AWS prices its P6e-GB200 UltraServer at $10.582 per B200 accelerator-hour, sold only through EC2 Capacity Blocks for ML and only in fixed 36- or 72-GPU blocks. That single rate survived a price hike that pushed every other AWS GPU Capacity Block higher, and it now sits within pennies of CoreWeave's own GB200 rate. Neither detail shows up when you just read the per-accelerator number off AWS's pricing page.
TL;DR: What AWS's GB200 NVL72 Rate Actually Costs
- Rate: AWS charges $10.582 per B200 accelerator-hour for P6e-GB200 UltraServers, the same on both block sizes, as of 7 Oct 2026.
- Block totals: roughly $381/hr for the 36-GPU block and $762/hr for the full 72-GPU rack, before any term discount.
- Reserved, not on-demand: these sell only through Capacity Blocks, booked up to six months ahead and non-interruptible once the window starts.
- The exemption: AWS's October 7, 2026 hike raised P6-B300, P6-B200, P5, P5e, P5en, and P4de rates, but left this one frozen.
- Spheron reserves GB200 NVL72 from a single 8-GPU node, with pricing confirmed within a business day instead of a rate card. Reserve GB200 NVL72 →
AWS GB200 NVL72 Cost: The $10.582 Per-Accelerator Rate on P6e UltraServers
AWS's P6e-GB200 UltraServer rate is $10.582 per B200 accelerator-hour, and it's the same number regardless of which of the two available block sizes you reserve, according to AWS's own Capacity Blocks pricing page. Multiply that per-accelerator rate by the GPU count in each block and the totals look like this:
| Block shape | GPUs | Per-accelerator rate | Hourly block total | 90-day Capacity Block total |
|---|---|---|---|---|
| u-p6e-gb200x36 | 36 | $10.582/accelerator-hr | ~$380.95/hr | ~$822,856 |
| u-p6e-gb200x72 | 72 | $10.582/accelerator-hr | ~$761.90/hr | ~$1.65 million |
The 90-day totals come from a sizing breakdown of AWS's Blackwell systems by The Next Platform, and they line up with the per-accelerator rate: you're paying for a fixed number of GPUs for a fixed number of hours, upfront, with no discount for partial utilization during the window.
This is the one AWS Blackwell rate that didn't move in AWS's most recent price adjustment. AWS's October 7, 2026 Capacity Block price increase raised rates on P6-B300, P6-B200, P5, P5e, P5en, and P4de, the standard (non-UltraServer) instance families. The P6e-GB200 UltraServer rate wasn't on that list. After the hike, standard P6-B200 Capacity Blocks run $14.208 per accelerator-hour and P6-B300 runs $16.146, both now above the $10.582 GB200 UltraServer rate that's stayed flat since launch. That's an inversion of the usual pattern, where rack-scale, NVLink-connected systems typically cost more per GPU than a standard multi-instance cluster, not less. For the full detail on what else moved in that hike and why, see our breakdown of the AWS Capacity Block price increases.
Pricing fluctuates based on availability and AWS's own rate updates. The AWS figures above are published rates checked as of 7 Oct 2026; check AWS's Capacity Blocks pricing page directly for the current rate before budgeting.
How the AWS Capacity Block Bill Actually Works
The $10.582 per-accelerator rate isn't the whole invoice. AWS's own Capacity Blocks pricing page splits the charge into two separate pieces: a reservation fee and an operating system fee. The reservation fee is the $10.582-per-accelerator rate itself, and it's charged upfront, at the moment you schedule the block, not metered out as the reservation runs. The operating system fee is the second, smaller piece, billed separately at per-second granularity while an instance is actually powered on (SLES is the one exception, billed in hourly minimums). Reserve a 72-GPU u-p6e-gb200x72 block for a 90-day window and the $761.90/hr reservation fee is due as a lump sum before a single GPU-hour of training actually runs; the OS fee is what varies with how much of that window you spend with instances live.
