Which companies benefit the most from AI Drug Discovery?

Investment research · As of October 4, 2026

Which companies benefit most from AI drug discovery? Start with who gets paid. A chip supplier sells computing capacity, a software company sells a research workflow, and a drug developer sells rights to a molecule—or keeps them through commercialization. Each can benefit from the same scientific advance on very different terms.

Our shortlist is Insilico Medicine, Schrödinger and Eli Lilly. Insilico has the strongest reported AI-native licensing economics among the public platforms we examined. Schrödinger offers the clearest established software exposure. Lilly is our preferred long-term pharmaceutical beneficiary because it combines proprietary data, development capacity and ownership of drug profits. This ranks business exposure; valuation would determine which stock offers the best return.

The number that best captures the distinction is 97%: approximately that share of Insilico’s first-half revenue came from discovery and pipeline activities. Even this direct AI beneficiary earns principally by developing and licensing drug assets. Its economics already look much more like biotechnology than a software subscription business. The percentage is calculated from approximately $103 million of discovery and pipeline revenue within $106 million of total H1 2026 revenue. [1]

Our organizing idea is simple: follow the rights, the receipts and the company’s scale. The biggest supplier can have the least concentrated exposure. A small platform can retain enormous upside while requiring shareholders to finance years of clinical risk.

The leading exposures earn their returns at different points in development. Sources: company disclosures; assessments are our judgments.

Insilico leads on licensing economics, with a clinical test still ahead

A discovery platform can collect research funding, an upfront licensing payment, development milestones and eventual royalties. The commercial partner typically takes specified drug rights and the obligation to fund subsequent work. For the platform, the crucial questions are how much it collects before success, how often it can repeat the transaction and how much upside it retains. Merck’s collaboration with Variational AI illustrates the allocation explicitly: Merck receives exclusive development and commercialization rights to resulting compounds. [2]

Insilico Medicine, listed in Hong Kong as 3696, is our strongest demonstrated AI-native licensing beneficiary. Its interim report records approximately $106 million of H1 2026 revenue, against $27 million a year earlier—growth of about 287%, calculated from the reported period totals. Discovery and pipeline activities supplied almost all the latest half’s revenue. [1]

That concentration is both the attraction and the risk. A valuable program can generate substantial revenue well before approval, but licensing receipts arrive unevenly. We would give Insilico credit for monetization while requiring repeat transactions before treating the first-half performance as a durable earnings base.

Its clinical evidence gives the licensing case substance. Rentosertib, a TNIK inhibitor developed for idiopathic pulmonary fibrosis, was tested in a randomized phase IIa study involving 71 patients. After 12 weeks, mean forced vital capacity rose 98.4 milliliters in the highest-dose arm and fell 20.3 milliliters with placebo. Subtracting the two means gives an unadjusted difference of about 119 milliliters. The comparison involved 18 high-dose patients and 17 placebo patients, with lung function assessed as an exploratory efficacy endpoint. [3]

That is a meaningful human signal, and it is the strongest counterweight to dismissing AI discovery as promotional language. It also leaves substantial work ahead. Insilico announced first dosing in September in a Chinese phase III study planned for 320 participants across 47 centers, with 52 weeks of treatment. The investment case now needs the short, small-study signal to survive a longer and larger test. [4]

Our preference for Insilico within the direct platform group rests on the combination: reported licensing revenue and controlled patient evidence. Either can weaken. A failed pivotal study would impair the retained-asset case; an inability to repeat partner receipts would challenge the platform’s commercial durability.

Schrödinger gets paid for the workflow across many drug programs

Insilico’s exposure depends heavily on individual drug assets. Schrödinger offers a different route: sell computational tools used across customers and programs, while retaining a separate discovery business. Its foundation combines physics-based simulation with newer AI capabilities, so the established software franchise is broader than generative AI alone. [5]

We prefer Schrödinger as the established software exposure because it already has a commercial engine. Q2 2026 software revenue was $32.5 million at a 71% gross margin. Multiplying the two yields approximately $23.1 million of software gross profit, before product development, selling costs, administration and other operating expenses. [5]

Software supplies an established revenue base; a single milestone accounted for a substantial portion of quarterly discovery revenue. Source: Schrödinger. [5]

The growth picture is less comfortable than the margin. Software revenue fell 10% from the prior-year quarter. Management attributed the decline primarily to its accelerated transition to hosted licensing; hosted revenue reached 47% of quarterly software sales, versus 30% over the trailing four quarters. Customer commitments and usage therefore matter alongside the timing of reported revenue. [5]

