Independent research institute · Pre-seed

Europe's community-run inference layer.

Hanseatic Alliance for Networked Sovereign Architecture

H.A.N.S.A. Institute is an independent research institute building a sovereign, auditable AI inference network from hardware Europe already owns — so access to AI doesn't depend on a handful of hyperscale datacenters, and your traffic never leaves the geographic boundary you choose. Public benefit first.

Fund & partner Read the thesis Run a node → Download for Windows

Here to back the mission? Funders, foundations, and research partners start at Fund & partner. Here to run AI on your own hardware? Builders and node-runners head to the network →.

Why this institute exists

AI is getting more capable and less accessible at the same time

The dominant paradigm is "bigger model, bigger datacenter." That concentrates cost, control, and energy in a few hands. We think the next decade of useful AI — chains of small, fast, specialized tasks — belongs on hardware people already own.

01 // DEPENDENCE

A few hyperscalers hold the keys

Access to AI runs through a small number of centralized providers. When one goes down, changes terms, or restricts a region, everyone downstream is stuck.

02 // COMPLIANCE

Regulated data can't just leave

GDPR and the AI Act demand inference that is auditable and jurisdiction-bound. Our network enforces geographic boundaries at the routing layer — country, EU, or world, by design.

03 // ENERGY

Centralized compute is heavy

The "more datacenters" answer carries a real energy and hardware footprint — while billions of capable consumer GPUs sit idle most of the day. Using what exists is cheaper for everyone, too.

What we stand for

Access over profit, by design

The institute exists to keep these commitments true as the network grows — even where they'd be inconvenient for a purely commercial operator.

🔓

Lower the entry barrier

One-command install. No API keys, no usage bills, no account needed to run it. If you have a gaming PC, you can run and serve real models today. Apple Silicon Macs work too (tested on one Mac so far).

🐝

A community swarm

A network owned by its participants, not a platform. Share spare compute and draw on others' when you need more — geo-scoped to rules you choose.

🌍

Out of the datacenter

Push inference to the edge — homes, labs, small operators — and study what becomes possible when compute is distributed instead of centralized.

🌱

Sustainable by design

Use hardware that already exists, run it cooler, and waste less. Lower energy per inference and longer hardware life are first-class goals, not afterthoughts.

🛡️

Security for the many

The same swarm that shares compute can carry verified, signed security updates to homelabs and small operators — keeping edge AI safe as the field moves fast, without needing a corporate IT department behind you.

Trust, stated honestly

Three trust tiers — we tell you exactly what each guarantees

Most "is it safe?" confusion disappears once the deployment tier is named. We publish our threat model and never let marketing outrun the architecture.

TierWhat it isConfidentiality today
Local One machine, fully offline. One command starts a private, OpenAI-compatible endpoint. Full
Exactly as private as your own machine — because it is your own machine.
Private swarm Nodes owned by one organisation — yours or ours under contract — behind your own boundary. High
Determined by your physical and network control, as with any on-premise system.
Public swarm Community nodes matched by the coordinator under each node's geo policy. Limited today
Send only non-sensitive workloads for now. Already in the node and being tested between machines: a filter that strips names, credentials, and identifiers from a request on your own machine before it leaves. Rolling out through the alpha: relay routing, so the serving node never learns who asked. Once both are in place, the worst case with a malicious operator is an anonymous prompt with the sensitive parts already removed. Node validation and confidential execution follow on the funded roadmap.
What holds today: cryptographically signed updates (nodes reject tampered software), answers from peers that carry a receipt signed by the node that served them, a coordinator designed to stay off the inference data path, geo-scoped routing, and encrypted transport. Rolling out through the alpha: relay routing for sender anonymity, a client-side PII firewall (already in the node, being tested between machines) that scrubs names, passwords, and IDs before a prompt leaves your machine, and local sensitivity routing that keeps flagged prompts on your own node by default. What comes later — node validation, confidential execution — is written down in our public threat model with a date and an owner, not hidden.
The name

Why "Hanseatic"?

H.A.N.S.A. stands for Hanseatic Alliance for Networked Sovereign Architecture — and the medieval reference is deliberate.

A 400-year proof that federations of equals work

From the 12th to the 17th century, the Hanseatic League — the Hansa — linked merchant cities across the North and Baltic seas: Lübeck, Hamburg, Gdańsk, Bergen, Riga and nearly two hundred more. It had no king, no capital, no state behind it. Every city kept its own laws and stayed fully sovereign — yet together they ran the most reliable trade network in Europe, built on shared routes, common rules, mutual protection, and reciprocity: you carried goods for the network, and the network carried for you.

