In September 2026 the median on-demand rate is $6.49 per GPU-hour across 21 providers, according to GetDeploying, with specialist clouds at $5.99 to $6.82 and the hyperscalers at $14 to $16. The cheapest verified in-stock rate is $3.75, spot runs about $4.40, and 36-month reserved commitments have been quoted as low as $2.25. One B200 costs about $4,700 a month at the median; an eight-GPU server about $38,000.

About $7.99 per GPU-hour at the median on demand, with specialist clouds from $7.10 and hyperscalers up to about $17.80. The reserved median is $5.71, and 48-month commitments have been quoted as low as $3.13. One B300 costs about $5,800 a month at the median; an eight-GPU server about $47,000.

The B300 is the Blackwell Ultra refresh of the B200: 288 GB of memory against 180 GB, 1,400 W against 1,000 W, the same 8 TB/s bandwidth, and about a dollar to a dollar and a half more per GPU-hour. It helps when a model or context does not fit on a B200; it does not help when the bottleneck is bandwidth.

For training, usually yes: a B200 costs about twice an H100 per hour but does roughly two and a half to three times the work, so it is cheaper per result at the medians, and reserved discounts widen the gap. For fine-tuning and inference that fits in 80 GB, the H100 at half the hourly price is often still cheaper per job.

For inference on a model that fits in 141 GB, the H200 at $4.40 is usually the cheapest option. For training, or inference that needs FP4 throughput or more than 141 GB, the B200 wins. The B200's FP4 precision roughly doubles inference throughput over FP8, which matters at scale.

The same GPU runs about $6 on specialist GPU clouds and $14 to $16 on the hyperscalers, which bundle networking, storage and their own margins. Spot and small providers go lower still. Supply is the other factor: with only 16 percent of B200 listings confirmed in stock, providers with capacity charge for it.

Yes. Eight B200s on a monthly commitment is a standard offer from specialist clouds and bare metal providers, at roughly $38,000 a month at the median, and it is often the fastest route to Blackwell capacity. We place single-server monthly deployments as well as reserved clusters.

An eight-GPU HGX B200 server draws roughly 14 kW, and an HGX B300 server 16 kW and up, from 1,000 W and 1,400 W per GPU respectively plus CPUs, memory, networking and fans. That is air-coolable at two or three servers per rack, but beyond what most facilities built for 5 to 15 kW racks can serve.

Over three years at full utilization, owning an eight-GPU B200 server and hosting it costs about $550,000, against about $890,000 to rent reserved at typical discounts and $1.36 million on demand. But the lowest quoted 36-month reserved rates come in near $470,000, cheaper than owning, without the capital or the facility.

Harder than the rate cards suggest. Only 16 percent of B200 and 11 percent of B300 listings are confirmed in stock, according to AIMultiple, and allocation varies by provider and month. Quoting several providers at once, and getting the delivery date in writing, is how deployments actually land.

Probably, following the pattern of every previous generation, though the timing depends on how fast Rubin ships and how strong demand stays. B200 on-demand rates rose about 16 percent in the year to September 2026 because supply lagged demand. A long reservation signed today should carry an upgrade path or an exit for that reason.

Nothing. The provider you choose pays us from its channel budget, the same way it pays its own sales team, and the rate is not marked up. If the best answer is a different GPU or the provider you already use, we say so.