Cloud GPU Pricing 2026: RTX 4090, A100 & H100 Hourly Costs

Reviewed September 12, 2026. The active snapshot below uses official provider pages and the SynpixCloud public product API. Historical rows keep their original dates. A missing quote means this review did not establish a reproducible comparable rate; it does not mean the provider has no pricing.

The pricing illusion is that every "$/hour" figure buys the same unit. It does not. A marketplace offer, a community-hosted GPU, a fixed single-GPU instance, a reserved accelerator, and a whole VM can all use hourly language while pricing different resources and risks.

Quick Answer

The September 2026 RTX 4090 references here span $0.34–$0.74/hour across distinct listings and tiers, not a guaranteed available market range. Compare memory, host configuration, and stock before comparing the number; A100 and H100 examples below use different memory sizes and billing units.

The lowest visible number is not automatically the lowest usable system. First match the workload to VRAM and topology; then compare the same pricing model.

How This Snapshot Was Reviewed

This update uses a strict source hierarchy:

1. Official provider pricing or GPU model pages.
2. Official provider documentation explaining the billing model.
3. SynpixCloud's public product API for its own inventory.

Marketplace and deployment-dependent numbers are observations, not reservations. We checked Vast.ai's public site and pricing documentation but did not capture a reproducible RTX 4090 offer; the old $0.35 marketplace reference is not carried forward as a September quote. For Google Cloud, select an actual VM shape and region before calculating a comparable price.

RTX 4090 Pricing: Marketplace, Tier, and Fixed Listing

The three RTX 4090 references answer different questions:

• Vast.ai shows a live marketplace. Hosts set prices, and compute, storage, and bandwidth can be separate cost components.
• RunPod publishes Community and Secure Cloud references. On September 12, its RTX 4090 model page showed $0.34/hr Community and $0.74/hr Secure; the main pricing page also listed RTX 4090 Pods at $0.74/hr. The console is the final deployment quote.
• SynpixCloud exposed a $0.48/hr standard 24GB RTX 4090 listing in the public API on September 12, 2026, with zero available in that observation. This is not the separately listed 48GB RTX 4090PLUS. Check the GPU market before topping up. Default access is SSH; application setup and actual network configuration need checking, not assumptions based on a headline rate.

For the RunPod-only search intent, use the RunPod RTX 4090 price guide. For the marketplace-versus-tier decision, use the Vast.ai vs RunPod comparison.

A100 and H100 Pricing: Do Not Flatten the Unit

Lambda

Lambda's public instance table, reviewed September 12, 2026, makes the unit relatively clear:

The A100 80GB number belongs to an eight-GPU instance even though the table expresses it per GPU-hour. The Lambda pricing guide keeps that configuration detail visible.

AWS Capacity Blocks

AWS Capacity Blocks are reserved accelerator capacity, not ordinary on-demand single-GPU listings. These September 12, 2026 references are for US East (N. Virginia):

AWS states that Capacity Block prices can change with supply and demand and are charged at the rate prevailing when the reservation is purchased.

Google Cloud A2 and A3

Google Cloud's A2 machine family includes A100 GPUs; A3 High and Mega include H100 GPUs. Do not conflate them with A3 Ultra, which uses H200. These are accelerator-optimized VM shapes, so compare the machine price, not a detached GPU-only rate.

There is another constraint hidden by a simple "$/GPU-hour" comparison: Google documents that a3-highgpu-1g, a3-highgpu-2g, and a3-highgpu-4g must use Spot or Flex-start provisioning. Use the official pricing calculator or console for the region and provisioning model you can actually deploy.

730-Hour Planning Math

These calculations use the September 12, 2026 snapshot and multiply hourly references by 730. They exclude storage, bandwidth, tax, failed-job retries, and idle-resource charges. They are not monthly discounts or offered AWS reservation durations.

Use the GPU cost calculator for your actual runtime rather than treating 730 hours as a default workload.

For a 40-hour week, the RunPod Community reference yields $13.60, Secure $29.60, Lambda single A100 40GB $79.60, and Lambda single H100 PCIe $131.60, before the same exclusions. Quote the whole multi-GPU instance when that is the purchasable unit.

Price History

Vast.ai history belongs on the provider comparison page because it is a marketplace observation, not a fixed list-price series.

Decision Reframe

Choose in this order:

1. Fit: Does the workload fit in 24GB, 40GB, or 80GB without spilling?
2. Topology: Is this one GPU, multiple independent GPUs, or an interconnected system?
3. Pricing model: Marketplace, community, fixed instance, reserved capacity, Spot, or Flex-start?
4. Full bill: Compute, storage, bandwidth, reservation commitment, and retry cost.
5. Availability: Can the exact configuration be launched when needed?

Only after those variables match does an hourly price comparison become meaningful.

---

Related pricing guides:

• Lambda Labs GPU Pricing 2026
• RunPod RTX 4090 Price 2026
• Vast.ai vs RunPod RTX 4090 Pricing
• RTX 4090 vs A100 vs H100
• GPU Cost Calculator

---

Official sources reviewed September 12, 2026:

• Vast.ai GPU marketplace
• Vast.ai pricing documentation
• RunPod RTX 4090 model page
• RunPod GPU pricing
• Lambda GPU instances and pricing
• AWS EC2 Capacity Blocks for ML pricing
• Google Cloud GPU pricing
• Google Cloud GPU machine types
• SynpixCloud public GPU product listings