That upfront structure is also why AWS describes Capacity Block pricing as dynamic, moving with fleet-wide supply and demand, even though the specific $10.582 GB200 UltraServer rate held still through the October hike that moved every other Blackwell and Hopper family on the list. A rate that hasn't changed yet isn't contractually frozen, it's just the number that happened to survive the most recent repricing. Check the live page before locking in a multi-month block rather than budgeting off a number in a blog post, this one included.
What You Actually Get: a 72-GPU NVLink Domain, Minimum Block Size, and Lock-In
The $10.582 rate buys access to a genuinely large NVLink domain, not a slice of one. A full u-p6e-gb200x72 UltraServer delivers 360 petaflops of dense FP8 compute and 13.4 TB of total HBM3e across its 72 GPUs, connected by up to 28.8 Tbps of Elastic Fabric Adapter (EFAv4) networking per UltraServer. That's a single coherent domain a model can span without crossing node boundaries.
Three constraints come attached to that number, and they matter more than the headline rate:
- Only two shapes exist. You reserve either the 36-GPU u-p6e-gb200x36 or the 72-GPU u-p6e-gb200x72. There's no smaller unit, so a team that needs 8 or 16 GB200 GPUs still books at least a half rack.
- One location. P6e-GB200 UltraServers are only available in the Dallas Local Zone (us-east-1-dfw-2a), an extension of US East (N. Virginia), per AWS's own launch announcement. There's no regional choice to shop against.
- Fixed windows, booked ahead. EC2 Capacity Blocks for ML reserve anywhere from a single instance up to 64 instances or UltraServers, and the window has to be booked up to six months in advance, per The Next Platform's sizing analysis. Once it starts, it's not interruptible and it's not cancelable.
For the full rack architecture and spec table, our GB200 NVL72 guide covers the hardware in depth rather than repeating it here. And if you're still deciding whether a rack-scale system is the right call at all versus a single Blackwell GPU, the H200 vs B200 vs GB200 comparison is the fork to check before committing to either an AWS UltraServer or a reservation anywhere else.
Why GB200 Isn't an On-Demand SKU Anywhere, Including Spheron
Every GB200 provider worth comparing sells it as a reservation of some kind, not a walk-up hourly rental. AWS calls its version a Capacity Block: you reserve a 36-GPU or 72-GPU UltraServer for a fixed future window, booked up to six months out, non-interruptible once it begins. Spheron's version is a quote: GB200 NVL72 capacity on Spheron reserves from a single 8-GPU node, pricing depends on quantity, commitment length, region, and networking, and the team confirms availability, region, and price within one business day of a form submission. Neither lets you spin up a GB200 instance the way you'd spin up an on-demand H100.
The one partial exception is CoreWeave, which has listed a GB200 NVL72 instance at $42.00/hr for 4 GPUs, a metered rate rather than a pre-booked block, which works out to $10.50 per GPU-hour, a figure we checked against CoreWeave's own price list in our GB300 vs GB200 pricing breakdown. That $10.50/GPU-hr sits about eight cents below AWS's $10.582 UltraServer rate. One is a metered instance rate and the other is a reserved-block rate, so they're not the same product, but the two numbers have effectively converged, and it's a reminder that a per-GPU rate alone doesn't tell you what kind of commitment you're actually buying.
CoreWeave's $42.00/hr rate was checked against its own price list on 16 Aug 2026 and can move independently of AWS's figures; confirm the current rate on CoreWeave's own pricing page before budgeting.
