China AI drug discovery ecosystem: China AI drug discovery is scaling rapidly, but the next winners will be defined by data, wet-lab feedback and clinical execution—not molecule generation alone.

Overview

This deep dive examines how AI is changing China’s drug-development ecosystem. Bernstein argues that early-discovery speed and licensing activity have improved sharply, but sustainable value will depend on whether companies can link AI models to proprietary biological and clinical data, wet-lab validation, manufacturing capacity and drug-development execution.

Core views

AI drug discovery has moved from a niche capability toward a more mainstream component of China’s biopharma industry. China AIDD companies expanded active pipelines from 47 assets in 2023 to 219 in 1Q-3Q26, with preclinical programs rising from 45 to 194 and clinical programs from two to 25. The report attributes the most visible benefit to faster early development: reported time from hit identification to first-in-human studies is about 22 months for AIDD companies, compared with 59 months for large-pharma benchmarks. This can allow more programs to enter the clinic within a given time, although reported clinical-execution improvements of 5-15% remain comparatively limited. Potential manufacturing-conversion and SG&A savings could also broaden the earnings opportunity for established pharma. Licensing has become an important form of external validation and monetization, but headline values should be interpreted carefully. Global AI-related licensing value reached $54 billion in 1Q-3Q26, above $31 billion in FY2025, with 27 transactions worth at least $100 million versus 21 previously. AI represented 25% of global licensing value, up from 3% in 2020. China-origin transactions accounted for 23% of cumulative deal count and 32% of value, suggesting larger average deal sizes, while AI-related deals contributed 25% of China-origin licensing value in 1Q-3Q26 after single-digit shares in 2022-25. However, 74% of the 98 transactions through 3Q26 were asset-only deals, and only 13% involved defined existing assets; collaborations to develop new assets represented 61% of deal count and 58% of value. Bernstein therefore notes that much announced value is milestone-dependent and reflects expectations of future delivery rather than realized clinical or commercial value. Upfront payments per asset have not consistently exceeded those of non-AI deals. The central constraint is that ligand discovery is not drug discovery. Foundation models and generative methods can identify binders and generate molecules faster, but a high-affinity ligand may fail because of solubility, pharmacokinetics, toxicity, manufacturability or other developability problems. Bernstein cites a simulation showing that a 20% reduction in development failure rates would generate greater savings in capitalized cost per approved drug than comparable improvements in speed or R&D spend. Yet available data do not show better advancement rates for AIDD: global AIDD companies had 19% Phase I and 7% Phase II progression rates, China AIDD companies 14% and 5%, versus 26% and 14% for global top-20 multinational pharma companies. Insilico Medicine showed 31% and 10%, respectively, but the report emphasizes that most AIDD programs are still early and evidence remains immature. The report describes a shift from protein-structure prediction toward de novo design, in which models generate novel sequences or molecular structures for a defined biological objective. But generated candidates still require computational checks for folding, target engagement and molecular behavior, followed by experimental validation. Anew Labs’ example reduces more than 3,000 generated nanobody designs to about 50 candidates for wet-lab testing; Insilico’s Chemistry42 applies iterative optimization against potency, selectivity and ADMET criteria. Bernstein’s conclusion is that model benchmarks alone are inadequate because downstream biological relevance and clinical outcomes matter more than isolated screening performance. Consequently, competitive advantage is shifting toward lab-in-the-loop systems. Public structural datasets and model architectures are broadly accessible, whereas experimentally validated, standardized proprietary datasets are not. In a design-test-learn loop, computational designs are screened in wet labs using assays such as BLI, SPR and ELISA; the resulting biological measurements are then fed back into the model for the next optimization cycle. AnewDesign states that repeated optimization can produce up to a 100-fold improvement in binding affinity and single-digit nanomolar KD values. Bernstein argues that companies combining AI design, automated experimentation and continuous learning can develop more durable data advantages than companies relying solely on increasingly large models. This transition also elevates CROs and CDMOs. AIDD players require wet-lab validation, process development, regulatory execution, manufacturing scale-up and eventually commercialization support. The report cites collaborations including Latent Labs with GenScript, Boltz’s use of WuXi AppTec’s OTS compound library, Earendil Labs’ 2026 strategic partnership with WuXi Biologics, and Insilico’s work with WuXi AppTec, Tigermed and JOINN. Bernstein views these providers as increasingly embedded in the AIDD value chain because their modality-specific data, experimental throughput and manufacturing capacity enable the feedback loops AI platforms require. Bernstein identifies three winner archetypes. Frontier AI companies can win through leading foundation models and generative-design capabilities, particularly when models are validated and improved through integrated wet labs. Scaled service and data providers—including CROs, CDMOs and technology platforms—can benefit as horizontal integrators through proprietary preclinical data, standardized workflows and manufacturing capacity; WuXi AppTec and WuXi Biologics are cited as coverage examples. Integrated biopharma companies can act as vertical integrators by combining AI with proprietary preclinical, clinical and commercial datasets, as well as ownership of development decisions and commercialization; Innovent, Hengrui and Hansoh are cited as examples. The report does not expect a winner-takes-all market dominated by model developers. For selecting beneficiaries, Bernstein advises concentrating on tangible outcomes rather than AI claims: proprietary-pipeline expansion, clinical progression, licensing validation, R&D productivity, data ownership and execution across the value chain. CSPC, Insilico, Earendil Labs and XtalPi are highlighted as active dealmakers, while established biopharma companies such as Hengrui, CSPC, Hansoh, Innovent and BeOne maintain far larger active-program bases than China’s AI-native companies. For integrated pharma, the relevant evidence is faster target-to-clinic progress, pipeline refresh, earlier termination of weak programs, licensing output and eventual clinical success.

