TensorDock Alternatives

Six TensorDock alternatives for renting GPUs, compared on who runs the hardware, what an hour really costs, and what happens when a machine goes away.

TensorDock rents GPU servers from independent hosts at low prices: H100s from $2.25 an hour, consumer cards from $0.12. The catch is in that model. When hosts pull their machines, the dashboard shows no GPUs; when a host goes down, your run stops. Since Voltage Park bought TensorDock in 2025, many users have also looked for a supplier they can count on.

The short answer: Thunder Compute if you want a simple GPU VM from one operator, Lambda for 8-GPU training nodes, Modal for serverless jobs paid by the second, RunPod for a wide GPU choice with serverless on top, Vast.ai if price comes first, and Voltage Park for on-demand H100 clusters.

Romain CochardCEO of Hack'celeration
Updated 20266alternatives compared5criteria each2026pricing checked

The ranking is editorial: a sponsored placement is always flagged on its card, and no affiliate link changes a score.

The honest take

Why developers leave TensorDock

TensorDock is a GPU marketplace: independent hosts, from tier 3 and 4 data centers to smaller operators, list their servers, set their price, and you rent a KVM virtual machine with root access, Windows included. It is cheap and flexible. Here is what pushes people elsewhere, each point with the fact behind it.

GPUs that are not there when you need them

On a marketplace, supply is whatever hosts list today. Users describe opening the deploy page and finding no available GPU of the model they wanted, and the public feedback board has a request titled 'the gpu selection is a barren desert'. A provider that owns its fleet, like Thunder Compute, Lambda or Voltage Park, does not have hosts who can withdraw machines.

A host that goes down takes your run with it

Reliability is a standard TensorDock holds hosts to, not a guarantee to you. When a machine fails, downtime compensation is paid in credits, not in lost work: an 8-GPU run lost at hour 6 of 10 costs you those 6 hours again. The compensation is a multiple of your hourly rate, so it shrinks with the cheap rates that drew you in.

The headline price is only the GPU

The $2.25 an hour on the H100 is the GPU alone. CPU, RAM and storage are chosen and billed separately, and a stopped instance keeps paying for its disk: the deploy page shows a small configuration at $0.049 an hour running and $0.005 an hour stopped, or $3.65 a month for a machine that does nothing.

Spot is still in beta, and storage keeps billing

Spot instances let you bid below on-demand, with host minimum bids around 50% of the on-demand price. They are labeled beta, and when you are outbid your storage is still billed at the host's standard rate. Vast.ai's interruptible instances do the same job on a larger, longer-running market.

Support on a 24-hour standard

TensorDock states a 24-hour response time standard for support. That is fine for a billing question, long when a training run is stuck. On its public feedback board, a user wrote in 2026 that they had been unable to use the service for almost six days and that the chat team had stopped answering.

New owner, open questions

Voltage Park bought TensorDock in 2025 and later merged with Lightning AI. The marketplace kept running, but its founder moved to a role at the parent company and new leadership took over. Teams that bet production work on a provider want to see where the product goes next, and some move while they wait.
1
Best for reliable, simple GPU VMs

Thunder Compute

3.9/5

Thunder Compute is a GPU cloud that runs its own fleet instead of listing third-party hosts. That is the main difference with TensorDock: you rent from one operator, so the machine you used yesterday is the same kind of machine you get today, and there is no host who can pull the hardware back. You get a VM with SSH, a choice of extra vCPUs and disk, and snapshots to save an instance and restart it later.

Its pricing is short and readable. An A100 80GB costs $1.09 an hour and an H100 SXM $3.20 an hour, with a few vCPUs included in the GPU price. Extra vCPUs cost $0.04 per vCPU-hour, the first 100 GB of disk per GPU is included, and anything above is $0.219 per GB per month, billed only while the instance runs. Data egress is $0.00 per GB, which matters if you move checkpoints and datasets out often.

How the price reads against TensorDock: on the H100, TensorDock's headline rate of $2.25 an hour is lower, but it is the cheapest host on a marketplace, and CPU, RAM and disk are billed on top. On the A100 80GB, Thunder's $1.09 an hour is below TensorDock's A100 SXM floor of $1.80. For a 100-hour fine-tuning month on one A100, that is $109 against at least $180 before storage.

Its limit: the GPU list is short (A6000, L40, A100 80GB, H100 SXM), there is no Windows desktop and no consumer card like the RTX 4090, and it is built for single-node work, not 64-GPU clusters.