That is the exact shape of the network we're building. Every node stays sovereign — your hardware, your rules, your geographic boundary. The coordinator is a rulebook, not a ruler: it matches compatible peers and enforces each node's policy, but is designed never to touch the inference traffic itself. What the Hansa proved with ships and kontors, we're rebuilding with GPUs and encrypted streams: coordination without central control.

The research thesis

An open testbed for distributed inference

Combining smaller models can rival or beat single large ones — Mixture of Agents, self-consistency, speculative ensembling. The research exists; what's missing is a real-world substrate to run it at scale. The institute builds and stewards that substrate.

Why distributed, why now

For agentic workloads — many small, fast, specialized calls — a single monolithic model call is like using a freight train to deliver a letter. Intelligent routing of the right small model to the nearest capable node is both cheaper and a genuinely open research question with no public testbed today.

H.A.N.S.A. turns distributed nodes into a unified inference fabric and exposes the knobs researchers need: ensemble methods, task-aware dispatch, latency-aware peer ranking, and opt-in anonymized telemetry to study it all honestly.

First measurement, on our own small benchmark (30 Oct 2026): given material too long to read directly, an agent that handed searching and counting to the node's tools answered 12 of 12 questions correctly, against 6 of 12 with plain file tools, and used about 32 times fewer tokens of its own.

Flagship research goal

GPU temperature-aware task routing

During long agentic runs, local silicon accumulates heat. Nodes already step aside when their GPU is busy or hot. With funding, the network will route tasks by each node's GPU temperature in real time, shifting work to cooler peers, and measure what that does to sustained operating temperature and hardware wear. No public testbed measures the thermal behaviour of agentic workloads on consumer hardware today; building one is the goal.

How the network works

Built on open standards, off the data path

The coordinator handles discovery and geo-policy only — it is designed to stay completely off the inference data path. Want the full technical walkthrough and the install? See the network page →

01 // INSTALL

Hardware → running model

One command detects your GPU and starts a private, OpenAI-compatible endpoint. Works fully offline. No account needed.

02 // CONNECT

Opt into the swarm

Choose your geo scope — country, EU, or world. The coordinator matches you with compatible peers under your policy.

03 // SHARE

Contribute & consume

Offer idle compute, draw on the network when you need more. Designed so the coordinator never sees inference data.

04 // STUDY

Research substrate

The same batch-splitting primitive that distributes work also enables multi-model ensemble research at network scale.

Timeline

From self-hosted MVP to a funded research program

MVP done. Hardening in Q4 2026. Community alpha in Q1 2027.

Done
Foundation & MVP
  • Angel pre-seed secured
  • One-command install → private offline node (double-click on Windows)
  • EU nodes forwarding under geo policy
  • Home-to-home P2P, no port forwarding
  • Signed updates, applied unattended on live nodes
  • Model files hash-checked before every start
  • Accounts and invite-only network joins
  • Steps aside when you game
  • One-click connection to Claude desktop
  • Public network map
Now · Q4 2026
Hardening
  • Security hardening before outside nodes join
  • PII firewall: testing between machines
  • Agent-harness trials: Hermes, goose, opencode
  • Academic partnerships
Q1 2027
Community Alpha
  • Invite-only EU node operators
  • Open node enrollment
  • Privacy layer: relay routing + PII firewall
  • GDPR attestation layer
2027
Research & Production
  • Ensemble inference (MoA)
  • Thermal-aware routing
  • Open ecosystem
  • Research publications
Who's building this

A founder who'd rather ship

Grant reviewers fund people. Here's the person behind H.A.N.S.A. — and what we'll build the team into with funding.

CS

Cezary Szałkowski

Founder

Founder. Four years at McKinsey & Company in internal technology, capped by taking a small data-automation project from pitch to working MVP. The real founder education came outside: stranded in Sri Lanka during the pandemic, he launched a distributed digital-fabrication startup to cut reliance on centralized imports — then, when travel reopened, passed its equipment to local young entrepreneurs so the work could carry on without him. A builder who ships, mostly alone, and says so plainly.

H.A.N.S.A. turns that same instinct on AI — replacing dependence on a handful of hyperscale datacenters with the compute people already own.

With funding, the first hires are a distributed-systems engineer and a security/research lead — to take the network from MVP to a trusted open testbed and keep edge nodes safe.
Open to grants & aligned partners

Help fund accessible, sovereign AI

We're raising to take H.A.N.S.A. from a working MVP to a funded research program and a live community network. We're looking for foundations, public research programs, and mission-aligned partners who care about AI access, European digital sovereignty, and sustainable compute.

Build the teamFirst engineering and research hires
Run the networkCoordinator, EU nodes, public dashboard
Open the researchEnsemble + thermal-routing studies, published
Email funding@hansa.institute

Aligned investors welcome too — if the mission fits your mandate, say hello.