Analysis framework

Bernstein compares AIDD pipeline growth, development timelines, clinical progression rates and licensing transactions with traditional pharma and global peers. It then links the limitations of molecule-generation models to the need for biological, translational and clinical data, and assesses which business models and value-chain positions can capture the resulting opportunity.

Methodology notes

• Industry AnalysisUpstream-Midstream-Downstream Transmission

  Drug-development value-chain analysis

  The report traces how AI-generated hypotheses require wet-lab testing, clinical development and manufacturing, explaining why value may accrue not only to model developers but also to CROs, CDMOs and integrated drug companies.

• Other

  Pipeline, clinical-success and licensing-deal benchmarking

  Bernstein compares pipeline counts, stage progression, development timelines, deal values, deal structures and upfront payments to distinguish early discovery momentum from demonstrated clinical value.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• WuXi AppTec

  Covered CDMO/CRO beneficiary of growing wet-lab validation, data-generation and development demand from AIDD companies.

  Strengths

  Cited as a scaled service and data provider with relevant capacity and value-chain integration.

  Comparison

  Grouped with WuXi Biologics as a horizontal integrator rather than a pure-play model developer.

  Risks

  AIDD demand must translate into sustained outsourced experimentation and development activity.

• WuXi Biologics

  Covered biologics CDMO positioned to support AI-designed biologics through end-to-end development and manufacturing.

  Strengths

  Earendil Labs collaboration covers cell-line, process and formulation development plus GMP manufacturing.

  Comparison

  Cited alongside WuXi AppTec among scaled service and data-provider examples.

  Risks

  Value realization depends on partner assets progressing through development.

• Innovent (1801.HK)

  Covered integrated biopharma example among potential AI-era vertical integrators.

  Strengths

  Can combine AI with proprietary preclinical, clinical and commercial data across the drug-development chain.

  Comparison

  Cited with Hengrui and Hansoh as an integrated biopharma beneficiary rather than a pure AIDD platform.

  Risks

  AI-related productivity improvements must translate into clinical and commercial outcomes.

• Hengrui (600276.CH)

  Covered integrated biopharma example and AI-platform partner of METiS TechBio.

  Strengths

  Large active-program base and end-to-end development capabilities.

  Comparison

  The report states established biopharma companies retain substantially larger active pipelines than China AI-native companies.

