Byte GooseByte Goose ComputeEarly access

More AI from
every GPU.

Rent GPU compute, lease out idle capacity, or lock a fixed-rate reserved cluster. One marketplace for training, agents and inference.

Rent a compute Compute deals Lease your GPUs

Featured deal

Reserved

32-node B300

256 GPUs · from March 2027

$4.45
/GPU-hr (3 yr)
288 GB
HBM3e / GPU
800 Gb/s
IB / GPU

Compute

One place to run your AI workloads.

Bring your code for training, agents or inference. Byte Goose places the work on compatible GPUs and handles execution.

01

Training & RL

Train and fine-tune models. Checkpoints persist so jobs resume after preemption or a node swap.

02

Agent sandboxes

A sandbox for every agent: run code, manage files, isolate execution per task.

03

Inference

Serve models through an API. Send requests, stream tokens, scale replicas with traffic.

BYO compute + liquid capacity

Run your work. Share your idle compute.

Bring your own machines: your jobs run there first. When GPUs sit idle, lend eligible capacity to other teams through the marketplace and earn credits or payouts.

  • Your workloads always take priority
  • Idle windows are filled with preemptible marketplace jobs
  • Per-GPU earnings and utilization dashboards
Your workloads Marketplace jobs Idle
GPU 0
GPU 1
GPU 2
GPU 3

Illustrative schedule: queued jobs fill free slots; existing workloads stay in place.

Marketplace

Two sides, one exchange.

Owners lease GPU capacity. Teams rent it. Byte Goose verifies supply, matches demand and manages the contract.

For teams renting GPUs

  1. 1Describe the GPUs, term and workload you need.
  2. 2We match you with verified providers and confirm capacity.
  3. 3Accept a fixed quote; pay by wire or card after confirmation.
  4. 4Get SSH, Kubernetes or managed job access on your start date.

For GPU owners

  1. 1Submit your hardware: GPU model, nodes, fabric, location.
  2. 2Byte Goose verifies the hardware, networking and security posture.
  3. 3Publish a listing with term rates and quarterly capacity.
  4. 4Get matched with renters and paid per confirmed lease.

Request lifecycle

SubmittedReviewingMatchedQuotedAcceptedScheduledActive

Initial cluster

  • 32 nodes · 256 NVIDIA B300 GPUs
  • 4 racks
  • Quarter starts: Mar · Jun · Sep · Dec

Per node

  • 8 GPUs · 288 GB HBM3e each
  • 128+ CPU cores · 2 TB DDR5
  • 30.72 TB local NVMe · direct liquid cooling

Fabric

  • NVLink 5 · 1.8 TB/s per GPU
  • XDR InfiniBand · 800 Gb/s per GPU
  • 1:1 non-blocking with SHARP

GPU compute deals

Configure a reserved cluster.

Fixed USD per GPU-hour. Longer terms and prepayment lower the rate. One node = 8 GPUs.

Contract term

Prepayment
Nodes4 / 32 · 32 GPUs

1 square = 1 node = 8 GPUs

Your estimate

3 years · Mar 2027 – Feb 2030

$3,745,690

Estimated term total

GPU-hour rate
$4.45
Upfront (25%)
$936,422
Remaining balance
$2,809,267
Monthly equivalent
$104,047
GPU-hours
841,728
Calendar hours
26,304
Browse all compute deals

Reference rates, exact calendar hours, excluding tax. Monthly equivalent is a comparison, not a billing schedule. A request is not a reservation.

Submit a request

Rent, lease or reserve.

Every request lands in the Byte Goose compute queue. We match renters with verified providers, confirm capacity, then send a quote. Nothing is charged at submission.

We only use your contact details to respond to this request.

Under the hood

How the marketplace works.

The technical model behind deals, listings, rentals and leases.

Placement

Jobs declare GPU type, count and memory. The scheduler bin-packs them onto compatible nodes, preferring NVLink-local placement for multi-GPU work.

Checkpoint & resume

Training state is snapshotted to persistent storage, so a job can move between nodes or providers and continue where it stopped.

Isolation

Each tenant runs in its own container or VM with dedicated GPUs, encrypted volumes and network policies between tenants.

Metering

Usage is metered per GPU-hour from node telemetry. Reserved terms bill the fixed rate; on-demand bills actual hours.

Fabric aware

Listings record interconnect (NVLink, InfiniBand, RoCE) so distributed training is only matched to non-blocking fabrics.

Verified supply

Providers pass hardware burn-in, bandwidth tests and KYC before listings go live. Capacity is held, not oversold.

Data model

gpu_providersOwners leasing capacity, with verification status.
gpu_listingsClusters or node pools: GPU model, nodes, fabric, region, availability.
gpu_listing_pricesDeal grid: USD per GPU-hour by term and prepayment.
gpu_capacity_blocksPer-quarter node inventory with held and reserved counts.
compute_requestsRent, lease and deal requests from the form below.
compute_request_eventsAudit log of each request status change.
gpu_leasesConfirmed contracts linking a request to a listing.

Byte Goose Compute is in early access. Requests are not reservations; capacity and pricing are confirmed before any order. Reference cluster specs and rates are approximate industry estimates.