There is no box for your AI

The comfortable reading is that pharma is slow and the boxes will catch up eventually. But the buyers have a defensible reason, and it shows up in the clinical data.

A BCG team, publishing in Drug Discovery Today, assembled the outcomes for AI-discovered molecules entering the clinic. In Phase 1 those molecules succeed at 80 to 90 percent, which the authors call substantially higher than historic industry averages. In Phase 2 the rate is roughly 40 percent, "albeit on a limited sample size, comparable to historic industry averages."

Phase 1 asks whether a molecule behaves in a human body (whether it is tolerated, whether it reaches exposure, whether it does what the chemistry predicted), and that is a design problem, which is exactly the thing computational design is measurably better at. Phase 2 asks whether hitting this target treats this disease. That is a biology problem, and no amount of molecular design improves your odds on it, because the target hypothesis was fixed before the engine ever ran. So a buyer pricing a molecule has no risk-adjusted reason to pay a premium for computational provenance, and should be paying instead for the target, the data and the stage.

And the published criteria say the same thing. AstraZeneca introduces its "5Rs" framework (right target, right tissue, right patient, right safety, right commercial) like this: "We assess all projects against our '5Rs' framework to ensure we take forward the most attractive opportunities, regardless of source." There is no sixth R for method. Lilly says it is "agnostic to the biologic basis of programs, so long as we have high conviction in the science." We read every published criteria list we could find (AstraZeneca, Roche, Genentech, Lilly, Merck, GSK, Sanofi) and discovery provenance appears in none of them.

And one further point cuts against the seller here. Counsel advising on AI-derived assets now describe a two-track diligence, scientific validation of the discovery plus legal validation of the AI claims, driven by the USPTO requirement that at least one human make a significant contribution to the conception of each patent claim. The buyer ends up wanting representations about who invented what, an allocation of rights over the data used and produced, and escrow protection against the platform company failing. That is friction a conventionally discovered molecule does not carry, though we should say that this rests on one law firm's account rather than on a survey of practice.

None of that tells you what to do instead, though, and the answer is not in the forms. It is in what the buyers actually paid for over the past three years, which sorts into three quite different transactions with three quite different prices.

Everything below rests on one dataset, so here is what it is. We assembled the China-to-West transaction record from January 2023 to July 2026 across four research streams (19 metabolic deals, 26 non-metabolic, 8 from Hengrui alone, plus oncology and platform sets that overlap these), checking each against company press releases and SEC or Hong Kong exchange filings, and recording upfront cash separately from headline biodollar totals throughout. Roughly thirteen further deals we know of could not be verified in the time available, and they are excluded rather than estimated.

Read that record and the deals sort into three kinds.

Door 1 is a named molecule against a named target, and upfronts there run from about $85 million (Hansoh to GSK, October 2023) to $800 million (SystImmune's EGFR×HER3 bispecific ADC to Bristol Myers Squibb, December 2023), with Innovent's two clinical assets to Takeda sitting above the range at $1.1 billion in cash. And this is the door a seller nobody has heard of can actually get through, as Eccogene did in 2023 on a cold approach.

Door 2 is a multi-target discovery collaboration where most or all of the targets are not named at signing, and upfronts run $350 million to $1.2 billion. In 2026 this is where the largest money in China-to-West licensing sat: CSPC to AstraZeneca at $1.2 billion in January, Innovent to Pfizer at $650 million in May, Hengrui to Bristol Myers Squibb at $600 million in May, Innovent to Lilly at $350 million in February. That is roughly $2.8 billion of upfront cash for programmes that were, with one exception, unspecified when the money was agreed (CSPC named one of its eight), and the Hengrui deal had not closed as of July, so $600 million of it has been announced rather than paid. If you are selling an AI discovery capability, Door 2 is where the money is, though as the next section shows it is not a door you can start at.

Door 3 is a pure platform or technology licence, and the numbers there are an order of magnitude smaller. Adagene licensed its SAFEbody masking technology to Third Arc for $5 million, to Exelixis for $11.0 million on execution, and to Sanofi for $17.5 million, all three figures taken from its own SEC filings. Sanofi paid BioMap $10 million for a protein-model collaboration, and its Insilico deal disclosed no clean upfront at all (only "up to $21.5 million" blended across upfront and target-nomination fees). MediLink licensed its ADC platform to BioNTech for $25 million. The largest we could source to a primary document was AstraZeneca's $110 million to CSPC. That is six clean disclosed upfronts, five of them at $25 million or below, alongside a great many deals in this category that disclose nothing at all.

Upfront cash, by what the buyer was actually buying

Each dot is one disclosed upfront payment. Log scale. China→West deals, January 2023 – July 2026.

Every company that walked through Door 2 had already sold named molecules to large Western buyers. Two of the four had sold to that particular buyer before, and two had not, which turns out to be the interesting part.

