A price is a fact about a seller, not about a chip — and no fact in AI infrastructure has more sellers quoting it further apart than the price of a GPU-hour. The object of this piece is deliberately narrow: one rack, the NVIDIA GB200 NVL72 — per NVIDIA's own specification page, 36 Grace CPUs and 72 Blackwell GPUs in a rack-scale liquid-cooled design acting as a single 72-GPU NVLink domain with 130 TB/s of scale-up bandwidth, 1,440 PFLOPS of NVFP4 tensor-core throughput (sparse), and 13.4 TB of HBM3E behind 576 TB/s of GPU memory bandwidth. One rack, one silicon generation, one NVLink topology. Three of the largest GPU landlords on earth publish a per-GPU-hour rent for it today, on pages anyone can open. They disagree by a factor of 2.6.
The three numbers, each from the seller's own public page (checked November 10, 2026). CoreWeave — the specialist neocloud that IPO'd on this exact business — lists GB200 NVL72 on-demand at $42.00 per hour per instance on both its North America and Europe pricing tables, and its own footnote defines the instance as two GB200 superchips, each with one Grace CPU and two Blackwell GPUs — four GPUs, which makes the rate $10.50 per GPU-hour (its published per-GPU "inference" rate matches at $10.50). AWS publishes EC2 Capacity Blocks for ML pricing with the u-p6e-gb200x72 UltraServer at $761.904 per hour ($10.582 per accelerator) for 72 × B200, US East (Dallas) Local Zone — and the table helpfully notes that reservation prices "are updated regularly based on trends in supply and demand," with the next scheduled update in January 2027. Azure does not print a page; it exposes a retail price API, and a direct query against prices.azure.com for the Standard_ND128isr_NDR_GB200_v6 SKU in East US returns $108.16 per VM-hour for Linux on-demand — the VM carries four Blackwell GPUs (192 GiB each) per Microsoft's size documentation, so $27.04 per GPU-hour. Microsoft's same documentation specifies that 18 of these VMs compose one 72-GPU GB200 NVL72 rack (1.4 ExaFLOPS FP4, 13.5 TB shared HBM3E, 130 TB/s NVLink). Same rack; three official rents.
How far apart is "the same price"?
Run the annualization (ours; arithmetic shown). One rack is 72 GPUs × 8,760 hours = 630,720 GPU-hours per year. At CoreWeave's $10.50: $6.62M per year. At AWS's $10.582: $6.67M per year. At Azure's on-demand $27.04: $17.06M per year. The gap between the Azure on-demand anchor and the AWS/CoreWeave band is roughly $10.4M per rack-year — which, at JLL's published 2025 construction cost of $10.7M per megawatt (see our modular-delivery analysis), is approximately the cost of building one entire megawatt of data center. Two buyers renting physically identical silicon pay a difference that funds a building.
The committed tiers are where the anchor starts to tell the truth. The same Azure API returns $606,388 for a 1-year reservation and $1,250,676 for a 3-year reservation on that 4-GPU VM. Per GPU-hour: on-demand $27.04 → 1-year $17.30 (−36%) → 3-year $11.90 (−31% again; −56% versus the list). Follow the decay and note where it lands: $11.90 is within 13% of AWS's capacity-block rate ($10.582) and CoreWeave's on-demand rate ($10.50). Azure's own accountants, pricing their own silicon against their own customers' commitments, converge on the same band the specialist clouds quote to anyone with a credit card. The list price is an anchor for the uninformed; the commitment tier is the market — the identical lesson our colocation analysis drew from Digital Realty's renewal spreads, where headline asking rents moved 1.5% in half a year while renewals repriced 25.4% in one quarter.
What does AWS's own table say about slices?
The inversion is printed on the same AWS page. The generational ladder for 8-GPU slices reads: H100 (p5) $5.970 per GPU-hour → H200 (p5en) $7.895 → B200 (p6-b200) $14.208 → B300 (p6-b300) $16.146. Now set the rack-scale GB200 against it: $10.582. The newest-generation 8-GPU B300 slice rents 53% above the full 72-GPU GB200 rack. The same-generation B200 slice rents 34% above the rack built from that very GPU. Renting slices of the newest silicon costs more per GPU-hour than renting the entire previous-generation NVLink domain — because what you are actually buying in the rack is not 72 chips, it is the NVL72 topology: the 130 TB/s domain that makes 72 GPUs behave as one. Capacity blocks price the topology as a unit; slices price the chip plus the inconvenience of not having the topology. Scale is not a volume discount here — it is a different product.
What is the utilization arithmetic?
Two ratios decide your real cost, and neither appears on any price page (ours). First, utilization: the rack bills whether or not it computes. On Azure on-demand, a rack running 60% of the hours pays an effective $45.07 per used GPU-hour; at 90% utilization, $30.04. The committed tiers amortize the same way — which is why a 3-year reservation at $11.90 against 60% utilization costs you $19.83 effective, more than the specialist's on-demand list at full utilization. Second, granularity: the three products are not perfectly fungible. AWS's capacity block is a scheduled reservation window (reservation fee paid up front, OS fee on top) — a slot you book, not an always-on rental. Azure's VM is a 4-GPU slice inside the NVL72 rack, with 4× 400 GB/s InfiniBand for scale-out — cross-VM work pays the network, not the NVLink domain. CoreWeave's instance is bare-metal-shaped. Before comparing dollars per GPU-hour, compare what the hour actually buys.
Who should care, and how. If you are buying in 2027 planning cycles, the discipline this piece argues for is the one from our rental and colocation analyses: record the commitment-tier price, not the list. A team quoting "GPU rental costs about $27 an hour" and a team quoting "$10.50–$10.58" are describing the same rack; the difference between them is not market conditions but whether their budget was built from anchors or from commitments. And if you are the kind of buyer who can commit — the 3-year Azure number at $11.90 is the hyperscalers quietly conceding that the specialists' posted band is the real market price.
What are the caveats?
Every number here is a published list or API rate, checked November 10, 2026, and each is a moving target. AWS states that capacity-block reservation prices update regularly with supply and demand (next scheduled update January 2027); Azure's API rates are retail list and enterprise agreements sit below them; CoreWeave's on-demand is the top of its own ladder (its reserved terms are cheaper and unlisted). We could not verify a public list price for Google's a4x (GB200) instances on an open pricing page — it is absent from this comparison by the same rule that governs every piece on this forum: if the number is not on a page anyone can check, we do not print it. Note also that AWS labels the GB200 rack's accelerators "B200" — the NVL72's GPUs are Blackwell-generation B200 dies in a rack-scale package, so the cross-SKU comparisons above are like-for-like at the silicon level. Finally, the three products differ in what surrounds the hour — scheduling model, OS fees, egress, the boundary of the NVLink domain — and none of those differences show up in a per-GPU-hour figure. The spread is real; the products are not perfectly identical; treat the 2.6× as a bound on how far a list price can drift from the market, not as a pure arbitrage quote.
Comments