Federated public AI infrastructure — pilot & rollout¶
A conditional concept for the federated public AI infrastructure, originally framed around Geneva 2027. The sequence below is retained for discussion; it is not the current implementation plan or a delivery commitment. The current development focus is Evidence-Gated Agents.
Illustrative MVP¶
A working Public AI Compute & Evidence pilot: one governed API, several public models, at least two independently operated nodes/providers, transparent quota rules, public model/node/evaluation cards, one constrained public-interest agent or evaluation use case, and a governance blueprint for expansion.
Minimum build:
- one gateway exposing 2–4 public models;
- broker for project identity, quota, data-class policy, and accounting;
- one inference/cloud node plus one independent research/HPC node;
- model, node, and evaluation cards; incident channel; monthly usage summary;
- one public-data evaluation or multilingual civic-information agent;
- admission criteria, prohibited-use baseline, conflict disclosure, appeal/redress path, and anti-capture rule.
Start with an Apertus/PublicAI-backed inference + evaluation MVP (PublicAI here = the Public AI Inference Utility, publicai.co — the live access layer, separately operated from the publicai.ch chapter) because it builds on the nearest live pattern. Apertus is the first anchor because the live open assets sit there today; the federation's rules remain model-plural. Add broker/evidence controls around the access layer, then federate a second independent node/provider, ideally with a second model family.
A small civic agent grid can run in parallel as the citizen-participation track, but it carries more reliability and coordination risk. It should not replace the inference MVP.
Success is not global scale. Success is proving that public compute can be shared under inspectable rules.
Development sequence¶
- 0–30 days — define and align: adopt working terms; treat Apertus/PublicAI as reference implementation; map candidate nodes; choose one public-data use case; draft
ComputeNode v0,PublicAIWorkload v0, and card templates. - 31–90 days — prototype: governed API pattern; accounts, quotas, logging profiles; one batch evaluation plus one constrained agent task; public cards and a lessons report.
- 3–6 months — federate: add a second independently operated node via adapter; test routing, failover, quota accounting, disconnect, incidents, and public reporting; keep sensitive data out of scope.
- 6–12 months — institutionalise: technical steering group, separate public-interest governance group, admission rules, contribution credits, and pilot-scale funding tied to the evidence deliverable.
Decisions to settle¶
| Decision | Default recommendation |
|---|---|
| First pilot | Apertus/PublicAI-backed inference + evaluation MVP; small civic agent grid in parallel |
| First nodes/providers | One PublicAI-style inference utility node + one Swiss research/HPC or sovereign-cloud node |
| First broker operator | PublicAI/SPIU-adjacent (publicai.ch: cooperative in formation, not yet registered) or Swiss/SNAI-adjacent neutral host; no new-institution claim |
| Identity scope | PKI/CA + signed SLAs for institutional nodes; Keycloak/OIDC for users; eduGAIN-style later |
| Contribution credits | Banded GPU-credit accounting later; do not block the pilot |
| Commercial providers | Allowed only with disclosure, portability, quota limits, and no control over governance or findings |
| Sensitive data | Excluded from the first public pilot |
Framing¶
Apertus/PublicAI is the nearest live anchor; the Charter adds broker/evidence controls. Node control stays local while identity, quota, policy, evidence, and accountability are shared. The first deliverable is small and testable.