A GPU pricing comparison across hyperscalers usually starts with the one number AWS, Azure, and Google Cloud all publish: an on-demand H100 rate. Buyers treat those three figures as a clean three-way match. They aren't, not once B200 enters the picture. AWS and GCP will each sell a plain 8-GPU B200 instance with nothing else attached to it. Azure won't. Its self-serve Blackwell VM, ND GB200 v6, bundles four B200-class GPUs with two Grace CPUs per VM, so a reader comparing "Azure B200 pricing" against AWS and GCP's B200 rates is comparing a different product, not just a different price. This post puts AWS P5, Azure ND H100 v5, and GCP A3/A4 side by side for both H100 and B200, including what each hyperscaler's reserved and committed-use ladder actually saves, what each one's spot tier publishes, and where the sticker rate stops being the real bill.
For the fuller market picture beyond just the big three, see the GPU cloud pricing comparison covering 5+ providers including neoclouds.
TL;DR: GPU Pricing Comparison Across Hyperscalers for H100 and B200
AWS P5 is the cheapest hyperscaler for H100 on-demand as of 23 Oct 2026, and only AWS and GCP sell a standalone B200 instance.
- H100 on-demand: AWS p5.48xlarge is about $6.88/GPU-hr, GCP a3-highgpu-8g about $11.06, and Azure ND96isr H100 v5 about $12.29, all as of 23 Oct 2026.
- B200 on-demand: AWS p6-b200.48xlarge runs about $14.24/GPU-hr and GCP a4-highgpu-8g about $4.28, as of 23 Oct 2026; Azure sells no standalone B200 VM.
- Reserved discounts: AWS, Azure, and GCP each cut H100 roughly 55-56% on a 3-year term as of 23 Oct 2026; GCP's A4 B200 gets none.
- Spheron: H100 runs $2.64/hr on-demand ($2.20/hr spot) as of 24 Oct 2026, with no published egress fee. See current GPU pricing.
H100 Pricing Head-to-Head: AWS P5, Azure ND H100 v5, GCP A3
All three hyperscalers sell the same underlying chip, NVIDIA's H100 SXM5, and all three sell it only as part of a fixed 8-GPU node. AWS's p5.48xlarge lists at $55.04/hr on-demand in us-east-1, or $6.88 per GPU-hour, per economize.cloud's AWS EC2 pricing tracker. GCP's a3-highgpu-8g lists at $88.49/hr on-demand in us-central1, or $11.06 per GPU-hour, per economize.cloud's GCP Compute Engine tracker. Azure's ND96isr H100 v5 lists at roughly $98.32/hr pay-as-you-go, or about $12.29 per GPU-hour, per Spheron's own ND H100 v5 breakdown, which ran the full SKU and regional analysis this post only has room to summarize.
| Hyperscaler | Instance | On-Demand $/hr (node) | On-Demand $/GPU-hr |
|---|---|---|---|
| AWS | p5.48xlarge (8x H100 SXM5) | $55.04 | ~$6.88 |
| GCP | a3-highgpu-8g (8x H100 SXM5) | $88.49 | ~$11.06 |
| Azure | ND96isr H100 v5 (8x H100 SXM5) | ~$98.32 | ~$12.29 |
Pricing fluctuates based on GPU availability. AWS and GCP figures above come from economize.cloud's pricing trackers checked 23 Oct 2026; Azure's figure is as last checked for Spheron's own ND H100 v5 breakdown. Verify current rates with each provider's own pricing calculator before budgeting.
On this table, AWS is the cheapest hyperscaler for H100 by a wide margin, GCP sits in the middle, and Azure is the most expensive, roughly 79% above AWS for identical hardware. None of these three figures include quota wait time. AWS P5 instances need a service quota increase from a default of 0 vCPUs before they'll launch at all, Azure's NDv5 SKUs carry their own 1-4 week quota wait in most regions, and GCP A3 availability varies by zone. The full AWS breakdown, including Capacity Block pricing and the EFA quota process, is in the AWS H100 pricing guide.
Spot narrows some of that gap, unevenly. AWS P5 spot capacity is scarce, as this blog's own AWS P5 breakdown found: when available it runs roughly 44% under on-demand, about $3.83 per GPU, with high interruption risk since reserved and on-demand buyers get priority for the same pool. Azure's ND96isr H100 v5 has a cheaper Low Priority tier, roughly $2.25 to $3.69 per GPU-hour by region, per this blog's own Azure breakdown, though several regions show zero NDv5 spot capacity on a given day. GCP doesn't publish a flat A3 spot rate either, so no hyperscaler here gives H100 spot buyers a number as reliable as its on-demand rate.
