Cowen 2026: AI Was the Whole Conversation — Here's What Cut Through
Cowen 2026 — I spent a few days at TD Cowen's fifth annual Tools/Dx "Revolution," a panels-driven conference with a deliberately long-horizon lens. If you compressed it into one sentence: AI was the whole conversation, and the more interesting question wasn't whether it matters but where the value actually lands. Panel after panel, the same pattern surfaced — the value in this industry is migrating off the assay (off the box, off the chemistry) and into the system around it.
A handful of through-lines stuck with me.
1. AI is a demand multiplier for the wet lab, not its executioner.
The most useful — and most contrarian — read of the week. With CRO and tools stocks having been sold as if AI were about to vaporize wet-lab demand, the room said the opposite: in TD Cowen's conference polling, by something like 23 to 1, attendees expected biopharma's AI build-out to increase wet-lab spending over the next few years, not shrink it. The mental model panelists kept reaching for was a flywheel — more AI requires more data to train, test, and validate, which requires more wet-lab work, which feeds better models. AI changes the shape of R&D; it doesn't delete the bench.
That is exactly the kind of second-order point the market keeps getting wrong, and it's why I keep arguing the tools layer is undervalued. Even on the clinical-R&D side, where the disruption fear is loudest, the room was measured — pegging AI-driven cost savings in the high-teens percent, and stressing that the near-term win is time and productivity, not the wholesale elimination of spend. A good deal more sober than the selloff implies.
2. The AI that wins is rebuilt into the workflow — and general models are eating the bespoke ones.
Scale and price were not what the room rewarded. Asked which data assets and tools actually matter to pharma, attendees prioritized product quality, data accessibility, workflow integration, and turnaround over raw size. The valuable asset isn't the biggest model or dataset; it's the one that's rich, multimodal, connected to providers, and actionable inside a real workflow. Customers are done buying point solutions — they want connected data fabrics and AI-ready infrastructure.
The other half of this story came from the foundation-model talks: the general-purpose models are quietly out-performing the bespoke medical ones, sometimes on a fraction of the data. In one demonstration, models played a sequential-diagnosis game — asking for information one question at a time, paying for each test they ordered, then committing to an answer — and beat physicians without ordering the entire hospital. The design was the point: not one model answering, but several agents with distinct roles (one tracking the differential, one asking the next question, one watching the budget), coordinated like a tumor board. AI is becoming infrastructure, not a feature. The losers will be the ones who automated their existing org chart. In diagnostics, workflow is still the product — AI just raises the stakes.
3. Clinical adoption is generalizing — and MRD is the tell.
The single most useful data point: roughly half of U.S. oncologists ordered an MRD test last quarter. The panels called MRD the most important growth driver in oncology Dx, still in the "first inning," with a long-term vision that ordering it becomes as automatic as ordering a CT. Read that against the 50% figure and the story isn't "MRD is here" — it's that half the addressable prescribers haven't started, on something many already treat as standard of care. The trend that matters is generalization, not new indications.
That maturity also sets up consolidation. The room overwhelmingly expects M&A to heat up, with MRD leaders as the prize — differentiated, expensive-to-replicate assets — and an interesting split on who buys: three in four expect Tools/Dx-MedTech acquirers, but a quiet one in four expect Big Tech to buy its way in as data becomes the asset. This is the adoption curve the publication data has been drawing, described in commercial language by the people living it.
4. The real moats are data and reimbursement — and only one of them made Cowen's list.
Two moats came up everywhere, and neither is the assay.
The first is data, and Cowen put it near the top: the value of diagnostic data is inflecting, with the most confident companies pitching the longitudinal dataset behind the test — exomes, cell-free DNA over time, pathology, outcomes — fused into models and sold into pharma's development loop. A telling detail: several operators noted the data business may need to court a different set of investors than the legacy test business to be valued properly.
The second moat barely registered in Cowen's official ten, and it's the one I'd put my chips on: reimbursement access. The systemic picture from the oncology sessions is genuinely alarming — coverage turnaround that used to take nine to twelve months now pushing past three years, growing backlogs, and review cycles that have eroded a decade of carefully built trust. Underneath it, a structural absurdity: Medicare cannot consider cost in a coverage decision, so a roughly $4,000 test that lets you safely de-escalate a $130,000-a-year therapy gets judged as if the downstream savings don't exist. This is the sick-care-to-healthcare gap in its most concrete form. The biology is no longer the bottleneck. The 60-day sit period is.
5. Longevity and multiomics: a bullish base case, a wildly divided room.
The longevity and screening panels were the most energetic — and the polling was the most spread out I saw all week. Asked how many multi-cancer early-detection tests would run across all vendors in five years, answers ranged from under one million to more than ten, landing around five on average. That sounds modest until you realize it's a 15x-plus ramp from today and actually above Cowen's own forecast. So not skepticism — genuine disagreement about magnitude on top of a bullish base case. The nuance underneath: single-cancer tests (clear path, established reimbursement, defined workflow) were the near-term favorites, while multi-cancer and multi-omics offer richer data at prices built for a different planet. For screening to reach asymptomatic populations, the same two gates kept coming up — cost has to fall toward $100, and the false-positive problem has to be solved — which is why the serious longevity players pair imaging with genomics and biomarkers rather than betting on a single modality. I remain constructive on early detection; the week was a reminder that conviction and timing are different things.
Underneath all of it, Cowen's closing theme and the omics panels agreed: the next wave of precision medicine runs on multimodal datasets — genomes, proteins, imaging, longitudinal monitoring — rather than single-analyte tests, and it's starting to spread from research into the clinic.
The through-line behind the through-lines.
Put it together and you get one picture. The instrument, the panel, the chemistry — the things we used to call the product — are necessary and increasingly commoditized. The defensible value is collecting in the layer around them: the workflow the AI redesigns, the longitudinal data that compounds, the reimbursement access that's brutally hard to win, and the clinical habit that turns a test into a standard. AI doesn't change that thesis; it accelerates it.
It's the same thing I took away from Roche's Diagnostics Day six weeks earlier, now said by an entire conference: in precision medicine, the infrastructure layer is starting to matter more than the molecule.
The rest — the parts that don't belong on a public page — we can cover over a beer.
Polling figures cited here are from TD Cowen's audience surveys at its 5th Annual Tools/Dx Revolution conference (June 2026); the reads and opinions are my own.