Why Anthropic Built a Wet Lab
When news broke that Anthropic quietly runs a wet laboratory in the San Francisco Bay Area, most coverage framed it as a business-expansion story: another AI company moving into biotech. That framing misses the real signal.
A company whose core asset is software — models, weights, GPUs — just spent hundreds of millions of dollars acquiring pipettes, robotic arms, and incubators. Anthropic's head of life sciences, Eric Kauderer-Abrams, confirmed to Reuters that the company is doing physical biology work in its own facility: "To do biology, the final test is still and will be for a while in real lab work. We absolutely are doing that today."
This is not diversification. It is the most prominent example yet of a structural constraint that will shape every "AI plus physical world" industry: the bottleneck is no longer generating ideas, but verifying them.
The Scissors Gap
AI has supercharged the design phase of science. In August, Anthropic published results from a protein design experiment: Claude Opus 4.8 and Mythos Preview designed protein binders against 15 targets, which contract research organizations Adaptyv Bio and Twist Bioscience independently synthesized and tested. Of 1,320 designs, 354 bound their targets — a hit rate of about 27%, and 14 of 15 targets yielded a confirmed binder. That is roughly double the industry's typical 10–15% hit rate.
But flip the same numbers around: even with a doubled hit rate, three out of every four candidate molecules still end up in the trash. Every discarded design consumed real synthesis capacity, real reagents, real technician hours. The faster the design engine runs, the more wet-lab throughput it burns.
The cost curve tells the story. According to reporting, Anthropic's orders to Twist Bioscience grew from roughly $25 million in 2025 to about $50 million in 2026, and are projected to reach $100 million next year. Outsourced verification was becoming both the biggest expense and the slowest link — external partners mean data feedback loops measured in weeks, and iteration speed determines how fast the models improve.
Building its own lab converts verification from a purchased service into an owned asset. Kauderer-Abrams put it plainly: "There's some things that we can do much faster in our own hands," with the goal to "operate at the largest possible scale."
Verification Economics: The Second Cost Curve
Here is the framework this move reveals: every "AI plus physical" industry runs on two cost curves. The generation cost — what it costs a model to produce a candidate. And the verification cost — what it costs to confirm that candidate in physical reality. AI collapses the first curve; it barely touches the second.
The ratio between the two curves determines the industry's structure. When generation gets cheap and verification stays expensive, verification capacity becomes the scarce asset — and whoever owns it captures the margin.
Anthropic's sequence over the past 18 months reads like a textbook execution of this logic: Claude for Life Sciences launched in 2025; a reported ~$400 million stock acquisition of drug-development startup Coefficient Bio in April 2026, with Novartis CEO Vas Narasimhan joining Anthropic's board; the Claude Science workbench in June; the Model Hardware Standard (MHS) in August, a protocol that lets AI agents control lab equipment — cutting device integration time from weeks or months to hours in early tests; and now the wet lab itself. At a June event, Kauderer-Abrams said Anthropic would run preclinical programs in areas traditional pharma finds financially unattractive: rare diseases, "undruggable" targets, bispecific and trispecific antibodies.
Three transferable corollaries follow from verification economics:
First, verification assets get revalued. Contract research organizations, lab-automation vendors, testing facilities — the owners of "slow links" sit at the top of the AI-era value chain, not the bottom.
Second, the competitive metric shifts from model performance to loop cycle time. A five-point hit-rate improvement matters less than halving the design-build-test-learn cycle. Data feedback speed, not benchmark scores, becomes the KPI.
Third, hardware interfaces are the new battleground. Lab instruments speak incompatible protocols; engineers used to spend weeks integrating one device. Whoever sets the standard — as MHS attempts to — controls AI's entry point into the physical world.
Where Else the Gap Bites
Apply verification economics to other AI frontiers and the map sharpens.
Robotics. Simulators can generate millions of training episodes overnight, but the sim-to-real gap means physics remains the only honest verifier. That is why Tesla and Figure race to build real-world data collection fleets rather than just GPU clusters — the robot lab is the wet lab of embodied AI.
Autonomous driving. Simulated miles are effectively infinite; real test fleets are not. Fleet size, not parameter count, sets the pace of commercialization.
Materials and energy. AI can predict millions of candidate structures, but each must be synthesized and measured. The companies pulling ahead are the ones pairing prediction models with automated synthesis facilities.
The counterexample proves the rule: MIT researchers found in 2025 that roughly 95% of enterprise generative AI investments had produced no measurable financial return. A common failure pattern: budgets concentrated entirely on the generation side, with nothing built for verification.
What To Do About It
If you are building an AI-science or "AI plus hardware" startup:
- Put verification loop cycle time at the center of your metrics — alongside, or above, model benchmarks.
- Secure verification assets early: partner, acquire, or build, based on iteration frequency and data sensitivity.
- Design the data feedback pipeline as first-class architecture, not an afterthought.
If you lead an established company in pharma, manufacturing, or materials:
- Revalue your verification infrastructure — it may be worth more than your balance sheet says.
- In partnerships with AI firms, data feedback terms are the core of the negotiation, not license fees.
- Be skeptical of vendors selling only generation: without a verification loop, their solution is probably one of the 95%.
If you invest in this space:
- Within any vertical, favor the verification-asset holders and the loop-cycle leaders.
- Track hardware-interface standards like MHS — the winner there may not be a model company at all.
AI has proven it can think fast. The next decade belongs to whoever makes the physical world keep up.