Roche is already building autonomous AI labs to discover drugs

On September 28, Roche confirmed something that just a few years ago would have sounded like science fiction: it has already begun building autonomous laboratories powered by artificial intelligence.

The news came out during its Pharma Investor Day. Aviv Regev, executive vice president of Genentech Research and Early Development, explained that Roche aims to move toward what it calls "AI independence" within research and development. KSL

And here "autonomous laboratory" means something far more interesting than installing a language model on a computer.

What is an autonomous laboratory, really?

The basic idea is to connect several technologies we normally see separately:

• Artificial intelligence to analyze information and propose decisions.
• Automation to control experimental processes.
• Robotics to physically interact with samples and equipment.
• Sensors and instruments that produce data.
• Systems capable of using those results to determine what to do next.

The result can become a cycle: propose, experiment, measure, learn, and experiment again.

King's College London, which curiously announced its own Autonomous Labs initiative today, describes precisely these systems as platforms where AI, robotics, sensors, and automation can run experiments, analyze results, and use the evidence obtained to decide what to test next. King's College London

That last step is the one I find truly important.

Automating a laboratory instrument has been done for decades. Giving the system the ability to use the result of an experiment and decide what the next one should be changes things quite a bit.

Roche has been putting the pieces in place for months

This announcement didn't come out of nowhere either.

In March, Roche introduced a massive artificial intelligence infrastructure built around NVIDIA technology.

The company added 2,176 NVIDIA Blackwell GPUs, bringing its infrastructure to more than 3,500 GPUs. Roche described it at the time as the largest hybrid AI Factory announced within the pharmaceutical industry. Roche

Roche: NVIDIA AI Factory

The infrastructure is distributed between the United States and Europe and was created specifically to accelerate the development of treatments and diagnostic solutions.

AI can be very good at generating hypotheses. To turn that into science, it needs data, computing power, and, eventually, to experiment on the physical world.

Roche is trying to connect those pieces.

Then came the robots

In July, another piece appeared.

ABB Robotics announced a global collaboration with Roche Diagnostics to develop robotic solutions for clinical laboratories.

Among the first applications are pathology slide handling and autonomous systems for moving materials within laboratories. The collaboration includes articulated robots and autonomous mobile handling systems. Cision News

It's important to distinguish two related but different initiatives here.

The collaboration with ABB is focused on clinical laboratories and diagnostics. Roche's current announcement refers to autonomous laboratories for research and drug discovery.

The technology converges, but the goals are not exactly the same.

AI is already involved in real decisions at Roche

Perhaps the number that best explains why Roche is taking this step is 40%.

According to data presented by the company, 40% of its pipeline decisions made between the fourth quarter of 2025 and the second quarter of 2026 had a traceable contribution from AI or computational tools. KSL

We're no longer talking about isolated experiments with ChatGPT.

Roche also expects Target Nexus, one of its AI tools, to contribute to 80% of its research portfolio decisions by the end of 2026. KSL

That shows an interesting transition.

First we used AI to analyze data.

Then it started making recommendations.

Now it's beginning to integrate into the process that generates new data.

Why does automating the experiment matter so much?

Let's think about a researcher who wants to determine which combination works best among hundreds or thousands of possibilities.

Each cycle involves preparing materials, setting up instruments, running tests, collecting results, analyzing them, and deciding what the next test will be.

AI can dramatically reduce the space of possibilities.

But if afterwards everything has to stop so a person can translate each of the model's decisions into a new series of experiments, we're still stuck with a bottleneck.

An autonomous lab tries to close that gap.

The cycle changes

In simple terms:

AI → hypothesis → robot → experiment → data → AI → next experiment.

That makes it possible to imagine systems running for much longer periods, exploring alternatives and adjusting experiments as results come in.

It doesn't mean AI can decide on its own which medication to give a person. Nor does it mean Roche has removed scientists from its labs.

We're talking about automating and speeding up parts of the research process.

And that difference matters enormously when we're talking about health.

Roche still keeps humans in the process

There's another part worth mentioning, because "autonomous lab" can lead to some pretty exaggerated headlines.

Roche itself has previously noted that implementing AI and machine learning in labs must maintain human oversight to preserve safety, transparency, and reliability of results. Diagnostics

Roche Diagnostics: AI and machine learning in the lab

Autonomy, at least based on what we currently know about the project, is related to the ability to automate work and experimentation cycles.

Not to removing scientists from the equation.

In fact, Roche continues to invest a considerable amount of money precisely in scientists.

A few days ago it inaugurated its new Innovation Center in Boston, which brings together researchers from Roche and Genentech along with data science and artificial intelligence specialists to work on drug discovery and development. Roche

There's also an economic reason

Pharmaceutical research is expensive.

Roche reported that around 2 billion Swiss francs in R&D savings will be reallocated toward productivity programs and initiatives. At the same time, the company aims to bring up to 20 new molecular entities to market by 2030. KSL

In addition, Roche reported that its phase III success rate during 2026 exceeds 80%, compared to 65% in 2025. KSL

We can't attribute that increase directly to AI. Correlation and causality are different things, and Roche didn't present those numbers as proof that AI is responsible for the increase either.

But they do help us understand the context.

The company wants to produce more useful research with more efficient processes, and it's making a fairly strong bet on AI, computing, and automation to achieve that.

This is Physical AI applied to science

There's a reason this topic seems far more interesting to me than the weekly launch of yet another chatbot.

Over the past few years, we've watched models learn to generate text, images, video, and code.

Agents added another layer: using tools and executing digital actions.

An autonomous lab takes another step because it connects that computational reasoning with physical actions and real experimental results.

The model can receive information from the world, make a decision, trigger an action through a machine, and then analyze what happened.

That comes quite close to what the industry is starting to call Physical AI.

And medicine is probably one of the places where that combination could generate the greatest impact.

My takeaways

I don't think the most important part of this news is that Roche has robots in its labs.

That already exists.

The important thing is trying to close the loop between reasoning, physical action, and experimental learning.

Roche already has an infrastructure of more than 3,500 GPUs, AI tools participating in research decisions, collaboration with robotics specialists, and centers where scientists, data, and AI work together. Now it's starting to connect those pieces through autonomous labs. Roche

If it works at scale, we could end up seeing AI agents whose job isn't booking a flight or modifying a file.

They could spend days testing hypotheses to find a molecule worth studying.

Conclusion

For a long time, we talked about AI as a technology locked inside our computers.

Autonomous labs show where the next stage might be headed.

AI analyzes. Robotics executes. Instruments produce evidence. The system learns from those results, and scientists can guide and supervise the process.

It still remains to be proven how much it can truly accelerate drug discovery, which parts of the process will achieve meaningful autonomy, and what controls will be necessary.

But Roche has already moved past the PowerPoint presentation.