AWS raised EC2 Capacity Block for ML prices for the second time in 2026, effective July 1: P6-B300 now runs $14.04 per accelerator-hour, P6-B200 runs $12.355, and P5, P5e, P5en, and P4de all went up roughly 20% alongside them, according to AWS's own Capacity Blocks pricing page. That's on top of a ~15% hike AWS pushed through on January 4. If your reserved GPU bill jumped twice this year and you're wondering whether that's just how AWS pricing works now, it is, and it's worth seeing what the same guaranteed capacity costs elsewhere.
This post breaks down both hikes, the full new rate card across every Capacity Block instance family, what a Capacity Block actually locks you into versus a Reserved Instance or spot, and a direct comparison against Spheron's live B200 and B300 rates. For the equivalent breakdown on AWS's H100 line, see our AWS P5 H100 pricing guide, and for the fuller cross-provider B200 picture, see B200 cloud pricing across providers.
Quick Answer: New AWS Capacity Block Rates at a Glance
AWS EC2 Capacity Blocks for ML now cost, per accelerator-hour in US regions: $14.04 for P6-B300, $12.355 for P6-B200, $6.865 for P5en, $5.97 for P5e, $5.191 for P5, and $2.214 for P4de. These are the rates published on AWS's Capacity Blocks pricing page after the July 1, 2026 increase, and they apply uniformly whether you reserve one accelerator or a full node.
Why AWS Raised Capacity Block Prices Twice in 2026
Capacity Blocks are AWS's reserve-ahead product for GPU instances: you commit to a fixed block of accelerators for a fixed time window, pay a fixed price upfront, and get a guarantee that the hardware will be there when your window starts. That guarantee is now getting more expensive on a schedule.
The January 4 Hike: P5e and P5en Go Up 15% Uniformly
AWS raised EC2 Capacity Block prices roughly 15% on January 4, 2026, over a weekend and without advance customer notification, according to The Register, which documented the increase on p5e.48xlarge and p5en.48xlarge. p5e.48xlarge went from $34.61/hr to $39.80/hr in most regions; p5en.48xlarge went from $36.18/hr to $41.61/hr. US West (N. California) saw a steeper jump, with p5e.48xlarge landing at $49.75/hr.
That January move mattered beyond the dollar figure. It was reported as the first time in AWS's roughly 20-year EC2 history that the company raised list prices on a GPU instance pricing model instead of cutting them, breaking a pattern customers had come to take for granted, according to ByteIota. Two decades of "AWS prices only go down" ended on a Sunday.
The July 1 Hike: P6-B300 and P6-B200 Go Up ~20%
Six months later, AWS did it again. Effective July 1, 2026, Capacity Block prices rose approximately 20% across P6-B300, P6-B200, P5, P5e, P5en, and P4de, per AI Weekly. All other EC2 pricing was left alone; this was specifically a reserved-GPU-capacity increase. The new headline numbers: $14.04/hr per accelerator for P6-B300, $12.355/hr for P6-B200.
Two hikes in six months, both landing on the same product line, is not noise. It's the new baseline for anyone budgeting reserved GPU capacity on AWS through the rest of 2026.
AWS's Explanation vs What Analysts Are Actually Seeing
AWS gave nearly the same explanation both times. On the July increase, an AWS spokesperson said: "EC2 Capacity Blocks for ML pricing vary based on supply and demand patterns, as described on the product detail page. This price adjustment reflects the supply/demand patterns we expect this quarter," per AI Weekly.
Corey Quinn, the cloud economist behind Last Week in AWS, didn't buy the supply-and-demand framing when AWS gave it after the January hike: "This was AWS updating the published base rates on their pricing page. That's a policy decision, not supply/demand," he said, per AI Weekly's reporting, which cites InfoQ's original coverage of the January increase. His point is straightforward: a demand-driven price would fluctuate by region and by moment, the way spot markets do. AWS instead moved the published base rate up by a flat percentage across every region at once in January, then did the same thing again in July. That's a rate card change, not a live market signal.
