Vast.ai Alternatives

Six Vast.ai alternatives for renting GPUs, compared on the same cards, the real price per GPU-hour, and what you lose when a machine goes away.

Vast.ai is often the cheapest place to rent a GPU: an H100 SXM from $1.53 an hour and an RTX 4090 from $0.31, billed by the second, with prices that update hourly. Not always: Nebius sells preemptible H100s from $0.79 a GPU-hour, and Thunder Compute's RTX A6000 at $0.35 is level with Vast.ai's lowest A6000 offer. The price comes from its model. Independent hosts list their machines, so quality, bandwidth and uptime change from one offer to the next, and an interruptible instance stops the moment someone bids higher.

The short answer: Lambda for 1 to 8 data-center GPUs that stay up, Thunder Compute for a cheap GPU VM from one operator, Hyperstack for per-minute VMs with no bandwidth fees, Modal if your GPU sits idle most of the day, RunPod if you want the closest thing to Vast.ai with a managed tier on top, and TensorDock if you want another marketplace with full VMs.

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 people leave Vast.ai

Vast.ai is a GPU marketplace: hosts, from certified data centers to individuals with a few cards at home, list machines and set their price, and you rent a Docker container on one of them. It lists 68+ GPU types across 40+ data centers, with prices set by supply and demand. The reasons people move are the other side of that model, each with the fact behind it.

A host can disappear mid-run

An on-demand rental cannot be interrupted by another renter, but it still runs on someone else's machine. If that host goes offline, your work stops with it. Users still described instances going offline at random in the middle of their work, with no warning, in 2026. On a training run without frequent checkpoints, one lost six-hour night on 8 GPUs at $1.53 costs about $73 of GPU time before you count the delay.

Interruptible means stopped, not slowed

Interruptible instances are 50% or more cheaper, and the docs are clear about the price of that discount: they can be stopped by higher bids, and the stop kills running processes. Anything not saved to disk is gone. That is fine for a batch job that resumes from a checkpoint and a bad deal for an inference endpoint people are waiting on.

The listed price is the cheapest offer, not the usual one

The H100 SXM at $1.53 an hour and the RTX 4090 at $0.31 are the lowest offers of the moment; the median offers on the same page are $1.73 and $0.47, and the figures update hourly. Prices are set by supply and demand, so the host with the best reliability score, the bandwidth you need and enough disk often costs more, and next week's offer may not match this week's.

Network speed varies by host

Data transfer costs vary by host and cover upload and download, and the speed does too. One user reported in 2026 downloading at about 1 MB/s from a machine that listed 2,010 Mbps. On a 200 GB dataset, that is the difference between minutes and more than two days.

Security depends on who hosts you

The docs say provider security varies significantly and that individual hosts may have less formal security measures. The Secure Cloud filter limits you to vetted data centers with ISO 27001 certification, but the cheapest offers are usually outside it. Teams with client data or a security review often want one operator they can audit instead.

A stopped instance may not come back

Stopping an instance releases its GPU. When you restart, the instance waits for that same machine, and the docs warn that the GPU may be reassigned to other users: if the restart is stuck for more than 30 seconds, someone else has likely rented it, and the wait can last indefinitely. Your data is still on that host, still billed, and you copy it to a new instance by hand.
1
Best for 1 to 8 data-center GPUs that stay up

Lambda

4.0/5

Lambda is a GPU cloud built for machine learning, and the opposite of a marketplace: every instance runs on hardware Lambda operates, with the Lambda Stack (drivers, CUDA, PyTorch) already installed and a large local SSD attached. You pick a GPU type and a count, from one card to an eight-GPU HGX machine, and you get the same configuration every time you launch it. Nobody can reclaim the machine because they want it back for themselves.

Prices are listed per GPU-hour. One H100 SXM costs $4.89 an hour, an H100 PCIe $3.79, an A100 40GB $1.99 and a GH200 with 96 GB of memory $2.59. The eight-GPU H100 machine comes to $4.59 per GPU-hour, so $36.72 an hour for the box, with 1,800 GiB of RAM and 22 TiB of local SSD. Above that, Lambda sells reserved clusters through its sales team.

Price order against Vast.ai, same card: on the H100 SXM, Lambda's $4.89 is more than three times Vast.ai's cheapest offer of $1.53. For 100 hours on one H100, that is $489 against $153. What the $336 buys is a machine that does not change hosts, a support desk behind the hardware and storage you do not have to vet. Lambda's cheapest single cards are the A6000 at $1.09 and the Quadro RTX 6000 at $0.69 an hour.