  Risks

  Model announcements alone are not sufficient evidence of R&D value creation.

• Hansoh (3692.HK)

  Covered integrated biopharma example among potential AI-era winners.

  Strengths

  Cited as an established biopharma with relevant pipeline scale and value-chain integration.

  Comparison

  Grouped with Innovent and Hengrui as vertically integrated beneficiaries.

  Risks

  Clinical execution and measurable R&D productivity remain the relevant tests.

• Asymchem (002821.CH / 6821.HK)

  Covered company rated Outperform.

• BeOne (ONC)

  Covered company rated Outperform.

  Strengths

  Listed among established biopharma companies with substantially larger active-program bases than China AI-native companies.

  Comparison

  Included among seven covered companies with a broadly stable 7-11% share of China active clinical-stage and approved assets during 2021-2026YTD.

• Kelun-Biotech (6990.HK)

  Covered company rated Outperform.

• Akeso (9926.HK)

  Covered company rated Market-Perform.

• CSPC (1093.HK)

  Covered company rated Market-Perform and highlighted as an active China AI-dealmaker.

  Strengths

  The report identifies four major AI deals and $1.48 billion of aggregate disclosed upfront payments in its selected 2023-2026YTD deal screen.

  Comparison

  Cited with Insilico, Earendil Labs and XtalPi as a dealmaking leader among Chinese companies.

  Risks

  Licensing commitments remain dependent on future asset delivery and milestones.

• Sino Biopharmaceutical (1177.HK)

  Covered company rated Market-Perform.

• Zai Lab (9688.HK)

  Covered company rated Market-Perform.

Key data

• China AIDD active pipeline47 assets in 2023; 219 assets in 1Q-3Q26Growth was led by preclinical assets, with clinical-stage progression also increasing.
• Preclinical programs at China AIDD companies45 in 2023; 194 in 2026YTDClinical programs increased from 2 to 25 over the same period.
• Hit identification to first-in-human timelinec.22 months for AIDD companies versus 59 months for large pharmaDemonstrates a reported early-development speed advantage.
• Global AI-related licensing value$54 billion in 1Q-3Q26Exceeded $31 billion in FY2025; AI represented 25% of global licensing value versus 3% in 2020.
• China-origin AI licensing contribution25% of China-origin licensing value in 1Q-3Q26Up from single-digit shares during 2022-25.
• AI deal composition74% of deal count and 68% of deal value were asset-only; defined existing assets were 13% of transactions and 10% of valueMost announced value remains linked to future asset development.
• Clinical progression ratesGlobal top-20 MNCs: 26% Phase I and 14% Phase II; global AIDD: 19% and 7%; China AIDD: 14% and 5%The report finds no significant AIDD success-rate advantage to date.

Impact & implications

The report sees AI as an ecosystem opportunity rather than a pure software or model-development theme. Its view is that the most durable benefits should accrue to companies that translate AI into validated biological insight, proprietary data, efficient development and commercial execution, while headline licensing values alone do not establish clinical or economic success.

Risks

• AIDD has not yet demonstrated consistently better Phase I or Phase II advancement rates than traditional pharma, and most programs remain early-stage.
• Models optimized for binding affinity may fail later because of toxicity, poor pharmacokinetics, solubility, manufacturability or other developability issues.
• Much AI-related licensing value is milestone-dependent and tied to future asset creation rather than defined existing assets.
• Public datasets and widely disseminated model architectures may make algorithmic capability alone difficult to defend.

What to watch

• Pipeline progression into clinical development and eventual clinical success, rather than molecule-generation volume.
• Licensing validation, including the mix of upfront payments versus milestone-dependent deal value.
• R&D productivity indicators such as target-to-clinic speed, pipeline refresh, early termination of weak programs and licensing output.
• Whether companies build proprietary biological and clinical datasets through automated wet labs, CRO/CDMO partnerships and closed-loop learning.
• Execution by CROs, CDMOs and integrated biopharma across development, manufacturing and commercialization.