⬡viabandwidthGPU Compute
Beyond the GPU-Hour 2026 GPU cloud pricing buyer's guide cover
$3.99
Runpod H100 SXM list rate
$4.29
Lambda 1x H100 SXM list rate
$917.70
Illustrative 230-hour compute bill
15%
Idle-time uplift in that scenario

The purchasing question

The cheapest GPU-hour is not necessarily the cheapest completed workload. A headline rate excludes the amount of billable time a job consumes and may exclude persistent storage, data movement, support, minimum commitments and tax.

01 · Cost metricStart with the completed workload

Record three figures separately: the provider's native quoted rate, the billable consumption under its start-and-stop rules, and the successful output produced at the required quality and deadline.

From list rate to useful compute

List rate$3.99per H100 SXM GPU-hr
Billed time230h200h productive + 30h overhead
Compute bill$917.70230h × $3.99
Effective rate$4.59per productive GPU-hour
Illustration only. Excludes storage, transfer, taxes and provider-specific billing rules.
Completion-cost equation. Total workload cost = GPU or instance charges + mandatory storage + data movement + managed-service and support fees + minimum commitments + applicable tax. Effective cost per completed workload = total workload cost ÷ successful completed jobs.

At Runpod's published $3.99 H100 SXM Pods rate, 230 billed hours for 200 productive hours produces a compute-only bill of $917.70, or $4.59 per productive GPU-hour. This is transparent arithmetic, not an observed Runpod workload or a claim about its idle behavior.

02 · Commercial structureWhat exactly are you buying?

The service boundary changes the meaning of every price.

01ON-DEMANDInstance or GPU while activeBest for bursts · watch idle runtime
02RESERVEDCapacity for a term or minimumBest for steady load · watch unused hours
03SPOTInterruptible computeBest for restartable jobs · watch retries
04MANAGEDAPI, endpoint or throughputBest for served output · watch token and idle fees

A 1-GPU instance, an eight-GPU HGX node, a reserved cluster and a serverless model endpoint are different products. Dividing a node price by eight is useful only when all eight GPUs, the interconnect, CPU/RAM, storage and minimum purchase are recorded.

03 · Pricing evidencePublished rates: four GPU families

Official published price observations

H100 SXM 80 GB
Runpod Pods / 1 GPU$3.99
Lambda / 1 GPU$4.29
H200 141 GB
Runpod Pods / 1 GPU$5.29
Crusoe HGX$4.29
B200 180 GB
Runpod Pods / 1 GPU$7.99
Lambda SXM6 / 1 GPU$6.99
A100 SXM 80 GB
Runpod Pods / 1 GPU$1.79
Crusoe SXM$2.30
USD per GPU-hour as displayed on provider pages. Not a matched market ranking.
Provider and sold objectGPUNative list rate
Runpod PodsH100 SXM 80 GB$3.99/GPU-hour
Lambda 1x instanceH100 SXM 80 GB$4.29/GPU-hour
Runpod PodsH200 141 GB$5.29/GPU-hour
Crusoe GPU instanceH200 141 GB HGX$4.29/GPU-hour
Lambda 1x instanceB200 SXM6 180 GB$6.99/GPU-hour
Runpod PodsB200 180 GB$7.99/GPU-hour
Runpod PodsA100 SXM 80 GB$1.79/GPU-hour
Crusoe GPU instanceA100 SXM 80 GB$2.30/GPU-hour

Rates rechecked on 2 November 2026. Provider selectors, region, availability, attached resources and tax can change the transactable configuration. Only the two 1-GPU H100 SXM examples form the worked comparison below; the remaining rows are deliberately directional.

04 · Total billWhat the actual invoice includes

Unknown must remain unknown. It must not silently become zero.

