OpenAI
Rated its newest model “High” for biology skill on its own scale. Offers a biology model only to labs it has checked and approved. In September it shelved a newer model after tests found it more deceptive.
← Market Storm
Market Storm · ARKG · 8 October 2026
In gene stocks, using the most AI isn’t the same as winning from it.
AI is making the thinking part of biology cheap. So who gets paid? Often it isn’t the company using the most AI. It’s the one that owns what AI can’t make.
The short version
AI got good at the thinking.
On short, well-defined tasks, AI can now read research, plan experiments and design proteins on a computer. In some tests it does about as well as experts.
A medicine still has to work in a person.
That part runs at the speed of cells, patients and trials. AI can’t skip it.
So value moves to what AI can’t make.
Patient data nobody else has. Lab work that proves an idea. Medicines already proven in people.
Six numbers to remember
- 21 of 24AI-found drugs (first spotted with AI’s help) passed the first stage of human trials, which checks safety. A typical drug passes about 52% of the time.
- 4 of 10passed the harder “does it work?” stage, against about 29% for a typical drug. Too few to tell skill from luck.
- 9 vs 5Twist’s scores out of 10: 9 for gaining if AI takes off, but only 5 for using AI itself.
- ≤0.2%of Eli Lilly’s sales, at most, goes to its AI lab with NVIDIA. Lilly still ranks first here.
- $93.2MTempus’s sales from its data business, April to June. About 24% of all its sales.
- $24 vs $8what investors pay for $1 of yearly sales at Twist, and at Tempus.
Start here
ARKG is one basket holding many gene companies
Before the AI part, the basics: what a share is, and what this fund owns.
A share (or stock) is a small piece of a company. People buy and sell shares, and the price moves every day.
ARKG is an ETF, short for exchange-traded fund. It holds shares in many companies at once, and people buy and sell it like a single share. ARK Invest runs it.
It holds about 33 companies that work with genes and biology. Some read DNA or make blood tests. Some use AI to design drugs. A few edit genes, rewriting DNA to treat disease.
If ARKG were 100 squares
Each square is 1% of the fund. Tests, lab tools and data fill more than half. The AI drug designers fill about 11 squares.
- Tests and patient dataTempus, Personalis, Guardant, Natera, CareDx27%
- Lab tools10x Genomics, Twist, Illumina25%
- AI drug discoveryAbsci, Schrödinger, Recursion11%
- Gene editingCRISPR Therapeutics, Beam8%
- MedicinesEli Lilly, Compass6%
- The other ~18 companieseach under 2.4% of the fund23%
The 15 biggest holdings, exactly
- 10x GenomicsLab tools10.38%
- TwistLab tools10.12%
- TempusTests and data8.40%
- CRISPR TherapeuticsGene editing5.28%
- AbsciAI drug discovery5.23%
- PersonalisTests and data4.91%
- GuardantTests and data4.62%
- NateraTests and data4.52%
- IlluminaLab tools4.51%
- CareDxTests and data4.51%
- Eli LillyMedicines3.33%
- SchrödingerAI drug discovery3.04%
- BeamGene editing2.95%
- RecursionAI drug discovery2.58%
- CompassMedicines2.44%
Two lab-tool makers sit at the top. On 5 October, 10x Genomics (machines that study single cells) and Twist Bioscience (DNA printed to order) made up about a fifth of the fund between them.
The question this article asks: which of these companies are best positioned if AI takes off? Here, “takes off” means AI that can run month-long research projects by 2028, and whole research programs by 2030. Chapter 8 puts rough odds on that.
The evidence
AI got good at the thinking part of biology
The big AI labs are pushing into biology, and the price of a good idea is falling fast.
AI agents on a test of real science work
The newest AI agents solve more than half the tasks on an early test of real science work.
Anthropic
Says its AI, Claude, now takes the lead on 26% of Anthropic’s own research tasks, up from under 1% in February. In one project, 950 Claude agents found a new enzyme system in 21 hours.
xAI
The maker of the Grok chatbot. This report found no biology product from it.
Google DeepMind
Released AlphaGenome Atlas, which predicts what 9 billion possible DNA changes do. It’s free for research.
