Affiliate disclosure: HostFleet may earn a commission if you sign up through links on this page. That never changes the analysis. Read the live HostFleet about page for methodology and affiliate-policy context.

Source-backed rate-card and lifecycle comparison; estimated monthly totals. This refresh uses official provider pricing pages, public vendor APIs, and HostFleet’s live GPU pricing dataset. It is not a capacity, latency, throughput, reliability, or current-stock benchmark. Fifteen selected H100 price anchors retain their September 28, 2026 provider checks. Nebius’s H100 rate and October 1 effective date were rechecked on October 1, 2026 against two official pricing surfaces. Verda’s selected H100 rate was refreshed on October 2, 2026 against its official pricing page and public instance API. Jarvis Labs’ one-GPU H100 SXM on-demand row was checked on October 11, 2026 against its official public pricing table. Massed Compute’s one-GPU H100 (80GB) and H100 SXM5 (80GB) catalog rows were checked on October 12, 2026. This is a set of targeted provider checks, not a new full-market or invoice verification. TensorDock’s two stop modes, storage-only release state, and restart constraints were checked September 28. Verda’s prepaid billing terms were rechecked September 25. Lambda’s termination, guest-poweroff, and filesystem rules were checked September 14. HostFleet did not measure invoices, refund timing, capacity, or state-transition timing.

Price anchors rechecked: Massed Compute October 12, 2026; Jarvis Labs October 11, 2026; Verda October 2, 2026; Nebius October 1, 2026; remaining 15 selected rows September 28, 2026
Full-dataset baseline: September 24, 2026; targeted Verda and Nebius patches October 1, Verda refresh October 2, and Jarvis Labs patch October 11
TensorDock lifecycle checked: September 28, 2026; no current fixed H100 rate was added
Currency: public USD on-demand list prices before tax
Comparison unit: one listed GPU-hour or a published per-second/per-minute equivalent
Boundary: a public rate does not prove inventory, quota, regional access, approval, or equal performance

H100 rental price per hour in 2026: 19 rates and variant costs

The cheapest selected H100 rate is now a tie: Hyperstack and Koyeb both publish $2.50 per hour. That does not make them interchangeable. Hyperstack’s row is a one-GPU PCIe VM whose stopped state remains billable. Koyeb’s row is a serverless application instance with public-preview scale-to-zero and a documented five-minute default idle period.

Massed Compute follows the $2.50 tie at $2.73/hour for its base H100 (80GB) VM. Its separately listed one-GPU H100 SXM5 VM is $3.14/hour; the $0.41/hour gap is a different catalog configuration, not a price change to the base row. Northflank lists $2.74/GPU-hour, and RunPod Secure Cloud lists $2.89/hour. Jarvis Labs is no longer near the price floor: its one-GPU H100 SXM on-demand row rose from $2.69 to $3.49/GPU-hour by October 11, a 29.7% increase. Northflank’s number is only a GPU component; CPU and memory are extra. These rows package different resources and lifecycle controls.

The September 28 recheck found 18 selected H100 rates unchanged from September 24. Novita’s official marketplace API still exposes product H100-80GB.22c150g with 22 vCPU, 150 GB RAM, a 60 GB container-disk quota, and $3.39/hour. Verda changed repeatedly: HostFleet observed its exact H100 offer at $3.282/hour on September 17, $3.416 on September 24, $3.555 on September 28, $3.627 on October 1, and $3.663 on October 2. The latest October 1–2 step adds $25.92 to a 720-hour estimate; the cumulative September 17–October 2 increase is 11.6%, or $274.32 for 720 hours.

Nebius’s previously scheduled increase is now effective. Its unified H100 SXM/NVLink VM moved from $3.85 to $4.50/GPU-hour on October 1, a 16.9% increase. That adds $468.00 to a 720-hour planning case and moves Nebius behind Modal and DigitalOcean in this rate-only ordering, into a tie with Fal’s $4.50/hour list rate.

The lifecycle still matters more than either ranking change. Verda prepays pay-as-you-go compute in 10-minute increments and returns unused terminated time in the next billing period; shutdown does not release the compute charge. Lambda has a different trap: guest shutdown or poweroff puts the instance into Alert and billing continues; only Lambda’s termination operation ends the compute meter. TensorDock makes the distinction explicit in its API: a normal stop keeps the GPU reserved at the running rate, while stop-and-release drops the VM to storage-only billing and makes the next start capacity-dependent.

Current H100 price-per-hour comparison

The table is sorted by normalized hourly rate and shows one selected product surface per provider. Massed Compute’s row is its base H100 (80GB) VM, not its separate H100 SXM5 SKU; the variant comparison below prices both. Per-second prices are multiplied by 3,600 and per-minute prices by 60. Product shape remains visible because a PCIe VM, an SXM machine, a serverless instance, and a managed inference deployment are not equivalent purchases.

