Proceedings · Session S-164 · filed November 17, 2026
AI & Emerging Tech in R&DSession paper
Pharma's $29.5 Billion AI Bet Faces a Seven-Year Wait for Proof
Pharma's top 20 firms signed $29.5bn in AI drug-discovery deals in 2025 with just $600m paid up front, and decisive clinical proof may not arrive until 2030.
By Rebecca Stone5 min read1,054 words
Summary
- Top 20 pharma AI drug-discovery deal value doubled to $29.5 billion in 2025; only $600 million (2%) was paid up front
- Venture and PE funding for AI-enabled drug discovery doubled from $4.1bn (2023) to $8.4bn (2025)
- Of 38 deals signed in 2023–2024, less than 1% of maximum milestone value had been paid by June 2026
- Average drug development cost rose to $2.67 billion in 2025 from $2.23 billion in 2024, per Deloitte
- McKinsey does not expect a critical mass of AI-drug clinical readouts until around 2030

Pharma's top 20 companies signed AI drug-discovery deals worth a combined $29.5 billion in 2025, but only $600 million — about 2% — changed hands up front, according to McKinsey. Of 38 deals signed in 2023 and 2024, partners had collected less than 1% of maximum milestone value by June 2026. The gap between headline value and cash delivered frames the central question for R&D managers: whether AI can shorten drug development enough to justify the spend, and when anyone will know.
How much money is moving?
Venture and private equity funding for AI-enabled drug discovery roughly doubled from $4.1 billion in 2023 to $8.4 billion in 2025, McKinsey reports. Headline deal values among top-tier pharma doubled in parallel. Meanwhile, Deloitte puts the average cost of developing a drug from discovery to launch at $2.67 billion in 2025, up from $2.23 billion in 2024 and roughly $2 billion in 2021. AI investment is rising faster than the cost curve it is meant to bend.
Among AI-enabled programs with sufficient self-reported and public data, first-in-human or equivalent preclinical milestones arrived 15% to 80% faster than large-pharma medians, McKinsey's September report found. That is a measured result from early-stage work, not from clinical development. Separating the two matters: no AI-native asset has yet produced the Phase 3 efficacy readouts that would settle the question.
Why is the feedback loop so slow?
Guang Yang, a partner at McKinsey, points to the multi-property problem. "Developing a successful drug candidate requires balancing many different properties, from potency and selectivity to safety and how the molecule behaves in the body," he said. "Different computational models can help predict and optimize these properties, but the real challenge is finding a molecule that meets the necessary criteria at the same time."
He described a recent biotech request: a small molecule that modulates a specific ion channel, beats existing compounds on selectivity, penetrates the brain, and shows a suitable pharmacokinetic and pharmacodynamic profile (how the body processes the drug and how the drug acts on the body). "They're essentially looking for a unicorn, a molecule that sits at the intersection of many demanding requirements," Yang said.
Molecule design attracts AI investment because experiments report back quickly: a model proposes an antibody, scientists synthesize it, and an assay shows within weeks whether it binds. "If you go through 1,000 iterations of feedback, the model will unquestionably improve," Yang said. The same fast signal exists in software engineering and competition math. In drug discovery, a binding result is one answer among many, and the decisive answer — whether a drug helps patients — takes years. "Imagine every time you tell the model it's wrong, you have to wait five years, versus just a month," Yang said.
What happens in the lab?
Trust and data hygiene limit adoption at the bench. Christian Baber, chief portfolio officer at the Pistoia Alliance, a life sciences collaboration organization, said chemists' distrust of large language models is a persistent issue. "It's quite difficult to build trust and really easy to lose it," he said. A confident error backed by an invented reference can end a scientist's willingness to use a system, because the models remain probabilistic even as performance improves.
One workaround the Alliance uses: have the LLM write deterministic tools, such as Python scripts that convert instrument data, then test and validate the script for reuse. "You use something unpredictable to generate something predictable, and that's what you use," Baber said. Regulators are more receptive to that approach.
Data-sharing practices also lag. Some labs share reports while keeping underlying data internal; even labs with FAIR data systems (findable, accessible, interoperable, reusable) keep data inside the group. Baber wants researchers to record why they ran an experiment, what they expected and what they concluded. Some companies now count only what reaches the electronic lab notebook, forcing conversations like: "I did 30 syntheses last month." "Your ELN says you didn't do any." "Well, then you didn't do them." What is missing, he said, "is the carrot for people."
Is culture the real bottleneck?
The Pistoia Alliance's January white paper, Elevating Organizational Change Management in Pharma and Life Sciences, drew on five structured sessions and survey data from seven pharmaceutical companies and two technology partners. Participants described "change theater" — elaborate communications and rollout plans that leave users' needs unresolved. Most respondents rated practices such as pre-rollout user preparation, clear accountability and leadership advocacy as unmet.
Some leaders "will quite often talk the talk but not walk the walk," Baber said, adding extra work without extra time. He compares digital transformation to a chemical reaction that needs energy up front before settling at a lower state. His alternative to incentives is simple: allowing time. He also warns against rewarding volume of candidates handed downstream: "If your carrot is how many things are thrown over the fence, all you'll do is barely clear the fence."
Yang sees companies running two tracks: decentralized tool experimentation by individual researchers, whose tools "may change every three months," and institution-wide bets that demand executive conviction years before results appear. Some companies are building supercomputing capability ahead of proven returns; others are holding back and risking a late start if the technology delivers around 2030; a third group shifts early risk to partners through milestone-tied payments.
What comes next?
McKinsey does not expect a critical mass of readouts for AI-originated drugs until around 2030; most AI-native clinical assets remain in Phase 1 or Phase 2, and roughly 90% of assets entering Phase 1 historically never reach approval. One-quarter of Phase 2 and Phase 3 failures stem from unpredicted safety issues — a prime AI target, though whether AI can meaningfully raise clinical success rates "remains an open question," Yang said. Trial operations offer a faster loop: enrollment data shows within about a month whether the right sites were chosen. "Right now, Phase 1 to approval takes roughly seven years. Imagine squeezing that down to four," Yang said. For now, AI may be increasing lab work through a "training bump" of large-scale data generation — and the industry will not know when the balance tips until the readouts arrive.
via sec.gov (Original)
Filed under
- ai-drug-discovery
- pharmaceutical-r-d
- milestone-based-deals
- laboratory-data-sharing
- organizational-change-management
More from Rebecca Stone
Show full bio
Market editor covering regulation, policy and their impact on research and development at Hypothesis Wire.
274 articles
References
- Nature asks if AI can restructure translational research pipeline
- Isomorphic Labs predicts AI-designed drugs as $5bn raise talks surface
- Anthropic’s $14.23B R&D budget nears Eli Lilly’s $14.6B
- AI Money Moves From Prediction Models to Research Infrastructure
- AI accountability, FDA shifts and a $1.15 billion deal