and terms	Actual use, renewal, support costs, and the customer’s ability to perform validation
Agreed research deliverables	Research work, data, or candidate results for a defined problem	Acceptance of delivery and scientific and experimental effort per additional project
Drug-asset license	Asset-use rights defined by field, territory, and obligations, with agreed handover content	Usable evidence, authority to license, downstream progress, and payment conditions

The Insilico–Lilly announcement illustrates a combination: exclusive worldwide development, manufacturing, and commercialization rights for preclinical novel oral therapeutics in certain indications, together with multiple research programs on Lilly-selected targets. The whole agreement cannot be called software sales, nor a sale of one existing asset. Licensing grants agreed use rights; it need not transfer ownership of all intellectual property, all training data, or the platform model. The full contract is not public, so those terms cannot be supplied by assumption. Source: March 29, 2026 disclosure.

The distinction changes customer economics. Software can serve multiple customers within permitted scopes, whereas the same exclusive rights to an asset cannot be licensed without limit. Research growth can add projects while adding delivery obligations. An asset deal can monetize accumulated work, but sustained growth requires new transferable projects or progress on existing conditions. Reuse of the model does not automatically make each asset’s revenue reusable.

2. Model scores must become evidence another team can use

The technical background is more specific than faster generative AI. Drug design balances activity at the target, selectivity against other targets, exposure, toxicity risk, synthesis, and formulation feasibility. A model may propose or rank candidates. If its ranking is mismatched to the experimental setting, a higher computational score can merely accelerate a wrong direction. The buyer needs to know which uncertainty has been reduced by data and which remains its responsibility.

A checkable handover has four steps. First, specify the target, indication, and required properties; otherwise different definitions of success cannot be compared. Second, link predictions to experiments with methods, controls, repeats, and original records, distinguishing computational, cellular, and animal results. Third, test whether the receiving team can reproduce the findings under its conditions, preserving failures and applicability limits. Finally, map data, materials, methods, relevant IP, and permissions to contractual promises. Each step resolves an interface. A best-case curve cannot replace the delivery package.

Evidence package is an analytical framework here, not a claim that these companies’ disclosed contracts contain an identical checklist. Diligence varies with development stage and transaction scope. Compound identity and synthesis routes, assay conditions, exposure and safety boundaries, and the origin of rights require project-specific assessment. A patent or exclusive license alone does not establish unrestricted freedom to operate against other rights. Reproducible evidence, traceable materials, and clear licensing authority let another team assess continued investment.

FDA explanations separate the scientific stages: preclinical studies address basic safety questions including toxicity, and do not replace studies in people. These US descriptions are not treated as a complete compliance standard for a global contract. They clarify why discovery, preclinical work, clinical research, and marketing review answer different questions. FDA: preclinical research; FDA: clinical research.

Isomorphic’s September 29 article describes vertical integration of data, models, software, and physical infrastructure. Its molecular-search demonstration is followed by selecting compounds, synthesis, and laboratory validation; it describes preclinical data as helping prepare clinical development. Computational search duration is therefore neither an end-to-end drug-development timeline nor evidence of clinical benefit or portfolio-wide success. The case supports a strategic reading that the company is connecting design with validation, without independently establishing how much clinical success AI causes. Source: company strategy and illustration.

Original transaction-mechanism illustration, not experimental or transaction-outcome data. Software access, agreed research deliverables, and drug-asset rights with an evidence package are parallel contract objects rather than an inevitable progression. Only the drug-asset license is expanded: evidence and intellectual-property diligence checks evidence, ownership and licensing authority; the contract defines licensed scope, territory and party obligations; the buyer undertakes downstream development, manufacturing and commercial execution according to the contractual allocation of responsibilities. The payment categories are assessed separately: upfront payment requires checking payment conditions and cash received; development, regulatory and commercial milestones depend on specified triggers; sales royalties depend on agreed sales bases and terms. A headline total cannot be treated directly as cash received or current revenue. Signing, handover and performance, and accounting revenue recognition are different assessments, with cash and revenue timing checked separately. Technical efficiency may add asset options, but realization depends on evidence, rights and downstream progress; faster discovery does not demonstrate clinical success. Arrows represent information and contract-analysis flow, not a company’s measured results or a guaranteed commercial outcome. — Original analytical illustration, not experimental data or a company’s realized transaction outcome. The contract objects are parallel alternatives; only the drug-asset license is expanded. Licensed scope, territory, obligations and downstream responsibilities follow the contract. Upfront payments, contingent milestones and sales royalties require separate assessment; a headline total is not directly cash or current revenue. Signing, handover and performance, and revenue recognition are different events. Technical efficiency does not establish clinical success.

3. Licensing redistributes work and potential upside

Why might a large pharmaceutical company license a preclinical program rather than only buy software? It is a capability-allocation question. Using a tool still requires organizing scientific questions, experiments, IP, and project management. Buying rights and a handover package for further development may bundle some early exploration into external delivery. The buyer compares this with internal projects, other licensing opportunities, and outsourced research, rather than solely with another model’s prediction score.