Provider and productH100 scopePublic list rateOfficial evidence and check date
Hyperstack1x H100 80 GB PCIe VM; 28 CPU, 180 GB RAM, local storage; Canada$2.50/GPU-hrHyperstack pricing, Sept. 28, 2026
Koyeb1x H100 80 GB serverless instance; 15 vCPU, 180 GB RAM, 320 GB disk$2.50/hrKoyeb pricing, Sept. 28, 2026
Massed Compute1x base H100 (80GB) on-demand VM; 20 vCPU, 128 GB RAM; storage shown as 1,250 with no unit$2.73/hrMassed Compute pricing, Oct. 12, 2026
Northflank1x H100 80 GB managed-cloud GPU component; CPU, memory, disk, and egress separate$2.74/GPU-hrNorthflank pricing, Sept. 28, 2026
RunPod PodsH100 PCIe Secure Cloud Pod$2.89/hrRunPod pricing, Sept. 28, 2026
Thunder Compute1x H100 80 GB PCIe base; 4 vCPU, 32 GB RAM, 100 GB persistent disk included$3.20/GPU-hrThunder pricing, Sept. 28, 2026
Lambda Cloud1x H100 PCIe VM$3.29/GPU-hrLambda GPU instances, Sept. 28, 2026
Novita AI instance1x H100 80 GB SXM; live API product H100-80GB.22c150g lists 22 vCPU, 150 GB RAM, 60 GB container-disk quota$3.39/GPU-hrNovita marketplace API, Sept. 28, 2026
Jarvis Labs1x H100 80 GB SXM on-demand instance; public row lists 16 vCPU and 200 GB RAM$3.49/GPU-hrJarvis Labs pricing, Oct. 11, 2026
Verda GPU instance1x H100 80 GB SXM5; selected configuration includes 30 CPU and 120 GB RAM; storage separate$3.663/hrVerda pricing and instance catalog, Oct. 2, 2026
ModalH100 allocated to a serverless container$0.001097/sec ($3.9492/hr)Modal pricing, Sept. 28, 2026
DigitalOcean GPU Droplet1x HGX H100; 20 vCPU, 240 GiB RAM, 720 GiB boot, 5 TiB scratch, 15,000 GiB transfer$4.41/GPU-hrDigitalOcean GPU pricing, Sept. 28, 2026
Nebius AI Cloud1x H100 SXM/NVLink VM; 16 vCPU, 200 GB RAM; eu-north1$4.50/GPU-hrNebius pricing and Compute pricing, Oct. 1, 2026
Fal custom deploymentH100 80 GB on-demand custom deployment; excludes the separately advertised “as low as” rate$4.50/hr listFal pricing, Sept. 28, 2026
RunPod ServerlessH100 PRO Serverless Flex worker tier; not an exact-card reservation$4.79/hrRunPod pricing, Sept. 28, 2026
Replicate private deploymentH100 managed model deployment$0.001525/sec ($5.49/hr)Replicate pricing, Sept. 28, 2026
Paperspace Machine1x H100 80 GB SXM5; 20 vCPU, 250 GB RAM, 50 GB SSD; approval may be required$5.95/hrPaperspace pricing, Sept. 28, 2026
CoreWeave InferenceSingle-GPU inference rate for inference-platform customers$6.16/GPU-hrCoreWeave pricing, Sept. 28, 2026
Baseten deploymentManaged H100 deployment; 80 GiB VRAM$0.10833/min ($6.50/hr)Baseten pricing, Sept. 28, 2026

The selected range remains $2.50 to $6.50 per listed GPU-hour across 19 product surfaces. The low end includes a serverless instance and an allocated VM. The high end includes managed inference. Treating the range as a performance ranking would be benchmark theater.

What the lowest-cost newer rows change

Koyeb ties for the lowest rate, with a documented idle tail

Koyeb’s official pricing page and instance reference both publish $2.50/hour for one H100 instance, checked September 28, 2026. The pricing card includes 15 vCPU, 180 GB RAM, and 320 GB disk. GPU availability is region-specific; the public rate is not a stock guarantee.

The operating model changes the cost calculation. Koyeb’s scale-to-zero documentation explicitly includes GPU instances and labels the feature public preview. A GPU service can set its minimum to zero. The default idle period is five minutes, and the service wakes on a supported inbound request.

At the current H100 rate, the nominal default post-traffic idle tail is an estimate of:

$2.50/hour × 5/60 hour = $0.2083

That is arithmetic from sourced inputs, not an invoice measurement. It excludes startup, model loading, active processing, storage, and any gradual multi-instance scale-down. A held internet connection prevents the service from becoming idle. HTTP/2 requests cannot wake a sleeping service; WebSocket has a separate limited wake path. Koyeb also counts the configured autoscaling maximum against organization quota even while the service is at zero.

The honest conclusion is narrower than “serverless means free when idle.” Koyeb can scale a GPU service to zero, but protocol choice, idle qualification, wake behavior, quota, and public-preview status all belong in the deployment design.

