AI Drug Discovery Deals: Read the Upfront, Not the Headline

AI Drug Discovery Deals: The Headline Is Not the Price

AI drug discovery deals have settled into a familiar shape. In October, Zealand Pharma agreed to work with Generate Biomedicines on AI-engineered protein therapeutics for obesity and metabolic disease, in a deal reported as worth up to nearly 1.2 billion dollars. The detail underneath tells a different story: 30 million dollars upfront, up to 35 million in early preclinical milestones, and the rest contingent on research, clinical, regulatory, and commercial success over many years, plus royalties. That is not a criticism of this particular deal, which follows a well-established pattern. It is a reminder that the headline number in an AI pharma partnership is a ceiling, not a price, and anyone evaluating these deals needs to read them that way.

Reading a Deal: Headline Terms and What They Mean

The language of deal announcements is designed to sound large. Each component means something quite specific.

What the Zealand and Generate Deal Shows

The structure of the deal is instructive. Generate will lead research through lead candidate selection using its AI platform, and Zealand takes responsibility for clinical development, regulatory submissions, and worldwide commercialisation. The deal covers two initial programmes with an option to add a third. That division of labour is typical: the AI company contributes discovery, the pharmaceutical partner contributes development muscle and carries most of the later-stage cost and risk.

It also follows the pattern of earlier high-profile AI partnerships, including Isomorphic Labs' agreements with Eli Lilly and Novartis, which paired upfront payments in the tens of millions with headline values well above a billion. Obesity, where demand for new mechanisms and formats is intense after the success of GLP-1 drugs, is an especially active area for these collaborations.

How AI Drug Discovery Deals Are Structured

Most pharma partnerships, AI or not, share risk through milestones. The partner pays a modest amount to start, then pays more as a programme proves itself. That makes sense for both sides. The AI company gets validation, cash, and a path to royalties. The pharmaceutical company limits its exposure until there is evidence the platform can deliver.

Why AI Drug Discovery Deals Lean on Milestones

The reason is attrition. Most drug candidates that enter clinical trials do not reach approval, and many fail in phase 2, when efficacy is first tested properly in patients. A deal worth 1.2 billion dollars if everything succeeds is worth a fraction of that once you weight each milestone by its realistic probability of being reached. AI may improve the odds at the discovery stage, but it has not yet been shown to change late-stage success rates at scale, a point we made in why phase 2 is the real test.

Testing the Platform Claim

Every AI discovery company says its platform makes discovery faster, cheaper, or better. The useful question is what evidence supports that for the specific modality and target class in the deal. In our view, the strongest signals are:

• Candidates actually delivered. How many development candidates has the platform produced, for whom, and how long did each take?
• Clinical-stage assets. Have any platform-derived molecules entered human trials, and with what results?
• Relevant modality experience. A small-molecule track record says little about protein or antibody design, and vice versa.
• Independent validation. Peer-reviewed results, external benchmarks, or partners who have publicly renewed collaborations.
• Clarity on what is AI. How much of the success came from models versus experienced medicinal chemists and biologists?
• Repeatability. Whether results hold across several targets or depend on one standout programme.

Diligence Questions Before You Sign

1. What is the probability-weighted value? Model each milestone with realistic success rates, not just the headline.
2. How much cash is committed in the first two years? Upfront, research funding, and near-term milestones together.
3. Who decides when a candidate is good enough? Selection criteria and decision rights shape both cost and conflict.
4. Who owns the data generated? Experimental results feed the platform, and that has value beyond this deal.
5. How is IP split? Background platform IP, programme-specific IP, and improvements need clear terms.
6. What exclusivity applies? Can the AI company work on the same target or pathway for a competitor?
7. What happens if the programme stalls? Termination, reversion of rights, and data return terms matter more than they look.

Pros and Cons for the Pharmaceutical Partner

• Pro: access to new capabilities. Partners gain discovery approaches they could not build quickly in-house.
• Pro: limited early exposure. Milestone structures keep the initial commitment small.
• Pro: more shots on goal. Several programmes across partners diversify discovery risk.
• Con: overstated speed. Faster candidate generation does not shorten clinical development, which dominates timelines.
• Con: data leakage risk. Partner data can improve a platform that later serves competitors unless terms prevent it.
• Con: integration effort. Handover between an AI company's research and internal development teams can be bumpy.

Real Scenarios Worth Thinking Through

These scenarios are illustrative, showing how AI drug discovery deals play out in practice rather than presented as verified case studies.

A mid-sized pharmaceutical company's board sees a headline valuing a proposed AI partnership above a billion dollars. Business development presents a probability-weighted model showing an expected value a small fraction of that, most of it in royalties a decade away. The board approves the deal, but with realistic expectations and a tighter focus on near-term deliverables.

A partner discovers late in negotiations that experimental data generated in the collaboration would be used to train the AI company's general platform. It negotiates restrictions on use of its target-specific data, and accepts general model improvements as a fair exchange.

A collaboration produces a candidate quickly, and both sides celebrate. Preclinical toxicology then raises concerns the platform did not predict. Because the contract defined clear decision rights and a path to back-up candidates, the programme continues rather than collapsing into dispute.

“After testing multiple AI models on coding, research, and business prompts, combined outputs produced more reliable results than any single model.” Internal multi-model evaluation, Talkory research team.

Data, IP, and Who Learns From the Collaboration

An AI discovery partnership is not only a deal about molecules. It is a deal about learning. Every assay result, structure, and failed candidate generated in the collaboration can improve the AI company's models, which then become more valuable to every future partner. That is not necessarily a problem, and it is part of why AI companies accept modest upfronts. But it should be priced and governed explicitly, with clear distinctions between general model improvements and target-specific data the partner wants protected.

Protein and biologics design raise additional considerations, including biosecurity, which we covered in AI protein design. As design tools become more capable, responsible-use provisions are worth including in collaboration agreements, not just in company policies.

Why Talkory Wins

Evaluating an AI discovery partnership involves scientific claims, deal structures, and market context, all of which are easy to read one-sidedly. Talkory runs the same question across GPT, Claude, Gemini, Grok, Perplexity Sonar, and Kimi K3 in one pass. A business development or diligence team can ask what evidence would support a platform's claims for a given modality, how comparable deals have been structured, or what risks a particular IP clause creates, and see whether six independent models agree. Where they diverge, that is where the team should dig into primary sources, talk to experts, or push harder in negotiation.

Final Verdict

AI drug discovery deals will keep arriving with headline values above a billion dollars, and most of them will look like the Zealand and Generate agreement: modest guaranteed cash, large contingent milestones, and royalties far in the future. That structure is rational given drug development risk. The mistake is reading the headline as the price. Weight each milestone realistically, focus on near-term committed cash, test the platform's track record in the relevant modality, and negotiate data, IP, exclusivity, and termination terms carefully. The deal value is a ceiling. The diligence decides what it is really worth.