Skip to content
ArkNet logoArkNet
Prepaid Compute Tokens

Buy token packages.
Metered by kernels.

ArkNet charges tokens for the work your program actually performs — measured directly from kernel execution, not wall-clock time. Start with fixed token packages to access distributed spot GPUs.

View packages Token billing docs
Spot Compute
Per execution
Tokens deducted per logical run
Metering
Kernel-based
Work units, not time limits
Mixed GPUs
Seamless
H100 → 5090 → A100 routing
Entry Level

Starter

For evaluation, testing, and early package pulls.

$10one-time
  • 25,000 Compute Tokens
  • Kernel-metered access
  • Standard Spot GPU routing
  • Dashboard usage tracking
Buy Starter
Most popular

Builder

For active development and recurring execution usage.

$50one-time
  • 150,000 Compute Tokens
  • Access to all GPU architectures
  • Priority network routing
  • Detailed execution receipts
  • Set max-burn limits
Buy Builder
Volume

Scale

For sustained workloads and team-level usage.

$150one-time
  • 500,000 Compute Tokens
  • Lowest cost per execution
  • Dedicated support channel
  • Advanced rate limits
  • Custom webhook integrations
Buy Scale
Token Burn Control

Hard caps per execution, enforced before work begins.

Each execution declares a maximum token burn. ArkNet pre-authorizes those tokens from your balance, runs the execution across the required GPUs, then deducts only the final kernel-metered cost. Unused tokens are instantly returned.

Set a max token limit
Your runner sets a hard token cap per execution. If your balance is too low, the execution never starts.
$ARKNET_MAX_TOKEN_BURN=2500
Status: 2500 tokens pre-authorized
Fail closed
Insufficient tokens → no execution → no surprise errors mid-run.
Predictable spend
Max token burn is the absolute upper bound.
Kernel-metered deduction
Nodes attest usage as kernel work on specific SKUs. A single execution may include multiple GPU segments; total token cost is the sum of metered work.
01
Pre-authorize
Reserve max tokens for the execution.
02
Execute
Run kernels on one or more GPUs (segments).
03
Receipt → Deduct
Deduct kernel-metered usage, un-reserve remainder.
Example Token Receipt
H100
1,240 Tokens
Attention + matmul
RTX 5090
610 Tokens
Post-processing
Tokens Deducted1,850 Total

A quick comparison

Legacy clouds bill by the hour. ArkNet consumes prepaid tokens per execution using kernel-metered work. “Eqv/hr” helps you sanity-check token conversion against legacy providers.

Cost Comparison
ArkNet Spot deducts tokens per execution based on generated receipts.
ProviderH100 (80GB)A100 (80GB)RTX 4090 (24GB)Billing model
ArkNet SpotToken-based
~$1.85 eqv / hr~$1.10 eqv / hr~$0.25 eqv / hrPer execution (Tokens)
AWS (p5.48xlarge)
$4.10 / hr$3.67 / hr—Per hour (Fiat)
Google Cloud (A3)
$3.95 / hr$3.40 / hr—Per minute (Fiat)
Lambda Labs
$2.49 / hr$1.29 / hr—Per hour (Fiat)

* “Eqv/hr” is a convenience projection showing token consumption rate on recent workloads. You are strictly billed tokens per execution.

Frequently Asked Questions

Straight answers on token billing, kernel metering, and spot routing.

Ready to deploy?

Grab a token package, set your max burn constraints, and scale effortlessly across the distributed spot network.

Buy tokens