A pricing detail worth knowing: one H100 SXM costs $0.30 more per hour than each GPU of the eight-GPU machine ($4.89 against $4.59). If your job can use several cards, the larger box is the cheaper rate per GPU. Its limit: there is no consumer card like the RTX 4090, and no interruptible tier to cut the bill on jobs that could survive a stop.

Standout features
  • Own fleet, no third-party hosts
  • 1x to 8x H100 and B200 machines
  • Lambda Stack preinstalled
+Pros
  • ✓Runs finish without a host vanishing
  • ✓Clear per-GPU-hour grid
  • ✓8x H100 cheaper per GPU than 1x
−Cons
  • ✗H100 more than 3x Vast.ai's lowest offer
  • ✗No RTX 4090 or consumer cards
  • ✗No interruptible discount
Lambda vs Vast.ai
CriterionLambdaVast.ai
H100 SXM$4.89/hr (1x)From $1.53/hr
Who runs itLambda's own cloudIndependent hosts
Cheap modeNoneInterruptible, 50%+ off
Verdict

Pick Lambda if a lost run costs you more than the hourly gap and you train on one to eight data-center GPUs. Pick Thunder Compute instead for a single A100 on a budget, and stay on Vast.ai for RTX 4090 work.

Try Lambda →
2
Best cheap GPU VM from one operator

Thunder Compute

3.9/5

Thunder Compute runs its own GPUs and rents them as virtual machines you reach over SSH or from VS Code. That removes the main job a Vast.ai user does before every rental: reading host reliability scores, checking bandwidth and hoping the offer is still there tomorrow. You choose a GPU, add vCPUs and disk if you want, and save the machine as a snapshot when you stop working.

The grid is short. An RTX A6000 costs $0.35 an hour, an L40 $0.79, an A100 80GB $1.09 and an H100 SXM $3.20, each with a few vCPUs included. Extra vCPUs are $0.04 per vCPU-hour. The first 100 GB of disk per GPU is included, and anything above costs $0.219 per GB per month, billed only while the instance runs. Snapshots cost $0.05 per GB per month, and data egress is $0.00 per GB.

Price order against Vast.ai, same card: on the H100 SXM, Thunder's $3.20 is about twice Vast.ai's lowest offer of $1.53, so it is not where you save on Hopper GPUs. Vast.ai's A100 SXM4 starts at $0.40, median $0.84. Thunder's A100 80GB at $1.09 is the price of one fixed operator: 100 hours of fine-tuning cost $109, with no host to vet. Keeping a 100 GB snapshot between projects costs $5 a month, and the disk itself costs nothing while the VM is off, the reverse of a stopped Vast.ai instance.

Its limit: four GPU types only, and no RTX 4090 or 5090 for image generation. Instances come with 1, 2, 4 or 8 GPUs and are never preempted, so there is no cheaper interruptible tier; lower rates mean reserving GPUs for a fixed term, on a quote from sales.

Standout features
  • One operator, no host lottery
  • Disk billed only while running
  • Free data egress
+Pros
  • ✓A100 80GB at $1.09/hr
  • ✓Snapshots at $0.05/GB/month
  • ✓Every add-on priced on one page
−Cons
  • ✗H100 about twice Vast.ai's lowest offer
  • ✗No consumer RTX cards
  • ✗Reserved rates only on a sales quote
Thunder Compute vs Vast.ai
CriterionThunder ComputeVast.ai
A100 80GB$1.09/hrVaries by host
Stopped instanceDisk not billedStorage still billed
Egress$0.00/GBSet by each host
Verdict

Pick Thunder Compute if you rent one A100 or A6000 at a time and are tired of choosing hosts. For H100 hours at the lowest price, Vast.ai stays cheaper; for an eight-GPU HGX box with 22 TiB of local SSD, look at Lambda.

Try Thunder Compute →
3
Best for per-minute VMs with no bandwidth fees

Hyperstack

3.9/5

Hyperstack is a GPU cloud run in data centers in Europe and North America. You rent virtual machines with one or more GPUs from a prepaid balance, attach volumes, and pay by the minute. It sits between the marketplaces and the large managed clouds: a fixed price per GPU type, published on one page, with on-demand, spot and reserved tiers.

