AWS runs more of the internet than any other cloud, and for a team that needs databases, queues, analytics and identity in one account, that breadth is the reason to stay. This article is about a narrower question: where to run AI and GPU workloads when EC2 is not the right home for them.
Three things push teams to look. GPU instance rates are high relative to specialist providers. Capacity for current-generation cards often means a quota request, a service limit increase, or a conversation with an account team. And the configuration overhead – VPCs, subnets, security groups, IAM roles, EBS volumes – is a lot of machinery for a workload whose actual requirement is “one H100 for six hours”.
One note on pricing before we start. AWS GPU instance pages publish specifications rather than rates, and on-demand pricing varies by region and instance family. We checked on 31 August 2026 and could not verify a comparable published hourly figure, so this article does not quote one – use the AWS pricing calculator for your region and compare against the figures below. Every competitor rate here was verified at source on 31 August 2026, and Runpod's own rates are pulled live from our pricing page.
What to look for in an AWS alternative
- Published, self-serve rates. Can you see the price without talking to sales? Several providers on this list publish rates for part of their fleet and gate the rest.
- On-demand versus contract. Some of the most attractive published figures require a 12, 24 or 48 month commitment, or are preemptible. Check which number you are reading.
- Billing granularity. Per-second, per-minute or per-hour. On bursty workloads this matters as much as the headline rate.
- The bottom of the range, not just the top. Plenty of providers sell H100s. Far fewer sell something sensible for a model that fits in 24GB.
- Scale-to-zero. Whether there is a serverless tier, or whether you pay for idle GPU hours between requests.
- Egress. Data transfer is where cloud bills surprise people, and it is the single clearest difference between AWS and most of this list.
- Compliance evidence. Not the badge on the marketing page, but whether SOC 2 reports, BAAs and DPAs are available for your security review.
The eight best AWS alternatives
1. Runpod
Best for: teams whose requirement is GPU compute, and who want the same primitives from prototype through to production.
Runpod is an AI developer cloud rather than a general-purpose one. There are no managed databases, no queues and no data warehouse. The argument is that a team running training or inference does not need the other two hundred services AWS sells, and should not pay the complexity cost of having them.
What is different:
- Per-second billing across Pods, Serverless and Clusters, at published rates, with no ingress or egress fees.
- A native serverless layer that scales from zero to thousands of workers on request volume, with sub-200ms cold starts via FlashBoot. Active workers remove cold starts entirely for steady traffic.
- A real bottom to the range. An RTX A5000 is {{gpu:rtx-a5000}}/hr and an L4 is {{gpu:l4}}/hr on Secure Cloud – and Community Cloud goes lower again, with an RTX 4090 at {{gpu:rtx-4090:community}}/hr.
- No quota requests. Inventory is visible in the console at deployment time across 31 global regions.
- Bring your own container from Docker Hub, GitHub Container Registry or Amazon ECR, alongside templates from the Runpod Hub.
- SOC 2 Type II certified, HIPAA and GDPR compliant, with reports, BAAs and DPAs available for security review.
Limitations: no ancillary services, so a full application still needs a database and object storage somewhere. The company is younger than AWS, which matters to some procurement processes. Runpod is a component of an architecture, not a replacement for an entire cloud account.
Pricing: H100 PCIe {{gpu:h100-pcie}}/hr, H100 SXM {{gpu:h100-sxm}}/hr, H200 {{gpu:h200}}/hr, B200 {{gpu:b200}}/hr, A100 PCIe {{gpu:a100-pcie}}/hr, L40S {{gpu:l40s}}/hr, A40 {{gpu:a40}}/hr. AMD MI300X is $2.39/hr on Secure Cloud (verified 31 August 2026). Full pricing.
2. Google Cloud Platform
Best for: teams anchored in data analytics, or who want TPUs.
If you are leaving AWS but still need a full cloud, GCP is the closest like-for-like. Its distinctive asset is hardware nobody else has: TPUs are Google's own accelerators, and for large-scale training on frameworks that support them they are a genuine alternative to NVIDIA rather than a marketing line. Vertex AI covers AutoML through managed deployment, and GKE handles GPU-backed Kubernetes with less friction than most.
Limitations: a narrower catalogue than AWS, a track record of deprecating smaller services, and quota increases that often route through an account representative. TPUs also require framework support – if your stack leans on CUDA-specific libraries, that migration is not free.
Pricing: published rates vary by region and machine type, with sustained-use discounts, committed-use contracts and Spot VMs. Check the GPU pricing page for your region.
3. Microsoft Azure
Best for: organisations already standardised on Microsoft – Entra ID, Microsoft 365, existing enterprise agreements.