Standout features
  • One operator, no third-party hosts
  • Free data egress
  • Snapshots to pause and resume
+Pros
  • ✓A100 80GB at $1.09/hr, below TensorDock's A100 floor
  • ✓Pricing page shows every add-on
  • ✓Free egress for checkpoints and datasets
−Cons
  • ✗Short GPU list, no RTX 4090
  • ✗No Windows desktop VMs
  • ✗Not built for multi-node clusters
Thunder Compute vs TensorDock
CriterionThunder ComputeTensorDock
SupplyOwn fleet, one operatorIndependent hosts
A100 80GB$1.09/hrA100 SXM from $1.80/hr
Egress$0.00/GBSet by each host
Verdict

Pick Thunder Compute if you left TensorDock because GPUs were missing or hosts went down, and you work on one GPU at a time. Pick RunPod instead if you need an RTX 4090 or serverless endpoints, and Lambda if you need eight H100s in one box.

Try Thunder Compute →
2
Best for 8-GPU H100 training boxes

Lambda

3.9/5

Lambda is one of the oldest GPU clouds for machine learning. Every instance runs on Lambda's own cloud, comes with the Lambda Stack (drivers, CUDA, PyTorch) and a large local SSD, and scales from one GPU to eight in a single machine, then to reserved clusters.

On-demand prices are listed per GPU-hour. A single H100 SXM costs $4.89 an hour, an H100 PCIe $3.79, an A100 40GB $1.99. The 8x H100 machine comes to $4.59 per GPU-hour, so $36.72 an hour for the box. The 8-GPU box is Lambda's cheapest H100 SXM rate on demand. Its 1-Click Clusters of 16 or more H100s for 2 weeks to a year are listed higher, at $6.16 to $5.54 per GPU-hour; reserved capacity, quoted by sales, is where Lambda sells its lowest prices.

How the price reads against TensorDock: Lambda is roughly twice TensorDock's $2.25 H100 headline. What you buy with the difference is a full 8-GPU node with 1,800 GiB of RAM and 22 TiB of local SSD on the 8x H100 configuration, from one operator. For an 8-GPU run of 10 hours, Lambda costs about $367; TensorDock would be about $180 in GPU time plus CPU, RAM and disk, if eight H100s are free on one host that day.

Its limit: there is no consumer card for hobby work, the lowest per-GPU rates need a reserved cluster negotiated with sales, and for a single GPU you pay well above the marketplaces.

Standout features
  • 1x to 8x H100 and B200 nodes
  • Lambda Stack preinstalled
  • Up to 22 TiB local SSD on 8x
+Pros
  • ✓Real multi-GPU nodes for training
  • ✓Clear per-GPU-hour price grid
  • ✓Reserved clusters when you outgrow on-demand
−Cons
  • ✗H100 about twice TensorDock's headline rate
  • ✗Lowest rates only on reserved clusters
  • ✗No consumer GPUs
Lambda vs TensorDock
CriterionLambdaTensorDock
1x H100 SXM$4.89/hrFrom $2.25/hr
Multi-GPU8x H100 in one nodeDepends on the host
OperatorLambda's own cloudIndependent hosts
Verdict

Pick Lambda if you train on eight GPUs and need the run to finish without a host vanishing. For one GPU at a time, Thunder Compute or RunPod cost much less.

Try Lambda →
3
Best for serverless GPU jobs in Python

Modal

3.9/5

Modal is not a VM rental. You write a Python function, decorate it with the GPU you want, and Modal starts a container when the function is called and stops it when it returns. There is no server to keep running between jobs, which is the opposite of the TensorDock model where you pay for a VM until you delete it.

Prices are per second. An H100 costs $0.001097 a second, about $3.95 an hour; an A100 80GB $0.000694 a second, about $2.50 an hour. CPU and memory are billed on top. The Starter plan costs $0 a month, includes $30 of compute each month and up to 3 seats. Team costs $250 a month with $100 of compute included and unlimited seats.

How the price reads against TensorDock: per hour, Modal's H100 is about 75% more than TensorDock's $2.25 headline. The math flips when the GPU sits idle. An inference endpoint that works 2 hours a day costs about $237 a month on Modal's H100 (2 h × 30 days × $3.95), against $1,643 for a TensorDock H100 VM left on all month (730 h × $2.25).

Its limit: you work inside Modal's Python SDK, not on a machine you SSH into; no Windows, no persistent desktop, and long training runs on one GPU cost more than on a plain VM. We compare its rivals in our Modal alternatives.

