AI is reshaping the underlying logic of drug discovery, and an investment opportunity similar to the semiconductor industry's "bottleneck trade" is quietly taking shape within the pharmaceutical supply chain.

On October 4, Freda Duan, a partner at top-tier dollar fund Altimeter Capital, recently published an article pointing out that the "bottleneck trade" in the AI drug discovery (AIDD) sector is taking shape, with a structure highly similar to the previous chain-style rally in semiconductors that extended from GPUs to HBM, networking, and power.

The collective entry of frontier AI labs is the most notable signal. Anthropic has formed a dedicated life sciences research team and established a wet lab, OpenAI has built the GPT-Rosalind model family designed specifically for biology and drug discovery, ByteDance's independently operated Anew Labs already has an internal drug R&D library, and Isomorphic Labs has explicitly stated it will push AI-designed drugs into human clinical trials.

At the same time, signals from the supply chain side are equally clear. DNA synthesis company Twist Bioscience (TWST) expects AI drug discovery orders to achieve triple-digit percentage growth for both fiscal 2026 and fiscal 2027. GenScript's AI drug discovery business doubled year-over-year in the first half of fiscal 2026, and channel research indicates its actual validated capacity has reached approximately 8,000 designs per day, with the potential to increase to 16,000 by the end of 2026.

This logic is highly similar to the bottleneck transmission path in the semiconductor industry. GPU scarcity drove sequential explosions in demand for HBM memory, networking equipment, power, and cooling systems.

In AI drug discovery, the bottleneck is likewise being transmitted down the industry chain: AI models generate designs → DNA synthesis → protein production → experimental validation → preclinical testing, and each link could become the next beneficiary node.

(Supply chain data shows that upstream discovery and early preclinical R&D segments have already begun to feel the impact of increased experimental volume.)

Frontier AI Labs Accelerate Deployment, Drug Discovery Enters a Paradigm Shift

In the article, Freda Duan describes the current AI drug discovery landscape as a genuine "regime change."

Anthropic has elevated life sciences research to a strategic level, and the first major project completed by its wet lab was using Claude agents to identify a previously uncharacterized enzyme system related to CRISPR-like DNA repeat sequences.

OpenAI's GPT-Rosalind model family covers medicinal chemistry, protein engineering, genomics, and experimental workflows.

ByteDance's former AI drug discovery department now operates independently under the name Anew Labs, integrating foundation models, wet lab validation, and its own drug pipeline, with ByteDance still holding a majority stake.

Isomorphic Labs has explicitly announced that its goal is to advance AI-designed drugs into human clinical trials.

It is worth noting that there are clear differences in the commercialization paths of these institutions.

Anthropic's current positioning is closer to the research and platform layer, and it has explicitly stated it does not intend to become a traditional commercial clinical-stage biotech company;

ByteDance's Anew Labs is closer to a biotech platform model with its own drug pipeline;

Isomorphic Labs sits between the two, with an explicit commitment to advance drugs into clinical trials itself.

This means the commercialization paths for AI drug discovery are diverging: licensing to large pharmaceutical companies, collaboration milestones plus royalties, or directly incubating independent companies. For the downstream supply chain, regardless of which path is taken, the incremental demand for experiments will be real.

Why AI Actually Increases Wet Lab Demand

AI improves efficiency, so why does it actually increase wet lab experimental volume? The answer lies in the fact that AI fundamentally changes the economics of hypothesis generation.

First, hypothesis generation approaches zero cost, and experimental validation becomes the new bottleneck. In traditional drug discovery, the cost of experts screening candidate molecules is extremely high, and only the highest-confidence candidates can enter the expensive validation process.

AI makes the design of candidate molecules almost free at the margin, greatly increasing the "chances to play," and the bottleneck shifts from the design side to the biological validation side.

Second, failed experiments are also valuable, as negative samples become training data. Anthropic provides a typical case: Claude designed 1,320 protein binders, Adaptyv converted these digital sequences into DNA, expressed the proteins, and tested binding, with only 354 successfully binding.

But the other 966 failures were not wasted—what does not express, what does not bind, what has poor affinity—these are all valuable negative labels used to train the next round of model iteration. Traditional drug discovery only asks "does candidate X work," while AI drug discovery additionally asks "what can this experiment teach the model."

Third, wet labs are transforming from making drugs to making training data. TWST has explicitly stated that its business is shifting toward generating structured experimental result data from AI-designed sequences that can be fed directly into models. GenScript's sequence-to-binding data workflow can return results within 4 to 7 days.