On demand, an H100 PCIe costs $2.50 an hour, an A100 80GB $1.35, an L40 $1.00 and an RTX A6000 $0.50. An H100 SXM costs $3.20, but it only comes as an eight-GPU VM, so the smallest SXM rental is $25.60 an hour. Spot VMs cut the H100 PCIe to $2.00, 20% below on-demand, and reserved H100 SXM capacity starts at $2.72. Hyperstack charges nothing for bandwidth or data transfer.

Price order against Vast.ai, same card: Vast.ai's lowest H100 SXM offer of $1.53 is less than half of Hyperstack's $3.20, and Vast.ai rents it one card at a time. On the A100, Hyperstack's $1.35 is a fixed rate you can plan on, where a Vast.ai offer may be gone tomorrow. The free bandwidth matters for the Vast.ai user who pays each host's transfer rate on every dataset download: pulling 2 TB of training data costs nothing extra here.

Its traps: if your balance runs dry and you do not recharge within one week, your resources may be permanently deleted. And its spot discount is modest next to Vast.ai's interruptible market, which runs at half the on-demand price or less.

Standout features
  • No bandwidth or transfer fees
  • Billed per minute
  • On-demand, spot and reserved tiers
+Pros
  • ✓H100 PCIe at $2.50/hr, one card
  • ✓Data-center hardware at fixed prices
  • ✓Spot H100 PCIe at $2.00/hr
−Cons
  • ✗H100 SXM only as an 8-GPU VM ($25.60/hr)
  • ✗Resources can be deleted a week after the balance runs out
  • ✗No consumer RTX cards
Hyperstack vs Vast.ai
CriterionHyperstackVast.ai
H100PCIe $2.50/hr, SXM $3.20/hrSXM from $1.53/hr
BandwidthFreeSet by each host
Cheap modeSpot, 20% off on H100 PCIeInterruptible, 50%+ off
Verdict

Pick Hyperstack if you want fixed data-center prices and move a lot of data. If you need a single H100 SXM rather than eight, look at Lambda or Thunder Compute; if price alone decides, Vast.ai remains lower.

Try Hyperstack →
4
Best for GPU work that comes in bursts

Modal

3.9/5

Modal is serverless: you write a Python function, mark which GPU it needs, and Modal starts a container when the function is called and stops it when it returns. There is no instance to keep alive, no host to choose and no machine billed while you sleep. That solves a different Vast.ai problem than the others on this list: not reliability, but paying for a rented machine that sits idle between requests.

Prices are per second. An H100 costs $0.001097 a second, about $3.95 an hour; an A100 80GB $0.000694, about $2.50; an L40S $0.000542, about $1.95; a T4 $0.000164. CPU and memory are billed on top. The Starter plan costs $0 a month, includes $30 of compute each month for up to 3 seats, and 1 TiB of network egress a month. Team costs $250 a month with $100 of compute included.

Price order against Vast.ai, same card: per running hour, Modal's H100 at $3.95 is about 2.6 times Vast.ai's lowest offer. The math turns when the GPU is idle. An endpoint that works 2 hours a day costs about $237 a month on Modal's H100, against $1,117 for a Vast.ai H100 at $1.53 left on for a 730-hour month. The $30 of monthly credit alone covers about 7.6 H100 hours.

Two traps to know. Functions can be preempted by default: Modal says interruptions are rare, but a long job still needs checkpoints. And Sandboxes, its isolated environments for running untrusted or generated code, pay 3 times the list price for CPU and memory, while their GPUs keep the standard rate. Finally, you work inside Modal's SDK, not on a machine you SSH into. If that lock-in is the worry, our Modal alternatives list the other serverless options.

Standout features
  • Per-second billing, scale to zero
  • $30 free compute every month
  • GPU functions in plain Python
+Pros
  • ✓No bill while nothing runs
  • ✓Free monthly credits
  • ✓1 TiB of free egress a month
−Cons
  • ✗H100 about $3.95/hr while running
  • ✗Functions preemptible by default
  • ✗No SSH machine of your own
Modal vs Vast.ai
CriterionModalVast.ai
H100About $3.95/hr, only while code runsFrom $1.53/hr while the instance exists
Idle cost$0 of computeStorage still billed
Free tier$30 compute/monthNone
Verdict

Pick Modal if your GPU work comes in bursts: inference, batch jobs, scheduled pipelines. If a GPU stays busy most of the day, a VM on Thunder Compute, Hyperstack or Vast.ai itself costs less.

