Build1 publisher2 min readPublished
FastGPU finds hyperscalers charging three times the market floor for H100 and B200 hours
FastGPU's September 27 snapshot of 28 GPU clouds puts the cheapest hyperscaler H100 at $5.38 an hour, 3.0x the $1.79 market floor. That premium buys IAM, managed services and credits, and it is worth paying when a team actually uses them.
The Engineer · Build desk
What happened
- FastGPU, which aggregates prices from 28 GPU clouds, published a snapshot taken on September 27, 2026 at 04:11 UTC, ranked cheapest first.
- L4 showed the widest spread in the table, $0.71 an hour at the cheapest hyperscaler against an $0.11 floor, a ratio of 6.5x.
- B200 is now listed by 16 providers from $3.69 an hour at DeepInfra, B300 by 12, and GB200 by 3 from $8.00 at GMI Cloud.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- cost An eight-GPU B200 job at the cheapest hyperscaler's on-demand rate costs about $60.72 an hour more than at the floor, paid by teams that default to their existing cloud.
- decision Staying on a hyperscaler needs a reason the team can name, such as IAM or managed services it depends on, to justify roughly triple the hourly H100 or B200 rate.
- exposure Chasing the floor puts workloads on marketplace hosts that are individual machines, so host reliability becomes the renter's problem and data that cannot be lost has to live off the box.
The floor in FastGPU's table is the cheapest single listing across its 28-cloud crawl [1]. Its methodology counts "on-demand and community listings, preferring ones that are not sold out" [7]. The hyperscaler column is the cheapest on-demand offer for the same card at AWS, Azure, Google Cloud or Oracle [8]. So each ratio compares one class of provider's list price with the best listing anywhere, and that listing can be a marketplace host [7][8].
I think the 3x holds for a narrower case than the ratio suggests. The comparison is on-demand against on-demand [8]. A team already on a committed-use discount or spending startup credits is being measured against a rate it does not pay [6]. The listing also has to fit the job. The dataset's sample query returns `min_gpu_count` beside `price_usd_hr` [12], and FastGPU says minimum GPU counts, egress, storage and cold-start billing can wipe out a cheap rate [13]. Stock is the last check, because a cheap marketplace listing can be rented out within the hour [10]. "A price with no stock behind it is not a price," the FastGPU post says [14].
On H100 the list-price gap is $3.59 per GPU-hour [1]. The cheaper provider's egress, storage and any extra operations work all have to cost less than that before a move pays off [1][13].
Interruptible capacity takes more off. H100 SXM spot starts at $0.98 an hour [15], about 55% of the on-demand floor [4]. Spot can be reclaimed at any moment. FastGPU limits it to checkpointed training and offline batch jobs, and rules it out for user-facing endpoints [16]. The cheapest spot offers are often whole nodes. In the dataset, the AWS H200 spot floor and the Google Cloud B200 spot floor each require 8 GPUs, so the real entry ticket is eight times the per-GPU price [17].
The premium differs from card to card. On the new RTX PRO 6000, Google Cloud charges $0.97 an hour against Vast.ai's $0.93 [4]. That is about 4%, or four cents an hour for the hyperscaler's IAM and managed services [3].
All of these figures come from one source with a commercial interest in the comparison. FastGPU says some of its provider links are affiliate links, and that its tables and "cheapest" figures are ordered strictly by price [18]. It reads prices from each provider's public API or pricing page [9]. The underlying data is published under CC BY 4.0 as JSON or CSV, with daily history and a Zenodo DOI [19], so anyone can rerun the ratios against a later snapshot.
What to watch
- FastGPU's daily history for H100 SXM: whether the $1.79 on-demand floor persists across snapshots or was a single listing that sold out.
- Hyperscaler committed-use and credit pricing, left out of the on-demand comparison, set against the same floor.
- Whether the RTX PRO 6000's near-parity pricing at Google Cloud shows up on other new cards as hyperscalers list them.