In some workflows, what customers care about may no longer be obtaining the physical protein itself, but rather feeding structured experimental results back into the model. TWST and GenScript are therefore increasingly approaching the role of "biological data foundries."

Supply Chain Data Confirms Demand Explosion, Capacity Enters Rapid Expansion Track

The article points out that beyond theoretical logic, trackable quantitative signals have already emerged at the supply chain level.

TWST expects AI drug discovery order growth to reach triple-digit percentages in FY26, with order growth in FY27 continuing at the same magnitude. GenScript's AI drug discovery business doubled year-over-year in the first half of fiscal 2026, and its platform publicly advertises industrial-grade validation capacity of more than 4,000 designs per day, equipped with a complete sequence-to-data workflow.

Channel research indicates actual progress is even faster. Current actual capacity is approximately 8,000 designs per day, with the potential to reach approximately 16,000 by the end of 2026.

From the supply chain path perspective, the beneficiary segments in order are: AI models → sequence design → DNA synthesis (TWST, etc.) → protein production (GenScript, etc.) → experimental validation → sequencing (ILMN, TXG, etc.) → automation equipment → preclinical testing (CRO).

In terms of market size, if AI is merely a more efficient R&D tool, the relevant spending pool is the $300 billion to $400 billion in total global pharmaceutical R&D expenditure per year. If AI truly expands the number of viable drug programs, the opportunity will be even larger, because it will also drive incremental demand for DNA synthesis, protein production, experimental validation, and preclinical work.

In addition, an approximate estimate can be made from the perspective of AI companies' own spending. If Anthropic achieves $80 billion in annualized revenue in 2026 and invests just 1% of that into AI drug discovery, that alone implies an annual investment volume of approximately $800 million.

How This Time Differs from 2020/21

Freda Duan cautions that investors familiar with this theme will instinctively recall the 2020 to 2021 AI drug discovery boom and the subsequent crash.

Exscientia's DSP-1181 molecule advanced from discovery to Phase I clinical trials in less than a year, but was subsequently terminated for failing to meet the Phase I evaluation criteria for the target mechanism. The failure at that time had two clear causes: a sharp tightening of the macroeconomic environment, combined with the brutal confirmation that "faster discovery does not equal better clinical biology."

But Freda Duan believes the 2026 backdrop is fundamentally different from 2020/21, as that cycle occurred before the emergence of GPT.

The core logic of 2020/21 was: use computation to reduce low-value wet lab screening, with AI as a search efficiency tool, aiming to reduce physical experimental volume and lock onto candidate drugs faster with fewer syntheses and fewer tests.

The core logic of 2026 is entirely different: use AI to greatly expand the hypothesis space, then use wet labs to generate real validation data. Its internal loop is: AI → experiments → data → better AI → more experiments. The sheer number of usable experiments created by AI makes physical biology itself the bottleneck.

Based on existing data, a 2024 study showed that the Phase I success rate for AI-discovered drugs is approximately 80% to 90%, higher than the historical industry average; the Phase II success rate is approximately 40%, roughly in line with the historical average, though the sample size remains limited.

How the Bottleneck Will Transmit Along the Chain, and Where the Trade's Risks Lie

Freda Duan emphasizes that DNA synthesis and protein production are currently the most direct bottleneck nodes.

As more candidate drugs are pushed into the preclinical stage, animal testing may become the next constrained link. Primate prices have already approached previous highs, CRO capacity remains persistently tight, and AI drug discovery pushing even more candidates into preclinical work will only intensify this pressure.

Analysis suggests the potential duration of this trade is relatively long, and it will be difficult for a fundamental logic termination to occur until Phase I/II clinical results emerge after 2028. Key milestones:

From 2026 to the end of 2027, Isomorphic's first batch of AI-designed drugs enters human clinical trials, and more AI-native programs advance to the IND-enabling stage;

From 2028 to 2030, clinical trial results begin to reveal whether AI-designed drugs are truly superior to traditional drugs.

In the near term, if experimental budgets stop growing, AI-generated designs cannot be translated into effective wet lab results, or capacity catches up with demand too quickly, the upstream "picks and shovels" trade will come under pressure.

In the long term, if AI drugs perform brilliantly in the discovery stage and Phase I but fail at normal failure rates in Phase II/III, the entire investment logic will face a fundamental challenge—which is why Phase II clinical results are so critical.

Quanquan Gu, a professor at the University of California, Los Angeles, summarized this logic succinctly:

Biology may become AI's next frontier for recursive self-improvement—designed in computers, tested in labs, with experimental feedback converted into semi-synthetic training data, improving models, and repeating the cycle.