The licensor faces a choice too. Earlier licensing may bring agreed payments sooner and place some downstream work with a partner. Retaining an internal program may preserve more later value but require capital, clinical execution, and risk capacity. Neither route is inherently better. Retained rights, continuing obligations, cost sharing, and payment triggers alter the result. A drug contract must not be assumed to transfer every risk to the buyer.

Isomorphic announced strategic research collaborations with Lilly and Novartis on January 7, 2024; its September 2026 strategy also emphasizes vertical integration and taking potential medicines toward the clinic. These paths can run in parallel. This does not establish abandonment of partnerships or completion of clinical validation. Its announced $2.1 billion Series B on May 12, 2026 is financing capital, separate from customer license revenue or realized drug sales; the stated uses include scaling the engine and pipeline. Earlier collaboration announcement; Financing announcement.

My trend judgment is conditional: if AI lowers the cost of proposing candidates faster than the cost of reliable validation and handover, competition will place greater weight on teams that repeatedly produce developable projects. Experimental operations, project selection, rights allocation, and downstream coordination may gain bargaining importance. The advantage lies in outputs that remain usable after handover, rather than merely keeping a model name secret. Buyers’ clinical, manufacturing, and commercial capabilities still affect where value accrues.

4. A headline total is an index of payment conditions

The March 29 disclosure makes Insilico eligible for a $115 million upfront payment, with development, regulatory, and commercial milestones potentially bringing total deal value to approximately $2.75 billion, plus tiered royalties on future sales. It does not establish that the amounts were all received or recognized as revenue, nor disclose independent prices per asset or every trigger. Calling the headline money already earned by AI omits the central uncertainty. Source: payment structure.

An original cash ledger makes the missing information explicit. Use discrete periods $t=0,\ldots,H$. Let $U_t$ be upfront-type cash actually received in the period, $M_j$ the fixed amount of milestone payment $j$, and $I_{j,t}$ an indicator that its agreed conditions have been met and payment is actually collected in that period. Conditions may be satisfied earlier; each one-off payment appears at most once. $R_t$ is sales royalty cash calculated and collected under the contract. Contract cash is then:

\mathrm{Cash}_t=U_t+\sum_{j=1}^{J}M_j I_{j,t}+R_t

Prospective economic contribution to the licensor must also subtract retained cash outlays paid in that period $C_t^{\mathrm{ret}}$ and consider timing. Let $D_t$ be deterministic discount factors and $\mathcal I_0$ current information. Amounts use a common currency and period convention, and the relevant expectations are assumed to exist:

V_H=\mathbb E\!\left[\sum_{t=0}^{H}D_t\bigl(\mathrm{Cash}_t-C_t^{\mathrm{ret}}\bigr)\,\middle|\,\mathcal I_0\right]

This is an analytical ledger over finite horizon $H$, excluding value beyond that horizon, not a valuation of an actual deal or company equity. It assumes nonoverlapping cash categories and identifiable fixed-amount milestone payments; refunds, installments, or variable payments may need extensions. Without contract details, actual costs, and credible probabilities, the equation cannot support a seemingly precise estimate.

The expected contribution of one potential milestone is:

\mathbb E[D_tM_jI_{j,t}\mid\mathcal I_0]=D_tM_j\Pr(I_{j,t}=1\mid\mathcal I_0)

A headline total leaves out trigger-and-collection probabilities, timing, and future retained cash outlays. Linearity of expectation does not require milestones to be independent; clinical stages can be related. Arbitrary industry-average transition rates do not establish the value of a particular contract. Clinical conditions, the buyer’s continued investment, contractual performance, and payment collection are different events.

At each stage the buyer economically reassesses remaining R&D and payment costs, potential value after approval, and alternative projects. Staged payments can align investment with emerging evidence. They do not establish a legal right to terminate freely, compensation terms, or required efforts; those cannot be inferred from an announcement. The structure allocates risk in stages rather than demonstrating that risk has disappeared.

5. Reusable technology does not ensure repeatable quarterly revenue

Insilico reported first-half 2026 revenue of $106.303 million: drug discovery and pipeline development contributed $103.131 million, or 97.0%, and software $2.697 million, or 2.5%. The first category combines different research and rights arrangements, so it is not a pure asset-license revenue line. Management attributes growth mainly to upfront payments and identifies new-deal negotiations and pipeline handover as timing factors. The interim financial statements were reviewed by Deloitte; the 2026 figures are unaudited. Source: printed pages 13, 25, and 55.

These are reported operating results, but a half-year of project revenue cannot simply be annualized into a stable run rate. One engine can support more projects; reuse addresses part of production, not equal scientific difficulty, licensing authority, or handover schedules. Faster handover of a large project may lift one period’s revenue without demonstrating that the same scale will recur every period.

Cash and accounting revenue also require separate treatment. Note 3 reports $105.805 million of group revenue recognized over time during the half year; the aggregate does not identify the recognition method for a particular Lilly contract. IFRS 15 links recognition to performance obligations and transfer of promised goods or services, either at a point or as performance progresses. Signing, receipt of an upfront payment, and recognition are not interchangeable terms. Printed page 63; Official IFRS 15 explanation.