Standout features
  • Per-second billing, scale to zero
  • $30 free compute every month
  • GPU functions in plain Python
+Pros
  • ✓No bill while nothing runs
  • ✓Fast cold starts for inference
  • ✓Free monthly credits to try it
−Cons
  • ✗H100 about $3.95/hr when it runs
  • ✗Locked into Modal's SDK
  • ✗No SSH VM or Windows
Modal vs TensorDock
CriterionModalTensorDock
BillingPer second, only while code runsPer second while the VM exists
AccessPython functionsFull VM, root, SSH
Free tier$30 compute/monthNone, $5 to start
Verdict

Pick Modal if your GPU work comes in bursts (inference, batch jobs, cron pipelines). If you keep a GPU busy most of the day, a VM on Thunder Compute or RunPod costs less.

Try Modal →
4
Best all-rounder for pods and serverless

RunPod

3.7/5

RunPod is the closest thing to TensorDock's model with more tooling around it. It sells GPU pods in two clouds: Community Cloud, where vetted third-party hosts list machines, and Secure Cloud, run in certified data centers. It adds serverless endpoints, templates for ComfyUI, vLLM or Jupyter, and network volumes you can attach to any pod.

Every price is per second with no commitment. An H100 SXM costs from $2.69 an hour on Community Cloud and $4.39 on Secure Cloud; an A100 SXM from $1.39 and $2.19; an H100 PCIe from $1.99 and $2.89. Container disk is $0.10 per GB per month, and a volume disk costs $0.10 while the pod runs and $0.20 while it sits idle.

How the price reads against TensorDock: Community Cloud is close to TensorDock's marketplace (from $2.69 against $2.25 on the H100), with a single support desk and templates. Secure Cloud costs about 60% more, but you know which kind of facility runs your job. The trap is the idle volume: 500 GB kept on a stopped pod is $100 a month at $0.20 per GB.

Its limit: pods are containers, not full VMs, so there is no Windows desktop and no custom kernel; support is the weak point that users mention most, which our RunPod review scores at 2.6 out of 5. The full grid is on our RunPod pricing breakdown, and if RunPod itself is what you want to replace, see the RunPod alternatives.

Standout features
  • Community and Secure clouds
  • Serverless endpoints
  • One-click templates
+Pros
  • ✓Wide GPU choice, consumer cards to B300
  • ✓Serverless and pods in one account
  • ✓Per-second billing
−Cons
  • ✗Containers, not full VMs
  • ✗Idle volumes billed at $0.20/GB/month
  • ✗Support often slow
RunPod vs TensorDock
CriterionRunPodTensorDock
H100 SXMFrom $2.69/hr (Community)From $2.25/hr
Machine typeContainer podFull KVM VM
ServerlessYesNo
Verdict

Pick RunPod if you want TensorDock-like prices with templates and serverless on top. Pick Thunder Compute if you want a plain VM from one operator, and Vast.ai if price is the only thing that counts.

Try RunPod →
5
Cheapest GPU marketplace

Vast.ai

3.6/5

Vast.ai is the other big GPU marketplace, and the one TensorDock users find first when they want a lower price. Hosts, from data centers to individuals, list machines; you filter by GPU, reliability score, location and price, then rent on-demand, interruptible or reserved.

Prices move with supply and demand. Live offers on the pricing page start around $1.67 an hour for an H100 SXM and $0.27 an hour for an RTX 4090. Interruptible instances are 50% or more below on-demand, and reserved instances go up to 50% off for a commitment. Storage and bandwidth are set by each host and shown on the offer.

How the price reads against TensorDock: Vast's lowest H100 offer ($1.67) is below TensorDock's $2.25 headline, and its cheapest RTX 4090 ($0.27 against $0.35) is lower too. The saving is real only if your job survives a host going away: for a 100-hour H100 run, $167 against $225, a $58 gap that one lost day of training wipes out.

Its limit: quality varies from host to host. The $1.67 is the cheapest H100 SXM offer of the moment: the median offer on the same page is $2.25 an hour, level with TensorDock's headline, and the RTX 4090 goes from $0.27 to a median of $0.47. Read the reliability score and the host's bandwidth before you rent.

Standout features
  • Lowest H100 and RTX 4090 offers
  • Interruptible bidding
  • Reliability score per host
+Pros
  • ✓Usually the lowest hourly rate
  • ✓Huge choice of consumer and data-center GPUs
  • ✓Filters for reliability and location
−Cons
  • ✗Host quality varies a lot
  • ✗A host can vanish mid-run
  • ✗Bandwidth varies by host
Vast.ai vs TensorDock
CriterionVast.aiTensorDock
H100 SXMFrom about $1.67/hrFrom $2.25/hr
ModelMarketplace with bidsMarketplace with fixed host prices
Cheap modeInterruptible, 50%+ offSpot, beta
Verdict

Pick Vast.ai if you liked TensorDock for its price and your jobs checkpoint often. If a lost host costs you more than the saving, take Thunder Compute or RunPod Secure Cloud.

