The funnel flipped
For thirty years the bottleneck was to invent a candidate molecule. AI removed exactly that constraint: generative models (RFdiffusion, AlphaProteo, ProGen, ESM) output meaningful sequences and structures in batches of 10⁴–10⁶ within hours, at near-zero cost apiece. What is scarce now is not design, but everything that comes after it — physically making, testing, de-risking, trialing in humans, manufacturing, and registering.
Design — expensive and slow
A candidate idea took years to produce: library screening, chemists' intuition, "make-and-test" cycles of weeks-to-months each. Design was bottleneck #1.
Design — nearly free and endless
AI generates candidates faster than they can be synthesized. The gap between "generated" and "physically tested" is on the order of ×1000 and growing. The bottleneck moved — it did not disappear.
Where exactly did it run off to?
The right framing is not "where is the one bottleneck" but "in what order do they flare up" as AI floods the top of the funnel. Below is a map of that cascade — and where the money sits on it.
The pipeline: from design to market
Eight stages of a biologic's journey. The tag shows how much each stage becomes a bottleneck under mass AI generation. Click a stage to reveal the figures.
Physical testing 10²–10³ per run vs 10⁹ generated (×1000 gap). Design hit rate 6%–100% per target, so you must test each. ~$5 / construct (DMX). Confirming function in vitro takes days-to-weeks and can't 'just' be made high-throughput.
Timeline/cost estimates are typical for a biologic (antibodies etc.); real values depend on the modality.
From a billion down to single digits
Each stage can pass orders of magnitude fewer than the previous one generates. The axis is logarithmic: each division is ×10. You can see where the pipe sharply narrows — and why even infinite design runs into the physics below.
Illustrative orders of magnitude (not exact per-program numbers). Top — how many candidates can realistically be generated/tested; bottom — how many drugs the whole industry brings to market per year.
Where it chokes first — and where the real ceiling is
Yes, testing drowns before everything else. But it is important to distinguish the nearest bottleneck (the wet lab) from the structural ceiling that no automation will lift quickly.
Near-term front — flaring now → 3 years
Test / MakeValidation (Test) — the prime candidate for choking. Physical assays handle hundreds–thousands of designs per run against billions generated. AI-design hit rates swing from 6% to 100% across targets — so you have to test each one; brute-force automation doesn't cut it.
Synthesis + expression (Make) — DNA is ordered in 2–10 days, protein in ~2 weeks. Costs fall fast, but 100–1000 genes from classic suppliers means weeks-to-months. A real bottleneck, but a shrinking one.
Structural ceiling — can't be lifted with code
Clinic / Mfg / FDAClinical trials — the final ceiling. 80%+ of trials miss enrollment timelines; 11% of sites enroll no patient. Patients and sites are finite and move at human pace.
Manufacturing (CDMO) — a capital wall. A new bio-plant takes 5–10 years and $0.7–6.5B. Bioreactor capacity grows ~8%/year against a ×2–3 surge in candidates.
Regulatory (FDA) — ~50 approvals a year, flat for a decade; in 2025 the FDA also cut ~3,500 staff. The Elsa AI tool speeds routine work, but not the approvals ceiling.
The key distinction
The wet lab can be widened with money and robots in 2–4 years (cloud labs, automation — already growing ~9–35% a year). The clinic, factories, and the regulator cannot be, at the same speed: there the constraint is live humans, physical capital, and institutions. So the truly scarce things won't be test tubes, but trial slots, factory capacity, and the regulator's attention.
It's not about speed — it's about predictive power
Even if testing became infinitely fast, it wouldn't double the number of drugs. Because the real scarcity isn't throughput, but signal: how well an early test result predicts success in humans.
This is Eroom's Law — the flip side of Moore's Law: over 70 years the cost of bringing a drug to market only rose, despite all the cheapening of chemistry and biology. The reason — at every stage we filter candidates with tests that correlate weakly with the clinic. Improving that correlation by even 0.1 is worth more than running 10× more molecules.
Fresh data confirms it: for AI-discovered drugs Phase 1 passes at 80–90% (vs 40–65% industry-wide) — because the molecules are cleaner. But Phase 2 is the same ~40% as everyone else: as soon as it comes down to real efficacy in patients, AI's edge evaporates for now.
AI drugs' success vs the industry, by phase
- AI drugs
- Industry
Source: Jayatunga et al., Drug Discovery Today, 2024 (67 AI molecules in the clinic).
