2026 will be remembered as the year pharma went all in on AI.
Since January, nearly every major pharmaceutical company has announced some version of the same thing. GSK signed a five-year foundation-model deal with Noetik. Eli Lilly built a co-innovation lab with NVIDIA. Merck committed a billion dollars to Google Cloud. This month, Novo Nordisk named Amazon Web Services its preferred AI partner. The dollar figures are staggering. And the messaging has been almost identical across all of them: get access to the right model, the right compute, the right discovery platform, and you’ll win the next decade of drug development.
Then Q2 earnings calls wrapped, and almost no one could point to a measurable return. A faster trial. A failure caught earlier. A drug that reached folks sooner because of what they’d invested. Executives talked about pilots and roadmaps. The announcements are loud but the silence on results is the real story.
I get why these deals get announced the way they do. The models are really incredible. But when everyone has access to roughly the same model capability, the model stops being an advantage. It becomes table stakes, like electricity. Nobody wins on having electricity. They win on what they build with it.
A lot of the conversation this year has moved toward the discovery platform or the algorithm on top of the model as the real differentiator. That’s closer. But I’d push it one level further. The real edge is the data the platform is trained and run on. A world-class algorithm trained on incomplete, fragmented, clinically thin data will still give you incomplete, fragmented, thin answers. That’s the layer none of these mega-deals have actually touched.
To understand why, think about what it actually looks like when a real person tries to navigate serious illness in this country.
Take someone managing a rare condition who sees six different specialists over four years. Her rheumatologist is in one system. Her neurologist is in another. Two of her earlier records exist at community clinics she stopped going to. By the time anyone tries to assemble a complete picture of what’s happened to her, half the story is missing or inaccessible. And she’s one of the simpler cases.
The US has no unique patient identifier. Providers have little regulatory incentive to share records across systems. A single person’s health history can be scattered across every hospital, lab, specialist, and health system they’ve ever seen. Insurance coverage varies by state and plan. Hospitals and their clinics inside the same health system often run on different EHR platforms that don’t talk to each other. Nobody has a complete picture of the folks these drugs are actually designed to serve.
Claims-based real-world data, the first data type I ever worked with as an intern at Genentech well over a decade ago, are still the primary input for most pharma real-world evidence work and inherit all of that fragmentation. Claims were designed for billing. They can tell you a treatment happened. They can’t tell you what led to it or why. The clinical context lives in physician notes, which are unstructured, inconsistent, and scattered across decades of care. Basic social factors like housing, income, and smoking status are frequently missing. Seventy-eight percent of health organizations say they can’t even locate social-determinant data in a patient’s record after it’s been collected.
Clinical trials are rigorous by design, and that rigor is also what limits them. They’re built around selected populations, defined observation windows, and specific endpoints. A person enrolled in a six-month trial doesn’t stop being a patient after month six. They get new conditions. They switch doctors. They move. With many rare diseases, patients might see seven different physicians before anyone lands on the right diagnosis. None of that shows up in a trial dataset. Most of it doesn’t show up cleanly in claims either.
The folks who are hardest to capture accurately in current datasets are usually the people a drug needs to work for the most. Complex conditions. Multiple diagnoses. Long, difficult care journeys. That’s where the data falls shortest, and that’s where the AI answers will be weakest.
I spent over a decade inside biopharma. I watched incredibly talented teams work hard with an incomplete picture of the people they were trying to help. That gap was always there. What’s new is the scale of investment now being built on top of it.
The next real competitive edge in this industry will come from the less glamorous work of actually fixing the data. Connecting fragmented records. Capturing the clinical detail sitting in unstructured notes. Building datasets that reflect the full reality of the folks who will rely on a new medicine to save their lives, including the ones that don’t fit neatly into a trial population.
That work is harder to announce. It doesn’t generate the same headlines. It’s also what determines whether these AI investments eventually deliver.
The only thread that runs through every part of the care journey is the patient. Every physician visit, every lab, every specialist, every prescription. When you stitch the data around that reality, with the patient’s knowledge and consent, the fragmentation starts to resolve itself. That’s a fundamentally different starting point than what most of the industry is working from today.
There are real people on the other side of this. A father whose daughter just got a life-changing diagnosis that should have come four years earlier, and has to reckon with what those four years might have cost her. A family sitting in a waiting room, wondering whether the new drug their mom just started is actually working. These are the people whose care journeys are scattered across systems. Getting the data piece right is how the industry actually reaches them, in time, with the right answer.
About Shashi Shankar
Shashi Shankar is Co-Founder and CEO of Novellia. He spent over a decade at Genentech and Roche in medical affairs, health economics, and digital health innovation.