Try Vast.ai →
6
Best for on-demand H100 clusters

Voltage Park

3.6/5

Voltage Park is the company that bought TensorDock in 2025, and it has since merged with Lightning AI. It runs its own fleet of NVIDIA HGX H100 servers and rents them on demand: from 1 GPU on Ethernet, or from 8 to 1,016 H100s connected by 3,200 Gbps InfiniBand, up in about 15 minutes, with no minimum term.

All three offers, on-demand on Ethernet, on-demand on InfiniBand and the long-term reserve (6 months or more, 32 to 8,000+ H100s), say 'contact for pricing', while one H100 for one hour starts at $1.99, without a contract. Managed Kubernetes and bare-metal access are available.

How the price reads against TensorDock: at $1.99 an hour, a Voltage Park H100 is below TensorDock's own $2.25 H100 headline, on the parent company's own hardware in Tier 3+ data centers rather than a third-party host's machine. For 100 H100-hours, that is $199 against $225.

Its limit: it only sells H100-class and Blackwell data-center GPUs, so there is nothing for a cheap RTX job; cluster prices are on quote; and it is built for training teams, not a single notebook.

Standout features
  • $1.99/hr H100, no contract
  • InfiniBand clusters up to 1,016 GPUs
  • Managed Kubernetes
+Pros
  • ✓Lower H100 rate than TensorDock's headline
  • ✓Large clusters without a long contract
  • ✓One operator, own hardware
−Cons
  • ✗H100 and Blackwell only
  • ✗Cluster prices on quote
  • ✗Less suited to hobby or single-notebook use
Voltage Park vs TensorDock
CriterionVoltage ParkTensorDock
H100From $1.99/hrFrom $2.25/hr
Clusters8 to 1,016 GPUs on InfiniBandDepends on the host
Consumer GPUsNoYes, from $0.12/hr
Verdict

Pick Voltage Park if you need H100s by the dozen for a training sprint. For one consumer GPU or a Windows VM, look at Vast.ai or RunPod instead.

Try Voltage Park →
At a glance

TensorDock alternatives compared

The six alternatives side by side, with the public on-demand price of each. The 'edge over TensorDock' column says what you gain by moving. Prices are per GPU-hour, before storage and taxes.

Best forEdge over TensorDockFree planTeam sizeVisit
1Thunder ComputeBest for reliable, simple GPU VMsSame VM workflow, one operator instead of a marketplace3.9/5A100 80GB $1.09/hr, H100 SXM $3.20/hr, billed per minute while running—1 to 20 peopleVisit →
2LambdaBest for 8-GPU H100 training boxesData-center 1x to 8x H100 and B200 with big local disks3.9/51x H100 SXM $4.89/hr, H100 PCIe $3.79/hr, A100 40GB $1.99/hr—Research and ML teamsVisit →
3ModalBest for serverless GPU jobs in PythonPay per second of GPU work, no idle VM3.9/5H100 $0.001097/sec (about $3.95/hr), A100 80GB $0.000694/sec (about $2.50/hr)✓Developers and AI startupsVisit →
4RunPodBest all-rounder for pods and serverlessTwo clouds (Community and Secure) plus serverless3.7/5H100 SXM from $2.69/hr (Community Cloud) or $4.39/hr (Secure Cloud)—Individuals to mid-size AI teamsVisit →
5Vast.aiCheapest GPU marketplaceLowest marketplace prices, interruptible bids3.6/5H100 SXM from about $1.67/hr, RTX 4090 from about $0.27/hr—Individuals and labsVisit →
6Voltage ParkBest for on-demand H100 clustersSelf-serve H100 clusters up to 1,016 GPUs3.6/5H100 from $1.99/hr without contract—AI startups and labsVisit →

Is your tool missing from this page? Email us: hello@hackceleration.com

Buyer's guide

How to choose a TensorDock alternative

GPU clouds come in three kinds. Marketplaces (TensorDock, Vast.ai, RunPod Community Cloud) list machines owned by many hosts: low prices, uneven quality. Operator clouds (Thunder Compute, Lambda, Voltage Park, RunPod Secure Cloud) run their own fleet: steadier, often pricier per hour. Serverless platforms (Modal, RunPod Serverless) bill only the seconds your code runs. The unit to compare is the GPU-hour: one GPU rented for one hour. On a VM it runs from the moment you start the machine until you stop it; on serverless only while a request or job is running.