The discovery segment grew to $23 million from $13.9 million, helped by a $10 million milestone associated with the Ajax Therapeutics acquisition. Full-year discovery revenue guidance was $65–75 million. The mix explains why stronger aggregate results can coexist with softer reported software sales: different contracts recognize revenue on different schedules. [5]

Lilly’s TuneLab partnership shows where Schrödinger can fit as AI models proliferate. The announcement identifies LiveDesign as a “priority interface” for participating biotech companies to access TuneLab workflows. Lilly supplies models and proprietary research knowledge; Schrödinger supplies the environment through which scientists use them. The arrangement uses federated learning to keep participating companies’ proprietary data separate and private. [6]

Our view is that a useful enterprise interface can remain valuable even when a customer controls the underlying models. The same arrangement also sets a limit on the thesis: owning the interface gives Schrödinger distribution and workflow relevance, while important scientific assets remain with pharma.

We would not turn that business preference into an unconditional stock recommendation. The company still funds drug research, and its clinical programs carry familiar biotechnology risks: development of SGR-2921 was discontinued in August 2025. The software case needs sustained customer growth and improving cash generation to support the broader enterprise. [7]

Lilly can capture value on both sides of the research contract

The software example leads to the larger question: who keeps the economics when discovery becomes more productive? Our preferred incumbent is Lilly. It can buy outside discoveries, develop internal candidates and supply models to other researchers while maintaining a large proprietary data base. TuneLab brings those roles together. [8]

Lilly says the initial TuneLab models incorporate proprietary data obtained at a cost exceeding $1 billion. Selected biotech partners receive access and, in Lilly’s words, “contribute training data,” improving the models available to the ecosystem. The exchange gives smaller companies access to capabilities while Lilly benefits from additional learning. [8]

We think that structure favors an incumbent with valuable experimental data and the resources to develop successful candidates. A model can become easier to access while the data, drug rights and commercialization capabilities remain economically important. Lilly’s advantage is its ability to combine those pieces.

Near-term cost savings are a smaller argument. Lilly reported $3.8 billion of R&D expense and $23 billion of revenue in Q2 2026. At that spending pace, our sensitivity produces $68–365 million of annual gross discovery efficiencies, using the assumptions below. The range is deliberately broad because the share of spending affected and the improvement achieved both remain uncertain. [9]

Even the upper sensitivity produces a modest benefit relative to Lilly’s overall revenue. Source: Lilly; discovery shares and efficiency gains are analyst assumptions. [9]

The calculation annualizes quarterly R&D, then applies an assumed early-discovery share and an efficiency gain to that portion. The resulting gross benefit equals approximately 0.07–0.40% of annualized company revenue. Implementation costs and reinvestment could absorb all of it, leaving net cash savings at zero or below during the investment period. [9]

Our Lilly thesis therefore rests principally on future medicines and development productivity. Faster or better discovery has much greater strategic value if it produces a differentiated drug than if it trims a small portion of the research budget.

Roche is a credible alternative. Its announced AI infrastructure included 2,176 on-premises GPUs and more than 3,500 combined on-premises and cloud Blackwell GPUs. Novartis offers another route through its Isomorphic Labs collaboration, which it subsequently expanded. These are concrete commitments to adoption; our preference for Lilly remains a judgment about its combined capabilities, rather than a measured league table of AI investment returns. [10] [11]

NVIDIA earns infrastructure revenue, but the theme is small at its scale

NVIDIA benefits earlier in the chain. Research organizations need computing capacity before they know whether their candidates will work. Roche’s deployment is a named example of that demand. The Lilly–NVIDIA co-innovation laboratory provides another: the companies announced up to $1 billion of joint investment over five years in talent, infrastructure and compute. [10] [12]

The commitment sounds large until it is placed beside NVIDIA’s existing business. Spread evenly, the full joint ceiling is $200 million a year. NVIDIA reported $89 billion of Data Center revenue in its latest quarter, equivalent to a $356 billion annualized pace. Even the deliberately excessive assumption that every dollar of joint lab spending becomes NVIDIA revenue would make it only 0.056% of that pace. [12] [13]

This comparison explains our judgment: NVIDIA is a genuine beneficiary, but AI drug discovery is a weak standalone reason to own the stock. The calculation sizes one project; the broader pharmaceutical computing opportunity extends beyond it.

