← All health signals
Infrastructure

Training Without the Data

Hospitals and rival drugmakers are learning to train shared models on records none of them will ever hand over, treating the data itself as too sensitive to move.

What is it?

Pharma firms and hospitals now train shared AI models on records that never leave the institution. Owkin’s deals with Sanofi and AstraZeneca show the method moving from prescreening into full drug pipelines, without partners ever pooling raw data. Whether it scales past one vendor’s relationships stays open.

Why now?

Institutions increasingly treat where data sits as a form of control — not handing records to rivals is itself the point. AI needs training scale no single hospital or drugmaker holds alone, so a method that trains without moving data fits the moment. Sovereignty over data is the default posture now.

Who’s building it?

Platform vendors like Owkin build the training method; pharma and hospital partners supply the data.

Federated-learning platformsPharma partnersHospital networksDiagnostics spinoutsData regulators
The full read
Why now, who is building it, where the market goes, and what stays unsettled.
Get weekly →
01 · The market
Who’s in it
Builders — training-method vendors, and the hospitals supplying the underlying data.
Buyers — pharma companies needing larger training sets than any one of them holds alone.
Gatekeepers — data-protection regulators and institutional governance boards approving each new partnership.
What’s at stake
Value is concentrating in the training method, not the data — whoever owns the protocol sits at the pipeline’s centre.
Institutional trust holds only while the data never leaves.
Regulatory approval gates every partnership before data moves.
What’s unsettled
Whether federated models match centralised training quality at real scale.
Whether the approach holds beyond pharma into finance and other sectors.
Whether one vendor’s expanding deals count as a market, or just a moat.
02 · How the demand matureda progression, not a timeline
Prescreening
A single narrow use case tests whether the training method works in practice.
Partnership
A funded spinout and a repeat license turn the pilot into a business.
Pipeline integration
Deals expand from prescreening into full drug-development workflows, proving the method scales across different partners.
Category
Other data-locked industries adopt the same architecture once pharma has proven it durable enough.
03 · Where the market goesthe forks, not a prediction
→Concentration or diffusion — one platform becomes the default rail for data-locked collaboration, or the method commoditises and spreads to rivals.
→Finance and other sensitive-data sectors adopt the same architecture, or the approach stays pharma-specific.
→Regulators eventually permit direct data pooling outright, removing the workaround’s whole reason to exist.
→Model quality catches up to centralised training, or federated learning stays a compromise.
The open question is whether one vendor’s expanding partnerships add up to a market — or whether this stays one company’s moat, however real the deals.
04 · The Digital Sovereignty deskadjacent markets, moving now
The Sovereign Model
Nations build their own AI infrastructure, locally.
Mistral raised $830M for a Paris data centre.
TechCrunch
Provable, Not Promised
Hardware proves data stayed private, not just promised.
Private compute, borrowed hardware.
InfoQ
Sovereign by Design
Rules that tie sensitive data to borders, not price.
The EU proposed four sovereignty tiers for public-sector cloud.
The Next Web
05 · The Wire
01
Owkin runs federated learning on cancer-centre and hospital data
MedCity News ↗
02
A pharma partnership extended.
MobiHealthNews ↗
03
A second licensing deal.
MobiHealthNews ↗
The panel on this signal
They agreed on
–Data-locked institutions can now train shared models without handing over any of the underlying patient records.
–Value is shifting toward the method, not the raw data.
–Multiple independent buyers, not just one vendor, are now adopting the approach.
The competition economist disagreed on
–The cited evidence traces primarily to one platform provider’s expanding set of partnerships; distinguishing that from a genuinely open, multi-vendor market with interoperable standards requires broader participation data that has not been shown.
Know more about the method →
✦How this signal was built
01 · Panel
A Delphi panel of 4 AI analysts, each reasoning independently.
→
02 · Debate
They challenge and revise each other’s reads over two rounds.
→
03 · Evidence
Every claim is tied to a cited, linked source on The Wire.
→
04 · Synthesis
A moderator writes it into one read, disagreement and all.