Same H100, three times the price: GPU cloud rates vary up to 6.5x across 28 providers

A price-tracking service that monitors 28 GPU cloud providers has published a live snapshot showing that rented accelerators still have no settled market rate: an identical card is routinely listed at roughly three times the price on one platform as on another, and the lowest offer can shift within the same day. The comparison, posted on dev.to by FastGPU on September 27, 2026, covers marketplaces, neoclouds, serverless platforms and the hyperscalers it tracks — AWS, Azure, Google Cloud and Oracle — ranked cheapest first.

The spread across providers

According to the dev.to post, the widest gap in the snapshot is on a small card: an NVIDIA L4 with 24 GB of memory rents at $0.11 per GPU-hour on Vast.ai while Google Cloud lists the same card at $0.71, a 6.5x difference. On data-center flagships the spread clusters between 3.0x and 3.6x:

• H100 SXM (80 GB): $1.79/hr from Cudo Compute against $5.38/hr at Google Cloud — 3.0x, with 22 providers carrying the card.
• H200 (141 GB): $2.60/hr at GMI Cloud against $7.91/hr at AWS — 3.0x across 18 providers.
• B200 (192 GB): $3.69/hr at DeepInfra against $11.28/hr at Google Cloud — 3.1x across 16 providers.
• B300 (288 GB): $4.89/hr at DeepInfra against $17.80/hr at AWS — 3.6x across 12 providers.
• A100 80GB: $0.89/hr at DeepInfra against $3.22/hr at Google Cloud — 3.6x.

One card defies the pattern: the RTX PRO 6000, a 96 GB workstation-class Blackwell part, rents at $0.93/hr on Vast.ai and $0.97/hr on Google Cloud, effectively at parity. FastGPU frames the hyperscaler premium as payment for the surrounding platform — identity and networking controls, managed services, committed-use discounts and startup credits — and argues it should be a deliberate choice rather than a default.

Spot rates drop lower, with strings attached

Interruptible capacity undercuts on-demand pricing again, per the post: H100 SXM from $0.98/hr, H200 from $1.84/hr and B200 from $1.77/hr. Two caveats follow. Spot capacity can be reclaimed at any moment, which suits checkpointed training and offline batch work but not a live endpoint. And the cheapest spot listings are often whole nodes: FastGPU's dataset shows the H200 spot floor on AWS and the B200 spot floor on Google Cloud each require eight GPUs, so the real entry cost is eight times the per-GPU figure.

Blackwell and consumer cards

The snapshot also indicates that NVIDIA's newest silicon is no longer exclusively a contact-sales product: B200 is listed on demand by 16 providers, B300 by 12 and GB200 by three, starting at $8.00/hr at GMI Cloud. On the AMD side, the MI300X rents from $1.71/hr at TensorWave, with no hyperscaler listing at snapshot time.

For small models, marketplace hosts keep consumer cards inexpensive: an RTX 4090 with 24 GB from $0.14/hr and an RTX 5090 with 32 GB from $0.21/hr, both on Vast.ai. FastGPU notes that 24 GB fits a 7B-parameter model in fp16 or a 13B model quantized to 4-bit, but cautions that marketplace machines are individual hosts, so reliability should be checked and anything irreplaceable kept off the box.

Open data, with a disclosure

The figures come from a continuously refreshed dataset FastGPU publishes under CC BY 4.0, downloadable as JSON or CSV without an API key, with daily history archived under a Zenodo DOI (10.5281/zenodo.22842387) and an MCP server exposing list_gpu_prices and match_workload tools for agent builders. Two caveats for readers: the post is first-party data from a company operating the comparison tool, and FastGPU discloses that some provider links on its site are affiliate links while stating its rankings are ordered strictly by price. The open license at least means the numbers can be independently checked.

Why it matters

For teams budgeting AI compute, the core lesson is that a GPU's hourly price is a property of the vendor, not the hardware — a several-fold spread on identical silicon means provider choice can matter as much as card choice. The fine print decides the real cost: whole-node minimums, egress, storage, cold-start billing and whether a listed price has stock behind it. And because marketplace floors shift as hosts list or delist machines, the durable artifact here is less the snapshot than the openly licensed live dataset, which teams can query directly or wire into their own tooling rather than trusting any single day's numbers.