Behind both hikes sits a simple capital story. Amazon committed roughly $200 billion to AI infrastructure capex for 2026, disclosed by CEO Andy Jassy on the February 2026 earnings call, with the bulk of it going into AWS data centers, power, chips, and networking, according to reporting on the announcement. Nvidia GPU supply hasn't kept pace with that spend, and AWS is now using its Capacity Block rate card as a lever on the gap. For more on how that supply squeeze is reshaping neocloud financing too, see our Nvidia neocloud backstop financing piece.
The New Rates: P6-B300, P6-B200, P5, P5e, P5en, and P4de Compared
Here's the full Capacity Block rate card as of July 1, 2026, per accelerator-hour, sourced directly from AWS's pricing page:
| Instance Family | GPU | US Regions ($/accelerator-hr) | Other Regions | 8-GPU Full Node ($/hr) |
|---|---|---|---|---|
| P6-B300 | B300 (Blackwell Ultra) | $14.04 | GovCloud $14.625 | ~$112.32 |
| P6-B200 | B200 (Blackwell) | $12.355 | GovCloud $12.870, Mumbai $12.355 | ~$98.84 |
| P5en | H200 | $6.865 | EU/APAC $6.241 | ~$54.92 |
| P5e | H200 | $5.97 | Consistent globally | ~$47.76 |
| P5 | H100 SXM5 | $5.191 | EU/APAC/Australia/Brazil $4.720 | ~$41.53 |
| P4de | A100 80GB | $2.214 | Consistent | ~$17.71 |
Full-node figures assume an 8-GPU block, which is how P5, P5e, P5en, P6-B200, and P6-B300 instances are configured (.48xlarge); P4de.24xlarge is also an 8-GPU shape. AWS says these reservation prices are "updated regularly based on trends in supply and demand," with the next scheduled update slated for October 2026, per the same pricing page. Budget for a third adjustment before year-end.
AWS list prices change without notice. The rates above reflect AWS's published Capacity Blocks pricing page as of 15 Jul 2026 and may have changed since. Check aws.amazon.com/ec2/capacityblocks/pricing for current figures.
Capacity Blocks vs Reserved Instances vs Spot: What You're Actually Locking In
A Capacity Block, a Reserved Instance, and a Spot Instance are three different products wearing the same "EC2 GPU" label, and confusing them is how budgets go sideways.
A Reserved Instance or Savings Plan discounts a flexible hourly rate over a 1 or 3-year commitment. You get a lower rate, but no guarantee of physical hardware availability on any given day; if AWS is capacity-constrained in your region, your reservation doesn't jump the queue for launch. A Spot Instance is the opposite trade: no commitment and the lowest price, but AWS can reclaim the instance anytime with minimal notice, and P5/P6 spot capacity is scarce enough in practice that it barely functions as a planning tier.
A Capacity Block sits in between, but closer to Reserved than Spot in what it buys you: you pay a fixed total price upfront for a fixed block of GPUs during a fixed, non-interruptible time window, as short as a few hours or as long as several weeks. What you're actually purchasing isn't a discount, it's a guarantee that specific hardware will be sitting there when your window opens, which is why AWS prices it at a premium over on-demand rather than a discount.
Why Capacity Blocks Cost More Than They Sound Like They Should
The premium is the point, not a bug. AWS is selling certainty in a market where Nvidia allocation is the binding constraint, and certainty has a price. But two back-to-back hikes on the same certainty product in six months means the cost of that guarantee is moving faster than most teams' budgeting cycles. A block you priced out in January at the 15%-higher rate is now another 20% more expensive on top of that, without your workload changing at all.
Capacity Blocks also stack the same hidden costs every AWS GPU instance carries regardless of billing mode: EBS storage for anything beyond ephemeral NVMe scratch, data egress at $0.09/GB, cross-AZ transfer for multi-node jobs, and a support tier if you need fast escalation on a reservation issue. Our guide to avoiding unexpected AWS costs walks through how those line items compound on top of the sticker rate.
