The Acceleration: Is AI Actually Changing the Economics of Hard Science?

1. Introduction

Most of the AI investment conversation lives in software — chatbots, agents, copilots. The more consequential question for frontier allocators sits somewhere less visible: can AI change the economics of hard science — the slow, expensive, failure-ridden business of discovering new drugs, new molecules, new materials?

That matters because scientific discovery is almost uniquely broken as an economic process. Bringing a single drug to market takes 10 to 15 years and roughly $2.6 billion, and about 90% of candidates that enter human trials fail. [1][2] Discovering and commercialising a new material has historically taken a decade or more. These are the timelines and failure rates that AI now claims to attack — not at the edges, but at the core.

The test is no longer hypothetical. In September 2026, two days after our Dubai event, Insilico Medicine dosed the first patient in the world's first Phase III trial of a drug whose biological target and molecule were both discovered by generative AI. [3] It is the clearest piece of evidence yet in a much larger question. This piece examines that question across the three scientific domains where AI has advanced furthest — and asks what a frontier allocator should actually believe.

Alex Aliper, President of Insilico Medicine, at the Future Investments Circle Dubai in September 2026.

2. The Pattern: Where AI Changes the Curve

To understand the investment case, it helps to understand exactly where in the scientific process AI intervenes — because it is not everywhere, and the distinction is the whole story.

Scientific discovery, simplified, runs in two halves. The first is discovery and design: choosing what to make — which biological target to pursue, which molecule to synthesise, which material to attempt. The second is validation: proving in the physical world that the thing actually works and is safe, through lab synthesis, animal testing, and — for drugs — human trials.

AI has made dramatic progress on the first half and far less on the second. It is extraordinarily good at searching vast combinatorial spaces, predicting structures and properties, and proposing high-quality candidates faster and more cheaply than human researchers. It cannot yet shortcut the physical validation that remains the real gate — the part where biology, chemistry, and materials physics assert themselves in ways no model has fully captured.

This is why the economic argument is subtler than "AI makes science cheap." The value of AI is not that it removes the expensive stage; it is that by producing better candidates, it means fewer of them fail at the expensive stage. In a business where cost is dominated by failure, improving the hit rate is worth more than improving the speed. Whether AI genuinely improves that hit rate, all the way through, is the question each domain below tests.

3. Proof Domain One: Drug Discovery

Drug discovery is where AI science is furthest along, and where the evidence is richest.

The field has scaled with startling speed. AI-designed drug programs in clinical development rose from roughly 24 in late 2023 to more than 200 by early 2026 — of which, by one count, 94 were in Phase I, 56 in Phase II, and 15 in Phase III. [4] Cumulative investment into AI drug discovery since 2019 is estimated at around $60 billion across roughly 175 clinical-stage programs. [5]

The lead evidence is Insilico Medicine. Its drug Rentosertib — for idiopathic pulmonary fibrosis, a progressive and fatal lung-scarring disease — is, by the company's description, the first drug with both an AI-identified target and an AI-generated molecule to reach Phase III. [3] The target, a protein called TNIK, was surfaced by Insilico's AI platform as a fibrosis driver not previously associated with the disease. [6] The Phase IIa trial produced encouraging signals: patients on the 60mg daily dose showed a mean improvement in lung function of 98.4 mL over 12 weeks, against a 20.3 mL decline on placebo. [7] The reported discovery cost was roughly $6 million, a fraction of the conventional figure. [1] Insilico is now public — a December 2025 Hong Kong listing, the largest biotech IPO there that year, with Eli Lilly and Tencent as cornerstone investors — and reported around $56 million in 2025 revenue, with partnerships spanning Lilly, Takeda, Menarini and others. [8]

The most rigorous measure of AI's clinical impact comes from a Boston Consulting Group analysis: AI-discovered molecules have passed Phase I at an 80–90% success rate, versus 40–65% historically — a genuine, large improvement in the early safety stage. But at Phase II, the same molecules succeed at roughly 40%, in line with the historical average. [4][9] That split is the single most important fact in the field: AI is demonstrably better at designing a safe, well-behaved molecule, and has not yet shown it can better predict whether a drug will actually treat a disease in a human body.