B200 Pricing Head-to-Head: What Each Hyperscaler Actually Publishes
This is where the three-way comparison breaks down, because only two of the three hyperscalers are selling the same thing. AWS's p6-b200.48xlarge (8x B200) lists at $83,170.94/month on-demand in us-east-1, equivalent to $113.93/hr for the node, or $14.24 per GPU-hour, per economize.cloud's AWS EC2 pricing tracker. GCP's a4-highgpu-8g (8x B200) lists at approximately $34.24/hr on-demand, or about $4.28 per GPU-hour, per Spheron's own Google Cloud A4 breakdown, which cross-checked that figure against independent GCP pricing trackers since Google doesn't publish a static A4 rate card. That puts GCP's B200 list price at roughly a third of AWS's for the same chip.
Azure doesn't have a row in this table, because it doesn't sell what AWS and GCP are selling. That's a rack-scale Grace Blackwell building block, architecturally closer to NVIDIA's GB200 NVL72 superchip design than to a plain GPU-only instance, and Microsoft doesn't publish a self-serve hourly rate for it the way AWS and GCP do for their B200 nodes.
| Hyperscaler | What it sells for Blackwell | GPUs per unit | Published on-demand $/GPU-hr |
|---|---|---|---|
| AWS | p6-b200.48xlarge, plain B200 node | 8x B200, no CPU bundle | ~$14.24 |
| GCP | a4-highgpu-8g, plain B200 node | 8x B200, no CPU bundle | ~$4.28 |
| Azure | ND GB200 v6, Grace+Blackwell VM | 4x B200-class GPU + 2x Grace CPU per VM | Not published |
Pricing fluctuates based on GPU availability. AWS's figure above comes from economize.cloud's pricing tracker checked 23 Oct 2026; GCP's figure is as last checked for Spheron's own Google Cloud A4 breakdown. Verify current rates with each provider's own pricing calculator before budgeting.
Spot only exists on one side of this table. AWS doesn't sell P6 B200 spot capacity in any meaningful volume. GCP does: Google lists Spot and Preemptible SKU groups for A4's instances, confirming the billing mechanism exists. Azure has nothing to compare here, since it has no standalone B200 VM to begin with.
A buyer who only wants raw B200 compute, with no CPU superchip attached, has exactly two hyperscaler options right now, not three. If Grace's unified CPU-GPU memory architecture is actually what your training job needs, Azure's ND GB200 v6 is the more relevant comparison, against AWS and GCP's own GB200 NVL72 offerings, not against their plain B200 nodes.
Reserved and Committed-Use Discounts Compared Across All Three
Every hyperscaler here sells a self-service 1-year and 3-year discount ladder for H100, and the discounts land closer together than the on-demand rates do. GCP's 1-year committed use discount brings a3-highgpu-8g to $61.38/hr (a 31% cut from on-demand), and the 3-year CUD drops it to $38.86/hr, a 56% cut, per the same economize.cloud tracker cited above. AWS's p5.48xlarge drops to roughly $38.53/hr on a 1-year Savings Plan (about 30% off) and roughly $24.77/hr on a 3-year reserved term (about 55% off), per this blog's own AWS P5 rate breakdown cited above. Azure's ND96isr H100 v5 drops to roughly $63.40/hr on a 1-year reserved term (about 35% off) and roughly $43.79/hr on a 3-year reserved term (about 55% off upfront), per this blog's own Azure ND H100 v5 breakdown, also cited above.
| Hyperscaler | On-Demand $/GPU-hr | 1-Year Discount $/GPU-hr | 3-Year Discount $/GPU-hr |
|---|---|---|---|
| AWS | ~$6.88 | ~$4.82 (~30% off) | ~$3.10 (~55% off) |
| GCP | ~$11.06 | ~$7.67 (31% off) | ~$4.86 (56% off) |
| Azure | ~$12.29 | ~$7.93 (~35% off) | ~$5.47 (~55% off) |
Pricing fluctuates based on GPU availability. Figures above were checked 23 Oct 2026 against economize.cloud's pricing trackers and this blog's own AWS and Azure rate breakdowns. Reserved and committed-use pricing can change independently of on-demand rates, so confirm current terms with each provider directly before signing a commitment.
Two things stand out. First, the percentage discount is nearly identical across all three at the 3-year mark, so a deep reserved commitment doesn't close the gap between AWS's cheap on-demand rate and Azure's expensive one; it scales all three down together instead. Second, none of these discounts apply to B200. GCP's A4 documentation states that standard committed-use discounts don't apply to that instance family; discounted B200 capacity on GCP runs through a separate AI Hypercomputer reservation path. AWS sells B200 commitment through Capacity Blocks, a fixed total price for a reserved window rather than a synthetic hourly rate, and that capacity can sell out weeks in advance.
A 3-year reserved term is also a real lock-in, not just a discount. You're committing to a specific instance type, in a specific region, for three years, in a market where a new GPU architecture ships roughly every 18 months.
