Why Alex Karp’s warning on frontier AI is crucial for
$LMND and
$TEM
I’ve been thinking about Alex Karp’s CNBC interview. He called the frontier AI business model “effing insane,” and not because the tech lacks capability. His concern is control.
Karp’s argument was extremely direct as always, when companies replace core workflows with closed models from OpenAI or Anthropic, ownership shifts. It accrues to the model provider through the weights and the IP baked into those systems. You pay for tokens and risk exposing the data and institutional knowledge that actually drives your edge.
That’s where Lemonade and Tempus look different.
Lemonade built its underwriting, pricing, and claims stack in house from the start. The execution just validates their model. Their loss ratio recently hit an all time low of 62%, and gross profit more than doubled YoY while growing 30% . Those gains came from proprietary models trained on Lemonade’s own first party claims and behavioral data. That dataset is exclusive to them.
Tempus is positioned similarly. They’ve assembled one of the largest libraries of clinical and molecular data in oncology. Tempus creates value by owning the full loop from sequencing to EHR to imaging to outcomes. The data compounds internally. Outsourcing that workflow to a third party model would mean outsourcing the moat.
Karp’s larger point is about governance. Who owns the data, where it resides, whether prompts are secure, and whether the vendor benefits more than the customer. Palantir frames this as “Ontology” or using AI while maintaining control. The same standard applies here. Lemonade and Tempus operate in insurance and healthcare, where data provenance and auditability are table stakes. Regulators and enterprise partners will choose the vendor that can demonstrate chain of custody.
The implication is clear in my opinion . Frontier models are powerful, but in regulated, data intensive sectors, the durable winners will be applied AI companies with closed, proprietary data loops. Lemonade and Tempus are applying AI to own their workflow, their data, and their unit economics.
If Karp’s thesis holds, that’s a structural advantage. You trust your own stack. You keep TEM or LMND’s data away from OpenAI or Anthropic.
Still an open question whether the market differentiates between owning your ontology and renting intelligence. We’ll see how that reprices.
youtu.be/0A3sGymV6kY?is=L9mV…