The enzyme system Claude found looks like the ones used for gene editing. What it does isn’t known yet, and outside scientists haven’t checked the work. Finding it took a day. Proving it is the slow part.
And the thing none of these labs has done yet: get a drug approved for sale. As of mid-2026, no AI-found drug had full approval from the FDA, the US agency that decides which medicines can be sold.
The evidence
But proof in people still takes years
A new drug is tested in people in stages. Phase 1 asks: is it safe? Phase 2 asks: does it work at all? Phase 3 tests it in a big group.
How sure can we be about AI-found drugs?
Each dot is one AI-found drug. The shaded band is where the true pass rate probably sits. If the typical rate falls outside the band, the difference is real. If it falls inside, we can’t tell yet.
Phase 1Is it safe?
21 of 24 AI-found drugs passed
The typical rate sits well outside the band. AI-found drugs really do pass the safety stage more often.
Phase 2Does it work?
4 of 10 AI-found drugs passed
The typical rate sits inside the band. With only ten results, we can’t yet tell skill from luck.
The first big tests are on the calendar
- Now
Generate Biomedicines, an ARKG company, is testing an AI-improved asthma antibody in about 1,600 patients, in Phase 3. An antibody is a protein that grabs one target in the body. Other drugs already work on this target, so a win would show AI improving a known kind of drug.
- Early 2027
The FDA decides on zasocitinib, a drug from Takeda, a big Japanese drug company. It was designed with computer simulations of molecules and with machine learning, software that learns patterns from data.
- 2029–30
The earliest likely approval for Insilico’s rentosertib, whose target and molecule were both found with AI. Its Phase 3 trial started in China in September 2026.
Some tests have already proven themselves. One cancer drug’s official FDA instructions now name Natera’s blood test. The test finds the patients who need the drug.
AI speeds up the thinking. The proving still runs at the speed of cells and patients.
The central finding
Using the most AI is not the same as winning from it
Each company got two scores out of 10: how much it uses AI today, and how much it gains if AI takes off. If those were the same thing, every company would sit in the top-right or bottom-left box. Many don’t.
The AI map
Filled chips are the four companies below. Faded chips were scored but not ranked.
Less AI, big gain
- Twist (uses AI 5/10, gains 9/10)
- 10x (uses AI 5/10, gains 7/10)
More AI, big gain
- Tempus (uses AI 8/10, gains 8/10)
- Lilly (uses AI 7/10, gains 7/10)
- Natera (uses AI 6/10, gains 7/10)
- Guardant (uses AI 6/10, gains 7/10)
- Butterfly (uses AI 6/10, gains 7/10)
Less AI, smaller gain
- GeneDx (uses AI 5/10, gains 6/10)
- CareDx (uses AI 4/10, gains 5/10)
- PacBio (uses AI 4/10, gains 5/10)
- Nurix (uses AI 4/10, gains 5/10)
- Adaptive (uses AI 4/10, gains 5/10)
- Intellia (uses AI 3/10, gains 6/10)
- Alamar (uses AI 3/10, gains 6/10)
- CRISPR (uses AI 2/10, gains 6/10)
- Ionis (uses AI 2/10, gains 5/10)
- Beam (uses AI 2/10, gains 5/10)
- Personalis (uses AI 4/10, gains 2/10)
- Compass (uses AI 1/10, gains 3/10)
More AI, smaller gain
- Recursion (uses AI 8/10, gains 6/10)
- Generate (uses AI 8/10, gains 6/10)
- Freenome (uses AI 7/10, gains 6/10)
- Schrödinger (uses AI 7/10, gains 5/10)
- Veracyte (uses AI 6/10, gains 6/10)
- Illumina (uses AI 6/10, gains 6/10)
- Absci (uses AI 6/10, gains 4/10)
The shovel seller
Twist
Uses AI today5/10
Gains if AI takes off9/10
No AI leader itself. But every protein an AI designs needs real DNA made to order before it can be tested, and Twist is one of the biggest makers of that DNA. Companies that design with AI are its fastest-growing customers.
On its August call, its managers confirmed an analyst’s math: AI orders of at least about $50 million in the year to September 2026, up from about $25 million, and about $100 million the year after.
All in on AI, still waiting
Recursion
Uses AI today8/10
Gains if AI takes off6/10
Uses AI as much as anyone. It builds its own AI models and even lets Tempus use one.