Massed Compute: the $2.73 H100 and $3.14 SXM5 are different VMs

Massed Compute’s official VM pricing catalog, checked October 12, 2026, lists a one-GPU H100 (80GB) VM at $2.73/hour with 20 vCPU and 128 GB RAM. It separately lists a one-GPU H100 SXM5 (80GB) VM at $3.14/hour with 14 vCPU and 122 GB RAM. Both appear in the public catalog, but that does not establish live stock or account eligibility. The catalog shows different storage quantities without an explicit unit in those rows, so this comparison does not relabel them as GB or treat them as equal storage bundles.

One-GPU catalog SKUPublished VM rate24 active hours720 active hours
H100 (80GB), the row in the 19-provider ranking$2.73/hr$65.52$1,965.60
H100 SXM5 (80GB), separate variant$3.14/hr$75.36$2,260.80
SXM5 minus base H100$0.41/hr$9.84$295.20

Estimate assumptions: one named VM is active for every modeled hour at its October 12 public USD list rate; 24 hours means one continuous day and 720 hours means 30 continuous days. The differences are rate subtraction and multiplication, not measured bills or throughput. They exclude tax, discounts, account-specific terms, and any separately charged resource. The Massed Compute billing overview, checked in HostFleet’s October 8 provider note, says it sums active-VM hourly rates, divides by 60, and debits the account each minute with micro-cent remainders carried forward. That documentation does not establish when initialization first becomes billable or how a partial minute at termination settles.

Buying decision: use the $2.73 row only for the base H100 (80GB) VM. If the workload requires the cataloged SXM5 variant, budget its $3.14 rate and verify the exact resource bundle, capacity, and measured application throughput before comparing value. The public catalog does not justify calling the base H100 PCIe, claiming the SXM5 variant is faster for this workload, or substituting one SKU’s rate for the other.

Jarvis Labs: the new $3.49 H100 rate changes the buying order

Jarvis Labs publishes $3.49/GPU-hour for a one-GPU H100 SXM on-demand instance, checked October 11, 2026 on its official pricing page. Its public row lists 16 vCPU and 200 GB RAM, and labels the rate on-demand rather than Spot or a term reservation. On-demand compute bills per minute. The former $2.69/hour anchor is no longer the public one-GPU on-demand rate. The increase is $0.80/hour, or 29.7%. At 720 continuously billable hours, the sourced-rate arithmetic moves from $1,936.80 to $2,512.80, a $576.00 difference. This is a cost sensitivity, not a measured invoice, capacity check, or performance comparison. It excludes retained storage, taxes, and any other charges or discounts.

Pausing stops compute billing while preserving data, according to the official SDK documentation. Paused data continues billing at $0.00014/GB-hour, verified from the Jarvis Labs FAQ on August 24, 2026. Pausing or deleting releases GPU capacity, so the same card and region are not guaranteed when the workload resumes.

For example, retaining 100 GB for 160 paused hours is an estimated $2.24:

100 GB × $0.00014/GB-hour × 160 hours = $2.24

The example isolates retained data and assumes no other charge. It does not claim current H100 inventory or a future resume success rate.

Northflank’s $2.74 is not an all-in VM

Northflank publishes an H100 component at $2.74/GPU-hour, checked September 28, 2026. Managed-cloud GPU use bills by the second once provisioned, but every workload also selects a CPU and memory compute plan. Persistent disk and network egress can add more.

The same pricing page lists CPU at $0.01667/vCPU-hour and memory at $0.00833/GB-hour, checked September 28, 2026. Northflank does not publish one model-specific minimum CPU/RAM plan for the H100, so HostFleet does not invent an all-in total. The correct formula is:

all-in compute rate = $2.74 GPU component
                    + selected vCPU × $0.01667/hour
                    + selected memory GB × $0.00833/hour

Disk, egress, tax, and other services sit outside that formula. Northflank’s managed GPU deployment documentation, checked August 31, 2026, also requires at least $50 in account credit before deployment; that is a funding prerequisite, not a quoted minimum charge.

Manual scale-to-zero is documented, but a service at zero instances is unavailable. The autoscaling docs describe configurable minimum and maximum counts, 15-second evaluations, and a five-minute downscale window; they do not establish automatic GPU scale-to-zero or request wake-up. Do not group this row with Koyeb merely because both products use managed application abstractions.

Thunder remains $3.20 after August’s one-cent move

Thunder Compute’s live pricing page and public pricing API agree on $3.20/GPU-hour for the one-H100 PCIe base configuration, checked September 28, 2026. HostFleet’s check of those same official surfaces on August 23, 2026 recorded the previous selected value of $3.19/hour.

The August change remains small but should not be hidden:

($3.20 - $3.19) × 720 hours = $7.20

The current 720-hour compute estimate is $2,304.00, up from $2,296.80 at the former rate. Thunder’s base still includes four vCPUs, 32 GB RAM, and 100 GB persistent disk. Its billing documentation says compute bills per minute while the instance runs and deletion stops instance billing.

The September 28 full-table verification found the other 18 selected H100 rates unchanged from September 24. Verda was the only selected H100 price that moved in that refresh. On October 1, the HostFleet dataset received targeted patches for both Verda at $3.627/hour and Nebius at $4.50/hour. Verda’s live price then moved again to $3.663/hour on October 2; that latest value is sourced directly from Verda and is not presented as a full-dataset refresh.

