Talaria
Proportional inference, kept under EU jurisdiction.
Talaria is an open European standard for proportional inference. It routes the AI work that does not need a frontier datacentre to a pool of consumer-grade compute already running in European homes. Encrypted end to end. The operator is architected not to receive data that identifies the individual a request relates to. Built so European model providers cut their per-token cost and keep every token under EU jurisdiction.
Talk to us about integrating Talaria →The compute is already in European households.Talaria puts it to work.
For European AI providers the frontier gap is not a quality gap. It is a compute economics gap. Talaria closes it with hardware that already exists, inside European homes.
Inference, not training, has become the cost bottleneck of European AI. European model providers do not lag the US frontier on quality. They lag on compute economics, because the compute is hosted abroad and every cent of it leaves the EU. That is a structural disadvantage and it is the one Talaria is built to close.
The substrate to close it already exists. Tens of millions of capable consumer devices sit in EU households, idle most of the day. Most real inference traffic is everyday work those devices can serve at the quality users expect. Talaria matches that traffic to the hardware that already sits in European homes.
A provider adopts it in stages. Add one open-source library, watch it run safely alongside your own infrastructure, then route the requests that are a good fit to the pool. Work that genuinely needs a frontier datacentre stays there. The rest runs on consumer hardware inside the EU. The system is conservative by design.
This is proportional inference. Not a marketplace and not a service. An open standard with a reference implementation, free to adopt forever, kept open by a social-interest lab rather than metered by a platform. The household earns. The provider pays a fraction. The token never leaves Europe.
The longer arc is European strategic autonomy. The same standard that serves EU providers today can serve the routine work of the wider agent economy tomorrow, pulling that compute into European hardware rather than out of it. Policy outcomes, delivered through products.
Most work runs close to home. The frontier handles the rest.
Talaria sorts inference by what it needs. The everyday majority runs on consumer-grade compute inside the EU. Only the work that genuinely needs a frontier datacentre goes there. When there is any doubt, the request stays with the provider.
European inference costs more, and the money leaves Europe
European model providers carry a real compute-cost disadvantage against the US frontier. The cause is not weaker models. It is that the compute is hosted abroad and every cent of it leaves the EU. The continent's answer so far has been a large datacentre build-out. That is one way to add capacity. It is not the only way. Most API traffic is everyday work that runs comfortably on the consumer-grade compute already sitting in European homes and idle most of the day. The capacity is sitting in European kitchens and home offices right now. Nobody has put it to work.
Lower cost, kept inside Europe
Route the work that does not need a frontier datacentre to consumer-grade compute and cut your per-token cost. The savings stay with the provider that did the work and they stay in the EU. You decide how to spend them. Wider free tiers, frontier access for power users, a price cut for end users, or a direct payout.
Quality and latency never regress
When the network cannot meet your quality bar, the request stays on your own infrastructure. Talaria is conservative by design, so it can only match or improve the end-user experience, never degrade it.
EU data sovereignty by construction
EU and EEA requests are designed to stay within their jurisdiction by default, as a property of how the system is built rather than a contractual promise alone. It is designed to strengthen your GDPR and AI Act procurement posture, with documentation your compliance team can review, served from hardware that is EU-resident by design.
Privacy by architecture
Talaria is architected so the operator serving a request does not receive data that identifies the individual it relates to. Privacy is a property of how the system is built. It is intended to support work in sensitive areas like health, law and finance, subject to each provider's own compliance review.
A transparent, accountable fee
Pricing is transparent and competitive, not a number a vendor invents. Surplus is structured to recirculate into the European AI ecosystem, not the platform. This is what hungry, not gluttonous means in practice.
Drop in, with no lock-in
Integrate through a single open-source library and a small block of configuration. No rewrite. Watch it run alongside your own infrastructure before you trust it with live traffic, and turn it off instantly with one flag. It integrates with standard monitoring tools your team already uses.
Open where openness earns trust. Confidential only where it must be.
- Open by default. The standard is openly licensed and the reference implementation is open-source. Adoption is free, forever.
- Stewarded by a social-interest lab. Phareau is a social-interest lab, not a platform and not a foundation. It stewards the open standard without extracting from it. Fee terms are designed to live in the standard and its governance, not in revocable terms of service, so they cannot be quietly changed against adopters.
- Hungry, not gluttonous. A steward-ownership structure keeps the company from drifting toward extraction, even after later funding. Surplus is built to recirculate into the European AI ecosystem.
- Confidential where it counts. The protocol is open because openness is how trust is earned. A small number of implementation details stay confidential, and we walk partners and evaluators through them directly, under NDA.
- Sourced, not asserted. Every quantitative claim names its source. We do not present an estimate as a measured fact.
The window is open
Three forces line up at once. Frontier agents have made routine, smaller calls a large and fast-growing share of AI demand, and today those calls pay frontier prices for work consumer hardware can do. Consumer-grade compute has become capable enough to run that everyday work at the quality users expect. And the EU funding and policy frame actively favours mission-driven open-standards ventures aligned with strategic autonomy. The window is open.
Owned by Europeans
Talaria keeps AI compute and the value it creates inside the EU. EU and EEA requests are kept within their jurisdiction by default, the audit posture is EU-resident by design, and providers inherit it into their own GDPR and AI Act compliance. The customers are European model providers and the substrate is European households. Phareau is a social-interest lab, hungry not gluttonous, with a steward-ownership structure that keeps the open standard from drifting toward extraction even after later funding. Paris is a hub, not a boundary, with teams across the Union. This is what European AI infrastructure can look like. Distributed, sovereign, and owned by Europeans.
If you run inference for European users, we should talk.
If you run inference for European users, we will show you the cost picture for your own traffic and walk your team through a no-risk shadow pilot. Adoption is free, forever.
Talk to us about integrating Talaria →