AWS Capacity Blocks GPU Rates Rise ~15% on October 7

Effective October 7, 2026, Amazon EC2 Capacity Blocks for ML charge more for reserved NVIDIA accelerators. AWS published a single new hourly rate per family across all available regions: P6-B300 $16.146 ($16.819 GovCloud), P6-B200 $14.208 ($14.801 GovCloud), P5en $7.895, P5e $6.866, P5 $5.970, P4de $2.546, P4d $1.696. Compared with the tables still on that same page today, those US-listed NVIDIA families move about 15%.

This is a run-cost event, not a market-color event. If you already planned a reserved GPU block for a training run, a fine-tune window, or a burst of H100/H200/B200/B300 capacity, the invoice changes. On-Demand and Savings Plans do not move with this notice. The reservation fee is locked when you buy, even if the block starts after October 7. That is the mechanism. Use it or pay the new list.

Industry recaps this week have called the move a fourth consecutive quarterly hike on reserved GPU blocks. AWS's own page does not number the series. It says reservation prices are dynamic, update with supply and demand, and were already scheduled to update in October 2026. We cite the official rates below and compare them to the tables AWS still shows as current. We do not treat the "fourth quarter in a row" label as something Amazon itself wrote.

This is the same NVIDIA-hour squeeze explainx.ai covered when Nebius raised GPU rental rates 20%: different vendor, different SKU, same pressure on anyone who rents accelerators instead of owning them.

TL;DR — what people are asking

The official October 7 rates (cite this page, not a recap)

AWS's announcement on the Capacity Blocks pricing page is the primary source. Quote it as written: reservation prices "are dynamic and are updated periodically based on supply and demand." Effective October 7, 2026, hourly rates per accelerator across all available AWS Regions will be the figures in the table below. "All other EC2 Capacity Block prices, as well as prices on fixed pricing models like On-Demand and Savings Plans, remain unchanged."

Instance-hour math is just eight times the accelerator-hour on the 48xlarge / 24xlarge shapes: a US p5.48xlarge that shows $41.528 today would be $47.76 at $5.970 × 8. A p6-b300.48xlarge that shows $112.32 today would be $129.168 at $16.146 × 8.

What did not change (read this before you re-quote)

AWS split the SKU on purpose. Capacity Blocks are a reservation fee charged up front when you schedule the block, plus an operating-system fee while instances actually run. Linux OS on these shapes is $0.000 in the OS table; RHEL, SLES, and Ubuntu Pro still add their own hourly line. That OS table is not the October 7 NVIDIA hike.

Three other buckets stay on the old numbers unless AWS updates them separately:

1. On-Demand accelerated instances — different product, different price list.
2. Savings Plans — AWS grouped them with "fixed pricing models" that remain unchanged.
3. Capacity Block families not named in the October 7 sentence — today's page still lists P6e GB200 UltraServers (Dallas Local Zone, $10.582 per B200 on the u-p6e shapes) and Trn1 / Trn2 Trainium blocks. Those sit under "all other EC2 Capacity Block prices remain unchanged."

If your finance sheet has one line called "AWS GPUs," split it. A team that lives on On-Demand P5s is not in this notice. A team that buys Capacity Blocks for a guaranteed H100 cluster is. Mixing those is how people either panic-buy the wrong SKU or miss a 15% reservation bump while staring at an unchanged On-Demand calculator.

The lock sentence is the other operational detail: "Your Capacity Block is charged at the prevailing rate at the time of purchase, even if the Capacity Block is scheduled to start after the price is updated." Capacity Blocks can be reserved up to eight weeks ahead (AWS's product page). A November training window purchased on October 6 is still billed at today's table. The same window purchased on October 8 is billed at the new list. Start date is not the rate date.

~15% is the US-table story — some regions move more

The ~15% figure is not a slogan AWS printed. It is the ratio of the October 7 all-region list to the current per-accelerator numbers in the US rows of the same page. $16.146 / $14.04, $14.208 / $12.355, $5.970 / $5.191, $6.866 / $5.97, $7.895 / $6.865, $2.546 / $2.214, and $1.696 / $1.475 all land at about 1.15.