Takeaway: the winner won't be whoever generates the most candidates, but whoever first builds a cheap predictive test — AI toxicity models, organs-on-chips, patient digital twins. That is the most valuable bottleneck relief of all.
Where the money flows — and where the bottleneck actually is
The market's core mismatch: capital pours into design (mega-rounds of $0.6–1B there), while the physical bottlenecks — validation and manufacturing — are underinvested. The higher and further left a stage sits, the sharper the "capacity shortage with cheap money coming in" — the opportunity zone.
Positions are the author's estimate from the collected data (2024–2026 investment vs bottleneck severity); bubble size ≈ how narrow it is. An illustration of the logic, not a market rating.
Six bets on bottleneck relief
Cloud and autonomous labs
Relieve Test/Make: "the wet lab as an API" — an AI model sends a design, a robot returns data. A direct cure for the main near-term bottleneck.
Adaptyv Bio (>10,000 proteins tested), Emerald Cloud Lab, Ginkgo, Trilobio ($8M), Lila Sciences ($435M)
Whoever fuses generation and validation into one loop captures the flow of AI-biotechs.
DNA synthesis and protein expression
"Printing" genes and producing protein at a speed and price that keep up with design. Direct upside from the flood of AI orders.
Twist Bioscience (5 AI clients = +$25M), Ansa Biotechnologies ($54.4M, DNA up to 50 kb), EvolutionaryScale
The "picks and shovels" layer — grows with every AI-biotech.
AI prediction of toxicity and ADMET
Attacks not throughput but predictive power — the most valuable relief of all. Cull the failure before the wet test and animals.
Simulations Plus, Schrödinger, Certara, ADMET-AI; ICH M7 already allows QSAR instead of the Ames test for impurities
0.1 more correlation with the clinic beats 10× more molecules (Eroom's Law).
Organs-on-chips and organoids
Replacing animals with human-relevant tests. A powerful regulatory tailwind: FDA Mod Act 2.0 + the 2025 roadmap.
Emulate ($82M Series E), CN Bio (working with FDA on antibodies), Hesperos, Mimetas, Quris-AI
Policy (FDA moving off animals) creates demand faster than method validation grows.
In-silico trials and digital twins
Compress the narrowest structural stage: fewer control-arm patients, faster enrollment, synthetic comparator arms.
Unlearn.AI ($50M, −25% control arm, qualified by EMA), Tempus (patient matching), QuantHealth
The only way to partly scale the clinic is to replace some patients with a model.
New-type biomanufacturing
The most underinvested bottleneck relative to its importance. Cell-free and continuous systems, re-shoring capacity for new modalities.
National Resilience (up to $825M), Debut Bio ($22M, cell-free), Cascade Bio; Ginkgo×OpenAI +40% to cell-free synthesis
Capital-heavy and unfashionable, so the shortage lasts longest — a premium for whoever builds capacity.
How the bottleneck migrates, 2026 → 2035
Relieve one bottleneck and the next one down the pipeline flares up. The most likely sequence:
The wet lab chokes. AI-biotechs hit the wall at validation and production. Explosive growth of cloud/autonomous labs and enzymatic DNA synthesis. The first bottleneck = your hypothesis.
Shift toward de-risking. Wet throughput partly solved — focus moves to predictiveness: AI-tox and organs-on-chips gain weight amid the FDA's move off animals. Quality of the cull, not volume, wins.
Clinic and factories become the limit. More candidates reach humans — but the pool of patients, sites, and CDMO capacity is finite. In-silico trials and new biomanufacturing are the scarcest resource.
The regulator and the health system. Amid a flood of filings, ~50 approvals/yr and paying for new drugs become the bottleneck. The question shifts from 'make it' to 'approve and pay for it'.
The whole thing in one sentence
AI turned protein design from a rare art into an endless stream. Further down the pipe, that stream meets the wet lab (relieved with money and robots over years), then live patients, factories, and the regulator (not quickly relieved by anything). So 'testing will choke' is right as the first wave; but the decade's strategic scarcity is predictive tests and clinical-manufacturing capacity, not molecule generation.
Sources — what this is based on
Compiled as a research map for assessing bottlenecks and opportunities at the intersection of bioengineering and AI. Research by Petr Ionov. Data — 2024–2026. Not investment or medical advice.