NextGen Insights: The Real AI Edge in Drug Development and How Investors Can Take Advantage of It

Why the sharpest value-add from AI in drug development sits downstream, and what makes those companies worth backing.

The money floods into the front of the pipeline, the platforms promising to surface the next molecule, because that is the part of the story that sells. But discovery is a lottery: for every 5,000 to 10,000 compounds screened, only a few hundred are worth advancing, and the odds get longer from there [2]. Someone has to play it. The question is whether it deserves the bulk of an AI investor’s chips, given it sits on the least predictable stretch of the whole process.

The harder, more valuable problem comes next. A molecule that survives discovery still must run the development gauntlet before a patient ever sees it: manufacturable, formulated to stay stable, scalable, and safe enough to file for a trial. Most never make it. Of the roughly 250 compounds that enter preclinical testing, only 5 to 10 reaches a first in- human trial [2]. Turning a promising molecule into one that can be produced is a bigger, nearer value inflection than sourcing molecule number 10,001, and unlike discovery it runs on physics and chemistry that behave predictably, which is where AI earns its keep.

Where the AI money goes, and where the opportunity lays

The mispricing in a sentence: AI dollars pile into the discovery-facing tools built to hand you a lead candidate, and barely trickle into everything downstream (manufacturing, formulation, and the grind of getting through trials and filings), where AI actually delivers [3].

Discovery AI pulls the money because it is legible, headline-ready, and sounds enormous. It also sits on the biology we still cannot predict, which makes it the most expensive bet and the least certain to pay off. The downstream layer offers no such headline. What it offers is structured problems on assets that already have value: by the time a candidate reaches process development, the question is not whether the biology works but whether the molecule can be made, stabilized, and scaled without burning a fortune. Those are engineering questions with measurable answers. None of this argues against funding new drugs. It is a narrower claim: for AI specifically, the return you can count on is downstream, and that case only gets easier as discovery-AI platforms raise enormous rounds and deliver few approvals while the quieter layer starts to get rewarded.

What you are actually investing in

So, what are you actually buying when you back an AI biotech company? Not a model or a molecule. You are buying a position in what Ron Adner calls a value architecture [1]: the whole system of pieces that must line up for a product to create value, and for the company to capture it. The algorithm or the lead compound is just the surface. What makes a company durable is the machinery around it, and whether that machinery keeps paying off across the next program and the one after.

That distinction is the whole game early on. A company whose only asset is a discovery model gets one shot, and a sharper model can leapfrog it next year. A company whose AI sits in the development layer owns something every molecule has to pass through. One is a bet on a single outcome; the other is infrastructure that compounds.

How to underwrite it

Score any AI application in biotech on three things.

Certainty. Can you trust it to work? Late-stage human efficacy is messy, high-variance biology, so a model’s output is hard to validate. Manufacturing and formulation behave like an exact science you can check against ground truth, so the ROI is legible. Structured beats unpredictable.

Position. A modest gain applied late, on an already-de-risked and valuable asset, moves more money than a dramatic gain on something still worth almost nothing.

Carryover. Does it serve one molecule or every program after it? Carryover separates a feature from a platform.

Run real research through that lens, and the pattern holds. Researchers at the University of Oklahoma, using bioreactor data from Wheeler Bio in Oklahoma City, built a model that predicts a cell line’s final antibody output from its first nine days in culture, trimming a two-week screening step to nine days, for a projected fifth off the cost per run and throughput rising from about 18 runs a year to 26 [5]. Elsewhere, models lift antibody titer by nearly half through better growth conditions [4], flag whether a molecule can even be manufactured before a dollar is spent making it [6], and pre-screen a drug’s stabilizing ingredients, the step that often decides whether a peptide is a convenient product or an awkward one [7]. Each scores high on certainty and carryover. None guesses what biology will do.

Who is already building it

This is not a thought experiment. Abinitio Bio builds foundation models for biomanufacturing that predict process outcomes, aiming to compress six-to-eighteen-month decisions into hours [9]. Persist AI pairs machine learning with a robotic lab for formulation, cutting long-acting-injectable development from about a year to two months, backed by Spero Ventures with Eli Lilly in the round [10]. Culture Biosciences runs cloud-connected bioreactors with an AI layer, closed a Series C in late 2025, and deepened a Cytiva partnership; an incumbent that size leaning in is its own signal [11].