The Hidden Costs Each Hyperscaler Adds That the Sticker Rate Hides
None of the per-GPU-hour numbers above are the real bill. That egress bill compounds with checkpoint downloads, cross-AZ transfer for multi-node training, and storage for model weights, none of which show up in the per-GPU-hour figure a hyperscaler leads with. Spheron's own egress cost teardown breaks the AWS, Azure, and GCP rate cards down to per-GB figures and shows exactly where a $12,000/month GPU bill turns into a $14,200 invoice.
Beyond egress, three other line items show up consistently across all three hyperscalers:
- Block storage. Checkpoint volumes and OS disks run on separate billed storage (EBS on AWS, Premium SSD on Azure, Persistent Disk on GCP), typically $40-80/month per node, more with multiple snapshots.
- Quota and support friction. All three require a capacity or quota request before an H100 or B200 instance launches, and timely help with that often means paying for a support tier too.
- Cross-zone networking. Multi-node training jobs spanning availability zones or regions pay a separate per-GB charge, distinct from internet egress.
Spheron's own pricing page describes a different structure for all of this: live marketplace rates aggregated from data center partners across multiple regions, billed per minute after a 20-minute minimum, with no published egress fee. That removes one whole category of surprise line item from the comparison, though it isn't a complete substitute for a hyperscaler account. Spheron carries no FedRAMP High authorization or HIPAA BAA the way AWS and Azure do, so regulated workloads with a mandatory certification requirement still need a hyperscaler regardless of price. The Spheron API docs cover how the per-minute billing and marketplace model work if you're evaluating it against a committed hyperscaler contract.
Why the Same H100 and B200 Cost Less Outside the Big Three
Every number in this post is a hyperscaler's own published list price, and even AWS's cheapest on-demand H100 rate sits well above what several neoclouds and marketplaces charge for the identical NVIDIA chip. That gap exists because AWS, Azure, and GCP are selling H100 and B200 bundled with a much larger platform: managed services, compliance certifications, a global network backbone, and enterprise support contracts that most GPU renters never touch. A marketplace that aggregates capacity from multiple data center partners carries none of that overhead, and it shows up directly in the per-GPU rate.
On Spheron specifically, H100 SXM5 runs $2.64/hr on-demand and $2.20/hr spot as of 24 Oct 2026, sold one GPU at a time rather than locked to an 8-GPU node, which matters if your job only needs one or two GPUs and would otherwise pay for six idle ones on AWS, Azure, or GCP. B200 runs $10.29/hr on-demand and $5.31/hr spot, also one GPU at a time, though on-demand B200 availability depends on live partner supply and has gone spot-only on at least one recorded date in 2026, so treat the on-demand figure as less guaranteed than AWS's or GCP's published B200 list price. Rent H100 on Spheron or check B200 availability to see which tiers are live right now.
Pricing fluctuates based on GPU availability. Spheron rates above are live as of 24 Oct 2026; the AWS, Azure, and GCP figures elsewhere in this post are each provider's own published list price and may have changed since. Confirm current rates directly with each provider before budgeting.
If quota waits, egress bills, and three-year lock-ins are the actual blocker standing between your team and an H100 or B200, Spheron's marketplace sells both by the single GPU with no quota ticket required.
Frequently Asked Questions
On published on-demand list rates, AWS P5 is the cheapest of the three at roughly $6.88/GPU-hr (p5.48xlarge), against Azure ND H100 v5 at roughly $12.29/GPU-hr and GCP A3 High at roughly $11.06/GPU-hr. Azure is close to 80% more expensive than AWS for the same H100 SXM5 hardware on a pure sticker-price basis. Spheron's H100 SXM5 runs $2.64/hr on-demand ($2.20/hr spot) as of 24 Oct 2026, undercutting all three without a reserved commitment.
Not as a standalone GPU instance. AWS sells B200 through p6-b200.48xlarge and GCP sells it through a4-highgpu-8g, both plain 8-GPU B200 nodes with no other hardware bundled in. Azure's self-serve Blackwell VM is ND GB200 v6, which pairs 4 Blackwell GPUs with 2 Grace CPUs per VM rather than selling B200 on its own, so 'Azure B200 pricing' isn't really a like-for-like comparison with AWS and GCP's plain B200 rates.
All three hyperscalers publish a 1-year and 3-year self-service discount ladder for H100. AWS's 3-year reserved p5.48xlarge rate cuts the on-demand price by roughly 55%, Azure's 3-year reserved ND H100 v5 rate also cuts by roughly 55%, and GCP's 3-year committed use discount on A3 High cuts by 56%. The three land within a few points of each other; GCP's A4 B200 instance is the exception, since standard 1-year and 3-year committed use discounts don't apply to it at all.
The sticker rate never includes data egress, and all three charge for it: AWS at $0.09/GB for the first 10 TB out, Azure at roughly $0.087/GB for the first 5 TB, and GCP in the $0.08-$0.12/GB range depending on tier. Add block storage for checkpoints, cross-AZ or cross-zone transfer for multi-node training, and quota or support fees on top of the per-GPU-hour rate before comparing it to a provider that bundles those in.