Try Modal →
5
Closest to Vast.ai, with a managed tier

RunPod

3.7/5

RunPod is the alternative most Vast.ai users try first, and the one closest to its model. It sells GPU pods in two clouds: Community Cloud, where vetted third-party hosts list machines, and Secure Cloud, run in certified data centers. On top come serverless endpoints, one-click templates for ComfyUI, vLLM or Jupyter, and network volumes you can attach to any pod.

Everything bills 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 RTX 4090 $0.34 on Community Cloud and $0.89 on Secure Cloud. Container disk costs $0.10 per GB per month; a volume disk costs $0.10 while the pod runs and $0.20 while it sits idle.

Price order against Vast.ai, same card: on the RTX 4090, Community Cloud at $0.34 is three cents above Vast.ai's lowest $0.31, so 100 hours of image generation cost $34 against $31. On the H100 SXM, $2.69 is $1.16 above Vast.ai's lowest offer, and Secure Cloud at $4.39 is closer to Lambda's price than to Vast.ai's. The trap is the idle volume: 500 GB kept on a stopped pod costs $100 a month at $0.20 per GB, double the running rate.

Its limit: pods are containers, not full VMs, Community Cloud still depends on hosts, and support is the weak point users mention most, which our RunPod review scores 2.6 out of 5. The full grid is in our RunPod pricing breakdown, and if RunPod later disappoints too, see our RunPod alternatives.

Standout features
  • Community and Secure clouds
  • Serverless endpoints
  • One-click templates for ComfyUI and vLLM
+Pros
  • ✓RTX 4090 at $0.34/hr, near Vast.ai
  • ✓Pods and serverless in one account
  • ✓Per-second billing
−Cons
  • ✗Idle volumes billed at $0.20/GB/month
  • ✗Community Cloud still depends on hosts
  • ✗Support often slow
RunPod vs Vast.ai
CriterionRunPodVast.ai
RTX 4090$0.34/hr Community, $0.89/hr SecureFrom $0.31/hr
H100 SXMFrom $2.69/hrFrom $1.53/hr
Managed tierSecure CloudSecure Cloud filter on hosts
Verdict

Pick RunPod if you want Vast.ai's RTX prices with templates and serverless on top. If what drove you away was hosts going down, pay for Secure Cloud or move to Lambda; if it was the bill, stay on Vast.ai.

Try RunPod →
6
Another marketplace, with full VMs

TensorDock

3.3/5

TensorDock is a GPU marketplace like Vast.ai, with one technical difference that matters: you rent a full KVM virtual machine, not a container. You get root, your own drivers and kernel modules, and Windows if you need a GPU desktop. Hosts list servers across 45 GPU models, from tier 3 and 4 data centers to converted mining rigs, and you can start with a $5 balance.

Headline prices: H100s from $2.25 an hour, A100 SXM4 from $1.80, RTX 4090 from $0.35, and consumer GPUs from $0.12. Each host sets its own price, so the same GPU shows several rates, and storage is billed on top of the compute price. There are no ingress or egress fees. Its docs still describe spot instances as a beta, but a pending request on its feedback board asks to bring them back, so check the dashboard first. Support works to a 24-hour response standard.

Price order against Vast.ai, same card: TensorDock is higher on both cards most people compare. The H100 at $2.25 sits $0.72 above Vast.ai's $1.53 offer, and the RTX 4090 at $0.35 is four cents above Vast.ai's $0.31, so 100 hours cost $35 against $31. You do not move to TensorDock for a lower bill, you move for the VM.

The trap that bites hardest: when your TensorDock balance reaches $0, your servers are automatically deleted, disk included, where Vast.ai first stops the instance and allows a short negative-balance grace period, sized on your average daily spend, before deletion. And because it is also a marketplace, the reliability problems that push people off Vast.ai are the same here. We cover what replaces it in our TensorDock alternatives.

Standout features
  • Full KVM VMs with root
  • Windows GPU machines
  • Start with $5
+Pros
  • ✓Real VM instead of a container
  • ✓Consumer GPUs from $0.12/hr
  • ✓No ingress or egress fees
−Cons
  • ✗Servers deleted when the balance hits $0
  • ✗Spot tier in beta, if offered at all
  • ✗24-hour support standard
TensorDock vs Vast.ai
CriterionTensorDockVast.ai
H100From $2.25/hrFrom $1.53/hr
MachineFull KVM VMDocker container
Balance at $0Servers deletedStopped, short grace period
Verdict

Pick TensorDock only if you need a full VM or Windows with a GPU on a marketplace budget. If you leave Vast.ai for reliability, TensorDock has the same host model; choose Lambda or Thunder Compute instead.