Two things to notice. The differentiations that matter are not “a bioprocess data platform,” since several exist; it is genuine prediction, an underserved segment, or proprietary data nobody else holds. And every one of these sits on a coast, while the biomanufacturing base and university research that feed the layer are spread far wider than the capital, a gap for an investor willing to look past the usual zip codes. Systems like these need a rare mix of bioprocess science, machine learning, and hard engineering, which is exactly what a strong university research ecosystem can produce when its labs are paired with working manufacturers rather than left to publish alone. A region with both, as the earlier university-and-manufacturer collaboration shows, can turn its own research into companies built close to the plants and the talent, instead of waiting to buy into rounds already priced on the coast.

It doesn’t stop at the bench

The same logic runs all the way to approval. Once a drug is in trials, a mountain of operational and regulatory work decides whether it reaches a patient, and most of it has nothing to do with biology: standing up trials, organizing trial data, running the statistics regulators expect, and drafting the filings. A single Investigational New Drug application runs to hundreds of source files and thousands of hours [12], and every month a drug waits instead of selling costs its sponsor tens of millions [13].

Mind which part AI can touch, because the certainty rule still holds. Predicting whether a trial succeeds is the low-certainty biology bet again; running the machinery around it (enrollment, data management, statistical programming, and regulatory drafting) is the structured work where AI is reliable, and the FDA is starting to bless it [14]. The companies are here: Weave Bio’s AutoIND compresses IND assembly [12]; Ritivel drafts INDs, study reports, and biologics license applications to FDA formatting from a company’s own data [13].

This layer already mints giants. Veeva is a roughly $35 billion company built on the regulated data-and-compliance backbone of drug development (trial documents, submissions, safety, and quality) with no drug discovery at all; Medidata sold for $5.8 billion on clinical-trial software [15]. The catch: those outcomes came from becoming the system of record, not from being a point tool, which is the bar a new entrant has to clear.

The real moat

Here the science changes the investment case rather than decorating it. In biomanufacturing, data is scarce; a company might have a few dozen runs of a process, not the millions of examples a large model wants. Counterintuitively, that is the moat. The models that win are not the biggest; they bake in known biology and use machine learning only for the gaps. The Oklahoma model beat a deep-learning approach precisely because it anchored in decades-old fermentation kinetics [5]. Compute can be rented and off-the-shelf models copied; a decade of proprietary process data, and the judgment to use it, cannot. It is also what incumbents cannot easily assemble: a drugmaker has deep data on its own processes [8], but an independent company pooling data across many clients sees patterns no single pipeline contains. The defensible asset is not the algorithm. It is the proprietary, cross-client data feeding it, a value architecture with carryover built in.

The takeaway

The instinct in AI-biology is to fund the moonshot, the model that designs the drug. That work has to be funded, and someone will win it. But set the pieces of this paper side by side, the attrition math, the certainty and carryover test, the companies already in the ground, and the data moat that cannot be bought, and a narrower claim holds: the opportunity is real, and it is already moving, the AI that makes drugs faster to manufacture, cheaper to formulate, and easier to move through trials and filings. It scores high on certainty, acts where assets are already valuable, and compounds across every program it touches. It is being built by people who know the science well enough to see where the value sits, and it rewards the investor who can read the biology and the operations of scaling a product at the same time. That investor is not alone: strategics and specialist funds are already writing into these rounds, and regulators have started qualifying the tools. What the trend still needs is capital convinced these companies can return it, because that conviction is what tells the next set of founders they can build a scalable company next to the plants and talent that already exist across the middle of the country, rather than assuming the only credible address is a GMP suite in Boston. That is where AI aligns with fund timelines and expectations in drug development, and where the private markets have to reach next.

