Nodus

Capacity disappears. Your workload should not.

Nodus is the infrastructure execution layer for AI workloads. Describe the work and the budget. We choose the infrastructure, price the finish instead of the hour, and carry progress across a reclaim.

Read the docs Open the console
Nodus selects the best available route for an AI workload 70B FINE-TUNE data · deadline · budget NODUS CANDIDATE B200 cluster · higher cost SELECTED H100 capacity · best fit RECOVERY PATH A100 capacity · ready
Routing is not choosing the cheapest machine. It is choosing the best path to a completed workload.

For teams running training, fine-tunes, evals, and batch inference

Start with one job you cannot afford to lose.

You already know which one. Long enough to get reclaimed, expensive enough that restarting hurts. Submit it as requirements, not as a preselected SKU.

  1. 01

    Describe the work

    Model, data, expected runtime, deadline, budget, and what happens if an instance disappears.

  2. 02

    Get a routing plan

    We evaluate fit, live capacity, total cost, and available recovery paths before the work starts.

  3. 03

    Run toward completion

    Nodus keeps making infrastructure decisions while the workload runs, not only at launch.

Built by engineers who did this at
  • Citadel
  • Google
  • Snowflake
  • Scale AI
  • Amazon
  • Uber

Product

Request in. Completion out.

Click a region. A workload is profiled, routed across live offers, leased, and run to completion. When capacity is reclaimed it is fenced and resumed, not restarted.

Click a region on the graph · or step through

A workload is profiled into an envelope, routed across live offers, leased, run to completion, and resumed through reclaim Workloads Envelopes Offers Bids Finish Inf Batch 70B Eval Agent 24GB 16GB 80GB 8GB vCPU A100 B200 H100 H100 vCPU b1 b2 b3 b4 b5 Done fence

01 · Profile

Compile intent into an execution envelope

Nodus does not take a machine order. The profiler compiles the workload into a versioned execution envelope: footprint, parallelism shape, deadline slack, residency, restartability. The envelope carries a continuity mode that decides what recovery may do later, and a confidence value routing scores rather than hides.

In
Workload revision and inputs
Out
Versioned execution_envelope: resources, topology, locality, continuity mode, confidence
Durable ack
Envelope row and outbox committed in one transaction

02 · Routing

Score the path that finishes

The router is pure. It never calls a provider. It scores the envelope against a freshness-bounded snapshot of normalized offers and supplier health, maximizing the probability of finishing on time and inside budget. Policy violations are rejects, not soft penalties. Capacity the work does not need is a routing failure even when it is available and cheap.

In
Envelope, offer and health snapshots, remaining budget and deadline
Out
Primary route, fallback graph, bid_ceiling, and the score breakdown behind them
Durable ack
Decision and its explanation persisted in the same transaction

03 · Bidding + Lease

Turn the decision into a held lease

Quote, bid, lease. The ceiling is remaining budget minus non-compute cost, recovery reserve, and fallback margin, so a reservation can never spend the money the workload needs to finish. The attempt is written with its idempotency key before the supplier is called. Capacity counts as held only once the lease ID and terms validate. A commitment layer, not a trading product.

In
Route plan and its bid_ceiling
Out
lease.status=reserved with a validated supplier lease ID and accepted terms
Durable ack
reservation_attempt row written before the provider call

04 · Execution + Checkpointing

Run it, and make progress portable

The controller reconciles desired against observed state, so restarts and duplicate messages are safe. A provider reporting “VM created” advances nothing: the run becomes running only after the runner proves its identity and emits a verified heartbeat. Checkpointing runs throughout the job, never after it. Parts upload to object storage, every checksum is verified, an immutable manifest is written, and a compare-and-set advances the latest-manifest pointer. Local disk is not progress.

In
Start request for a generation under the reserved lease
Out
A running generation, plus verified manifests in provider-independent storage
Durable ack
Runner identity proof and heartbeat; commit verifies every SHA-256, then advances the pointer

05 · Recovery

Fence, re-route, resume at generation N+1

Reclaim, heartbeat loss, provisioning failure, and budget risk all enter the same saga. The signed event is stored before anything acts on it, and the old generation is fenced before anything replaces it. At most one generation may write durable output, so a partitioned instance cannot overwrite committed progress. The workload re-enters routing against current offers and resumes as generation N+1 from the latest verified manifest. A marginally cheaper offer is not a reason to disturb a healthy run.

In
Signed reclaim or failure signal, persisted before it is acted on
Out
Generation N fenced; generation N+1 restored under a new lease, or an honest failed outcome carrying the reason
Durable ack
Fencing token in Postgres; the new generation is active only after a verified heartbeat