Try TensorDock →
At a glance

Vast.ai alternatives compared

The six alternatives side by side. The reference point is Vast.ai's cheapest live offer: H100 SXM from $1.53 an hour, RTX 4090 from $0.31. The 'edge over Vast.ai' column says what you get by moving, since on price alone Vast.ai is hard to beat. Spot capacity is the exception, such as Nebius's preemptible H100 from $0.79. Prices are per GPU-hour, on demand, before storage and taxes.

Best forEdge over Vast.aiFree planTeam sizeVisit
1LambdaBest for 1 to 8 data-center GPUs that stay upOwn fleet, 1x to 8x H100 machines that do not change hosts4.0/51x H100 SXM $4.89/hr, A100 40GB $1.99/hr—ML and research teamsVisit →
2Thunder ComputeBest cheap GPU VM from one operatorOne operator, disk free while the VM is off, free egress3.9/5H100 SXM $3.20/hr, A100 80GB $1.09/hr—1 to 20 peopleVisit →
3HyperstackBest for per-minute VMs with no bandwidth feesFixed data-center prices per minute, no bandwidth fees3.9/5H100 PCIe $2.50/hr, A100 $1.35/hr—Startups to mid-size teamsVisit →
4ModalBest for GPU work that comes in burstsPay only the seconds a GPU works, no idle instance3.9/5H100 about $3.95/hr, A100 80GB about $2.50/hr, per second✓Developers and AI startupsVisit →
5RunPodClosest to Vast.ai, with a managed tierRTX 4090 within cents of Vast.ai, plus Secure Cloud and serverless3.7/5H100 SXM from $2.69/hr, RTX 4090 $0.34/hr (Community)—Individuals to mid-size AI teamsVisit →
6TensorDockAnother marketplace, with full VMsFull KVM VM with root or Windows, not a container3.3/5H100 from $2.25/hr, RTX 4090 from $0.35/hr—Individuals and labsVisit →

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

Buyer's guide

How to choose a Vast.ai alternative

A GPU cloud rents you a graphics card in someone's data center by the hour, so you can train or run AI models without buying the hardware. They come in three kinds. Marketplaces (Vast.ai, TensorDock, RunPod Community Cloud) list machines owned by many hosts: lowest prices, uneven quality. Operator clouds (Lambda, Thunder Compute, Hyperstack, RunPod Secure Cloud) run their own fleet at fixed prices. Serverless platforms (Modal) start a GPU only when your code runs. Crusoe and other larger clouds aim at teams signing for clusters rather than someone leaving Vast.ai. CoreWeave sells its HGX H100 as an 8-GPU instance at $49.24 an hour, Nebius charges $4.50 per H100-hour, and DigitalOcean, which absorbed Paperspace, $4.41. AWS, Google Cloud and Azure sell the same GPUs inside a full cloud, which pays off when your data and credits already live there. Production platforms (Northflank, Baseten) deploy your model as a service, with autoscaling, logs and metrics. Northflank lists an H100 at $2.74 an hour and can run in your own cloud account (BYOC).

The unit is the GPU-hour: one GPU rented for one hour. Billing per second, per minute or per hour only changes how the last fraction is rounded. On-demand means nobody can take the machine while you pay; interruptible or spot means you pay less and accept being stopped when demand rises. Storage is billed per GB per month, and the question that changes the bill is whether it still runs while the instance is stopped. Egress is what you pay to move data out.

Before you compare hourly rates, demand four answers: the price of the same GPU model (an H100 PCIe is not an H100 SXM), what a stopped machine costs, what happens to your disk when the balance hits zero, and who owns the hardware. On Vast.ai itself, supply depends on the card. Its pricing pages rate availability high for the RTX 4090, RTX 5090 and A100 SXM4, and low for the H100 SXM and PCIe. 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 credit covers about 7.6 H100 hours, so a weekly fine-tune or a demo can run for nothing. A forgotten Vast.ai instance at $1.53 an hour costs more than that over one weekend.

One GPU most of the day

Take a plain VM on a fixed price. For 100 hours on an A100 80GB, Thunder Compute costs $109 and Hyperstack $135. On a marketplace, check what the same 100 hours cost with the host you would actually pick, not the cheapest one.

Multi-GPU training

Rent the machine, not the cards one by one. Lambda's 8x H100 at $36.72 an hour gives one node with 22 TiB of local SSD. Finding 8 free H100s on one marketplace host is luck, not a plan.

