Fragmented compute
Laptops, workstations and servers sit idle while inference demand remains concentrated in a few centralised providers.
Phase 0 protocol and reference implementation
Federated AI compute,
without pretending the internet is one GPU.
FACF is an open protocol and reference architecture for routing eligible AI workloads across independently operated compute nodes — with explicit policy, short leases, metering and bounded fallback.
Whole-workload schedulingOne complete request or job runs inside one provider cell.
Human-controlled capacityProviders choose models, hours, power limits and concurrency.
Fail-closed policyUnknown identity, trust or data class never expands access.
The practical architectureCoordinate independent cells.
Coordinate independent cells.
Keep execution local.
The WAN control plane discovers capacity and places work. Local runtimes such as Ollama or vLLM execute it. FACF does not stretch tensor parallelism across consumer internet links.
Workload ownerApplication or agentOpenAI-compatible request or asynchronous job
→
Control planeGateway + brokerIdentity · policy · scheduling · leases · metering · fallback
→
Provider cellFACF agent + runtimeAllowlisted model · local limits · outbound authenticated channel
Why nowUseful capacity already exists.
01 / DISCOVER 02 / CONTROL 03 / MEASURE
Useful capacity already exists.
The coordination layer does not.
Different constraints
Nodes vary by hardware, model, trust, geography, availability, cost and network quality. Policy must come before placement.
Verifiable service
Availability, latency, throughput and token counts need evidence before marketplaces, payouts or capacity claims.
Evidence, not hype
The open-weight record made this possible.
Local inference wasn't always capable enough to route seriously. The research trail is short and specific.
- Transformer (2017) — made attention-based models parallelizable enough to serve on a single GPU.
- Chinchilla (2022) — smaller models trained on more tokens beat larger ones.
- LLaMA (2023) — a 13B open-weight model matched GPT-3 on most benchmarks.
- LoRA (2021) and QLoRA (2023) — fine-tuning moved from a cluster to a single GPU.
- Mistral 7B and Mixtral (2023) — small and sparse-MoE open models matched or beat models several times their size.
- DeepSeek-R1 (2025) — put frontier-comparable reasoning behind open weights.
Trust by design
Clear boundaries beat vague promises.
The first network is deliberately narrow: manually approved EU providers, public or non-sensitive text workloads and approved open-weight models.
- Provider agents open outbound authenticated connections
- The broker never receives arbitrary shell access
- Request bodies are not logged by default
- Providers can pause, drain or stop sharing immediately
- Fallback stops after client-visible streaming begins
Open protocol. Managed network.
Adoption and sustainable stewardship can reinforce each other.
Protocol schemas, the community agent, conformance fixtures and baseline reference components are intended to remain open under Apache-2.0. Cloud Computing Oy may operate the official managed network, verified-provider programme, enterprise control plane, billing, support and service-level commitments.
Pilot pathStart small. Measure everything.
PHASE 0 PHASE 1 PHASE 2
Start small. Measure everything.
Earn the right to scale.
Protocol simulator
Scheduling, leases, providers, privacy-safe metering, Ollama boundary and fallback orchestration.
Verified federation
Three to five approved provider cells, production identity, transport, persistence and conformance.
Managed service
Customer gateway, provider programme, audited accounting and commercial service guarantees.
HAVE CAPACITY
Run a provider cell
You have an idle EU-based GPU — a workstation, a homelab box, a spare rack unit — with Ollama or vLLM already running. Phase 1 needs two to three more manually approved provider cells for a 30-day pilot: text inference only, open-weight models you choose, hours and power limits you set. No token, no marketplace, no unattended access to your machine.
HAVE WORKLOADS
Route a pilot workload
You want to test an OpenAI-compatible gateway backed by independently operated capacity instead of one centralised provider. The pilot measures time-to-first-token, throughput and success rate, and falls back before you notice a failure. Non-sensitive text inference only — no production data, no SLA yet. That is what the pilot is for.
This is a manually approved pilot, not a self-service signup. Expect a short screening conversation before anything connects.