What the Same Guaranteed Capacity Costs on Spheron
Spheron doesn't sell fixed-window reservations the way AWS Capacity Blocks do. Instead, B200 and B300 are available as standing on-demand and spot inventory, priced per GPU and billed per minute, so you get the guaranteed-capacity outcome (the hardware is there when you launch) without pre-buying a specific future window or committing to a fixed block size.
B200: AWS Capacity Block vs Spheron On-Demand and Spot
| Metric | AWS P6-B200 Capacity Block | Spheron B200 On-Demand | Spheron B200 Spot |
|---|---|---|---|
| $/hr per GPU | $12.355 | $7.37 | $2.69 |
| 8-GPU node, $/hr | ~$98.84 | ~$58.98 | ~$21.52 |
| 8-GPU node, monthly (720 hrs) | ~$71,165 | ~$42,469 | ~$15,496 |
| Commitment | Fixed block, non-interruptible | None, per-minute billing | None, per-minute billing |
| Savings vs AWS Capacity Block | Baseline | ~40% cheaper | ~78% cheaper |
For live B200 availability and rates, see B200 GPU rental on Spheron.
B300: AWS Capacity Block vs Spheron On-Demand and Spot
| Metric | AWS P6-B300 Capacity Block | Spheron B300 On-Demand | Spheron B300 Spot |
|---|---|---|---|
| $/hr per GPU | $14.04 | $9.02 | $5.81 |
| 8-GPU node, $/hr | ~$112.32 | ~$72.16 | ~$46.48 |
| 8-GPU node, monthly (720 hrs) | ~$80,870 | ~$51,958 | ~$33,466 |
| Commitment | Fixed block, non-interruptible | None, per-minute billing | None, per-minute billing |
| Savings vs AWS Capacity Block | Baseline | ~36% cheaper | ~59% cheaper |
Spheron aggregates capacity from 5+ providers on the backend, which is part of why on-demand availability holds up without requiring you to pre-buy a fixed window. For the full 288GB vs 192GB spec and throughput breakdown between the two GPUs, see our B300 vs B200 cost-per-token guide.
Pricing fluctuates based on GPU availability. The prices above are based on 15 Jul 2026 and may have changed. Check current GPU pricing → for live rates.
A team running an 8-GPU B300 job full time for a month goes from roughly $80,870 on an AWS Capacity Block to roughly $51,958 on Spheron on-demand, a difference of about $28,912 in a single month with no fixed reservation window required. That gap is what's pulling reserved-capacity budgets off AWS right now.
How to Move Reserved AI Workloads Off AWS Without Losing Guaranteed Access
The reason teams buy Capacity Blocks in the first place is to avoid the moment where they need GPUs and AWS doesn't have any. Moving off doesn't mean giving up that guarantee, it means getting it a different way:
- Confirm your actual reservation window. Most Capacity Block buyers overshoot: they reserve a week of B200s because the training run might run long, then pay for six idle days. On-demand billing per minute removes that overshoot tax entirely.
- Check spot tolerance for checkpointed jobs. If your training pipeline checkpoints regularly, spot pricing (roughly 59-78% below the AWS Capacity Block rate on B200/B300) is worth testing before defaulting to on-demand.
- Validate NCCL and networking config outside EFA. Multi-node B200/B300 jobs built around AWS's
aws-ofi-ncclplugin and EFA-specific interface naming need standard NCCL over InfiniBand or TCP/RDMA on non-AWS infrastructure; this is usually a config change, not a rewrite. - Re-check total cost, not just the hourly rate. Factor out AWS egress and cross-AZ charges, which don't exist as a line item once you're off AWS; see hyperscaler hidden costs compared for the fuller breakdown.
- Keep a reliability comparison in the decision, not just cost. Reserved capacity exists partly as insurance against outages as much as against scarcity. Our AWS outage cost breakdown covers what a single major incident costs teams that had no fallback provider.
Documentation for standing up GPU instances outside AWS's console flow is in the Spheron docs.
FAQ
How much did AWS raise Capacity Block prices in 2026?