Verda moved to $3.663, and its 10-minute rule changes cash flow

Verda’s pricing page and public instance catalog both publish $3.663/hour for the selected one-GPU H100 SXM5 instance, checked October 2, 2026. The selected fixed shape includes 30 CPU and 120 GB RAM; storage is separate. HostFleet observed the exact rate at $3.282/hour on September 17, $3.416/hour on September 24, $3.555/hour on September 28, and $3.627/hour in the October 1 dataset patch.

Across the September 17–October 2 movements, the exact increase is:

($3.663 - $3.282) / $3.282 = 11.6%

For 720 continuously billable hours, that changes the compute estimate from $2,363.04 to $2,637.36, a $274.32 increase. The October 1 dataset rate of $3.627 implied $2,611.44 for 720 hours; the October 2 live rate adds $25.92. Lambda’s $3.29 PCIe VM is now $268.56 lower over 720 hours, and Novita’s $3.39 catalog row is $196.56 lower. Those gaps do not override product differences, inventory, region, topology, or lifecycle behavior.

Verda’s official Pricing and Billing documentation, checked September 25, describes pay-as-you-go usage as prepaid in 10-minute increments, with unused time on a terminated instance refunded in the next billing period. At $3.663/hour, one nominal 10-minute block is $0.6105. A planning example that terminates after seven minutes allocates $0.42735 to elapsed compute and $0.18315 to the unused three minutes returned later:

10-minute prepayment = $3.663 / 6 = $0.6105
seven-minute compute = $3.663 × 7/60 = $0.42735
unused three-minute portion = $3.663 × 3/60 = $0.18315

This is arithmetic from the public rate and refund rule, not a measured invoice. It excludes storage, tax, and any other resource. The distinction is cash flow rather than a claimed 10-minute minimum charge: the unused terminated portion is returned later, not necessarily absent from the current billing period.

Verda’s lifecycle documentation says shutdown does not stop compute billing; deletion is the release action, and retained storage continues separately. Cost automation should therefore delete, confirm the instance is gone, and reconcile the later refund instead of treating an in-guest shutdown as completion.

TensorDock: stop can mean full-rate or storage-only

TensorDock is intentionally not a twentieth row in the ranked H100 table. Its marketplace hosts set separate GPU, CPU, RAM, and storage prices, and the public cloud-GPU page still labels its example rates as last updated July 24, 2024. That is useful product documentation but not a current, reproducible H100 list price. Adding an old or inventory-dependent number would make the ranking look more complete than the evidence allows.

The lifecycle documentation is current enough to expose a cost boundary that matters on any TensorDock GPU:

Provider actionDocumented compute treatmentWhat remains
Stop without disassociate_resources=trueGPU remains reserved; TensorDock says the VM bills at the same rate as when runningCompute and storage remain allocated
Stop with disassociate_resources=trueGPU is released; the VM enters a storage-only stateNetwork storage continues billing
Restart a released Core Compute VMCapacity-dependent; it can return on another server in the same locationSame-location GPU availability is required
Restart a Distributed Compute VMTied to the original serverRestart depends on that host and resource availability

These are sourced rules checked September 28, not measured invoice results. TensorDock’s cloud GPU page says host-priced GPU, CPU, RAM, and storage are billed per second. Its Core Compute introduction describes network storage and same-location restart behavior. The public API collection documents the disassociate_resources stop parameter plus account balance and aggregate hourly_cost fields.

One API caveat matters for automation. TensorDock’s own StoppedDisassociated response example retains static compute_price and total_price quote fields. Those fields describe the VM quote; they do not prove that compute is actively metered after release. A billing test should reconcile state against an otherwise idle account’s aggregate hourly_cost and balance instead of reading one static VM field as an invoice.

The practical rule is explicit: use stop-and-release when the goal is to end GPU billing, then verify storage-only account spend. That saves compute but gives up guaranteed possession of the GPU. A normal stop is the reservation-preserving choice and should be budgeted at the full running rate.

Lambda: guest poweroff keeps billing, while terminate destroys local data

Lambda publishes $3.29/GPU-hour for its one-GPU H100 PCIe instance, rechecked September 28, 2026. The rate is expressed hourly, but Lambda says On-Demand Cloud usage is billed in one-minute increments. Billing begins when the instance launches and passes health checks and ends when the instance is terminated. The public documentation does not unambiguously define partial-minute rounding or whether pre-health-check launch time can later appear on a bill, so this guide does not invent those details.

The dangerous distinction is between an operating-system power command and a provider termination:

  • Lambda currently documents launch, restart, and terminate actions; there is no pause or suspend state.
  • sudo shutdown -h now and sudo systemctl poweroff do not terminate or suspend the instance. Lambda says they put it into Alert, and billing continues.
  • The cost-control action is termination through Lambda’s console or Cloud API. Automation should confirm that the instance disappears from the running-instance list and alert when it remains in Alert or another unexpected state.