Azure is the other hyperscaler move, and the case for it is usually commercial rather than technical: an existing Microsoft relationship, committed spend to draw down, or identity already living in Entra. Azure Machine Learning is a mature managed platform, and the OpenAI service is a differentiator for teams that want those models with an enterprise contract behind them.
Limitations: the operational surface is comparable to AWS, so this is a lateral move rather than a simplification. GPU capacity in a given region can require quota requests. If your reason for leaving AWS was complexity, Azure will not fix it.
Pricing: published through a calculator rather than a flat rate card, and varies by region and VM family. We could not verify a comparable self-serve hourly figure on 31 August 2026. See our Azure alternatives piece for a fuller comparison.
4. Lambda
Best for: teams who want a maintained ML software stack and a straightforward rate card for large NVIDIA cards.
Lambda is a GPU specialist with a genuine asset in Lambda Stack, its maintained distribution of CUDA, cuDNN, PyTorch and TensorFlow. If you have ever lost a day to a driver mismatch, that is worth something. It also sells physical workstations and servers, which is unusual in this set.
Limitations: the fleet is narrow – five GPU models, with nothing below $0.79/hr and no consumer cards, so a model that fits in 24GB has nothing to rent. There is no serverless tier, so no scale-to-zero. Clusters start at 16 GPUs on a two-week minimum arranged through sales rather than self-serve. Instance access is described in their own wording as first-come.
Pricing (verified 31 August 2026): B200 $6.69, H100 SXM $3.99, A100 SXM 80GB $2.79, A100 40GB $1.99, V100 $0.79, per GPU per hour, plus tax. For comparison, H100 SXM here is {{gpu:h100-sxm}}/hr and A100 SXM is {{gpu:a100-sxm}}/hr.
5. Crusoe
Best for: large training runs where energy sourcing and vertically integrated data centers matter.
Crusoe builds and operates its own data centers with an explicit focus on power sourcing, which is a real differentiator for organisations with sustainability commitments to answer for. The top of its fleet is current – GB200 NVL72, B200, MI355X – and it has built out managed AI services including serverless inference, serverless fine-tuning and self-serve dedicated endpoints. Like Runpod, it charges nothing for ingress or egress.
Limitations: the newest hardware is Contact Sales rather than self-serve, so you cannot price a GB200 or B200 run without a conversation. The published fleet bottoms out at an L40S – there are no consumer or small data center cards.
Pricing (verified 31 August 2026): H200 $4.29, H100 $3.90, A100 80GB SXM $2.30, A100 80GB PCIe $2.00, L40S $1.50, AMD MI300X $3.45, per GPU-hour on demand. Storage runs $0.06–$0.10 per GiB per month depending on type. GB200, B200 and MI355X are Contact Sales.
6. Verda (formerly DataCrunch)
Best for: European data residency, confidential computing, and self-service multi-node clusters.
Verda – which rebranded from DataCrunch, moving from datacrunch.io to verda.com – is a European provider with an unusually complete product set for its size. Serverless containers scale to zero, Instant Clusters give self-service 16x to 64x GPU capacity over InfiniBand without a sales call, and confidential computing variants offer hardware-attested inference. It is SOC 2 Type II and GDPR compliant, and spot pricing runs about 50% below on-demand across the fleet.
Limitations: storage is $0.20 per GiB per month across all tiers, which is expensive next to most of this list – and note the unit, since GiB and GB are not the same. Reserved discounts are modest at short terms: 2% at one month, 3% at three, 4% at six. There are no consumer GPUs and no AMD.
Pricing (verified 31 August 2026): GB300 $8.62, B300 $7.50, B200 $6.11, H200 SXM5 $4.00, H100 SXM5 $3.25, A100 SXM4 80GB $1.79, A100 40GB $1.29, RTX PRO 6000 $1.89, L40S $1.37, RTX 6000 Ada $1.04, RTX A6000 $0.61, V100 $0.17, per hour on demand.
7. DigitalOcean
Best for: teams who want a small, comprehensible platform, and the widest current AMD range on this list.
DigitalOcean built its reputation on being a cloud you can understand in an afternoon, and GPU Droplets extend that to AI workloads. It acquired Paperspace, bringing that GPU platform in-house. It has gone further on AMD Instinct than anyone else here – MI300X, MI325X, MI350X and MI355X are all available – which matters if your stack runs on ROCm.
Limitations: no serverless GPU tier, so no scale-to-zero. And a real trap: powered-off GPU Droplets are still billed, because the resources stay reserved. You have to destroy the instance to stop charges.
Pricing (verified 31 August 2026, following a price change effective 1 August): H100 $4.41, H200 $4.47, L40S $1.57, RTX 6000 Ada $1.57, RTX 4000 Ada $0.76, AMD MI300X $2.59, per GPU per hour on demand. Billed per second with a five-minute minimum.
8. Vultr
Best for: broad geographic coverage from an independent provider, if you can commit to a term.