But it still needs a drug to clearly work in people. Its results in people so far come from small trials. The next updates are due on 2 November and in the first half of 2027.
The steady giant
Eli Lilly
Uses AI today7/10
Gains if AI takes off7/10
A giant drug company with the biggest AI setup here: an AI lab with NVIDIA, the chip maker. The two plan to put up to $1 billion into it over five years.
That’s at most $200 million a year, roughly 0.2% of Lilly’s sales, before NVIDIA’s share. Lilly ranks first because of its proven medicines and its cash, not the lab.
Both at once
Tempus
Uses AI today8/10
Gains if AI takes off8/10
A cancer-test and patient-data company that both uses AI and gains from it. It builds AI tools from its own patient records and sells them to drug companies such as AstraZeneca.
AI labs can’t copy records like those from the internet.
The gene editors use almost no AI. Their value rests on medicines that work in people. AI can’t copy those, but it doesn’t add much to them either, so they score lower on gains.
If anyone can rent the same AI, the edge belongs to what can’t be rented.
Data. Lab work. Proven drugs.
The evidence
Five other forces, all pushing the same way
Better AI isn’t the only thing acting on these companies. Five other forces are at work too, and they reward the same things.
Borrowing got expensive
The 10-year Treasury yield, what the US government pays to borrow for ten years, sets the tone for many other loans. It passed 5% in mid-September and was about 5.3% on 2 October. The Fed raised its rate on 16 September.
Companies that spend more than they bring in while they wait for proof feel this first.
Private data beats the model
Drug companies can take free AI models, train them on their own data and keep the result. The data is the part nobody else has.
That means competition for Tempus: its closest rival, Caris, grew its sales 45% last quarter, about twice Tempus’s pace. Caris isn’t in the fund.
AI labs go straight to drug companies
Novo Nordisk, one of the world’s biggest drug companies, works with Anthropic’s Claude directly. Lilly, which is in the fund, works with OpenAI on new antibiotics.
Beyond Lilly, this report found direct AI-lab links with only three fund companies: 10x, Twist and PacBio. None made any payments public.
China makes new drugs fast and cheap
Big drug companies often buy the rights to drugs that outside developers found. In 2025, 40% of those came from China, up from under 30% in 2024.
China 40%Everywhere else 60%
Like AI, this makes new drug ideas cheaper. Gene editors, and tests paid for in the US, face less of it than ordinary pills.
Big drug companies face a patent cliff
A patent is the legal right to be the only seller of a drug. About $300 billion of drug sales lose that protection by 2030.
So big drug companies buy smaller ones to refill their shelves. What they want most is a medicine already proven in people.
Each of these rewards the same things: data nobody else has, proof in people, and the cash to wait for it.
The evidence
What today’s prices already expect
A company can be well positioned and still have a high price. Some of these shares have already risen a lot.
+96%ARKG for the year, by 5 Oct
−8.8%in one day, 6 Oct. Twist fell about 19%, with no company news
+79%left for the year
By early October, Twist and 10x shares were each worth nearly six times their price at the start of the year.
One way to see what a price expects is price-to-sales: take a company’s total market value, what all its shares are worth together, and divide it by a year of sales.
What investors pay for $1 of yearly sales
Each coin is a dollar. A dollar of Twist’s sales costs about three times what a dollar of Tempus’s does, so far more growth is already counted in Twist’s price.
- Twist~$10.5–11B ÷ $456M, year ended 30 Sep$24
- Natera~$58B ÷ $2.88B$20
- 10x Genomics~$10.0B ÷ $620M$16
- Eli Lilly~$1.03T ÷ $86B$12
- Illumina~$41B ÷ $4.62B$9
- Tempus~$13.0B ÷ $1.6B$8
Analysts at banks also publish targets: their guess of where a share price will be in about a year. In early October, Twist’s and Natera’s prices were already above their analysts’ average guess. Those guesses often lag big moves. They show mood, not truth.
The verdict
Who’s best positioned, and why
Each company got five scores out of 10, combined into one. The tiers sort companies by why they score as they do, not by score alone.