For the same eight-useful-hours-plus-160-unattended-hours scenario used elsewhere in this guide, the compute-only exposure is:

useful compute = $3.29 × 8 hours = $26.32
unattended compute = $3.29 × 160 hours = $526.40
one-week compute = $552.72

Estimate assumptions: one H100 PCIe instance remains billable for 168 hours at the September 28 list rate, with no storage, tax, network, or other resource charge. This is planning arithmetic, not a measured Lambda invoice. Correct termination after eight hours limits the compute portion to $26.32; guest poweroff does not.

Termination has a data consequence. Lambda says all local, non-filesystem data is irrecoverably destroyed when an instance is terminated. Data that must survive needs to be copied before termination to an attached Lambda filesystem or another durable store.

A Lambda filesystem has its own lifecycle:

  • it survives instance termination and continues billing for as long as it exists, even while unmounted;
  • billing is per GiB used per month in one-hour increments, with no minimum storage period and no ingress or egress charge;
  • the public billing page’s $0.20/GiB-month figure is explicitly an example that might not reflect current pricing, so HostFleet does not use it as a current rate; the actual price appears during authenticated filesystem creation;
  • it must be selected when the instance launches, must share the instance’s workspace and region, and cannot be attached to an already running instance; and
  • it cannot be deleted until attached instances are terminated and the filesystem is detached. Lambda exposes no user-set usage quota, and hidden .Trash-* data can remain billable until permanently deleted.

The safe automation sequence is therefore copy, verify, terminate, confirm detachment, inspect, delete. Treat compute release and persistent-storage cleanup as separate operations.

Nebius: use the cloud stop control, not Linux shutdown

Nebius now publishes a unified $4.50/GPU-hour price for its one-H100 SXM/NVLink 1gpu-16vcpu-200gb VM in eu-north1, rechecked October 1, 2026. Both its pricing overview and detailed Compute pricing documentation show the former $3.85/GPU-hour rate and the new rate effective October 1. The current price includes the prescribed 16 vCPU and 200 GB of RAM. Persistent disks and other retained resources are separate.

The increase is $0.65/GPU-hour, or 16.9%:

($4.50 - $3.85) / $3.85 = 16.9%
720-hour increase = $0.65 × 720 = $468.00
730-hour increase = $0.65 × 730 = $474.50

These are arithmetic planning values, not measured invoices. The public rate does not prove live stock, quota, regional eligibility, provisioning success, performance, or an SLA.

The provider’s lifecycle documentation draws a sharp billing boundary:

  • GPU, vCPU, and RAM accrue only while the VM is Running, in one-second billing units. A Stopped VM has no compute charge.
  • A normal provider stop can spend up to 60 seconds in graceful termination. The public documentation does not identify whether the stop-side billing cutoff is the stop command, entry into Stopping, or arrival at Stopped; HostFleet does not invent that timestamp.
  • Linux shutdown and halt inside the guest are not cost controls. Nebius treats them as VM failure, automatically reboots the instance, and continues charging. Use the console, SDK, or nebius compute instance stop.
  • Deleting stops compute billing when the delete command is sent. VM-managed disks are deleted with the VM; standalone disks persist and keep billing.
  • Stopping keeps VM, GPU, CPU, RAM, and storage quota occupied. It does not reserve physical restart capacity: a later start can still fail with Not enough resources.
  • Local SSD data is erased on stop or delete. Persistent disks, snapshots, and shared filesystems remain chargeable by allocated size while the VM is stopped.

For the same eight-useful-hours-plus-160-unattended-hours scenario used in HostFleet’s GPU cloud cost calculator, leaving this H100 VM Running produces:

useful compute = $4.50 × 8 hours = $36.00
unattended compute = $4.50 × 160 hours = $720.00
one-week compute = $756.00

A successful provider-level stop reduces the compute part of those remaining 160 hours to zero. It does not make retained storage free. As a deliberately small planning example, 32 GiB of Network SSD retained for 160 hours costs about $0.50 at the current $0.071/GiB per 730 hours rate:

32 GiB × $0.071 × 160/730 = $0.498

The 32 GiB allocation is an exposed assumption, not a claimed H100 boot-disk minimum; the selected image may require more. The resulting scoped total would be about $36.50 for eight hours of H100 compute plus that retained disk, before tax, traffic, snapshots, shared filesystems, or other resources.

The stop-side transition cutoff remains sourced but not measured. A bounded ledger-reconciliation experiment has been designed, but it is account-gated and has not been run. The practical rule does not depend on that missing timestamp: automate the provider’s stop or delete operation, observe the final state, and alert if the VM returns to Running.

What one continuously allocated H100 costs for 30 days

These estimates multiply the October 12 Massed Compute base-H100 rate, the October 11 Jarvis Labs rate, the October 2 Verda rate, the October 1 Nebius rate, and the remaining 15 September 28 sourced rates by 720 hours. Modal, Baseten, and Verda use their unrounded native rates. The estimates assume one named product stays billable continuously. They exclude separate CPU/RAM, storage, IP, network, tax, support, commitments, extra replicas, and operational work. They are not quotes or performance comparisons. TensorDock remains excluded because it does not expose a current fixed H100 rate suitable for this arithmetic.

