Technical Documentation

Autonomous Remote Cluster Architecture,
layer by layer.

Arca Grid is built on six load-bearing layers: pinned runtimes, signed releases, fast-sync overrides, autonomous allocation, thermal routing, and a custody-free payout plane. None of them are optional, and none of them are hand-tuned at runtime.

Layer 01

Containerized Environment

Pinned CUDA 12.0 runtime with production-verified tuning.

Every worker is shipped as a sealed OCI image with a pinned CUDA 12.0 runtime, driver-floor enforcement, and a curated kernel-tuning manifest. There is no in-place driver upgrade path — replacement workers ship with a new tagged image or they do not ship at all.

CUDA 12.0OCI-signedDriver-floor pinned
Layer 02

Immutable Infrastructure

No runtime hot-swaps. Every release is cryptographically signed and tag-pinned.

Workers and orchestrators never patch themselves in place. Each release is built reproducibly, signed with our release key, and pinned by tag in the allocator's manifest. Rollback is a tag flip — never a hand-edit on a running host.

Cosign verifiedTag-pinned manifestReproducible builds
Layer 03

Automated Fast-Sync Overrides

Bootstrapped chain archives eliminate server-hop sync delays.

Cold-starts pull a verified archive snapshot from the nearest mesh edge instead of replaying the chain peer-by-peer. The snapshot is hash-checked against the signed release manifest before the worker is allowed to attach to a payout address.

Edge snapshotsManifest-checkedSub-minute cold start
Layer 04

Autonomous Spot-Market Allocation

The allocator locks the cheapest qualified worker before your container ships.

A continuous bidder walks the mesh's spot inventory, filters on driver floor, thermal headroom, and uplink class, then locks the cheapest match against a signed lease. Margin gating happens server-side; the public API never sees the underlying provider quote.

Signed leasesMargin-gatedProvider-blind
Layer 05

Hard-Capped Thermal Routing

Workers are migrated off hosts that breach per-class thermal ceilings.

Telemetry samples are graded against a per-class thermal ceiling. A breach drains the worker, signs a migration receipt, and rehydrates on a healthy host without losing the active payout binding. No silent throttling.

Per-class ceilingsDrain-and-migrateReceipted moves
Layer 06

Custody-Free Payout Plane

Payout addresses are injected as env, never persisted to host disk.

Wallet addresses are passed to the container as USER_WALLET at boot, kept in memory for the lease duration, and discarded when the worker drains. Arca Grid never takes custody of the underlying funds.

No host persistenceMemory-onlyCustody-free

Compliance posture

  • · No host-disk persistence of payout addresses or operator credentials.
  • · Every release tag is reproducible from the public source manifest.
  • · All worker→orchestrator traffic is mTLS-pinned on the mesh edge.
  • · Margin and provider pricing are strictly server-side and never returned over the public API.

Need fleet-grade orchestration on top of this stack?

Open the Fleet Console →
© 2026 Arca GridArca Grid · Autonomous Remote Cluster Architecture