Briefing 01 — Thesis

A protocol for distributed autonomous manufacturing.

ARCNet turns every participating facility into a training signal — and every training run into new capability for every facility.

Autonomous Resource is building the coordination layer between physical production and frontier-scale AI. ARCNet is the open protocol; the foundation model is trained on Frontier and Lux under ARC's DOE allocation.

1.7ExaFLOPS

Frontier peak (FP64)

40M GPU-hr

Allocation envelope (Y1)

12Pilot sites

Initial cohort

Q32026

Mainnet target

§ 01 · Protocol

A shared substrate for machines that make things.

Ingress · Physical sites

Plant A · CNC cell0x7F·A4

Plant B · Robotics line0x7F·B1

Plant C · Additive farm0x7F·C9

Plant D · Inspection0x7F·D3

Plant E · Assembly0x7F·E6

Signals: telemetry · process · outcome

Protocol core

ARCNet

SIGNALS/s12,418

ACTIVE NODES247

Egress · Capability

Process policiesπ.v142

Design priorsΦ.v87

Safety envelopesσ.v31

Schedulingω.v55

Anomaly detectionα.v22

Returned to every participating site

§ 01.2 · Why it matters

The same protocol, read differently.

Yield, uptime, and policy updates — as a service.

Participating sites stream anonymized process telemetry and receive back continually-improving policies: motion plans, setpoints, inspection thresholds, and scheduling. Local edge agents run fully offline; ARCNet only reconciles when connectivity allows.

01
Plug-and-play edge agent
Vendor-agnostic bridge; ships with PLC, OPC-UA, and ROS adapters.
02
Revenue share on contributed signal
Operators earn credits proportional to signal utility as measured in downstream eval.
03
Zero-knowledge IP protection
Process data is hashed and aggregated; no raw geometry or recipes leave the site.

A real-world dataset at industrial scale.

ARCNet exposes a versioned, benchmarked dataset of process → outcome pairs across heterogeneous physical systems — the first public corpus for manufacturing-scale foundation models.

01
Open evaluation suite
Reproducible benchmarks with held-out sites and seeded perturbations.
02
Paper-ready data cards
Provenance, consent, and licensing encoded per slice.
03
Co-authorship pathway
Flagship model release co-authored with cohort labs and facilities.

A compounding moat, not a model bet.

The durable asset is the protocol and the data-commons it creates. Every additional site expands coverage, sharpens the foundation model, and increases the value of incumbency for every other participant.

01
Network-effect economics
Marginal site adds non-trivially to capability ceiling, not just scale.
02
Subsidized training cost
Frontier / Lux allocation removes the dominant pre-training line item.
03
Strategic positioning
Sits between physical-AI silicon and verticalized SaaS, owning neither bet alone.

Re-industrialization as a compute program.

ARCNet translates national compute investment directly into industrial capacity. The protocol is open; the data commons is auditable; the capability flows to domestic manufacturing first.

01
Allocation utilization reporting
Quarterly public report on Frontier / Lux utilization and resulting capability deltas.
02
SMB on-ramp
Edge-agent subsidy program for facilities under 250 employees.
03
Supply-chain telemetry
Opt-in visibility into critical-component production without firm-level disclosure.

§ 02 · Supercompute

Training on Frontier and Lux, under ARC's DOE allocation.

The largest open compute envelope ever pointed at physical AI.

ARC's partnership secures a multi-year allocation on Oak Ridge's Frontier and the Lux AI supercomputer. ARCNet's training, evaluation, and continual-update loops run inside that envelope — with spill-over capacity reserved for participating research partners.

Host

Oak Ridge Leadership Computing Facility

Systems

Frontier (HPE Cray EX) · Lux (DOE AI testbed)

Allocation window

2026 Q1 — 2028 Q4

Primary workload

ARCNet Foundation Model pre-training & continual update

Governance

Joint steering: ARC × OLCF × cohort PIs

Frontier supercomputer at Oak Ridge National Laboratory

Nodes allocated

4,096/ 9,472

Current throughput

612PFLOPS

Step / 24h

1.4Mtok/s

§ 03 · Foundation Model

One model, every physical process it has ever seen.

01 / 06

Multi-modal process representation

Unifies vision, force, motion, and outcome signals into a single tokenized process stream — the substrate the model trains on.

02 / 06

Cross-facility generalization

Learns policies that transfer across heterogeneous hardware, recovering from distribution shift on unseen sites.

03 / 06

Outcome-grounded objectives

Training loss tied to yield, tolerance, energy, and cycle-time — not just next-token likelihood.

04 / 06

Continual update loop

Every production shift becomes part of the next training window — stable, audited, and reversible.

05 / 06

Safety & envelope constraints

Hard constraints are encoded as differentiable envelopes; violations are first-class training signal.

06 / 06

Edge-distilled deployment

Specialist heads compiled down to edge accelerators for latency-critical loops at each site.

Diagram · Loop

01 · Factories

Physical signal

INGESTED↗ 4.2 TB/d

02 · ARCNet

Protocol aggregation

NODES247 live

03 · Frontier / Lux

Training & eval

THROUGHPUT612 PFLOPS

04 · Return

Capability delta

DEPLOYevery shift

→ closed loop; every return tightens the next ingest

§ 04 · Roadmap

From testnet to continental-scale production AI.

Q4 · 2025

Protocol v0 & edge agent

Reference edge agent, signed telemetry schema, first 3 pilot sites on closed testnet.

Shipped

Q1 · 2026

DOE allocation active

Frontier + Lux envelope turned on; first large-scale pre-training run on mixed-site corpus.

Shipped

Q2 · 2026

Stakeholder testnet

12-site cohort online; continual-update loop live; foundation model v0.4 public benchmarks.

In progress

Q3 · 2026

Mainnet & governance

Open onboarding; contributor credit system live; joint steering committee chartered.

Planned

2027

100-site continent scale

Cross-vertical coverage (machining, additive, electronics, biomanufacturing); second-gen model.