Twice. AWS raised EC2 Capacity Block for ML prices by roughly 15% on January 4, 2026, then by roughly 20% again on July 1, 2026, covering P6-B300, P6-B200, P5, P5e, P5en, and P4de. All other EC2 pricing was unchanged both times.
What does a P6-B300 Capacity Block cost now?
$14.04 per accelerator-hour in US regions as of July 1, 2026 (AWS GovCloud runs $14.625). An 8-GPU p6-b300.48xlarge block works out to roughly $112.32/hr for the full node.
What does a P6-B200 Capacity Block cost now?
$12.355 per accelerator-hour in US regions and Mumbai as of July 1, 2026 (AWS GovCloud runs $12.870). An 8-GPU p6-b200.48xlarge block works out to roughly $98.84/hr for the full node.
Is a Capacity Block the same as an AWS Reserved Instance?
No. A Reserved Instance or Savings Plan discounts a flexible hourly rate over a 1 or 3-year term with no fixed start or end time. A Capacity Block reserves a specific block of GPUs for a specific, fixed time window, bought in advance at a fixed total price, non-interruptible, and non-cancelable. You're not buying a discount, you're buying a guaranteed slot.
Why did AWS say it raised Capacity Block prices?
AWS says Capacity Block pricing "varies based on supply and demand patterns" and that each adjustment "reflects the supply/demand patterns we expect this quarter." Cloud economist Corey Quinn pushed back on that framing after the January hike, calling it "a policy decision, not supply/demand" since AWS updated the published base rate uniformly across every region rather than letting prices move with actual regional demand. The same uniform-rate pattern repeated in July.
How much cheaper is Spheron than an AWS B200 or B300 Capacity Block?
At current live rates, Spheron B200 on-demand runs about 40% below the AWS P6-B200 Capacity Block rate per GPU, and Spheron B300 on-demand runs about 36% below the P6-B300 rate. Spheron spot pricing is cheaper still: roughly 78% below AWS B200 Capacity Block pricing and 59% below B300.
Locked into an AWS Capacity Block window that got more expensive twice this year? Spheron runs the same B200 and B300 hardware on-demand with per-minute billing and no fixed reservation window.
Spheron B300 instances → | Compare GPU pricing → | Get started on Spheron →
Frequently Asked Questions
Twice. AWS raised EC2 Capacity Block for ML prices by roughly 15% on January 4, 2026, then by roughly 20% again on July 1, 2026, covering P6-B300, P6-B200, P5, P5e, P5en, and P4de. All other EC2 pricing was unchanged both times.
$14.04 per accelerator-hour in US regions as of July 1, 2026 (AWS GovCloud runs $14.625). An 8-GPU p6-b300.48xlarge block works out to roughly $112.32/hr for the full node.
$12.355 per accelerator-hour in US regions and Mumbai as of July 1, 2026 (AWS GovCloud runs $12.870). An 8-GPU p6-b200.48xlarge block works out to roughly $98.84/hr for the full node.
No. A Reserved Instance or Savings Plan discounts a flexible hourly rate over a 1 or 3-year term with no fixed start or end time. A Capacity Block reserves a specific block of GPUs for a specific, fixed time window, bought in advance at a fixed total price, non-interruptible, and non-cancelable. You're not buying a discount, you're buying a guaranteed slot.
AWS says Capacity Block pricing 'varies based on supply and demand patterns' and that each adjustment 'reflects the supply/demand patterns we expect this quarter.' Cloud economist Corey Quinn pushed back on that framing after the January hike, calling it 'a policy decision, not supply/demand' since AWS updated the published base rate uniformly across every region rather than letting prices move with actual regional demand. The same uniform-rate pattern repeated in July.
At current live rates, Spheron B200 on-demand runs about 40% below the AWS P6-B200 Capacity Block rate per GPU, and Spheron B300 on-demand runs about 36% below the P6-B300 rate. Spheron spot pricing is cheaper still: roughly 78% below AWS B200 Capacity Block pricing and 59% below B300.