Ai2’s collaboration with Providence Swedish Cancer Institute includes a plan as intriguing as the research itself: deploying AutoDiscovery in Providence’s own cloud, operated by the institution’s research team. Protected research and clinical data would be handled within that managed environment. Ai2’s announcement ↗
This might look like a research institution adopting another useful AI tool. But when we consider how scientific AI becomes more valuable, the arrangement takes on a broader significance.
The study began with public breast cancer data. Researchers investigated an immune-related signal in lobular cancer using another patient dataset and tumor specimens. This was not a clinical trial demonstrating immunotherapy’s efficacy, but an example of developing a computational lead into a question that further research could address. The collaboration’s findings ↗
Exclusive data were not a prerequisite for getting started. Public data provided the starting point. Going further required people who could judge what was biologically meaningful, along with the materials needed to test it.
Ai2’s retrospective describes statistically surprising candidates that offered little reason for further investigation in their biological or clinical context. Researchers’ knowledge helped redirect the search. Ai2’s research retrospective ↗
This involvement could change the role of research institutions using scientific AI.
Researchers do more than assess a system’s answers. They decide which questions to pursue, provide materials for validation and adjust the search in light of the results. Institutions carrying out this work could become development partners, helping establish concrete uses for scientific AI. Model providers cannot always take a research project through to completion on their own.
Public computing initiatives such as Empire AI offer another perspective on this relationship. The Simons Foundation’s call for basic-science proposals requires lead investigators to be affiliated with consortium member institutions, while allowing outside researchers to join under specified conditions. Providing computing capacity and organizing the teams that use it are connected tasks. The official funding call ↗
No single company or laboratory needs to possess all the computing resources, specialist knowledge, samples and validation methods a project requires. Collaborations between organizations with different contributions could help scientific AI reach more researchers.
Seen this way, the next business opportunities look different. Alongside selling access to models, providers could support deployment at research institutions, establish suitable environments for data use and help arrange collaborations with partners able to test the findings. There may be a market for the work needed to make research possible in practice.
This is not a prediction that large hospitals and research institutions will necessarily have an advantage. A small group’s understanding of a particular phenomenon, or its distinctive measurement methods, could make it indispensable to a collaboration. What matters is not size alone, but what a group can contribute to testing an AI-generated lead.
In future partnership announcements, the substance will be more revealing than the names of the systems being deployed. Who chooses the research questions? Who tests the findings? How do the results inform the next investigation? If deployment within an institution merely adds operational burden, this view will need to be reconsidered.
The next business opportunity in scientific AI may extend from finding promising leads to securing the people and materials needed to test them—and supporting the research that follows.