Tight budget, patient jobs

Stay on Vast.ai interruptible at half the on-demand price or less, and checkpoint every 30 minutes to storage you control. Delete instances you no longer use, since a stopped one keeps billing disk.
  • Is the hourly price for the same GPU model and memory size?
  • Is the price for the GPU alone, or with CPU, RAM and disk?
  • What does a stopped instance cost per GB of disk?
  • What happens to your data when the balance reaches zero?
  • Does the provider own the hardware or list hosts?
  • Is egress free, capped or billed per GB?
  • Can you snapshot an instance and restart it elsewhere?
  • Do you get a VM with root, or a container?
  • Is there an uptime SLA, and what does it pay back when a run is lost?
FAQ · 9 questions

Vast.ai alternatives · FAQ

  • Is Vast.ai reliable enough for production?

    Only with the right settings. Vast.ai itself labels on-demand instances best for production and interruptible ones best for batch training, since an interruptible instance can be stopped by a higher bid. On-demand still runs on a host's machine, so for anything users wait on, add the Secure Cloud filter: its security docs recommend Secure Cloud certified providers only, vetted data centers with ISO 27001 certification and Tier 3/4 standards.

  • How much do Vast.ai prices change?

    Often: Vast.ai states that its prices reflect the offers available right now and update hourly, so the rate you saw yesterday is not a quote. If you need a price that holds, convert an on-demand instance to a reserved one, paid upfront, for up to 50% off on 1, 3 or 6 month terms. It keeps the same priority as on-demand, so nobody can outbid you, but the machine is still a host's.

  • Which Vast.ai alternative can keep GPUs in Europe?

    Nebius is the clearest pick: its H100s run in eu-north1, in Finland, at $4.50 per GPU-hour on demand. On Vast.ai itself, the search filters include the host's location, so you can keep a rental inside Europe. Hyperstack has a Norway region, NORWAY-1, but its public flavor list places only RTX A4000 cards and CPU machines there. Ask for the region of the exact card before you sign a data-residency clause.

  • How do I close my Vast.ai account and delete my data?

    Destroy every instance and volume first, since a stopped instance keeps billing its disk until it is destroyed. If you own a team, transfer it or it is deleted with the account. Then ask the Support Chat to complete the deletion, and request a refund of unspent credit there too: spent credit is never refunded, and credit bought with crypto is not refunded at all.

  • Is a production platform like Northflank or Baseten a Vast.ai alternative?

    Yes, if you serve a model rather than train one. These platforms deploy your code as a service and scale it with traffic, with logs and metrics built in, so nobody watches an instance at night. You pay for that: Baseten bills an H100 at about $6.50 an hour, Northflank at $2.74, against $1.53 for Vast.ai's lowest offer.

  • What happens when my Vast.ai balance runs out?

    Your instances are stopped automatically and the GPU is released to other renters, which interrupts whatever was running. The disk is kept and still billed, even while your balance is negative, for a short grace period sized on your average daily spend; without a saved payment method, instances and data are then scheduled for deletion. A credit card on file is charged to cover the negative balance, which protects the instances.

  • Why is my Vast.ai instance so slow to download data?

    Because bandwidth belongs to the host, not to Vast.ai. Each offer lists its own upload and download speed and its own transfer price, and the real speed can fall far below the listing: a 2026 user saw about 1 MB/s on a machine advertising 2,010 Mbps. Run a speed test in the first minute, and destroy the instance if it is slow.

  • Can I run multi-node distributed training on Vast.ai?

    Yes, within limits. Vast.ai can build an overlay network for instances on machines in the same physical cluster, so NCCL sees one private LAN; instances on different hosts stay isolated behind NAT. RoCE or InfiniBand takes a cluster request, currently limited to A100, H100 and H200 machines, and private networking works only on Docker-based templates, not VMs. Without that step, Lambda's 1-Click Clusters start at 16 H100s, for 2 weeks to a year, at $6.16 per GPU-hour.

  • How do I move my work off Vast.ai?

    Copy first, then destroy. Push checkpoints and datasets to object storage or straight to the new provider with rsync, save your environment as a Docker image or requirements file, and check that the copy opens before you delete the instance. A stopped instance keeps its disk, but restarting it to copy data can stall if its GPU went to another renter, so copy while the instance still runs.

Hack'celeration Lab

Get the next breakdown in your inbox

No spam. Unsubscribe anytime.