Reserving GB200 Capacity on Spheron vs an AWS UltraServer Commitment
Put the two reservation paths side by side and the difference isn't the rate, it's the shape of the commitment you're signing up for:
| Factor | AWS P6e-GB200 UltraServer | Spheron GB200 NVL72 |
|---|---|---|
| Minimum unit | 36-GPU half rack | Single 8-GPU node |
| Published rate | $10.582/accelerator-hr | None; quote-based |
| How you book | EC2 Capacity Blocks, up to 6 months ahead | Form submission, confirmed within 1 business day |
| Location choice | Dallas Local Zone only | Pricing depends on region and networking requirements |
| Interruptible | No, once the window starts | Set by the reservation terms |
AWS wins on transparency: you can read $10.582 off a public page and know exactly what a 36- or 72-GPU block costs before you commit to anything. Spheron wins on granularity: an 8-GPU GB200 reservation doesn't force you into paying for a half rack you don't need, and quantity, commitment length, region, and networking requirements all factor into the quote rather than being averaged into one flat number. What Spheron doesn't offer is a rate card to comparison-shop against before you fill out the form, and like AWS, it's still a reservation product, not an hourly on-demand GB200 rental; a team wanting zero-commitment rack-scale access by the hour doesn't have that option on either platform. Spheron's broader reserved-commitments catalog covers 8 to 512+ GPUs with 24-48 hour quote turnaround across B300, B200, H100, H200, A100, and several other SKUs, but GB200 isn't listed on that general form; it routes through its own dedicated reservation page instead. For how reservation pricing compares across more of the market, including AWS's own fixed rate card, our breakdown of reserved GPU capacity bidding covers Capacity Blocks, resale markets, and quote-based reservations side by side.
The AWS rate in the table above is published and checked as of 7 Oct 2026; Spheron's is quote-based and varies by quantity, term, region, and networking. Check AWS's Capacity Blocks pricing page and submit a GB200 reservation request for current figures on each side.
A GB200 commitment is a six- or seven-figure decision either way. Whether a published $10.582 rate or a one-business-day quote fits your timeline better, map out your GPU count and term before you book.
Frequently Asked Questions
$10.582 per B200 accelerator-hour, the rate AWS publishes for both its P6e-GB200 UltraServer shapes, the 36-GPU u-p6e-gb200x36 and the 72-GPU u-p6e-gb200x72, sold through EC2 Capacity Blocks for ML in the US East (Dallas) Local Zone. That works out to roughly $381/hr for the half-rack 36-GPU block and roughly $762/hr for the full 72-GPU rack, before any reservation-length discount.
There's no single published rack price, because no provider sells a full rack as one line item at an hourly rate anyone can compare before committing. AWS's $10.582 per-accelerator figure is the clearest published number. CoreWeave has listed a GB200 NVL72 instance at $42.00/hr for 4 GPUs, which works out to $10.50 per GPU-hour. Spheron reserves GB200 capacity by quote rather than a fixed rate card, with the team confirming availability, region, and price within a business day.
Not in the walk-up sense. AWS sells P6e-GB200 UltraServers exclusively through EC2 Capacity Blocks for ML: you reserve a fixed block of 1 to 64 instances or UltraServers for a fixed future time window, booked as much as six months ahead, and the reservation is non-interruptible once it starts. The accelerator-hour rate is how AWS prices that reservation, not evidence that you can spin one up instantly like a standard EC2 instance.
No. AWS's October 7, 2026 Capacity Block increase raised rates on P6-B300, P6-B200, P5, P5e, P5en, and P4de, but left the P6e-GB200 UltraServer rate untouched at $10.582 per accelerator-hour. After that hike, the standard (non-UltraServer) P6-B200 Capacity Block rate rose to $14.208 per accelerator-hour and P6-B300 to $16.146, both now above the frozen GB200 UltraServer rate.
No. AWS splits the charge into a reservation fee and an operating system fee, per AWS's own Capacity Blocks pricing page. The reservation fee, the $10.582-per-accelerator rate, is charged upfront at the moment you schedule the block. The operating system fee bills separately at per-second granularity while the instance is actually running, with hourly minimums for SLES. The headline rate buys the reserved capacity itself, not the complete invoice.