Product shapeRate used720-hour compute estimate
Hyperstack H100 PCIe VM$2.50/hr$1,800.00
Koyeb H100 serverless instance$2.50/hr$1,800.00
Massed Compute base H100 (80GB) VM$2.73/hr$1,965.60
Northflank H100 component only$2.74/hr$1,972.80 plus CPU and memory
RunPod Secure Cloud H100 PCIe Pod$2.89/hr$2,080.80
Thunder Compute H100 PCIe base$3.20/hr$2,304.00
Lambda H100 PCIe VM$3.29/hr$2,368.80
Novita H100 SXM instance$3.39/hr$2,440.80
Jarvis Labs H100 SXM instance$3.49/hr$2,512.80
Verda H100 SXM5 instance$3.663/hr$2,637.36
Modal H100 container$0.001097/sec$2,843.42
DigitalOcean HGX H100 Droplet$4.41/hr$3,175.20
Nebius H100 SXM/NVLink VM$4.50/hr$3,240.00
Fal H100 custom deployment$4.50/hr list$3,240.00
RunPod Serverless H100 tier$4.79/hr$3,448.80
Replicate private H100 deployment$0.001525/sec$3,952.80
Paperspace H100 SXM5 Machine$5.95/hr$4,284.00
CoreWeave Inference H100$6.16/hr$4,435.20
Baseten managed H100 deployment$0.10833/min$4,679.86

A 720-hour total is a sensitivity case, not a forecast. It is appropriate only when the product remains billable for all 30 days. Use the GPU cloud cost calculator when the real question is useful time versus billable time.

The off switch can reverse the rate ranking

Hourly price matters less when the wrong lifecycle action leaves the GPU meter running.

ProductCompute-billing boundaryImportant residual
HyperstackHibernation deallocates the flavor; a merely stopped VM remains billableDisks, IPs, and volumes can remain billable
KoyebEligible public-preview services can reach zero after the idle policyWake protocol and idle conditions matter; no GPU cold-start number is claimed
Jarvis LabsPause stops compute billingRetained data bills; released GPU capacity may not return
Massed ComputeActive VM total is debited per minuteConfirm the exact stop/delete action before automation
NorthflankGPU bills by the second once provisioned; manual zero makes the service unavailableCPU, memory, disk, and egress are separate
RunPod PodsStop or terminate according to the required persistence modelStorage can continue; container-disk data can be erased
Thunder ComputeDeleting the instance stops instance billingConfirm snapshot and retained-disk handling
VerdaDelete the instance; shutdown does not stop compute billing; pay-as-you-go prepays 10-minute blocks and refunds unused terminated time next periodRetained storage remains separate; refund timing changes cash flow
TensorDockStop-and-release with disassociate_resources=true; a normal stop keeps the GPU reserved at the running rateStorage continues; restart is capacity-dependent and product-mode-dependent
Lambda CloudTerminate through Lambda’s console or Cloud API; guest shutdown/poweroff enters Alert and keeps billingTermination destroys local data; filesystems survive and bill until deleted
NebiusStop through the console, CLI, or SDK; guest shutdown/halt auto-reboots and remains billablePersistent storage bills; local SSD is erased; quota stays occupied; restart capacity is not guaranteed
DigitalOcean GPU DropletsDestroy the Droplet; powering it off does not stop billingReserved resources continue charging while powered off
Paperspace MachinesPower off stops compute billingStorage, public IPs, and add-ons can continue
Modal and RunPod ServerlessWorkers can return to zeroStartup, idle windows, and warm settings determine billable allocation

Lifecycle sources retain their own verification dates in the Sources section. This table does not imply that unlisted products lack cleanup controls; it highlights the boundaries with specific checked documentation.

A generic scheduler that calls “stop” is not portable cost control. Record the exact state transition that releases compute, the data consequence, and the resource that remains chargeable.

Choose product shape before hourly price

Self-managed capacity

Hyperstack, Jarvis Labs, Massed Compute, RunPod Pods, Thunder, Verda, TensorDock, Lambda, Novita, Nebius, DigitalOcean GPU Droplets, and Paperspace Machines are relevant when the buyer wants a VM, instance, or Pod and accepts responsibility for the image, inference server, authentication, rollout, health checks, logs, and cleanup. TensorDock belongs in this product-shape discussion but not the fixed-rate ranking because its public marketplace pricing is host-variable and its public example rates are stale.

Compare PCIe with SXM/NVLink/HGX, not just “H100.” Check included CPU, RAM, local and persistent storage, region, account approval, and release behavior. A low rate is unusable when the available topology or product shape does not fit.

Serverless and managed application instances

Koyeb, Modal, RunPod Serverless, and Northflank expose different application-level deployment models. Koyeb documents GPU scale-to-zero in public preview. Modal and RunPod publish serverless GPU rates with their own lifecycle controls. Northflank publishes a GPU component and does not establish automatic request-waking from zero.

