Some links in this article are affiliate links. We earn a commission at no extra cost to you when you purchase through them. Full disclosure.

Choose GPU hosting by the model you actually load, the capacity you can obtain and the operational responsibility you can accept. An application calling a hosted model API usually needs ordinary application hosting instead.

This is a research-based procurement comparison, not our benchmark or a guaranteed live capacity check. Provider order is retained; the old “GPU-first is always cheaper” tier ranking is removed.

Disclosure: We may earn a commission if you sign up through links on this page.

GPU-first providers

RunPod

Pods and serverless endpoints are different buying decisions. Pod pricing documentation directs buyers to the deployment console for the selected GPU quote. Compute and storage billing differ; stopped resources can still incur storage costs. Serverless billing includes billable worker phases, so “pay only for successful requests” is not a safe cost assumption.

For a short training job, check the available GPU/VRAM and persistence policy. For variable inference, test cold starts, minimum/active workers and spend limits. Do not assume that selecting a product guarantees capacity.

Evaluate RunPod for variable inference or short training jobs

No promotional-credit promise is made; verify the selected deployment and billing before funding an account.

Lambda

Lambda’s current pricing matrix separates instance GPU count, hardware variant and cluster commitments. On the retrieved US page, a one-GPU H100 PCIe 80GB instance is listed at $3.29 per GPU-hour; H100 SXM configurations differ. Taxes are additional. A multi-GPU cluster’s per-GPU rate is not the total server bill.

This can be a training or sustained-compute shortlist item when its available instance fits the workload. It is not established here as the cheapest provider.

CoreWeave: limited Vera Rubin production capacity

CoreWeave announced limited production cloud availability of NVIDIA Vera Rubin NVL72 on September 30, 2026. Cognition, the company behind Devin, is its first disclosed production customer. This is select-customer capacity with onboarding and capacity planning, not evidence of unrestricted on-demand availability or a published hourly price.

CoreWeave reports that Cognition measured up to 4.8x total token throughput for SWE-2 inference versus GB200 NVL72 and 3.8x output throughput for its reinforcement-learning workload. These are customer benchmarks published by the provider, not our independent tests or universal gains for every model, token mix and concurrency level. Request a workload-matched evaluation before using them in a cost forecast.

The NVL72 system already includes Vera CPUs. Do not confuse that with a separately orderable CPU-only cloud offering: standalone Vera CPU availability through CoreWeave remains a separate, unconfirmed procurement question. NVIDIA’s Vera CPU announcement describes partner availability, not a guarantee that any particular CoreWeave CPU SKU can be purchased now. For the production improvement layer, see our agent-observability comparison.

General clouds and dedicated servers

DigitalOcean

GPU Droplet documentation lists H100 on-demand at $4.41/hour and 8× H100 at $35.28/hour. GPU billing continues when a Droplet is powered off; destroying it ends the compute reservation. Spot capacity has a separate variable price and interruption implications.

This is useful to evaluate alongside an existing DigitalOcean application stack. It does not prove equivalent region, interconnect, CPU/RAM or availability to the Lambda quote above.

Compare DigitalOcean GPU Droplets for an existing DigitalOcean application stack

Vultr

Use the GPU provisioning documentation and current pricing to select a region and configuration. No numeric hourly price, global-stock promise or SLA percentage is asserted here without a matched quote/contract.

Evaluate Vultr for a region-matched GPU instance

Check current regional capacity and the product-specific service terms before purchase.

IONOS Bare Metal: a different category

IONOS Bare Metal is an existing CPU-server alternative, not evidence of a GPU SKU. Confirm any accelerator explicitly before considering it for a model that needs GPU memory. A CPU workload and GPU inference are not interchangeable merely because both offer dedicated hardware.

AWS, Google Cloud and Azure

Keep these on the shortlist when existing networking, data governance or purchasing agreements matter. Do not assume any has the best stock, highest price or a guaranteed SLA for all instance families. Obtain a region-specific SKU quote and service terms.

Build a like-for-like quote

Item	Required comparison
Accelerator	Exact GPU variant, VRAM, GPU count and interconnect
Instance	CPU/RAM, local disk, persistent volumes and network
Location	Region, quota and actually available capacity
Billing	On-demand versus commitment/spot; minimum billing and idle resources
Reliability	Contractual service scope, exclusions, support and recovery
Workload	Your model, quantization, concurrency and measured latency/throughput

The retrieved Lambda and DigitalOcean H100 figures are quote examples, not a hardware/region-matched savings benchmark. RunPod requires its live console quote; Vultr requires a selected configuration. We remove the old April rate matrix rather than filling gaps with guessed prices.

Promotional credit is one-time account eligibility, not a lower recurring rate. No $5 sign-up-credit promise is retained.

Inference, training or owned hardware?

For inference, compare the cost of warm capacity, startup, autoscaling and storage at your traffic pattern. For training, compare total job duration, checkpoint recovery and GPU interconnect requirements. For an existing cloud customer, include data movement and engineering overhead before switching.

Owned hardware adds acquisition, electricity, maintenance and downtime. The former two-month payback and universal six-to-eight-hour break-even claims are not supported. Use a workload-specific total-cost model; serverless versus dedicated GPU explains the billing distinction.

Availability must be checked in the console or with sales at purchase time. Off-peak timing, spot discounts and prepaid terms are not universal guarantees. Compliance certifications do not automatically make your application compliant.

Before purchase

Confirm capacity, data retention, backups, interruption/recovery and the exact contract. Test a representative job rather than applying a provider/customer benchmark to every model.

Continue with GPU memory planning, inference cost planning, vLLM serving, self-hosted inference, cloud hosting pricing and AI Deployment & Hosting.

Pricing/product scope checked October 1, 2026. No independent performance or account-identity test is claimed.