Other suppliers occupy useful positions without offering equally concentrated exposure. AWS became Novo Nordisk’s preferred cloud provider and strategic AI partner. Siemens acquired Dotmatics, whose whole software business was expected to generate more than $300 million of fiscal 2025 revenue at an adjusted EBITDA margin above 40%. Thermo Fisher’s NVIDIA collaboration targets scientific instruments and laboratory performance. [14] [15] [16]

We regard AWS, Siemens, Thermo Fisher and Danaher as secondary beneficiaries. They can earn from research activity and laboratory investment, but the available economics give us less reason to rank them above the shortlist on company-level AI-discovery exposure. Their broader businesses remain the principal investment cases. [15] [14] [17] [18]

Generate offers a pivotal-stage option; Recursion and Absci need more commercial proof

The attraction of a smaller platform is that a successful drug can matter enormously to the whole company. The cost is concentration: clinical spending arrives before commercial success, and shareholders may have to finance the gap.

Generate Biomedicines deserves a prominent speculative slot. Trading as GENB following its 2026 IPO, it had entered two phase III severe-asthma studies with its generatively designed anti-TSLP antibody, GB-0895. Q2 collaboration revenue was approximately $6.3 million. The central exposure is therefore the retained drug and its pivotal program, with collaboration receipts providing a much smaller current commercial base. [19]

We place Generate ahead of many earlier-stage platform stories as a more immediate pivotal-development case. That is a clinical-stage distinction. Its trials require funding, and a dependable valuation is necessary before translating the asset position into an expected shareholder return.

Recursion combines automated experiments, proprietary data and partnered and owned programs, including the acquired Exscientia business. Yet its reported H1 2026 revenue was approximately $14 million, down about 58% from $34 million a year earlier. Recognition of previously deferred partner payments also means reported revenue and fresh cash receipts can diverge. [20]

There is encouraging patient evidence. Recursion reported a median 43% reduction in total polyp burden among 12 efficacy-evaluable patients receiving REC-4881 in its open-label TUPELO study. The molecule was licensed from Takeda, which makes this a case of identifying and developing an application for an existing compound. The patient signal matters; its open-label design and small sample determine how much weight we give it. [21] [22]

Absci has retained antibody upside, including ABS-201, but its first-half partner revenue was approximately $0.5 million. Lilly’s participation in its equity financing supplies capital and industry backing; the investment enters shareholders’ equity. Our commercial assessment therefore remains more cautious than the scientific optionality might suggest. [23]

Recursion and Absci could benefit disproportionately from clinical success. We nevertheless put both below the leading names on demonstrated monetization. For either to move up, we want controlled efficacy, repeat partner cash receipts and funding that carries planned work through the next meaningful decisions.

Human efficacy will decide whether the ranking changes

The strongest argument against our shortlist is that it gives too much weight to receipts today. If AI materially improves clinical efficacy, companies retaining drug rights could create far more value than established software vendors. Insilico’s randomized study, Generate’s pivotal programs and Recursion’s early patient data make that possibility serious. [3] [19] [21]

Our restraint comes from where drugs historically fail. In the BIO, QLS and Informa study of 2011–2020 development programs, only 28.9% of phase II programs advanced to phase III, the weakest clinical transition. Better candidate design must eventually improve performance at that human-efficacy hurdle to support a broad claim of superior development economics. [24]

The early AI cohort is promising but immature. A published analysis of selected AI-native companies reported 21 successes among 24 molecules completing phase I, and four successes among ten completing phase II. Those small, selected groups support continued investment and testing. They are insufficient for us to underwrite an industry-wide approval-rate advantage against an unmatched historical population. [25]

Private competition also matters. Isomorphic Labs reported a $2.1 billion financing in May 2026, alongside its major pharmaceutical relationships. That capital can fund substantial competition for public platforms; Alphabet shareholders receive only indirect exposure to the business. Scientific leadership need not accrue to the most accessible listed pure play. [26] [27]

Our ranking would change with specific evidence. A persuasive, prespecified rentosertib phase III result would strengthen Insilico’s lead; failure would weaken it. Successful Generate pivotal results, accompanied by funding through regulatory filing, would push it toward the top of the direct-beneficiary list. For Schrödinger, we want sustained customer commitments and hosted usage followed by stronger cash generation; two consecutive weak reporting periods after accounting for the licensing transition would challenge our preference. [4] [19] [5]

For Recursion and Absci, our funding threshold is at least two years of planned operations without reliance on an uncommitted equity raise, alongside controlled efficacy and repeat partner cash. For pharma and suppliers, stronger earnings claims require repeatable productivity gains after implementation costs or discovery revenue large enough to affect consolidated growth.

The broader thesis would change with a prospectively tracked, indication-matched record of better phase II and III outcomes, including failures and withdrawals. Until that evidence arrives, we favor demonstrated receipts, useful workflows and ownership of development capabilities. The decisive advance will be a better result in patients—and the decisive investment question will be who owns it.