NVIDIA Blackwell GPU Cloud Pricing: H100 to B300 and GB300 Costs

The market for renting NVIDIA's data-center GPUs has quietly become one of the most consequential price battlegrounds in enterprise technology. As of late September 2026, aggregate trackers covering dozens of providers show on-demand rates for the workhorse H100 starting as low as $1.30 per GPU-hour, while the newest Blackwell Ultra parts — the B300 and the rack-scale GB300 — command medians near $7.88 and list prices as high as $18 per GPU-hour51513. For any organization deciding where to run training or inference, the spread between the cheapest neocloud and a hyperscaler allocation can now exceed a factor of two for the same silicon, making procurement strategy as important as architecture choice.

What the GPUs Are, and What They Cost to Rent

The current NVIDIA lineup spans three generations. The Hopper-era H100 (80GB HBM3) and H200 (141GB HBM3e) remain the mainstream workhorses. Blackwell brought the B200 with 192GB of HBM3e and roughly 8 TB/s of memory bandwidth, and Blackwell Ultra extended it with the B300, which pairs 288GB of HBM3e per GPU with 8 TB/s of bandwidth and FP4 support on fifth-generation Tensor Cores1210. At the top of the stack, the GB200 and GB300 NVL72 racks fuse 72 Blackwell GPUs and 36 Grace CPUs into a single NVLink domain aimed at frontier-scale training and very large model serving13.

Across tracked marketplaces, the H100's median on-demand rate sits around $3.38 per GPU-hour, with the cheapest listing at $1.305. The H200 — a memory upgrade rather than a new architecture — runs from about $2.09 to a median near $4.40, and specialized trackers put its market median around $3.82 per GPU-hour52. The B200 spans roughly $3.75 at the low end to a median of $6.79 across more than thirty providers, with hyperscaler allocations on constrained capacity reportedly reaching $16 per hour or more527. The B300's picture is the clearest sign of where pricing is heading: a September 2026 comparison of 13 providers with priced on-demand configurations puts the median at $7.88, the cheapest listing at $6.50, and reserved contracts down to $5.99 or even $4.25 per GPU-hour15. Spot instances, when available, dip below $415.

The rack-scale systems are a different market entirely. CoreWeave, often the reference point for GPU infrastructure, lists its GB200 NVL72 at roughly $10.50 per GPU-hour, versus $6.16 to $6.30 for its HGX H100 and H200 offerings and $8.60 for the B2004. GB300 capacity is scarcer still: the lowest listed on-demand price found is $18 per GPU-hour from Oracle Cloud, with most configurations available only on a reserved, request-a-quote basis13. A single GB200 NVL72 rack costs roughly $2 million to $3 million to buy outright, with some quotes running to $3.4 million1.

Why Inference Economics Are Driving the Premiums

The reason the newer GPUs command such steep premiums is not raw FLOPS — it is memory. The B300 carries 288GB of HBM3e against the H100's 80GB, which means trillion-parameter models can be kept resident with fewer partitions and less inter-GPU communication overhead10. For high-concurrency reasoning inference, that translates directly into more tokens served per dollar, and providers market the B300 explicitly on that basis10. The H200 occupies a similar logic: it rents at a modest premium over the H100 in purchase terms, but commands materially higher cloud rates because its 141GB memory pool lets larger models run on fewer GPUs per job13.

Benchmark-driven analysis puts the effective economics in sharper relief. SemiAnalysis's AI Cloud TCO model rates the B300 at roughly $2.26 per chip-hour at hyperscalers, $2.52 at neoclouds and $3.00 at the retail tier — rates that, combined with measured throughput, produce cost-per-million-token figures for inference workloads11. Those wholesale-style tiers are well below the sticker on-demand prices, which suggests that much of what enterprises pay at list is margin and risk premium rather than the underlying cost of the silicon. One wholesale broker quotes B300 clusters at $7.00 to $12.00 per GPU-hour, with hyperscaler Blackwell rates starting at $12.00 on-demand and wholesale deals saving up to 30%16.

Purchase Prices Explain the Rental Spread

Buying outright tells a parallel story. A new H100 card costs roughly $25,000 to $40,000 depending on variant and reseller, and used units have traded as low as $8,20012. The B200 carries an estimated street price near $40,000 to $45,000 per GPU, with a full 8-GPU HGX B200 server around $390,000 and a DGX B200 appliance near $500,000 to $515,000, though OEM 8-GPU systems have been quoted from $280,00017. The B300 is pricier per card at approximately $53,000, with DGX B300 systems clustering between $300,000 and $350,000 at the low end — and European configurations running from roughly €535,000 to over €1.4 million once RAM, support tiers and software bundles are included149. Notably, NVIDIA has not published a fixed list price for the DGX B300, leaving integrator quotes to define the market, and lead times stretch eight to twenty weeks9.

The implication for enterprises is straightforward: at roughly $7.88 per GPU-hour for a B300 versus about $53,000 to buy, the break-even point is on the order of thousands of hours — far beyond what an inference workload with elastic, spiky demand would ever burn on a single card. That is the core case for elastic pay-as-you-go capacity: bursty inference workloads can rent spot B300 time at under $4 an hour when available, while steady-state training with high utilization may justify ownership or long reservations15.

Where the Reporting Diverges

The sources do not fully agree, and the disagreements are themselves informative. GMI Cloud's pricing table, which reflects a provider's own marketed rates, lists H100 from $2.00, H200 from $2.60, B200 from $4.00 and GB200 NVL72 from $8.00 per GPU-hour — noticeably below the multi-provider medians3. Thunder Compute and Tech-Insider put on-demand B300 between $7.10 and $17.80, with managed DGX stacks as high as $18 and a 48-month reservation driving the floor to $3.13712. Cyfuture's survey of specialist GPU clouds found a tighter band of $6.94 to $8.55, with RunPod lowest at $6.94 and spot near $3.6714. Single-provider pages like VESSL's $7.38 reserved rate and inference.sh's $7.91 low with offers running to $72.06 for premium configurations bracket the same territory108.

The most likely reading of this divergence: the neocloud market has genuinely compressed prices toward the cost of capital and power for older parts, while hyperscaler rates bundle managed services, availability guarantees and enterprise support — and GB300 rack capacity remains scarce enough that spot and on-demand listings frequently sell out, as trackers note1315. Notably, some listings at the extreme low end (Fal.ai's $4.49 custom B300 pricing, for instance) are unverified or sold through negotiated deals, so headline floor prices should be treated skeptically15.

The Strategic Takeaway

NVIDIA's cadence of announcements — Hopper, Blackwell, Blackwell Ultra, and the NVL72 rack systems — has produced a tiered rental market in which each memory-bandwidth jump resets the price of inference. The H100 is now a commodity, renting near the cost of older A100-adjacent tiers; the H200 is a pragmatic bridge for long-context workloads; the B200 and B300 are the inference performance-per-dollar frontier; and the NVL72 racks are a niche for frontier training that only a handful of cloud platforms can even provision3413.

For enterprises, the practical playbook emerging from the data is threefold: benchmark candidates on cost per million tokens rather than per-hour rates11; arbitrage the spread between on-demand, reserved and spot tiers, where spot B300 capacity at under $4 an hour can undercut hyperscaler on-demand by 75%157; and treat rack-scale systems as capacity to be contracted months in advance rather than rented on demand. In a market where the same GPU can vary threefold in price between providers, the cost-effective platform is less a brand name than a function of workload shape, contract length and how aggressively a buyer is willing to shop.