Before comparing hourly rates, check three things: what is included in the GPU price (vCPUs, RAM, disk), what you pay when the machine is stopped, and what happens to your data if the host fails. Then do the math on your real usage, not on the headline. For the wider market, see our ranking of the best GPU cloud providers, the best inference providers if you only serve models, and our AI infrastructure picks.

A few hours a week

Serverless wins. Modal's $30 of monthly credits cover about 7.5 H100 hours (30 ÷ 3.95), so a weekly fine-tune or a small demo can run for free. A VM you forget to stop costs more than that in a weekend.

One GPU most of the day

Take a plain VM. On an A100 80GB for 300 hours a month, Thunder Compute is $327 (300 × $1.09); RunPod Community Cloud on an A100 SXM from $417 (300 × $1.39); TensorDock from $540 for an A100 SXM (300 × $1.80), before CPU and RAM.

Multi-GPU training

Rent the node, not the GPUs one by one. Lambda's 8x H100 at $36.72 an hour gives one node with 22 TiB of local SSD; Voltage Park scales the same job to InfiniBand clusters. A marketplace host with 8 free H100s is a lucky find, not a plan.

Tight budget, patient jobs

Vast.ai interruptible instances at 50% or more below on-demand, with checkpoints every 30 minutes to storage you control. Price the restarts in: losing 30 minutes on 8 GPUs at $1.67 is about $7, losing 6 hours is $80.
  • Is the hourly price for the GPU alone, or with CPU, RAM and disk?
  • What does a stopped instance cost per GB of disk?
  • Does the provider own the hardware or list hosts?
  • Is billing per second, per minute or per hour?
  • Are egress and data transfer free?
  • Can you snapshot an instance and restart it on another machine?
  • Do you get a VM with root, or a container?
  • How fast does support answer when an instance is stuck?
FAQ · 8 questions

TensorDock alternatives · FAQ

  • Is TensorDock still working?

    Yes. The site still deploys on-demand GPUs, with H100 SXM5 listed from $2.25 an hour and consumer GPUs from $0.12, and Voltage Park said the marketplace would keep operating as usual after the purchase. What varies is supply: what you can rent depends on which hosts are online that day.

  • Who owns TensorDock now?

    Voltage Park, a GPU cloud operator, bought TensorDock in 2025. Voltage Park's profits go to its parent, the nonprofit Navigation Fund, and Voltage Park itself later merged with Lightning AI, the company behind PyTorch Lightning.

  • Why does TensorDock show no GPUs available?
    Because the GPUs belong to hosts, not to TensorDock. When hosts take machines offline or every unit of a model is rented, it disappears from the deploy page. TensorDock publishes its live stock through a public API endpoint (dashboard.tensordock.com/api/stock), so you can script a check before a run instead of refreshing the page, or move to an operator that runs its own fleet.
  • What happens when my TensorDock balance runs out?
    Your servers are deleted. TensorDock bills from a prepaid balance it draws down while a server is deployed, and when the balance reaches $0 the servers are automatically deleted, disk included. Keep a buffer above your weekly spend, or copy checkpoints off the host every day, so an empty balance does not cost you the data.
  • Can I run a Windows VM with a GPU somewhere other than TensorDock?

    Rarely as simply. TensorDock lets you pick Windows 10 at deploy time, with at least 80 GB of disk for remote desktop. RunPod and Modal run containers, and Thunder Compute and Lambda offer Linux images, so a Windows GPU desktop usually means a hyperscaler such as AWS or Azure.

  • How do I move my work off TensorDock?

    Copy your data first, then switch. Push checkpoints and datasets to object storage (S3, R2 or B2) or to the new provider's volume with rsync, export your environment as a Dockerfile or requirements file, and only then delete the TensorDock instance, since a deleted VM's disk is gone.

  • Does TensorDock pay me back when a host goes down?
    Only in credits, as a multiple of your hourly rate. A measured outage on a top host is credited at 10 times the hourly rate per hour of downtime, on a standard host at 5 times, and unexpected data loss at 40 times the hourly rate, with a minimum of $5 and a cap at 100% of your previous month's spend. None of it restores the lost work, so save checkpoints off the host every 30 minutes on a long run.
  • Which TensorDock alternative is best for Stable Diffusion or ComfyUI?

    RunPod or Vast.ai. Both rent RTX 4090 class cards (from $0.27 an hour on Vast.ai) and RunPod has ready-made ComfyUI templates. Thunder Compute, Lambda and Voltage Park focus on data-center GPUs, which is overkill for image generation.

Hack'celeration Lab

Get the next breakdown in your inbox

No spam. Unsubscribe anytime.