Works Cited

1. Ron Adner, Winning the Right Game: How to Disrupt, Defend, and Deliver in a Changing World (MIT Press, 2021). Value architecture and value proposition.
2. Drug-development attrition funnel: for every ~5,000–10,000 compounds screened, ~250 enter preclinical testing, 5–10 reach a first-in-human trial, and ~1 is approved (standard PhRMA / DiMasi funnel; figures approximate). The “~90% fail” figure applies only to candidates already in clinical trials and excludes preclinical attrition. D. Sun et al., “Why 90% of clinical drug development fails and how to improve it?” Acta Pharmaceutica Sinica B 12(7):3049–3062 (2022), DOI 10.1016/j.apsb.2022.02.002.
3. The largest share of AI-biopharma venture funding goes to drug-discovery platforms (AI-drug-discovery investment near $1.7B in 2024 and growing), while the development and manufacturing tooling layer draws a small fraction. EY life-sciences AI analyses and sector funding trackers (2024–2025). Directional; based on industry funding estimates rather than a single audited figure.
4. J. Richter et al., “Machine Learning-Powered Optimization of a CHO Cell Cultivation Process,” Biotechnology and Bioengineering (2025), DOI 10.1002/bit.28943. An artificial neural network trained on historical and new CHO runs raised final mAb titer to 4.65 g/L from 3.15 g/L (~48%) in validation.
5. Penghua Wang, Deepika Verma, Yuk Chiu, John Klier & Chongle Pan, “Luedeking-Piret regression for multi-step-ahead forecasting and clone selection in monoclonal antibodies biomanufacturing,” Communications Engineering 4, 220 (2025), DOI 10.1038/s44172–025–00547–7. University of Oklahoma researchers using Wheeler Bio (Oklahoma City) bioreactor data; 5.85% mean error predicting day-16 titer from day-9 data; the paper projects a ~20% cost reduction and a rise from ~18 to ~26 runs per year from a shortened assessment.
6. P. G. Hawkins et al., “PROPERMAB: an integrative framework for in silico prediction of antibody developability using machine learning,” mAbs 17(1):2474521 (2025), DOI 10.1080/19420862.2025.2474521 (Regeneron).
7. “Machine learning driven acceleration of biopharmaceutical formulation development using Excipient Prediction Software (ExPreSo),” Computational and Structural Biotechnology Journal 27:4517–4525 (2025), DOI 10.1016/j.csbj.2025.10.026. Trained on 335 FDA-approved protein and peptide drug products. (Authors are affiliated with Leukocare AG, a formulation-services company.)
8. S. Panjwani et al., “Artificial intelligence and machine learning-assisted digital applications for biopharmaceutical manufacturing,” Biotechnology Progress (2026), Wiley/AIChE; major drugmakers report applying AI to compress manufacturing and CMC timelines.
9. Abinitio Bio, foundation models for biomanufacturing (Y Combinator, Spring 2026; abinitio-bio.com). Founder/CEO Daniel Mukasa. The “Echo” model and the six-to-eighteen-months-to-hours claim are the company’s own.
10. Persist AI, AI-plus-robotics formulation “Cloud Lab”; $12M Series A led by Spero Ventures with Eli Lilly & Company, Y Combinator, and others (May 2025); reported identifying an optimal long-acting-injectable formulation in ~2 months on a pharma project.
11. Culture Biosciences, cloud-connected bioreactors with the AI-enabled Console platform; Series C closed December 2025 (Northpond Ventures, S32, and others); Cytiva (a Danaher company) named exclusive global distributor of the Stratyx 250 (August 2025).
12. Weave Bio, AI-native regulatory automation platform; AutoIND accelerates IND submissions that traditionally rely on hundreds of source files and thousands of hours; $10M seed co-led by Innovation Endeavors and Magnetic Ventures (2024).
13. Ritivel, AI workspace drafting INDs, clinical study reports, and BLAs to FDA formatting (Y Combinator). The ~$45M-per-month cost of launch delay is the company’s own cited estimate.
14. FDA qualified AIM-NASH (developed by PathAI) as its first AI drug-development tool, via the Drug Development Tool qualification program, to standardize liver-biopsy scoring in MASH clinical trials (December 8, 2025; fda.gov). Industry analysts (e.g., McKinsey) estimate AI could materially shorten development timelines.
15. Veeva Systems, cloud software for life-sciences regulatory, clinical, quality, and safety operations, roughly $30–40B market capitalization on ~$3.2B annual revenue (2026); Medidata (clinical-trial software) acquired by Dassault Systèmes for $5.8B (2019). Illustrative that the regulated operational layer of drug development produces multi-billion-dollar outcomes without drug discovery.