The serverless GPU pricing matrix is the better companion when scaling policy matters more than allocated capacity. Use the Modal pricing guide and RunPod pricing guide for provider-specific billing boundaries.

Managed inference deployments

Fal, Replicate, CoreWeave Inference, and Baseten offer more opinionated inference surfaces. Higher hourly equivalents may make sense when rollout controls, autoscaling, observability, serving infrastructure, or support replace engineering work. CoreWeave’s public rate is specifically for inference-platform customers, not a self-serve one-GPU VM quote.

The Replicate pricing guide shows why setup, idle, and failed deployment work can matter more than prediction runtime.

Size the workload before renting an H100

An 80 GB H100 is not automatically the right purchase because it is newer. Start with model weights, quantization, runtime overhead, KV-cache demand, context length, batch size, and required throughput. HostFleet’s Llama 70B VRAM guide exposes the memory calculation without pretending to benchmark every stack.

If an A100 fits the model and the workload does not need H100-specific throughput or features, compare the A100 rental price guide. The older card can be the better infrastructure decision after a bounded workload test.

H100 buying checklist

  1. Fix the exact hardware requirement. Record PCIe versus SXM/NVLink/HGX, memory, topology, and minimum GPU count.
  2. Identify what the rate buys. Separate complete VM totals, GPU components, Pods, serverless workers, and managed deployments.
  3. Confirm account eligibility and capacity. Check region, quota, approval, credit prerequisites, and live inventory.
  4. Model billable time. Include startup, model loading, active work, retries, idle windows, scale-down, and warm minimums.
  5. Test the off switch. Verify whether stop, pause, hibernate, scale-to-zero, delete, or destroy ends compute billing.
  6. Add omitted resources. Price CPU/RAM, storage, IPs, network, tax, support, and retained data.
  7. Run a bounded deployment test. Measure provisioning, readiness, throughput, failure recovery, billed duration, and cleanup before production commitment.

Verdict

Koyeb and Hyperstack share the lowest selected public H100 rate at $2.50/hour, verified September 28, 2026. Koyeb is the more elastic documented shape, with public-preview GPU scale-to-zero and a five-minute default idle period. Hyperstack is an allocated PCIe VM whose stopped state remains billable. Equal hourly numbers do not mean equal bills.

Massed Compute follows the $2.50 tie at $2.73/hour for its base H100 (80GB) VM; its separate SXM5 configuration is $3.14/hour, or $295.20 more over 720 active hours. Northflank’s $2.74/hour is a GPU component, not an all-in workload price. Jarvis Labs’ one-GPU H100 SXM on-demand rate is now $3.49/hour, up 29.7% from $2.69. That adds $576.00 to a 720-hour single-GPU estimate, moving Jarvis behind RunPod, Thunder, Lambda, and Novita in this rate-only ordering. Jarvis still offers per-minute compute and a pause action that stops compute while retained data continues billing; current stock and successful resume were not checked.

Thunder’s selected rate remains $3.20/hour, only $7.20 above its former rate over a 720-hour estimate. Verda’s September 17–October 2 increases move its exact H100 rate from $3.282 to $3.663/hour, adding $274.32 to the 720-hour estimate and moving the current total to $2,637.36. Verda’s 10-minute prepayment is not a 10-minute minimum charge in this comparison because the provider says unused terminated time is refunded in the next billing period; deletion and refund reconciliation are both part of cost control.

Lambda documents a separate risk: guest shutdown or poweroff does not end billing. At $3.29/hour, confusing guest poweroff with provider termination can leave $526.40 of avoidable compute in the 160-hour example, and termination then requires a deliberate plan for local data and separately billed filesystems.

TensorDock adds a third pattern: its normal stop preserves the GPU reservation at full rate, while stop-and-release leaves storage-only billing and makes restart capacity-dependent. It is useful lifecycle evidence, but TensorDock still does not belong in the numeric ranking until a current, reproducible H100 offer can be sourced.

Nebius’s October 1 increase is now live: $4.50/GPU-hour, up 16.9% from $3.85. The 720-hour estimate rises from $2,772.00 to $3,240.00, a $468.00 increase. Nebius now ties Fal’s list rate and sits above Modal’s hourly equivalent and DigitalOcean’s H100 Droplet rate in this rate-only table. Nebius still differs from both: it is a unified VM rate with prescribed CPU and RAM, and its provider-level stop ends compute billing while retained storage can continue.

The defensible buying order is hardware fit, deployable product shape, current eligibility, billing lifecycle, complete cost, and only then hourly rate.

Choose the next planning step

  • Still validating whether an H100 fits the workload? Use the Llama 70B VRAM guide to account for weights, quantization, context, KV cache, and concurrency before treating an 80 GB card as the requirement.
  • Already know the card and expected run pattern? Put startup, active work, retries, idle retention, and warm capacity into the GPU cloud cost calculator instead of extending the hourly headline rate across every calendar hour.
  • Need an endpoint that can genuinely return to zero? Compare request-driven lifecycle rules in the serverless GPU pricing matrix before treating a VM or Pod rate as application-serving cost.

