Tempus brings AI to oncology: models trained on petabytes of clinical data, with the (still unproven) promise of lowering mortality

By Zendoric · September 8, 2026.

Eric Lefkofsky, founder and CEO of Tempus, explained on Fox Business's "The Claman Countdown" how his company trains what he calls oncology "foundation models": AI systems built from hundreds of petabytes (storage units equivalent to millions of billions of bytes) of real-world clinical data from cancer patients. According to Lefkofsky, these models detect patterns invisible to the clinical eye in order to fine-tune therapies, calibrate dosages and, ultimately, reduce patient mortality.

The interview offers no clinical outcome figures, peer-reviewed studies or hospital adoption data: it is a business television appearance, not a scientific paper. That is worth keeping in mind before taking the promise as a done deal.

As sector context, Tempus is one of the most visible companies at the intersection of genomics, clinical data and artificial intelligence applied to oncology. Founded by Lefkofsky —also known for having co-founded Groupon— the company combines genomic sequencing of tumors, clinical records and medical imaging to help oncologists choose treatments better matched to each patient's molecular profile. Tempus went public on the Nasdaq in 2024, a step that places it under the scrutiny of quarterly results and growth expectations, something that does not always sit well with the long timelines and caution inherent to medical validation.

The underlying idea —foundation models trained on massive volumes of real clinical data— is the same pattern we already see in other AI labs working on biomedicine: instead of an algorithm designed for a specific task, a generalist model is built over the domain (oncology, in this case) and then specialized for concrete uses, from choosing drugs to adjusting doses. It is a sensible strategy in theory, but its real value can only be demonstrated through clinical trials and auditable publications, not through a CEO's word on a television set.

Our reading: this is exactly the kind of AI application that underpins the brightest part of our underlying thesis. If the promise holds —and there are serious signals in the industry that computational genomics and AI models built on massive clinical datasets really are improving diagnostic accuracy— we are talking about a concrete step toward what interests us most about the long term: moving closer to treating cancer not as a generic disease but as thousands of distinct molecular diseases, each with its optimal therapy. That is, in essence, beginning to eradicate one of the most lethal causes of death in human history.

But the short term demands the same rigor we always apply: distinguishing demonstrated capability from corporate narrative. An interview on a business channel is legitimate marketing —Tempus needs to convince investors and hospitals— not clinical evidence. AI-driven precision medicine has spent years promising revolutions that arrive more slowly and more unevenly than announced, constrained by data quality, unequal access to these technologies between large and small hospitals, and the regulatory slowness inherent (and necessary) to anything that touches people's health. The underlying optimism holds, but it is earned trial by trial, not headline by headline.