The same H100 costs $1.19, $6.88, or $11.06 tonight
Tonight the two biggest clouds on earth joined the panel, and the price of one chip now spans nearly an order of magnitude on a single screen. AWS will rent you a single H100 SXM 80GB on-demand in Virginia for $6.88 an hour - instance p5.4xlarge, one GPU, no commitment, read from AWS's own price feed. Google Cloud's on-demand H100 costs $11.06 per GPU-hour in Iowa, and it comes in exactly one size: eight at a time, $88.49 for the block. At the other end, IO.NET's marketplace feed lists the same H100 SXM at $1.19 per GPU-hour. Same chip, same 80GB of HBM3, one spec sheet. The range from $1.19 to $11.06 is 9.3x.
Two honest caveats before the comparison. The $1.19 floor was not rentable when we verified it tonight - IO.NET's live feed shows every H100 on the network hired out, so that number is a queue, not a machine. And the cloud prices are on-demand rates: nobody running steady production workloads pays them, because commitment discounts sit one negotiation away. The panel tracks on-demand hourly because it is the only rate every provider publishes and the only one that lets you leave after an hour. Read what follows as the market's sticker prices, not its contracts.
What the premium buys
The clouds are not selling chips. They sell adjacency: an H100 inside your VPC, next to the data you already keep in S3 or GCS, under the compliance program your auditors already approved, wired to a fabric built for multi-node training. Google's choice to sell the H100 only as an 8-GPU block is the tell - the A3 product is the interconnect, and the GPUs come with it. AWS's p5 ships with its own fabric for the same reason. If your training corpus already lives in their object storage, the premium can be the cheaper option before a single GPU-hour runs: both clouds bill data leaving their network by the gigabyte, and moving terabytes to a cheaper provider costs real money on the way out.
What it doesn't buy
The chip is the same chip. A single-node job on portable data - fine-tune a model, run a batch of inference, render overnight - sees zero benefit from a VPC, a fabric, or an audit trail, and the market middle is crowded with honest machines. QuickPod's live range opens at $2.50, Zettabyte publishes an on-demand floor at $2.51, gpu.ai's community tier lists the H100 SXM at $3.24, all verified tonight. Against those, AWS at $6.88 is 2.1x the gpu.ai community rate and Google at $11.06 is 3.4x. Against the sold-out DePIN floor the multiples run to 5.8x and 9.3x. Nothing about the premium makes one GPU finish the job faster.
| Model | Specialist floor | AWS | Google Cloud |
|---|---|---|---|
| H100 SXM 80GB | $1.19 (IO.NET, sold out) | $6.88 (single GPU) | $11.06 (8-GPU blocks only) |
| H200 141GB | $3.29 (gpu.ai NVL) | $7.91 | $10.60 (8-GPU blocks only) |
| A100 80GB | $1.23 (gpu.ai) | $3.43 (8-GPU blocks only) | $5.07 (single GPU) |
| L4 24GB | - | $0.80 | $0.71 |
The premium is a price, not a law
The H200 row is the interesting one. AWS rents the 8-GPU H200 block at $7.91 per GPU-hour; Google charges $10.60 for its own 8-GPU H200 block. Two clouds, same chip, 34% apart from each other - and the specialist floor (gpu.ai's H200 NVL, live tonight) sits at $3.29, which makes AWS 2.4x the floor and Google 3.2x. The cloud premium is not a constant of nature. It is a negotiated number that moves per model, per region, and per quarter, which means it can be compared, waited out, and occasionally beaten.
Occasionally beaten is not a figure of speech. AWS's single H100 at $6.88 undercuts VoltageGPU's in-stock H100 at $6.95 - a specialist, today, charging more than AWS for the same GPU. Google's L4 at $0.71 beats AWS's $0.80, and its RTX PRO 6000 Blackwell at $4.50 is nearly half Hinode's $8.05 from-floor, because on workstation-class cards the clouds bought silicon in bulk while the specialists are renting you someone's desktop. The premium is a per-model fact, not a per-provider identity. The only way to see it is row by row, which is the entire point of putting the clouds on the same panel as the floor.
The buyer's checklist
- If the data already lives in S3 or GCS, price the egress before the GPU. Moving terabytes out of a cloud can erase the specialist gap before the first hour runs.
- Need exactly one cloud H100: AWS's p5.4xlarge is the only single-GPU on-demand H100 at a hyperscaler. Google's door opens at eight.
- Single-node job on portable data: the specialist floor wins. Nothing about a VPC makes one GPU faster.
- Multi-node training is what the premium actually pays for - the fabric, not the chips. If the job does not need the fabric, do not pay for it.
- Check stock before trusting a floor. The $1.19 H100 was sold out when we verified it tonight. A price you cannot rent is advertising.
- Running steady-state: everything on this page is the sticker price. Commitment discounts at the clouds and the specialists change the math below every number here, which is exactly the lever the sticker price exists to anchor.
The market used to hide this spread across a hundred pricing pages. Tonight it sits on one screen: the same chip at $1.19, $6.88, and $11.06, each price honest about what it includes. The premium is real and so is the floor. The only mistake is paying one while shopping for the other.