B// BACKBONE BY AGENTLED / INFRASTRUCTURE
LOCAL / WAITING

01 AgentLed infrastructure / first node

BACKBONE

Backbone is AgentLed's private AI inference infrastructure, assembled from hardware you already own. No rack. No cloud.

16 GB / LOCAL 32 LAYERS / WAITING WORKING PROTOTYPE
Next / 02 BACKBONE STARTS RECRUITING
OOM / 00:00:11

One laptop runs out of memory. Backbone starts looking for capacity already in the building.

02 Scavenge

Anything with
RAM gets a job.

Backbone discovers three retired phones, two mini PCs, and a dusty router. Idle devices stop being stranded assets and start behaving like one private machine.

FOUND +46 GB across five rooms
DISCOVERY / mDNS

Five heartbeats answer. Existing capacity becomes useful capacity.

03 Unified

The quiet
machines wake.

Backbone folds Apple silicon into a low-power memory bank: a Mac mini under the TV, a Studio from the office, and another laptop with 3% battery and 64 gigabytes to spare.

M1 MAXM2 PROM3 MAX
PRIVATE / LOCAL ONLY

176 GB of unified memory. Model weights and data stay inside the building.

04 Accelerate

Then the fans
start turning.

CUDA towers take the hot path. Backbone combines their tensor cores with the fleet's pooled memory instead of renting another cloud GPU by the hour.

GPU 0 / RTX 409096%
GPU 1 / RTX 309091%
COST / UP TO ~90% LOWER

Inference runs on hardware already owned and already paid for.

05 Turbulence

The network
has other plans.

Wi-Fi stalls. A phone thermal-throttles. Someone starts streaming a movie. Backbone measures the damage, preserves the chain, and routes around real life.

PHONE_02THERMAL
CUDA_02TIMEOUT
ROUTERJITTER
NODE LOST / REROUTING

Self-healing infrastructure keeps serving while individual nodes disappear and rejoin.

06 Shard

One model.
Eleven places.

Backbone splits thirty-two transformer layers into contiguous shards and places each one by memory and capability. An overloaded pair ejects two layers; neighboring machines catch them.

EMBED ATTN EXPERT OUTPUT
PLACEMENT SOLVER / ACTIVE

The model stops being a file. It becomes a route through the fleet.

07 First light

It answers.

The live Backbone prototype pools 270 GB across 11 heterogeneous nodes and sustains 22.7 tokens per second. Models and data never leave the local network.

BACKBONE / PRIVATE INFERENCE 22.7 TOK/S
YOU

Are you there?

FLEET

TTFT 842 MS11 NODES2 RECOVERED

Backbone prototype fleet inventory

Backbone is AgentLed's private AI infrastructure layer. This working prototype has eleven nodes and 270 GB of pooled memory. During network turbulence two nodes fail, routes recover, and all thirty-two model layers are redistributed before the first response.