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. All 19 selected H100 price anchors were rechecked against live vendor sources on September 14, 2026. Lambda’s termination, guest-poweroff, and filesystem rules were also checked September 14. HostFleet did not measure invoices or state-transition timing.

Price anchors rechecked: September 14, 2026, for all 19 rows below
Dataset baseline: September 10, 2026
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 public rates checked

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.

Jarvis Labs follows at $2.69/GPU-hour, Massed Compute at $2.73/hour, Northflank at $2.74/GPU-hour, and RunPod Secure Cloud at $2.89/hour. Northflank’s number is only a GPU component; CPU and memory are extra. The other rows package different resources and lifecycle controls.

The September 14 bounded verification kept all 19 selected H100 price anchors unchanged. It did catch one specification change: Novita’s live H100 API record now lists 22 vCPU and 150 GB RAM rather than the prior 16-vCPU, 128-GB shape. HostFleet’s full 148-cell dataset retains its September 10 verification date because this was an H100-only check. The main addition is a Lambda cost-control boundary the rate card cannot show: guest shutdown or poweroff puts the instance into Alert and billing continues; only Lambda’s termination operation ends the compute meter.

Current H100 price-per-hour comparison

The table is sorted by normalized hourly rate. 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 product	H100 scope	Public list rate	Official evidence and check date
Hyperstack	1x H100 80 GB PCIe VM; 28 CPU, 180 GB RAM, local storage; Canada	$2.50/GPU-hr	Hyperstack pricing, Sept. 14, 2026
Koyeb	1x H100 80 GB serverless instance; 15 vCPU, 180 GB RAM, 320 GB disk	$2.50/hr	Koyeb pricing, Sept. 14, 2026
Jarvis Labs	1x H100 80 GB SXM on-demand instance; public row lists 16 vCPU and 200 GB RAM	$2.69/GPU-hr	Jarvis Labs pricing, Sept. 14, 2026
Massed Compute	1x H100 80 GB on-demand VM; 20 vCPU, 128 GB RAM; storage shown as 1,250 with no unit	$2.73/hr	Massed Compute pricing, Sept. 14, 2026
Northflank	1x H100 80 GB managed-cloud GPU component; CPU, memory, disk, and egress separate	$2.74/GPU-hr	Northflank pricing, Sept. 14, 2026
RunPod Pods	H100 PCIe Secure Cloud Pod	$2.89/hr	RunPod pricing, Sept. 14, 2026
Thunder Compute	1x H100 80 GB PCIe base; 4 vCPU, 32 GB RAM, 100 GB persistent disk included	$3.20/GPU-hr	Thunder pricing, Sept. 14, 2026
Verda GPU instance	1x H100 80 GB SXM5; selected configuration includes 30 CPU and 120 GB RAM	$3.25/hr	Verda pricing, Sept. 14, 2026
Lambda Cloud	1x H100 PCIe VM	$3.29/GPU-hr	Lambda GPU instances, Sept. 14, 2026
Novita AI instance	1x H100 80 GB SXM; live API record lists 22 vCPU, 150 GB RAM, 60 GB container-disk quota	$3.39/GPU-hr	Novita marketplace API, Sept. 14, 2026
Nebius AI Cloud	1x H100 SXM/NVLink VM; 16 vCPU, 200 GB RAM; eu-north1	$3.85/GPU-hr	Nebius pricing, Sept. 14, 2026
Modal	H100 allocated to a serverless container	$0.001097/sec ($3.9492/hr)	Modal pricing, Sept. 14, 2026
DigitalOcean GPU Droplet	1x HGX H100; 20 vCPU, 240 GiB RAM, 720 GiB boot, 5 TiB scratch, 15,000 GiB transfer	$4.41/GPU-hr	DigitalOcean GPU pricing, Sept. 14, 2026
Fal custom deployment	H100 80 GB on-demand custom deployment	$4.50/hr	Fal pricing, Sept. 14, 2026
RunPod Serverless	H100 PRO Serverless Flex worker tier; not an exact-card reservation	$4.79/hr	RunPod pricing, Sept. 14, 2026
Replicate private deployment	H100 managed model deployment	$0.001525/sec ($5.49/hr)	Replicate pricing, Sept. 14, 2026
Paperspace Machine	1x H100 80 GB SXM5; 20 vCPU, 250 GB RAM, 50 GB SSD; approval may be required	$5.95/hr	Paperspace pricing, Sept. 14, 2026
CoreWeave Inference	Single-GPU inference rate for inference-platform customers	$6.16/GPU-hr	CoreWeave pricing, Sept. 14, 2026
Baseten deployment	Managed H100 deployment; 80 GiB VRAM	$0.10833/min ($6.50/hr)	Baseten pricing, Sept. 14, 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 14, 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.

