Charterbeta
Runs on spot, edge, on‑prem & cloud

The resilient compute fabric for AI, data & media

Charter turns unreliable hardware into a reliable supercomputer. We shard any job into tiny micro‑leases executed by stateless runners. If a node vanishes, the lease is reclaimed and progress continues from async checkpoints. Train models, index corpora, transcode video, or backfill tables—on whatever GPUs/CPUs you can find.

Exactly‑once semantics · Resume mid‑task · No container restarts
Three commands, end-to-end
# 1) Guided setup: Salad integration, auth, and model/preset
charter init

# 2) Create aggregator & deploy workers with your choices
charter deploy

# 3) Live losses; exports final adapter/weights on completion
charter monitor
PythonContainersWASM (planned)GPU & CPU aware

One fabric, many workloads

A unified abstraction for the jobs modern companies actually run.

AI Training & Fine‑tuning
LoRA and full‑parameter training that survives preemptions and heterogenous GPUs.
Inference & RAG Pipelines
Embed, index, route and generate across fleets. Exactly‑once ingestion with retries & dedup.
Data & Analytics
ETL, backfills, feature stores and periodic jobs without fragile orchestration.
Media & Simulation
Video transcode, rendering, simulation & HPC workloads on mixed CPU/GPU pools.

Platform building blocks

Compose just what you need—start with the hosted coordinator, bring your own runners.

Coordinator
Assigns micro‑leases, applies deltas, checkpoints progress and enforces exactly‑once semantics.
Runners
Stateless executors that can join/leave anytime across clouds, edge PCs and on‑prem clusters.
Data plane
Streaming I/O, artifact store and provenance to replay or audit any job output.

Why Charter?

A general‑purpose, fault‑tolerant execution layer for workloads that actually fail in the real world.

Async checkpointing
Progress is persisted continuously so preemptions only re‑do tiny slices, not whole epochs.
Micro‑leases
Split work by tokens, items, frames or time so heterogeneous nodes contribute efficiently.
Stateless runners
Runners can join/leave anytime. No orchestration restarts or fragile long‑lived containers.
Multi‑runtime
Run Python functions, containers, or WASM (planned) with a single abstraction.
Spot‑first economics
Turn cheap/preemptible GPUs & CPUs into a reliable pool without babysitting jobs.
Live observability
Track throughput, active leases, cost and provenance in real time.

How it works

A lightweight coordinator/runner pattern that thrives on flaky, heterogeneous nodes.

Define tasks
Wrap Python functions or containers. Declare resources (GPU/CPU/memory) and idempotency keys.
Distribute leases
Coordinator hands out micro‑leases to runners on demand. Stalled leases are reclaimed.
Stream & checkpoint
Results and deltas stream back; checkpoints happen asynchronously for fast resume.
Why not just k8s/cron/queues?
Traditional schedulers assume long‑lived nodes and coarse jobs. Charter embraces asynchrony and fine‑grained leases so unreliable, heterogenous resources contribute without forcing restarts or manual babysitting. Use Charter alongside k8s—Charter focuses on reliability and progress, not pod plumbing.
Works anywhere
Cloud spot, on‑prem, student labs, edge devices or community GPUs. Mix them freely. If a node vanishes, the lease returns to the queue and another runner picks it up.

From early feedback

Themes we’re hearing as we refine the platform.

“Give me per‑item leases with exactly‑once delivery — retries kill my pipelines today.”
“Hosted coordinator + BYO runners would save us months of infra work.”
“I want cost + throughput per job and provenance I can audit later.”

Simple, usage‑based pricing

Start free. Scale across any compute you can reach.

Developer
$0
  • Up to 10k leases/mo
  • 1 project · CLI + SDK
  • Community support
TeamPopular
$49/mo + usage
  • 100k leases/mo included
  • Hosted coordinator
  • Project roles + SSO
Platform
Usage‑based
  • Autoscale + metrics
  • Secure runner tokens
  • WAF/VPC peering

FAQ

Quick answers to the most common questions.

Is this only for AI training?
No. Charter is a general‑purpose execution layer for AI training & inference, data/ETL, media and simulation/compute jobs.
Do I need identical GPUs?
No. Lease sizes are flexible (tokens/items/frames/time) so heterogeneous nodes contribute proportionally.
What if a runner dies mid‑lease?
The lease expires, work returns to the queue, and another runner continues from the last async checkpoint.
How does this relate to k8s?
Use Charter within or alongside Kubernetes. Charter focuses on progress guarantees, idempotency and resilience rather than pod lifecycle plumbing.

Unify every GPU/CPU you can reach into one reliable fabric

Join the early access list for Charter. We’ll reach out with setup instructions and the hosted coordinator beta.

No credit cardCancel anytimeBuilt for spot & edge