Tempus AI: When the Data Layer Sits Inside the Clinic

Structural read, not clinical or investment advice. The author is a product strategist applying the 10-layer framework, not a domain expert in oncology, pathology, or healthcare regulation.

The setup. Tempus AI went public in June 2024 as one of the first AI-native medical companies at scale. The story is usually told as "AI in oncology." Through the layers it's something more specific.

L1, clinical & molecular data. Tempus's structural asset is a multi-modal oncology dataset: sequencing, imaging, clinical records, outcomes, assembled through hospital partnerships over a decade. This is L1 in its hardest form: not scraped, not synthetic, not easily replicable. Every new hospital integration deepens it.

L3, vertical foundation models. On top of the data, Tempus trains oncology-specific models (genomic interpretation, treatment response prediction). General foundation models from OpenAI or Anthropic cannot reach this layer without the L1 underneath, which is the structural point.

L8, clinical workflow & reimbursement. This is the layer most outside observers underweight. The output has to land inside an oncologist's decision moment, with billing codes, regulatory clearance, and liability assigned. L8 in regulated medicine is slow, expensive to build, and durable once built. It is also the layer that protects the stack against L2/L3 commoditization from above, a frontier model can match the prediction, but cannot ship it into the clinic without re-doing the L8 work.

How the layers behave differently here.
• L-1 and L1 matter more than in software. Sequencing instruments, imaging hardware, and patient-consented data flows are the real moats. Compute is not where the scarcity sits.
• L2/L3 compress more slowly. Regulatory clearance ties a model version to a specific use case. You can't ship a new model weekly the way a SaaS company can.
• L8 is the moat, not the wrapper. In software, L8 is often the unbuilt layer everyone is racing toward. In regulated medicine, L8 is the layer that takes a decade to build and is hardest to replicate.

The structural read: Tempus owns a defensible L1 + L3 + L8 stack inside oncology. The contested question is whether the stack expands horizontally (other disease areas) faster than newer entrants (PathAI, Paige.AI, generalist labs partnering with foundation-model providers) can assemble comparable data + workflow in adjacent verticals.

What to watch. Whether Anthropic or Google's medical-LLM efforts pair with hospital systems directly, which would attack L3 from above, and whether Tempus's L8 workflow density holds when that happens.

Public filings (Tempus S-1, post-IPO reports). Numbers approximate as of May 2026.