AI roughly doubles early-stage (Phase I) success by designing better molecules — but Phase II, where a drug must prove it treats disease, is unchanged. Source: BCG analysis, Drug Discovery Today / Clinical and Translational Science.

4. Proof Domain Two: Protein and Biologics Design

If drug discovery is the proof domain, protein design is the frontier that could go further — and it carries the field's most famous name.

AlphaFold, Google DeepMind's protein-structure-prediction system, won its creators the 2024 Nobel Prize in Chemistry and has been used by nearly 190,000 researchers in the UK alone. [10] Its commercial successor, Isomorphic Labs, is the best-funded company in the entire field: a $2.1 billion Series B in May 2026 led by Thrive Capital — the largest financing yet in AI drug discovery — with backers including Alphabet's GV, Temasek, Abu Dhabi's MGX sovereign AI fund, and the UK Sovereign AI Fund. [11] Its drug-design engine, IsoDDE, reportedly more than doubled AlphaFold 3's accuracy on the hardest protein-ligand prediction cases. [11]

And yet Isomorphic illustrates the field's central tension as sharply as Insilico does. Despite being the best-capitalised company in AI drug discovery, it remains entirely preclinical: no human patient has yet been dosed with an Isomorphic-designed compound, and its first clinical trial — originally targeted for end of 2025 — has slipped to the end of 2026. [11][12] By one estimate, its $2.1 billion round accounted for roughly 87% of all disclosed AI-drug-discovery venture capital in the first half of 2026 — meaning the sector's headline funding figure is dominated by a single company that has not yet entered the clinic. [4] That concentration is itself a risk signal worth reading: enormous capital is betting on a capability that remains, for now, unproven in humans.

The sector's record-funding headline is driven overwhelmingly by a single, still-preclinical company — a reason to read "record investment" with care. Source: [4].

The broader biologics wave is real nonetheless. Earendil Labs raised $787 million in March 2026 for AI antibody and biologic design, backed by Sanofi and Pfizer's venture fund; Xaira Therapeutics, Generate Biomedicines, and Chai Discovery (a $400 million round at a $3.8 billion valuation in July 2026) round out a crowded, well-funded field. [13][5]

5. Proof Domain Three: Materials Science

The third domain is the one most directly connected to everything else we cover — because new materials underpin better batteries, chips, solar panels, and grids.

The signature breakthrough is DeepMind's GNoME, which in 2023 predicted 2.2 million new crystal structures, of which roughly 380,000 were flagged as stable candidates — an expansion of the known catalogue that one analysis equated to roughly 800 years of manual discovery. [14][15] Microsoft followed with MatterGen, a generative model for designing materials to target properties, in early 2025. [16] In December 2025, DeepMind went further into the physical world, announcing with the UK government its first automated materials-science laboratory, pairing its Gemini model with robotics designed to synthesise and test hundreds of materials per day from 2026. [17]

But materials science also offers the field's most instructive cautionary tale — and the clearest illustration of the discovery-versus-validation gap. The celebrated A-Lab experiment at Berkeley, which claimed in 2023 to have autonomously synthesised 41 new compounds in 17 days, was the subject of a Nature correction in January 2026 after independent researchers questioned how many of the compounds were genuinely novel. [16] The episode captures the field's real bottleneck: predicting a stable material computationally is now cheap and fast; actually synthesising it, confirming it, and proving it useful remains slow, physical, and hard. The investment response has rightly shifted toward closed-loop automated labs — AI planning, robotic execution, and instrument characterisation in a continuous feedback cycle — because that is where the validation gap actually gets closed. [16]

The market reflects the optimism: generative AI in materials science is projected to grow from about $1.1 billion in 2024 to $11.7 billion by 2034. [18] As with the other domains, the question is whether the predictions translate into validated, manufacturable reality.

In every domain, AI has transformed computational discovery while physical validation remains the bottleneck. Sources: [3][4][11][14][16].