That ratio is not universal if your current row is already below the US table. AWS currently shows p5.48xlarge at $4.720 per H100 in several Asia Pacific, Europe, and South America regions versus $5.191 in US East/West. The October 7 sentence is a single P5 rate of $5.970 "across all available AWS Regions." $5.970 / $4.720 is about 26%, not 15%. p5en.48xlarge shows the same pattern: $6.865 in US rows versus $6.241 in several Europe and Asia Pacific rows, with a new all-region $7.895 (about 26% from $6.241).

If you run Capacity Blocks in Tokyo, Mumbai, London, Stockholm, or Sao Paulo on P5/P5en, do not copy a US 15% into the spreadsheet. Re-quote from the announced all-region list. P6-B300 and P6-B200 current tables are already flat across listed commercial regions ($14.04 and $12.355), so those families really do look like a clean ~15% plus the GovCloud premium.

This is also why "fourth consecutive quarterly hike" is a weak planning input even if recaps are directionally right. You cannot budget off a percentage meme. You budget off $/accelerator-hour × hours × GPU count, using the family and region you actually reserve.

Worked cost: one week of reserved H100s vs B300s

Use AWS's own billing shape: reservation hours × instance hourly rate, paid up front, plus OS only for hours you actually run.

Example A — one p5.48xlarge (8 × H100) for 7 days (168 hours), Linux, US table.

• Today: 168 × $41.528 = $6,976.70 reservation.
• From Oct 7: 168 × ($5.970 × 8) = 168 × $47.76 = $8,023.68.
• Delta: about $1,047 for that one node-week, ~15%.

Four nodes for the same week is four times that: ~$4,188 extra versus today's US Capacity Block table. That is a fine-tune or eval burst many product teams actually buy, not a frontier pretrain.

Example B — one p6-b300.48xlarge (8 × B300) for 7 days, Linux, US table.

• Today: 168 × $112.32 = $18,869.76.
• From Oct 7: 168 × $129.168 = $21,700.22.
• Delta: about $2,830 per node-week.

Example C — same week in a current $4.720/H100 region, if the all-region $5.970 list applies as written.

• Today P5 48xlarge: 168 × ($4.720 × 8) = 168 × $37.76 = $6,343.68.
• From Oct 7: still $8,023.68.
• Delta: about $1,680, ~26%.

None of these examples include RHEL ($1.8432/hour on the 48xlarge OS table) or idle waste. Capacity Blocks bill the reservation for the window you reserved. If you terminate early, you do not get the unused reservation hours back; AWS's own worked example on that page charges the full reserved instance-hours and only the OS hours you used. The hike makes idle reserved hours more expensive, which is an argument for tighter windows, not for buying a six-month block "just in case" at the new rate.

A100s are in this list — "first time this year" is not on AWS

Secondary recaps have said A100s are included in a quarterly Capacity Blocks hike for the first time in 2026. Verify against the AWS page, then stop. The October 7 list does include P4d $1.696 and P4de $2.546. Current tables still show those as 8 × A100 (p4d.24xlarge at $1.475/accelerator in N. Virginia, Ohio, Oregon; p4de.24xlarge at $2.214 in Oregon and N. Virginia).

What the live page does not contain is a history of prior 2026 Capacity Block updates. We cannot honestly assert from that URL that A100s were skipped in January or July. We can assert they are in this hike. If you still run production fine-tunes on reserved A100s because the job fits 40GB/80GB and the queue is cheaper than H100, this update hits you. If you already left P4, it is a rounding line.

P4d at $1.696 is still far below P5 at $5.970. The relative gap stays huge. The mistake is treating "old GPU" as "immune to reserved-block inflation." It is not, as of this notice.

Trainium is the family AWS did not put in the sentence

Trainium versus GPU architectures is a compiler and ecosystem choice, not a price-hack. It is still the first alternative to check on this page, because Trn1/Trn2 appear in the current Capacity Blocks tables and do not appear in the October 7 NVIDIA list.

Today's examples on the pricing page:

• trn1.32xlarge — 16 × Trainium — $9.532 per instance-hour ($0.596 per accelerator) in N. Virginia, Ohio, Oregon, Mumbai, Melbourne, Sydney, Stockholm.
• trn2.48xlarge — 16 × Trainium2 — $35.7608 ($2.235 per accelerator) in US East (Ohio).
• trn2.3xlarge — 1 × Trainium2 — $2.235 in Melbourne and Sao Paulo.