Jarvis Labs puts an SXM instance near the top

Jarvis Labs publishes $2.69/GPU-hour for a one-GPU H100 SXM on-demand instance, checked September 14, 2026. Its public row lists 16 vCPU and 200 GB RAM. On-demand compute bills per minute.

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 14, 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 14, 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 14, 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.

No selected H100 rate moved in the September 14 bounded check. That is useful evidence of rate-card stability, not evidence of available capacity or stable invoice totals.

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 14, 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 14 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 publishes a unified $3.85/GPU-hour price for its one-H100 SXM/NVLink 1gpu-16vcpu-200gb VM in eu-north1, rechecked September 14, 2026. That price includes the prescribed 16 vCPU and 200 GB of RAM. Persistent disks and other retained resources are separate.

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 = $3.85 × 8 hours = $30.80
unattended compute = $3.85 × 160 hours = $616.00
one-week compute = $646.80

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 $31.30 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 each sourced September 14 rate by 720 hours. Modal and Baseten 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.

Product shape	Rate used	720-hour compute estimate
Hyperstack H100 PCIe VM	$2.50/hr	$1,800.00
Koyeb H100 serverless instance	$2.50/hr	$1,800.00
Jarvis Labs H100 SXM instance	$2.69/hr	$1,936.80
Massed Compute H100 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
Verda H100 SXM5 instance	$3.25/hr	$2,340.00
Lambda H100 PCIe VM	$3.29/hr	$2,368.80
Novita H100 SXM instance	$3.39/hr	$2,440.80
Nebius H100 SXM/NVLink VM	$3.85/hr	$2,772.00
Modal H100 container	$0.001097/sec	$2,843.42
DigitalOcean HGX H100 Droplet	$4.41/hr	$3,175.20
Fal H100 custom deployment	$4.50/hr	$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.

Product	Compute-billing boundary	Important residual
Hyperstack	Hibernation deallocates the flavor; a merely stopped VM remains billable	Disks, IPs, and volumes can remain billable
Koyeb	Eligible public-preview services can reach zero after the idle policy	Wake protocol and idle conditions matter; no GPU cold-start number is claimed
Jarvis Labs	Pause stops compute billing	Retained data bills; released GPU capacity may not return
Massed Compute	Active VM total is debited per minute	Confirm the exact stop/delete action before automation
Northflank	GPU bills by the second once provisioned; manual zero makes the service unavailable	CPU, memory, disk, and egress are separate
RunPod Pods	Stop or terminate according to the required persistence model	Storage can continue; container-disk data can be erased
Thunder Compute	Deleting the instance stops instance billing	Confirm snapshot and retained-disk handling
Verda	Deletion is required; shutdown does not stop compute billing	Retained storage remains separate
Lambda Cloud	Terminate through Lambda’s console or Cloud API; guest shutdown/poweroff enters `Alert` and keeps billing	Termination destroys local data; filesystems survive and bill until deleted
Nebius	Stop through the console, CLI, or SDK; guest `shutdown`/`halt` auto-reboots and remains billable	Persistent storage bills; local SSD is erased; quota stays occupied; restart capacity is not guaranteed
DigitalOcean GPU Droplets	Destroy the Droplet; powering it off does not stop billing	Reserved resources continue charging while powered off
Paperspace Machines	Power off stops compute billing	Storage, public IPs, and add-ons can continue
Modal and RunPod Serverless	Workers can return to zero	Startup, 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, 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.

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

Fix the exact hardware requirement. Record PCIe versus SXM/NVLink/HGX, memory, topology, and minimum GPU count.
Identify what the rate buys. Separate complete VM totals, GPU components, Pods, serverless workers, and managed deployments.
Confirm account eligibility and capacity. Check region, quota, approval, credit prerequisites, and live inventory.
Model billable time. Include startup, model loading, active work, retries, idle windows, scale-down, and warm minimums.
Test the off switch. Verify whether stop, pause, hibernate, scale-to-zero, delete, or destroy ends compute billing.
Add omitted resources. Price CPU/RAM, storage, IPs, network, tax, support, and retained data.
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 14, 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.

Jarvis Labs is next at $2.69/hour, with per-minute compute and a pause action that stops compute while retained data continues billing. Massed Compute follows at $2.73/hour. Northflank’s $2.74/hour is a GPU component, not an all-in workload price.

Thunder’s selected rate remains $3.20/hour, only $7.20 above its former rate over a 720-hour estimate. More importantly, Lambda documents that 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.

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

Sources

Official pricing sources below were rechecked September 14, 2026.

Operating-boundary sources:

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; checked August 23, 2026
Thunder billing — running compute and deletion boundary; checked August 22, 2026
Verda lifecycle — shutdown, deletion, and retained storage; checked August 15, 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, updated September 10, 2026
HostFleet full-source verification note — /opt/hostbot/data/ai-hosting/notes/2026-09-10-gpu-pricing-full-verification.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 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.