Anthropic just did something software companies don't do: it bought pipettes.
Anthropic has confirmed it operates a wet lab — a physical biology facility — in the San Francisco Bay Area. Eric Kauderer-Abrams, the company's head of life sciences, verified the lab's existence and explained why: "We believe that to do biology, the final test is still and will be for a while in real lab work."
The popular assumption is that AI labs are software companies. They train models, sell tokens, and let customers handle the physical world. Anthropic's move breaks that assumption, and it exposes a structural shift that most coverage of "AI for drug discovery" has missed.
The thesis is simple: when intelligence becomes cheap, the bottleneck moves to verification. The companies that own the means of physical verification — labs, instruments, robots — will set the pace of AI-driven science. Anthropic is not building a nicer chatbot for biologists. It is buying a seat at the bench.
Where the speed actually dies
Start with the numbers. In an August experiment, Anthropic had Claude Opus 4.8 and Mythos Preview design protein binders against 15 targets. Contract manufacturers Adaptyv Bio and Twist synthesized the candidates and ran independent validation. Binders were confirmed for 14 of the 15 targets, with hit rates of 22.6% and 26.7% — well above the 10% to 15% typical of traditional design.
That is a genuine improvement. But read it the other way and the problem appears: roughly three out of four designed candidates still failed. Every failure consumed real synthesis capacity, real reagents, real weeks of lab time.
The cost of that gap shows up in Twist's order book. Its revenue from Anthropic grew from roughly $25 million in 2025 to about $50 million in 2026, and is expected to reach $100 million the following year. Doubling every year is a fine business for Twist. For Anthropic, it means the most valuable part of its science pipeline — the part that turns predictions into knowledge — sits in someone else's facility, on someone else's schedule.
This is the classic full-loop problem. AI accelerated the design half of biology by orders of magnitude. The experimental half still runs at the speed of pipettes and incubators. When one side of a loop accelerates and the other does not, the slow side becomes the whole system's speed. Anthropic's answer is to own the slow side.
The interface problem, and the standard that solves it
Owning a lab is not the same as letting a model run it. Laboratory instruments speak dozens of incompatible software dialects. Engineers routinely spend weeks, sometimes months, integrating a single device into a unified system.
Anthropic attacked this with the Model Hardware Standard (MHS), released in August. It is a protocol that lets AI agents control physical equipment: devices report their state, their executable operations, and their safety boundaries. In early testing, Anthropic says, device integration time collapsed from weeks or months to hours or minutes.
The early deployments are telling. At Genentech, Claude coordinated a liquid handler, a robotic arm, and a microplate reader simultaneously — adjusting parameters from results and diagnosing hardware faults on its own. At the University of Washington's Baker lab, Claude Code read qPCR curves in real time and directed a robotic arm to transfer sample plates.
Notice what these are. Not "AI suggests, human operates." A model closing the loop across multiple physical instruments. That is a different category of system from anything else in the AI industry's product line.
The Accumulation-to-Infrastructure framework
Put the pieces in sequence and a pattern emerges — call it the Accumulation-to-Infrastructure shift. AI companies enter a scientific domain in three phases: tools first (sell software to practitioners), then accumulation (acquire data, customers, and domain expertise), then infrastructure (own the physical means of verification).
Anthropic's last twelve months read like a checklist. In 2025 it launched Claude for Life Sciences. In April 2026 it acquired Coefficient Bio, a drug-development-savvy startup, in a stock deal reported at about $400 million, and Novartis CEO Vas Narasimhan joined Anthropic's board. On June 30 it launched Claude Science, a research workbench bundling more than 60 tools across genomics, structural biology, and cheminformatics. Customers already include Genentech, Bristol Myers Squibb, and Novo Nordisk. Now the wet lab. Job postings seek protein and nucleic acid characterization experts, and one states the ambition outright: "Our goal is to speed up progress in the life sciences by an order of magnitude."
Phase three is where the moat lives. Tools get commoditized. Datasets get replicated. But a robot-run wet lab, integrated with a frontier model through a proprietary hardware protocol, is a vertically closed loop that competitors must rebuild from zero — OpenAI and Google DeepMind do not own pipettes either. Whichever lab closes the design-test-learn loop internally first gets a compounding data advantage: every failed experiment feeds its models, and better models design better experiments.
There is a personal footnote that matters more than it looks. Dario Amodei has written that his father died of a disease whose treatment arrived a few years too late. Kauderer-Abrams framed the same drive in corporate terms: "By far, we see the biggest opportunity for that in the life sciences, and that is motivating everything that we're doing." Anthropic targets rare diseases and "undruggable" targets — bispecific and trispecific antibodies — exactly the areas big pharma deprioritizes on financial grounds.
The boundary, and why it is load-bearing
Anthropic has drawn one line with unusual clarity: it will not run clinical trials and will not commercialize therapies. "We're not competing with pharma and biotech companies that make their business in bringing drugs to market," Kauderer-Abrams said. The concern is real — pharma customers may worry that molecules screened in an Anthropic lab end up training Anthropic models.
The line also tracks the industry's hard wall. More than 100 AI-designed or AI-assisted drugs have entered clinical trials worldwide, yet as of the end of 2025 not a single fully AI-discovered drug had won FDA approval. AI dominates the early funnel; Phase II and III trials run on biology that no model has yet mastered, and a decade of validation stands between a molecule and a patient.
So the honest read of Anthropic's strategy is: own preclinical discovery, rent the rest of the road.
What to do with this
- If you run an AI company: ask where your verification bottleneck is. If your model's output cannot be cheaply tested, your competitors' real moat is not their model — it is their test loop. Owning or partnering deeply with the means of verification is a product decision, not an operations detail.
- If you are in biotech or pharma: the tool vendor you buy from is becoming a discovery competitor. Push hard on data-boundary terms now; in two years your negotiating position will be worse.
- If you build lab automation or robotics: the Anthropic-style hardware protocol is a preview of where instrument interfaces are heading. Interoperability with agent-controlled workflows is about to become a purchasing criterion.
- If you are an investor: the "AI drug discovery" category is bifurcating into software vendors and vertically integrated discovery shops. Price them differently — the second group carries capex and biology risk the first never touches.
The wet lab is small. The direction is not. When the company famous for warning about AI risk starts conducting physical biology experiments under its own roof, the signal is that AI's next frontier is not more intelligence. It is more reality.