Vultr operates in more than 20 locations with both cloud GPU instances and bare metal, a clean console and a solid API. Its Blackwell and Grace Hopper capacity is more current than most providers its size, and it carries AMD Instinct alongside NVIDIA.
Limitations: the headline rates need reading carefully. The attractive H100 figure is 24-month contract pricing, the B200 figure is a 48-month contract, and the AMD figures are preemptible instances that can be reclaimed. There is no serverless tier.
Pricing (verified 31 August 2026): genuine on-demand rates are A100 PCIe 80GB $2.397/hr, L40S $1.671/GPU/hr, A40 $1.712/hr, GH200 $1.990/GPU/hr. The H100 at $1.990/GPU/hr requires a 24-month commitment. For comparison, A100 PCIe here is {{gpu:a100-pcie}}/hr, L40S {{gpu:l40s}}/hr and A40 {{gpu:a40}}/hr.
How they compare
| Provider | H100 on demand | Cheapest GPU | Serverless | Egress |
|---|---|---|---|---|
| Runpod | {{gpu:h100-pcie}}/hr PCIe | {{gpu:rtx-a5000}}/hr Secure, lower on Community | Yes, scale to zero | None |
| AWS | Not published as a flat rate | Varies by region | Yes, via SageMaker | Charged |
| Google Cloud | Varies by region | Varies by region | Yes, via Vertex AI | Charged |
| Azure | Calculator only | Varies by region | Yes, via Azure ML | Charged |
| Lambda | $3.99 SXM, plus tax | $0.79 V100 | No | Not published |
| Crusoe | $3.90 | $1.50 L40S | Managed inference and fine-tuning | None |
| Verda | $3.25 SXM5 | $0.17 V100 | Yes, scale to zero | Not published |
| DigitalOcean | $4.41 | $0.76 RTX 4000 Ada | No | Included allowance |
| Vultr | $1.990, 24-month term | $1.671 L40S | No | Included allowance |
All competitor figures verified 31 August 2026. Runpod figures are live. AWS is included for shape rather than price, since it does not publish a comparable flat rate.
How to choose
If you need a full cloud, GCP or Azure are the realistic destinations, and you should move for a specific reason – TPUs, an existing Microsoft agreement, a data platform – not for simplicity. Neither is simpler than AWS.
If you want current hardware at scale and can talk to a sales team, Crusoe and Verda both have credible Blackwell fleets, and Verda will sell you a multi-node InfiniBand cluster without a call.
If GPU compute is the entire requirement, a specialist is cheaper and faster to get running. The differences that tend to matter are the bottom of the range, whether there is scale-to-zero, and whether egress is charged – because those three decide the bill on a real workload more often than the H100 sticker price does.
Most teams end up running both. Leaving application infrastructure on AWS and moving GPU workloads to a specialist is a common pattern; check egress in both directions before you split, because that is usually where the saving gets eaten.
FAQ
Which AWS alternative is cheapest for GPU workloads?
It depends which card you need. At the top of the range Verda's H100 SXM5 at $3.25 is the lowest verified on-demand figure in this set. At the bottom, Verda's V100 at $0.17 and Runpod's Community Cloud rates are the cheapest ways to get a working GPU. The more useful question is usually total cost rather than hourly rate: per-second billing, scale-to-zero and zero egress change the bill more than a few cents an hour on the sticker price.
Can I move an EC2 GPU workload without rewriting it?
If it runs in a container, largely yes – that is the normal case, and every provider here takes standard Docker images. What does not move is anything bound to AWS services: S3 paths, IAM roles, SageMaker pipelines, EFS mounts and CloudWatch instrumentation all need replacing. Containerised training and inference are the easy migration; a deeply integrated SageMaker pipeline is not.
What about AWS Trainium and Inferentia?
They are a genuine cost lever if your workload adapts to them, and nothing on this list is an equivalent – they are AWS-specific silicon. The trade is portability: code tuned for Neuron does not move to another provider. Worth modelling before committing, since it is the strongest technical argument for staying.
Do any of these have a free tier?
Not for GPUs, in any meaningful sense. AWS's twelve-month free tier covers CPU instances and storage, not GPU hours. Several providers offer startup credits – AWS Activate among them – but you should assume GPU time is paid time everywhere on this list.
How much can moving GPU workloads off AWS actually save?
It depends on your utilisation pattern, and anyone quoting a single percentage is guessing. The structural savings are per-second billing on bursty work, scale-to-zero on inference with uneven traffic, and no egress charges if you move data frequently. If your GPUs run near full utilisation around the clock, a reserved instance on AWS may well compete. If they sit idle between jobs, the gap is large.
Get started
Runpod offers on-demand GPUs with no minimum spend and no quota request, billed by the second. See current pricing, or deploy a Pod to test your own workload.