Sources

Official pricing sources below were rechecked September 28, 2026, except Nebius (October 1), Verda (October 2), Jarvis Labs (October 11), and Massed Compute (October 12). Nebius and Verda were checked against two official pricing surfaces; Jarvis Labs against its public one-GPU on-demand table; Massed Compute against its one-GPU H100 and SXM5 catalog rows. This does not advance other providers’ verification dates.

Operating-boundary sources:

  • TensorDock cloud GPU pricing and billing, Core Compute introduction, and public API collection — host-variable per-second resource pricing, reservation-preserving stop, stop-and-release, storage-only state, restart constraints, and account meter fields; checked September 28, 2026
  • Lambda billing, instance lifecycle, console controls, filesystems, and data import/export — one-minute instance billing, health-check and termination boundaries, guest-poweroff Alert behavior, local-data destruction, filesystem persistence, attachment limits, and cleanup requirements; checked September 14, 2026
  • Nebius Compute pricing, VM lifecycle, stop/start controls, storage types, and quotas — running-only compute billing, guest-shutdown reboot trap, delete cutoff, retained storage, local-SSD loss, quota retention, and restart-capacity boundary; checked September 10, 2026
  • Koyeb scale-to-zero — GPU inclusion, public-preview status, idle conditions, wake protocols, and default period; checked August 28, 2026
  • Koyeb autoscaling — quota and gradual scale-down behavior; checked August 28, 2026
  • Jarvis Labs FAQ and SDK documentation — per-minute billing, pause, storage, and released capacity; checked August 24, 2026
  • Northflank managed GPU deployment and autoscaling — separate compute plan, billing start, credit prerequisite, and scaling boundaries; checked August 31, 2026
  • Hyperstack states and billing — stopped and hibernated lifecycle; checked August 13, 2026
  • Massed Compute billing — active-VM per-minute debit and micro-cent carry; rechecked October 8, 2026
  • Thunder billing — running compute and deletion boundary; checked August 22, 2026
  • Verda Pricing and Billing — prepaid 10-minute increments and next-billing-period refunds for unused terminated time; checked September 25, 2026
  • Verda lifecycle — shutdown, deletion, and retained storage; checked September 25, 2026
  • DigitalOcean Droplet pricing documentation — powered-off billing and destruction; checked August 20, 2026
  • Paperspace Machine limits — H100 approval boundary; checked August 21, 2026
  • HostFleet GPU pricing dataset — /opt/hostbot-v2/src/data/gpu-pricing.json, full-table baseline September 24, 2026, with targeted Verda and Nebius patches from October 1 and Jarvis Labs H100 patch from October 11
  • HostFleet full-source verification note — /opt/hostbot/data/ai-hosting/notes/2026-09-24-gpu-pricing-full-verification.md
  • HostFleet Jarvis Labs price-event note — /opt/hostbot/data/ai-hosting/notes/2026-10-11-jarvis-on-demand-price-change.md
  • HostFleet Massed Compute one-GPU variant note — /opt/hostbot/data/ai-hosting/notes/2026-10-08-massed-compute-h100-variant-cost.md
  • HostFleet TensorDock lifecycle note — /opt/hostbot/data/ai-hosting/notes/2026-09-28-tensordock-stop-release-storage-boundary.md
  • HostFleet Verda price-change notes — /opt/hostbot/data/ai-hosting/notes/2026-09-21-verda-gpu-price-change.md, /opt/hostbot/data/ai-hosting/notes/2026-09-22-verda-gpu-price-change.md, and /opt/hostbot/data/ai-hosting/notes/2026-10-01-verda-gpu-price-change.md
  • HostFleet Koyeb scale-to-zero note — /opt/hostbot/data/ai-hosting/notes/2026-08-28-koyeb-gpu-scale-to-zero-limits.md
  • HostFleet Northflank autoscaling note — /opt/hostbot/data/ai-hosting/notes/2026-08-31-northflank-gpu-autoscaling-billing-boundary.md
  • HostFleet Thunder Compute pricing note — /opt/hostbot/data/ai-hosting/notes/2026-08-22-thunder-compute-gpu-pricing.md
  • HostFleet Nebius stop/delete evidence note — /opt/hostbot/data/ai-hosting/notes/2026-09-10-nebius-gpu-vm-stop-delete-boundary.md
  • HostFleet Nebius October 1 price-event note — /opt/hostbot/data/ai-hosting/notes/2026-10-01-nebius-gpu-price-increases.md
  • HostFleet Lambda termination/storage evidence note — /opt/hostbot/data/ai-hosting/notes/2026-09-14-lambda-cloud-termination-storage-boundary.md

Need self-managed H100 capacity? These are labeled affiliate links; the source citations above remain direct. RunPod signup (affiliate) and DigitalOcean GPU signup (affiliate) support HostFleet at no extra cost to you. Re-check the exact card, region, rate, storage, and shutdown behavior before purchase.