Summary: Oncology AI startup Triomics closed a $22 million Series B on May 27, 2026, with Battery Ventures leading and earlier backers Nexus Venture Partners, Lightspeed, and Y Combinator each putting money in again. For product teams the takeaway: once general models commoditize, differentiation stops being “can it summarize” and becomes pipeline depth (HL7, FHIR, faxes), EHR integration (Epic Marketplace), and verifiable output (citations on every field).
Oncology AI startup Triomics closed a $22 million Series B on May 27, 2026, with Battery Ventures leading and earlier backers Nexus Venture Partners, Lightspeed, and Y Combinator each putting money in again. The raise comes two years after the $15 million Series A of May 2024, when Triomics was still all-in on a single product for clinical trial matching. Its work has one concrete core: models purpose-trained for oncology, turning cancer records that can run a thousand pages into structured answers embedded in the clinical workflows a center already has.
This round says something about the market. Frontier models keep improving, yet hospitals pay for vertical AI — tuned to their specialty, plugged into the tools they already run, and open to audit. CEO Sarim Khan’s numbers back that up: enterprise customers have quadrupled in a year, and annual recurring revenue has grown tenfold.
Generic AI Chokes on Thousand-Page Charts
Khan is blunt: patients living longer means files that keep swelling — “we have seen medical records [with] thousands of pages of information.” OncoLLM, the company’s core engine, reads the whole record in one pass — progress notes, pathology, imaging, molecular data, faxes — not page by page. The choice matters clinically: a biomarker result buried on page 400 changes what page 12 means, and page-by-page summarizing discards exactly those cross-references. On formats, the platform takes HL7, FHIR, CCDA, PDF, and TIFF, connects natively to Epic, OncoEMR, and iKnowMed, and holds an Epic App Marketplace listing.
Models trained on oncology data, the company says, beat generic AI agents at this work — the reason institutions like Memorial Sloan Kettering and Yale Cancer Center run them.
Three Product Lines for Three Heavy Workflows
Three workflows eat the most staff time in a cancer center, and Triomics builds a product for each:
- PRISM handles clinical trial matching: every scheduled patient is run against every active trial each day, yielding a ranked worklist backed by citations. Per an ASCO 2025 abstract written with an academic medical center: roughly 30% more accruals, 95% accuracy, and 40% more matches
- Symphony covers visit preparation: it pulls disease status, treatment timeline, and biomarkers into a pre-visit snapshot, drafting notes in the physician’s own workflow. Customers report pre-charting time down 80% on new visits and 45% on follow-ups
- Harmony runs cancer registry and quality abstraction: NAACCR, SEER, COC, and QOPI submissions are generated field by field, each one citing its source — 96% accuracy, 75% faster abstraction, 60% faster case finding, per a preprint
Deployments and Operating Volume
The roster goes beyond MSKCC and Yale Cancer Center: the site also lists Mount Sinai (its AI trial-matching platform went live in January 2026) and the Regenstrief Institute (real-world evaluation started February 2026). On volume: more than 120,000 charts are read and over 65,000 patients are screened each month, with upwards of 10 NCI-designated cancer centers among customers — 4 of the top 10 US cancer centers by the company’s own count.
These are operating numbers, not demo statistics — nobody reads 120,000 charts a month without being wired into live scheduling and the EHR.
This Is Not a Scribe Company
Nearest are the AI scribes — Abridge and Microsoft’s Nuance — which condense doctor-patient conversations. Triomics works the other side: the mass of historical records sitting there before any conversation starts, delivered as structured, cited, auditable fields rather than narrative notes. Every answer maps back to the line in the record it came from — non-negotiable in registry submissions, where one wrong field forces a resubmission. The two categories should eventually collide, as scribes reach deeper into the chart and vertical platforms reach toward the visit itself; for now the boundary is clean.
Vertical AI, Executed by the Book
Two rounds bracket the story — $15 million in May 2024, $22 million now — and the method is the classic vertical-AI sequence: pick a specialty with long, fragmented, compliance-heavy records, then sell the model, the integrations, and the workflow as one package. For product teams the takeaway: once general models commoditize, differentiation stops being “can it summarize” and becomes pipeline depth (HL7, FHIR, faxes), EHR integration (Epic Marketplace), and verifiable output (citations on every field). That combination is what healthcare AI budgets will chase over the coming year.