How the combined score is built
- Gains if AI takes off 35%
- Its technology works 20%
- Uses AI today 15%
- How long the business lasts, cash included 15%
- How much growth the price already expects 15%
Best positioned
Lilly is the steady giant: proven medicines, a big AI lab, and the cash to pay for what AI finds. Tempus owns patient data AI can’t copy, and is on probation until more cash comes in than goes out.
- Eli Lilly7.4
- Tempus7.3
Well placed, but priced for big growth or still proving itself
Good businesses where the price already expects a lot, or where the AI payoff is still being shown.
- Twist6.9
- Veracyte6.9
- Natera6.7
- Illumina6.6
- Guardant6.3
- 10x Genomics6.0
AI platforms still waiting for their big proof
Generate is furthest along, with its Phase 3. Nurix uses less AI than the rest, but its value also rests on proof still to come.
- Schrödinger6.2
- Generate6.0
- Freenome5.8
- Recursion5.5
- Nurix4.9
- Absci4.3
AI isn’t the main story
Some score well. Their value just doesn’t depend on AI.
- CRISPR Therapeutics6.3
- Intellia5.8
- CareDx5.7
- Adaptive5.5
- Ionis5.4
- Beam5.1
- Personalis4.6
- Compass4.3
That’s why CRISPR Therapeutics, in Tier 4, outscores 10x Genomics, in Tier 2: CRISPR scores well, but not because of AI.
The scores are judgment calls. To test them, the weights were shuffled at random 10,000 times. Lilly came first in 59% of tries, and Tempus in 24%. Ranks five to fourteen sit within about a point of each other, so read them as a group, not an order.
The verdict
Three ways AI could go, and why even the fastest takes years
The ranking asks what happens if AI takes off. How likely is that? These are rough odds for 2027–2030: a judgment, not a measurement.
15%50%35%
15% · AI stalls
AI stays good only at short tasks. Or the money for building AI dries up, or a biosecurity scare (fear of AI helping someone make a dangerous germ) locks things down. Companies that already make money would hold up best.
50% · AI steady
AI would handle week-long tasks, with people checking, by 2028, and month-long tasks around 2029–30. Labs and clinics would stay the slow step.
35% · AI takes off
AI agents would run month-long projects by 2028 and whole research programs by 2030. The slow step would become how many robot labs exist to test their ideas.
The first version of this research said 20% stall and 30% takeoff. The second review moved five points to takeoff, because Claude now leads 26% of Anthropic’s own research tasks and OpenAI says it met its goal of an AI “research intern.” One thing held it back: the OpenAI model shelved after safety tests.
Even fast AI would pay slowly here
- 1
Share prices move first.
- 2
Then orders for lab tools and data.
- 3
Drug sales last, and not before about 2030.
The verdict
What would change the picture
One number would settle the big question: the next large count of AI-found drugs in Phase 2.
Today there are ten results. Suppose 30 or more AI-found drugs finish Phase 2, and at least half succeed. That would show AI picking drugs that work in people, not just designing them faster, and the companies that design drugs with AI (Generate, Recursion, Absci and Schrödinger) would move up.
Until then, dates to watch
- 2 Nov 2026
Recursion’s next results in people. A drug clearly working would be a strong sign for the companies that design drugs with AI.
- Early Nov 2026
Tempus’s cash flow for July to September. Weak cash flow would move it to Tier 2.
- Nov 2026
Twist’s forecast for next year: how big a part of its sales AI orders become.
- 31 Dec 2026
Tempus’s main AstraZeneca agreement runs to this date, according to its latest filing. No renewal, or more cash going out than coming in at year-end, would move Tempus to Tier 2.
- Any time
Anthropic’s next update. Claude leading more than half of its own research tasks would raise the odds of a takeoff. More AI models shelved for safety reasons would point to the “steady” future.
- Any time
Another Fed rate rise while the 10-year yield stays above 5% would squeeze every company spending more than it brings in, whatever the AI news.
Questions nobody can answer yet
- Does AI help pick drugs that work in people, or only design them faster?
- Who keeps the money when AI labs go straight to drug companies?
- How fast is AI really improving? The newest numbers come from the labs themselves, grading their own work. Outside checks lag behind.
What this means for you
Two habits worth keeping
You don’t need to own a single share for this to be useful. Both habits work far beyond stocks.
