Kyunghyun Cho, who co-developed the attention mechanism and the GRU that helped make modern large language models possible, is starting a frontier AI lab — for health. Ortet announced its launch on September 29, 2026, backed by a $500 million commitment from Thoreau, a health technology platform building compute, rules, and research infrastructure. The lab's mission: a patient-centered foundation model that learns from every patient, for every patient.

The pitch is a direct answer to the fragmentation that has defined health AI. Most healthcare AI tools today do one narrow task on one narrow dataset — a radiology classifier here, a billing assistant there — and never learn from the relationships between a patient's biology, treatment, outcomes, and the system around them. Ortet wants to invert that: one shared intelligence foundation, continuously improved and adapted across scientific discovery, clinical care, health operations, and financial administration.

The team#

The founding roster reads like a frontier-lab team that wandered into a hospital. Cho, co-founder and CEO, is the Glen de Vries Professor of Health Statistics at NYU and spent five years building an AI platform for lab-in-the-loop molecular design — co-founding Prescient Design, then establishing Genentech's frontier research team after its acquisition. Endpoints News notes three of the co-founders come out of Roche's Genentech.

He is joined by Keunwoo Choi, Ph.D., co-founder and Chief AI Officer (previously AI systems at Upstage, Spotify, ByteDance, and Genentech); Henri Dwyer, co-founder and CTO, who built and managed what Ortet calls the largest GPU cluster in life sciences at Genentech after engineering roles at Instagram and Lyft; Jeff Hammerbacher, co-founder and chairman, founder and CEO of Open Athena and previously founder and chief scientist at Cloudera; Elman Mansimov, Ph.D., co-founder and chief science officer, a pioneer of neural text-to-image generation who worked on post-training and agentic AI platforms at AWS; and Claudia Shi, Ph.D., a founding member of the frontier research team who worked on AI R&D at Meta Superintelligence Labs.

AI-assisted genomics research
Image: UC San Diego Health / Newswise — Ortet's bet is that one foundation model can learn across biology, treatment, and outcomes, rather than splitting them across narrow tools.

What the $500M buys#

The commitment funds compute and data infrastructure, the training and deployment of health-focused models, and hiring. Ortet says it has already secured its initial GPU cluster and is building the technical foundation to move research into production at scale — and that it will work with customers across the health ecosystem from launch, grounding its models in real-world needs. Goldman Sachs advised on the transaction.

“Patient health should improve with every interaction, but today its data, knowledge, and infrastructure remain too fragmented for that learning to happen at scale,” Cho said. “Ortet is building at every layer of the AI stack so that advances in research can translate into systems that understand the complete context of a patient and support better decisions across the continuum of care.”

Physician using a digital tablet in a hospital
Photo: Vecteezy — Ortet plans to work with health-ecosystem customers from day one to ground its models in real-world needs.

Thoreau's platform bet#

Thoreau, the backer, describes itself as a technology platform that creates, invests in, and acquires health platforms across provider, payer, employer, life sciences, and government markets. Its initial platform partnerships already include Ensemble — which manages over $55 billion in net patient revenue — and Penelope Health, a real-time health policy intelligence platform. Ortet is its founding AI lab investment, and Thoreau says its mission is to simplify the administration of health and extend patients' lives in the U.S. and globally.

Why this matters#

The announcement lands as health AI sits between two eras. The narrow-task era delivered real tools but real fragmentation: every model retrained per task, none of them compounding. Ortet is the most direct attempt yet to run the foundation-model playbook — train once, on everything, then adapt everywhere — on health, led by one of the researchers who made the current AI era possible.

The stakes are not just technical. A foundation model organized around the individual patient, spanning biological, clinical, operational, and financial dimensions, is a bid to own the layer where health data becomes decisions. Watch for the first customer announcements: a $500M lab with its own GPU cluster and a real-world deployment plan has an unusually short runway to prove the foundation-model approach beats the narrow-tool incumbents.

XXX