6. The Bear Case: What AI Has Not Changed

A credible investment view has to hold the hype at arm's length, and the evidence for scepticism is substantial and consistent across all three domains.

The defining fact remains: across roughly $60 billion and ~175 clinical-stage AI-discovered drug programs, there have been zero FDA approvals. [5] Analysts place roughly 60% odds on the first approval arriving in 2027 or 2028 — meaning the entire thesis still awaits proof at the only milestone that ultimately matters. [19]

The cautionary cases are real and recent. Recursion Pharmaceuticals — one of the most prominent AI-biotech names — saw its stock fall roughly 60% from 2024 to 2025 and had two lead programs stumble in Phase II readouts. [19] BenevolentAI went through three restructurings after its lead drug missed its primary endpoint. [19] Exscientia's AI-designed OCD candidate, once celebrated for compressing the discovery cycle from years to months, was discontinued in clinical development — a clean demonstration that faster discovery does not guarantee therapeutic success. [20] The chemist Derek Lowe and investors such as Bessemer's Andrew Hedin have made the same cautionary point: most clinical failures stem from efficacy and safety problems that only surface in human testing, and AI has not yet changed that. [21]

The common thread, across drugs, proteins, and materials, is the discovery-versus-validation gap. AI has genuinely transformed the half of science that happens in silico. It has not yet transformed the half that happens in the physical world — and that second half is where most of the cost, the time, and the risk still live.

The first approval is expected in 2027–2028. Until then, the entire thesis awaits proof at the only milestone that counts. Sources: [5][19].

7. What This Means for Allocators

The honest synthesis is neither "AI has cracked science" nor "it is all hype." It is that AI has measurably improved the economics of discovery while leaving the economics of validation largely intact — and that distinction should drive how capital is positioned.

Revenue and partnerships beat narrative. The most de-risked companies are those that have crossed from pure computational promise into real revenue, pharma partnerships, and — critically — clinical-stage assets. Insilico, with a Phase III drug, public-market discipline, and paying pharma partners, sits in a different risk class from a preclinical company valued in the billions on the strength of a platform alone. The Isomorphic concentration — one preclinical company absorbing most of the sector's capital — is a caution, not a template.

The picks-and-shovels layer is the most durable. As in defense, space, and firm power, the enabling infrastructure captures value regardless of which specific drug or material wins: the AI platforms themselves, the lab-automation and closed-loop robotics that close the validation gap, the proprietary datasets, and the compute. A company that sells the shovels earns whether or not any individual dig strikes gold.

2027–2028 is the real test. The field is approaching its first genuine verdict. With 15-plus AI-discovered drugs in Phase III and the first approvals expected in 2027–2028, the next 24 months will convert the thesis from projection to evidence — in either direction. This is patient capital: a position to build before the proof, underwritten with clear eyes about the binary ahead.

This was, in essence, the conversation at Future Investments Circle's Dubai event on September 7th, held alongside the AIM Congress. Our longevity panel — convened with our sister platform, Longevity Investors — brought Dr. Alex Aliper, President of Insilico Medicine, into a discussion of exactly this question: whether AI is genuinely rewriting the economics of biotech, or simply accelerating its earliest and easiest stage. The evidence, as ever at the frontier, is real, early, and not yet finished.

The field is approaching its first real verdict. Sources: [3][4][11][14][19].

8. Conclusion

The most important test of artificial intelligence is not taking place in a chat window. It is taking place in fibrosis wards in China, in robotic materials labs in Britain, and in the preclinical pipelines of the best-funded startups in biology.

The verdict so far is genuinely split. AI has changed the economics of scientific discovery — faster, cheaper, better candidates, and a real improvement in early-stage success. It has not yet changed the economics of scientific validation — the slow, physical, failure-prone business of proving those candidates work, where zero approvals still stand as the field's defining number.

For allocators, that makes AI-for-science one of the most interesting frontiers available: a thesis with real, measurable early progress, a clear and imminent test ahead, and a wide gap between the companies that have proven something and those merely promising it. The acceleration is real. Whether it reaches the finish line is the question the next two years will answer.