If you rent GPUs by the hour, here is the one-line state of the market in late October 2026: the floor stopped falling. After a summer where every monthly map showed the Blackwell floor cracking lower, this month the cheapest rates across all three workhorse cards are flat-to-firmer — and the spread worth watching has moved inside the specialty market, not between it and the hyperscalers.

If you read one line: a reliable on-demand H100 is ~$1.99–2.90/GPU-hr, an H200 ~$3.14–5/hr, and a B200 is either ~$3.75 or ~$8 depending on whether you mean the cheapest in-stock rate or the typical one — and the decision that saves you the most money still isn't which card, it's whether your GPU is busy enough to rent one at all.

The price map (published on-demand rates, late October 2026)#

Every figure below is a published on-demand rate, gathered the last week of October 2026 from public pricing pages and multi-provider trackers (getdeploying, Ornn, gpuperhour, Spheron, and vendor pages). GPU pricing moves weekly with supply, region, and commitment — read this for the shape of the market, then confirm on the provider's own page before you commit a dollar.

September's headline was "the B200 floor cracked below $4." October's is quieter and more important: it didn't crack any further. The era of "wait a month, it'll be cheaper" just paused.

What changed since September#

If you read last month's map, here's the delta in three lines:

  1. The floor stopped falling. H100's reliable on-demand floor held near $2 (and the brand-name cheap number, ~$2.50–2.90, is a touch firmer). The B200 in-stock floor is flat at ~$3.75. For the first monthly map all year, the cheap end didn't get cheaper.
  2. H200 firmed. Neocloud floors moved up from ~$2.30 in September to ~$3.14–3.99, and the transacted index sits at $5.24. If you were waiting for the H200 to drift down to the H100's price, that's not happening this month.
  3. The spread moved inside the specialty market. The gap that defined the summer was specialty-vs-hyperscaler (still a real 5–7×). The sharper gap now is floor-vs-median within the neoclouds: a B200 in-stock marketplace rate (~$3.75) versus a named neocloud's on-demand list (~$8) is a ~2× difference — and the cheap number comes with availability and trust caveats the expensive one doesn't.

The decision tree hasn't changed — but there's a new column#

For a year the question was two-sided: specialty cloud or hyperscaler. In October it's three-sided, and the middle is new:

And the oldest rule still beats all three sticker prices: a rented GPU bills 24/7 whether it's busy or not. Below roughly 40–50% duty cycle, a per-token API is cheaper than the cheapest B200 on this page. Work the break-even in Rent a GPU or Call an API? before you rent anything, and if you've decided to rent, reserved vs on-demand is the next number to run.

The through-line ties back to this week's funding race: the capital pouring into frontier labs is being spent on compute commitments measured in the hundreds of billions, and that demand is exactly what's holding the floor up. When the money is buying every available Blackwell, the price of renting one stops falling. That's the paused floor, in one sentence.