If you searched "cheapest GPU rental" and got ten different numbers for the same card, that's not noise — it's the actual shape of the market. "Cheapest" means three different things, and they differ by 2–4x on the identical GPU. Here's the one-screen answer, then how to pick the one you actually want.

**The cheapest reliable on-demand rate right now (November 2026):**

**The cheapest sticker (marketplace/spot, with caveats): H100 asks from ~$1.74 (Vast.ai) and spot near ~$1.03 (Spheron); H200 asks from ~$2.25–2.76; B200 single-provider floors as low as ~$3.69**. Real prices — but no availability guarantee.

**And the cheapest option is often *not renting at all.*** More on that below, because it's the one that actually saves you money.

"Cheapest" is three markets wearing one word#

Every figure here is a published rate gathered at the end of October 2026 from public pricing pages and multi-provider trackers (getdeploying, Ornn, Uvik, Singularity, Spheron). They move weekly with supply, region, and commitment — read them for the shape, then confirm on the provider's own page before you commit a dollar.

The confusion in the search results comes from blending three distinct markets:

  1. Marketplace / spot floor — independent hosts and preemptible capacity. The lowest sticker (H100 ~$1.74 ask, ~$1.03 spot), the least certainty. The instance can be reclaimed or simply vanish mid-run.
  2. Reliable named on-demand — a neocloud with an SLA and a support channel. This is the honest "production" number, and it's usually ~2x the floor. On a B200 that's the difference between a ~$3.69 marketplace quote and Lambda's ~$5.50–6.99 on-demand list.
  3. Market median — what a provider picked at random costs. Uvik's October index puts medians at ~$4.20 (H100), ~$5.70 (H200), ~$8.19 (B200). The gap between the floor and the median is simply the tax you pay for not shopping.

Above all three sit the hyperscalers (AWS, Azure, GCP) at roughly $7–14/hr for the exact same card. You're paying 3–6x for the platform, the compliance posture, and the capacity guarantee — not the silicon. For most founders, a neocloud is the right starting point, not a hyperscaler.

"Cheapest" is a trap question. The honest version is "cheapest that I can actually get, reliably, for my workload" — and that answer is almost always the reliable-on-demand number, roughly double the headline floor.

The number that beats every price on this page#

The single biggest lever isn't which card or which provider. It's whether you should rent a GPU at all.

A rented GPU bills 24/7 whether or not it's doing work. A per-token API bills only for tokens. Below roughly 40–50% duty cycle, the API is cheaper than the cheapest B200 here — so for bursty, spiky, or low-volume workloads, the genuinely cheapest option is not renting. Rent metal when you have steady high-volume load, need data isolation, or want a fine-tuned model always warm. Work the break-even in Rent a GPU or Call an API? before you reserve anything; once you've decided to rent, reserved vs on-demand is the next number to run.

Why today's cheapest may be the cheapest for a while#

Through mid-2026, the rental floor fell most months — the rational move was to wait a month for a lower number. Last month's map noted the floor had gone flat. This month the direction turned slightly firmer, and the reason is supply, not demand.

In late-October earnings, Microsoft and Google declined to raise their AI capex, and Microsoft said delayed third-party chips will push some data-center capacity into the first half of 2027 — the detail we unpack in this week's Founder's Wire. When the companies building the data centers can't get chips on schedule, the capacity that would have driven rental prices down doesn't arrive, and the floor holds or firms.

The practical read: stop treating "wait for a lower price" as a strategy. If you know you'll need steady capacity into 2027, the cheapest reliable rate you can lock today — and a reservation on top of it — is probably a better deal than the on-demand number you'll be quoted next spring. Shop the floor, but don't bank on it falling.