The quote has five cost boundaries

01COMPUTEGPU / VM / node
02DATAstorage / checkpoints
03MOVEMENTegress / cross-region
04SERVICEsupport / platform
05CONTRACTminimums / tax / exit
Missing in the rate card does not mean free in the contract.
  • 01Compute: when does the meter begin, pause and stop?
  • 02Storage: what persists after the instance terminates, and at what rate?
  • 03Movement: which ingress, egress, cross-zone and interconnect rules apply?
  • 04Service: which platform, scheduler and support charges are bundled?
  • 05Contract: what remains payable if utilization is zero?

Lambda's billing documentation says running on-demand instances are billed even when unused and that filesystems continue billing while they exist. It also states that filesystem ingress and egress are not charged.

05 · Worked buyWhen the cheaper GPU-hour loses

If both H100 offers require 200 billable hours, Runpod's listed rate produces $798 in compute charges and Lambda's produces $858. But at Lambda's $4.29 rate, the break-even runtime is just under 186.01 hours for the same accepted workload.

When a higher rate can win

RUNPOD200 h$3.99/h$798.00
LAMBDA186 h$4.29/h$797.94
Same 1-GPU H100 SXM. Equal quality, no extra fees. Times are scenarios, not measured speeds.
TestRunpod exampleLambda example
Same billed time200 h × $3.99 = $798200 h × $4.29 = $858
Break-even runtime200 h at agreed qualityLess than 186.01 h at the same quality
Unpriced obligationsStorage, network and tax: unknownStorage, network and tax: unknown

No benchmark supports the 186-hour premise. It is a break-even illustration, not a provider performance claim. Buyers must measure or contract the completed result.

06 · Buyer worksheetMake two quotations comparable

The downloadable eight-page field guide includes a two-supplier worksheet covering GPU variant and count, sold object, topology, host resources, region, commitment, billing increments, storage, data movement, support, tax, measured billable hours and all-in completed-workload cost.

Required fieldSupplier ASupplier B
Seller / actual infrastructure operator________________________
Official quote ID / issue & expiry dates________________________
GPU model, variant, VRAM, count / MIG________________________
Sold object: GPU / VM / node / endpoint________________________
Topology, network fabric, CPU/RAM________________________
Included / persistent storage________________________
Region, residency, start date / SLA________________________
Native price, currency, unit, tax status________________________
On-demand / spot / reserved; duration________________________
Billing increment, idle, pause, retry rules________________________
Minimums, deposits, exit, renewal________________________
Egress / cross-zone / data-movement fees________________________
Support, platform and extra service charges________________________
Proven billable hours / accepted output________________________
All-in cost / completed workload________________________
Qualification gate. Require model and memory variant, physical GPU count or partition, sold object, CPU/RAM/storage and fabric, region, term, tax treatment, billing granularity and dated official source. If an essential field is missing, label the comparison directional or not comparable.

07 · Findings and methodWhat the evidence can support

A large ledger is not a single price series

872candidate domains
339explicit GPU providers
4,511heterogeneous rates
60matched offers

This is a buyer guide, not a market census or benchmark. It uses official public list rates and illustrative arithmetic. It does not claim tested throughput, current stock, identical bundles, negotiated pricing or supplier-specific idle behavior.

Historical context only: ViaBandwidth's private 60-offer Q2-to-October same-offer cohort included 41 unchanged prices, 15 increases and four decreases across 10 provider domains. That narrow cohort cannot establish market-wide price inflation.

viabandwidth is the verified directory of GPU compute, colocation and carrier networks.

Public list prices are dated observations, not signed quotations or evidence of available stock. Worked runtimes are illustrative arithmetic, not measured supplier performance. Buyers must confirm the complete configuration, billing rules, region, availability, tax and contract terms before purchase.

© 2026 viabandwidth (Steven Higashi). All rights reserved. The figures and charts in this paper may be quoted with clear attribution to viabandwidth.

Cite as: ViaBandwidth, Beyond the GPU-Hour: The 2026 GPU Cloud Pricing Buyer's Guide, viabandwidth.com/gpu/guides/gpu-cloud-pricing-comparison.

viabandwidth.com · [email protected]