Take the two that had. CSPC sold AstraZeneca a preclinical Lp(a) molecule in October 2024 for $100 million, and eight months later AstraZeneca paid $110 million for access to CSPC's AI discovery platform. Fifteen months after the first deal came the $1.2 billion package (eight programmes, the AI peptide platform, and a monthly-dosing technology), with a fourth and smaller agreement following in July 2026. Innovent's relationship with Lilly runs much longer, back to a 2015 collaboration and a sintilimab licence expanded globally in 2020, well before Lilly paid $350 million in February 2026 for a discovery collaboration.

But the other two had no such history with the buyer who eventually wrote the cheque. Hengrui out-licensed named molecules to five Western buyers between August 2023 and March 2025, to Aiolos, Merck KGaA, Kailera, IDEAYA and MSD, before GSK paid $500 million across a set of agreements in July 2025 for a named COPD asset plus options over as many as eleven further unnamed programmes. Bristol Myers Squibb followed in May 2026 with $600 million for thirteen programmes and no molecule named. Neither had licensed a Hengrui molecule directly, though the connections are not quite nil: GSK acquired Aiolos Bio in January 2024, and Aiolos held a Hengrui asset, while four of the thirteen BMS programmes are assets flowing the other way, from BMS into Hengrui for Greater China.

And Innovent did the same thing twice over, having sold Roche a named DLL3 ADC in January 2025 and Takeda two named clinical assets in October for $1.1 billion in cash, before Pfizer (a buyer Innovent had never out-licensed anything to) paid $650 million the following May for twelve programmes with no targets specified. Pfizer's Chief Oncology Officer described that one as combining "Innovent's discovery and early clinical development" with Pfizer's global capability, which credits the seller's discovery function rather than any particular molecule.

So the rule is a disjunction rather than a sequence. A Door 2 buyer had either bought from you before, or could watch you selling to someone else serious, and either one appears to do. Which means track record substitutes for relationship, and a company with neither is not yet having this conversation.

And it is worth marking the limit of this, because it holds only for companies whose product is molecules. Chi-Med sold discovery capability to Eli Lilly in 2007 and to Johnson & Johnson in 2008, four years before its first molecule went west, and it did so as a loss-making small-cap rather than from any position of strength. Companies whose product is the platform sell the platform first, and this piece is describing someone else's problem.

Eccogene is a private, pre-revenue Shanghai company which in November 2023 had, by the explicit statement of its own Hong Kong listing prospectus (repeated three times, in three separate sections), no in-licensing or out-licensing relationship with anyone at all. Its last financing round had priced the whole company at RMB 1.09 billion, or roughly $155 to $160 million.

On 8 November 2023 it licensed one oral GLP-1 molecule, ECC5004, to AstraZeneca at Phase 1, for a $185 million upfront payment, and the prospectus describes that relationship beginning as preliminary discussions rather than as anything warmer. Two years later the company still employed 57 people.

So the upfront for that single molecule exceeded the entire value of the company that owned it, and six weeks later, on 18 December, AstraZeneca took Eccogene's whole Series C at a $498 million post-money valuation (roughly 3.2 times the previous round, and priced by the buyer itself). That is what one named molecule against a named target was worth, at Phase 1, from a fifty-seven-person company with no Western partner and no platform story.

If you are a drug company that uses AI, and that describes most of the companies reading this, then your computational capability is cost structure. It should show up in how fast and how cheaply you reached a candidate, and it should not be your third slide.

So start at Door 1, because it pays best to a seller nobody knows yet. Of the 43 mandates on our board, 38 accept preclinical or earlier, and in the metabolic stream of our record, where we have a clean stage split, 12 of the 19 deals signed at Phase 1 or before (Eccogene among them, at $185 million). Buyers will take development risk, and rather more of it than most sellers assume. What they will not take is vagueness about what the molecule is for.

Then treat Door 2 as the reward rather than the plan. The $2.8 billion that moved through it in 2026 went to CSPC, Hengrui and Innovent, three companies with long public records of selling molecules to large Western buyers, and that record is the thing you are actually building when you do a Door 1 deal.

Door 3 is worth using when the engine genuinely is the product, which for most drug companies it is not. The economics there are research funding and per-target milestones rather than an upfront on a molecule, the counterparties are often mid-caps and NewCos rather than large pharma, and you will be competing against well-capitalised Western AI-discovery firms. It is a real business. It is a different one from yours.

And read every headline number as a stage measurement. Across our record the multiple between headline and upfront cash tracks development stage closely: about 2.5 times for an approved product, 8 to 11 times at late clinical, roughly 18 times at Phase 1 or Phase 2-ready, and 25 to 73 times at preclinical.

A headline number measures stage, not value

Ratio of announced headline value to upfront cash, by stage at signing.

The same gap shows up on the platform side, where Recursion discloses in its own annual report more than $500 million in upfront and progress-based milestone payments received to date against "the potential for over $20 billion in additional milestones before royalties" (a different measurement from the stage multiples, being everything banked against everything promised, but pointing the same way). Which means a headline number on its own tells you roughly where a deal sits in development, and very little about what it is worth.