When a new tool makes one step cheap, ask what becomes scarce.
AI is making the thinking part of biology cheap. So the waiting moves to the parts AI can’t do alone: data nobody else has, lab work that proves the idea, patients in trials.
Using a tool a lot is not the same as being paid for it. In a gold rush, the shop selling shovels gets paid whether or not anyone finds gold.
A good story and a high price are two separate questions.
A company can be exactly where the future is heading and still cost more than that future is worth. Check the story and the price one at a time.
Receipts
How this article was made
A frontier AI model wrote this with a team of AI agents, each with its own job, from tracking what the fund owns to arguing against the conclusions.
Then it sent a second team of agents back over the work to find factual errors and try to prove the conclusions wrong. Their corrections were applied, and a final review checked the facts, the plain language and how the page reads on a phone.
21 AI agents worked on it in all. Most finance websites blocked the agents from opening pages directly, so some figures were confirmed through search results for the pages listed below.
What the fact-check found
- 10 held up
- 6 partly true or out of date
- 4 wrong, and corrected
The corrections that changed the most
Corrected
Natera was in the top tier for the wrong reasons
The first version said Natera brought in more cash than it spent and had little legal risk. Its own report to regulators shows a $67 million loss from April to June, under standard accounting rules.
Natera is appealing a court ruling that it owes Guardant, another fund company, about $290 million over false advertising. Its market value had also climbed to about 20 times its yearly sales, and its share price sat above analysts’ average guess. It moved to Tier 2.
Corrected
Tempus’s “$200 million in a quarter” was deals signed, not money in
The $200 million was bookings: deals signed that pay out over several years. Sales from Tempus’s data business that quarter were $93.2 million.
The cash it burned in the first half of the year also grew, to $80.8 million from $61.5 million. Its main AstraZeneca agreement runs to 31 December, according to its latest filing, though its managers say the AstraZeneca work runs into 2027. It stays in Tier 1, on probation.
Corrected
A typical drug passes Phase 2 about 29% of the time, not 40%
The first version compared AI drugs (4 of 10) with a typical rate of about 40%, and called them the same. The industry figure for 2011–2020 is about 29%. So AI drugs look a little better, but ten results are still too few to tell.
The “What would settle it” test was rewritten to match.
Partly true
Lilly is a weight-loss company with a big AI lab, not an AI company
Lilly’s AI work is real. It runs an AI lab with NVIDIA, budgeted at up to $1 billion over five years.
But the prices it actually gets for its drugs fell about 13%. Early on, its new weight-loss pill, Foundayo, sold about one-fifth as much as its main rival’s pill. This report found no AI-found Lilly drug in human trials.
It stays first, relabeled as the steady giant.
Corrected
Rentosertib’s Phase 3 started in September, not July
Insilico gave the first Phase 3 patient a dose of rentosertib on 9–10 September 2026, in China. An earlier claim said July. Insilico says AI found both the drug and the target it aims at.
Partly true
This year’s biggest gains were measured before a hard fall
The first version said Twist was up 495% and 10x up 473% for the year, each nearly six-fold. Those figures were from early October, before ARKG fell 8.8% on 6 October, from $56.77 to $51.75.
Twist fell about 19% that day. After the drop, Twist was still up more than 400% for the year, and ARKG about 79%.