Assess persistence by following the same project cohorts: signing, handover or performance, later conditions, collection, and recognition. Software subscriptions need separate renewal, activity, and support-cost evidence. These businesses can support each other, but aggregate growth alone mixes changes in contract composition, recognition timing, and technical productivity.

6. Multiple projects can share the same failure mechanism

A common asset-business argument is that a platform can generate many programs and smooth revenue through a portfolio. Project count does not prove independence. Let $X_i$ be random net cash contribution over a common horizon from project $i$, with finite second moments. Portfolio variability is:

\operatorname{Var}\!\left(\sum_{i=1}^{N}X_i\right)=\sum_{i=1}^{N}\operatorname{Var}(X_i)+2\sum_{i<k}\operatorname{Cov}(X_i,X_k)

Even with similar individual risks, projects relying on the same unvalidated biological mechanism, chemical issue, or partner budget can have positive covariance. Different indications, mechanisms, stages, and partners may offer some diversification, but their number alone does not establish it. This is a general risk identity, not an estimate of these companies’ correlations.

The quality of the project stock matters too. Counts that omit discontinued and failed programs create survivorship bias. Revenue concentrated in a few handovers does not become stable solely because a platform can scale. Group projects by stage, rights, and R&D budget, retain exits and their reasons, and distinguish producing more usable assets from initiating more exploration.

7. A pharma deal is commercial evidence, not an AI causal experiment

A deal with a major pharmaceutical company is evidence that both parties are willing to transact on agreed terms. It may also reflect a scarce target, chemistry and patent quality, pipeline strategy, bargaining conditions, or existing drug-development expertise. Deal size alone cannot isolate the incremental contribution of AI. A strong selected example may reflect selection of both programs and displayed outcomes.

Testing a change in R&D economics requires cohorts with comparable stages and tasks, including discontinued work, failures, and added validation costs. Record the starting point and acceptance endpoint, with synthesis, experiments, rework, and outsourcing. Account for team changes, target difficulty, and validation standards, or use prespecified comparisons to reduce confounding. Faster discovery, more transferable projects, and better clinical outcomes are three separate claims requiring separate evidence. No drug-design or clinical experiment was run here, and no company success probabilities are estimated.

A different path is plausible. If customers can use open or commercial models effectively and own validation, IP, and clinical capabilities, an independent AI vendor may be better positioned to sell tools or research services. If usable candidates become plentiful while downstream validation remains expensive, pharma buyers may gain selection and bargaining power. Technical efficiency need not translate proportionately into vendor revenue or profit; value depends on which scarce capability it still controls.

8. What would overturn the judgment over 6–24 months?

The direction worth tracking is a stronger role for project cohorts, handover quality, and staged realization in evaluating AI platforms commercially. This is not a prediction that a particular company must succeed. The following is a proposed observation protocol. Undisclosed measures remain unknown; a partnership list cannot substitute for them.

Window	Public evidence to follow	How it changes the judgment
6–12 months: handover	Signing, performance or handover, collection, and recognition for the same cohort, including discontinued, terminated, and delayed work	Many deals without verifiable post-handover progress weaken repeatable usable-asset production
6–18 months: later validation	Partner-reported progression and realized milestones, matched by stage, target, and rights; trial starts separate from results	More candidates with more rework or unimproved later outcomes weaken technical translation
12–24 months: persistent economics	New and existing program contributions, repeat collaborations, handover and continuing R&D costs, concentration by project and partner	Dependence on exceptional deals or correlated failures prevents portfolio scale from establishing steady growth
Ongoing: substitutes	Software renewals and use, customer validation capabilities, delivery and costs of comparable internal or external projects	If general tools enable comparable delivery, an independent platform or asset supplier’s premium can narrow

The hinge is the interface between technical capability and a tradable result. Proposing molecules, validating and handing them over, licensing defined rights, and progressing toward payment each need evidence. AI may improve some steps. Choosing software, research collaboration, or asset licensing determines how the improvement enters customer budgets and who continues to bear risk.

Sources

Building a new path to make medicines with AI — Isomorphic Labs · 2026-09-29 · Accessed 2026-10-09
Global pipeline licensing and AI drug discovery collaboration with Lilly — Insilico Medicine, HKEX announcement · 2026-03-29 · Accessed 2026-10-09
Interim Report 2026 — Insilico Medicine, six months ended 2026-06-30, printed pp. 13, 25, 55 and 63 · 2026-08-26 · Accessed 2026-10-09
Isomorphic Labs kicks off 2024 with two pharmaceutical collaborations · 2024-01-07 · Accessed 2026-10-09
Isomorphic Labs secures $2.1 Billion funding to scale its AI drug design engine · 2026-05-12 · Accessed 2026-10-09
IFRS 15 Revenue from Contracts with Customers — IFRS Foundation, official overview · Accessed 2026-10-09
Step 2: Preclinical Research — US Food and Drug Administration · Accessed 2026-10-09
Step 3: Clinical Research — US Food and Drug Administration · Accessed 2026-10-09