Sources 49
Company filings with regulators 9
- Natera — Form 10-Q, quarter ended 30 Jun 2026
- Eli Lilly — second-quarter 2026 results, 8-K exhibit (5 Aug 2026)
- Tempus AI — Form 10-Q, quarter ended 30 Jun 2026
- Illumina — second-quarter 2026 results, 8-K exhibit 99.1 (30 Jul 2026)
- Veracyte — second-quarter 2026 results, 8-K exhibit 99.1 (30 Jul 2026)
- Recursion — second-quarter 2026 results, 8-K exhibit 99.1 (5 Aug 2026)
- Twist Bioscience — fiscal third-quarter 2026 results, 8-K exhibit 99.1 (3 Aug 2026)
- 10x Genomics — second-quarter 2026 results, 8-K exhibit 99.1 (6 Aug 2026)
- Guardant Health — Form 10-Q, quarter ended 30 Jun 2026
Company and AI-lab pages 22
- Tempus — second-quarter 2026 results (30 Jul 2026)
- Tempus and Recursion — extended data license and new license to Recursion’s RNA foundation model (21 Sep 2026)
- Natera — second-quarter 2026 results (6 Aug 2026)
- NVIDIA and Lilly — co-innovation AI lab for drug discovery (12 Jan 2026)
- Generate Biomedicines — second-quarter 2026 earnings release (6 Aug 2026)
- Freenome — FDA approves SimpleScreen CRC; Abbott to commercialize in the US (27 Jul 2026)
- Takeda — FDA accepts zasocitinib NDA under priority review (14 Sep 2026)
- Insilico Medicine — first patient dosed in GENESIS-IPF-3, rentosertib’s Phase 3 (10 Sep 2026)
- Caris Life Sciences — second-quarter 2026 results (5 Aug 2026)
- Basecamp Research — Trillion Gene Atlas with Anthropic, Ultima Genomics, PacBio and NVIDIA (18 Mar 2026)
- Anthropic — launch post for its newest Claude model, with the science-test scores (22 Sep 2026)
- Anthropic — Measurements for understanding the pace of AI development inside frontier labs (Sep 2026)
- Anthropic — Claude for Life Sciences (20 Oct 2025)
- Anthropic — Claude accelerates protein design and analytical chemistry (18 Aug 2026)
- Anthropic — How Claude is uplifting biomolecular modeling (17 Sep 2026)
- OpenAI — safety card for its newest model, rated “High” for biology (Sep 2026)
- OpenAI — Research acceleration: the view inside OpenAI (Sep 2026)
- Anthropic — Claude discovers a novel enzyme system with CRISPR-like repeats (23 Sep 2026)
- Anthropic — Novo Nordisk customer story (Claude by Anthropic)
- OpenAI — its biology model for approved labs only (16 Apr 2026)
- Google DeepMind — AlphaGenome Atlas: a map of every possible DNA letter change in the human genome (8 Sep 2026)
- Eli Lilly — Lilly collaborates with OpenAI to discover novel medicines to treat drug-resistant bacteria (25 Jun 2024)
Market and research data 9
- StockAnalysis — ARKG holdings (snapshot of 5 Oct 2026)
- Investing.com — ARKG price and historical returns (6 Oct 2026 close)
- BIO, Informa and QLS — Clinical development success rates 2011–2020 (Feb 2021)
- Federal Reserve — FOMC statement (16 Sep 2026)
- Federal Reserve Bank of St. Louis (FRED) — 10-year Treasury yield, daily (DGS10)
- MarketScreener — Twist Bioscience analyst consensus (Oct 2026)
- MarketBeat — Natera stock up 4% to ~$406 vs a ~$331 consensus target (24 Sep 2026)
- Quiver Quantitative — ARKG historical prices (2025 year-end close $28.97)
- Macrotrends — Twist Bioscience stock price history (6 Oct 2026 close $166.97)
Analysis and news 9
- Jayatunga et al. (BCG), Drug Discovery Today — How successful are AI-discovered drugs in clinical trials? (2024)
- Stocktwits via Yahoo Finance — Cathie Wood’s ARKG is crushing her other ETFs (early Oct 2026)
- Investing.com — Why is Twist Bioscience stock plunging today? (6 Oct 2026)
- CNBC — OpenAI abandons plan to release upcoming model as safety concerns escalate (28 Sep 2026)
- Fierce Pharma — Lilly’s obesity pill Foundayo ‘disappointed’ in Q2 as injectables boom (Aug 2026)
- The Motley Fool — Twist Bioscience fiscal Q3 2026 earnings call transcript (Aug 2026)
- PwC — US pharma and life sciences deals 2026 midyear outlook (2026)
- Pharmaceutical Executive — Deepening ties: why China is becoming big pharma’s most essential R&D partner (Jun 2026)
- OncLive — FDA approves adjuvant atezolizumab for MRD-positive bladder cancer, with Signatera as the companion test (15 May 2026)
Market Storm is research, not investment advice. It is the output of an AI research method applied to public information, and it may contain errors. Nothing here is a recommendation to buy or sell any security. The author may hold positions in companies covered. Do your own research.
Get the next Market Storm in your inbox
One email when a real market moment triggers a new report.