$0.596 versus $5.970 is not an apples-to-apples FLOPS comparison. Neuron, XLA, and model support still gate whether a training job can move. For jobs that already run on Trainium in AWS, this notice is a relative-price move: NVIDIA reserved hours go up, Trainium reserved hours stay on the current table until AWS says otherwise. Re-read the page on October 7. "All other" is a legal remainder, not a promise that Trainium is frozen for the rest of the year.

GB200 UltraServers (P6e) are the other remainder. If your cluster is an UltraServer Capacity Block in Dallas, this NVIDIA P6-B200/B300 sentence is not automatically your invoice. Confirm the UltraServer row, not the 48xlarge row.

Same squeeze as Nebius — different invoice line

In September, Nebius raised NVIDIA GPU rental rates 20%, the second hike since May 2026. That was a rental list price at a GPU cloud. This AWS notice is a Capacity Blocks reservation fee at the largest hyperscaler GPU catalog most enterprise ML teams already have an account in.

What matches:

• You pay more per reserved NVIDIA hour for the same silicon generation.
• Supply/demand language from the vendor, not a charity discount.
• On-paper "more GPUs shipping in 2026–2028" does not mean your reserved hour got cheaper this quarter.

What does not match:

• Nebius: rental. AWS here: prepaid Capacity Block, OS billed separately, On-Demand explicitly carved out.
• AWS gives you a purchase-time lock you can use as a calendar tactic. Nebius coverage did not give a clean grandfathering rule.
• AWS published per-accelerator numbers you can put in a cell. Treat recap percentages as a check against those cells, not as the source.

The broader 2026 stack is still RAM and HBM inflation plus Stanford's accelerator bill-of-materials split. Reserved cloud GPUs and local DRAM are not the same market, but they are the same scarcity story hitting two budgets: the team that rents H100s for a week, and the team that tried to buy 128GB of DDR5 for a local box. Throughput-per-dollar on next-gen racks can offset sticker shock later. It does not cancel an October 7 Capacity Block delta on this generation.

Checklist before October 7

1. If you already planned a Capacity Block (dates, family, region, node count) buy it before October 7. The start can be after the 7th. The rate is the purchase timestamp. AWS still allows start dates up to eight weeks out; do not wait for a meeting on the 8th to click purchase if the window is already approved.
2. If you do not have a planned window, re-quote. Pull current vs October 7 accelerator-hours for your family and region. US ~15% is the default headline; P5/P5en outside the US table may be closer to ~26% if the all-region list is applied as written.
3. Consider Trainium only if the job already fits Neuron (or you have time to port). AWS has not listed Trn1/Trn2 in this hike. That is a verified remainder on today's page, not a reason to rewrite a CUDA training stack over a weekend.
4. Do not confuse On-Demand with Capacity Blocks. Unchanged On-Demand is not a discount on reserved blocks. Unchanged Savings Plans do not hedge a Capacity Block you have not bought.
5. Do not confuse UltraServers with 48xlarge P6. P6-B200/B300 48xlarge rates are in the hike. P6e GB200 UltraServer rows are not in the announced sentence.
6. Tighten the reserved window. You pay reservation hours whether the GPUs are sweating or sitting. A 15% higher reservation fee is a 15% higher tax on slack.
7. Re-run rent-vs-API-vs-local. Token and workflow cost work and cloud vs local open-weight tradeoffs still apply. A Capacity Block is for guaranteed cluster hours. If the job is actually an agent loop on an API, this hike is not your line item — unless your inference vendor is about to pass reserved-GPU inflation through.

What people are asking

Can I extend a block I already own and keep the old rate?

AWS documents that you can extend Capacity Blocks, but the October 7 notice is about new reservation prices, not a published grandfathering table for extensions. Treat an extension quote as a new purchase until your console shows otherwise. If the extension is priced at purchase time, buying the extension before October 7 is the conservative move.

Why would AWS raise reserved GPUs and leave On-Demand alone?

Capacity Blocks are the product you use when you must have the GPUs on a date. That demand is inelastic relative to "maybe we will grab a P5 if the region has spare." AWS says these reservation prices move with supply and demand. Leaving On-Demand and Savings Plans unchanged is consistent with "the scarce, date-certain SKU re-prices; the interruptible or commitment SKU does not, this round." It is also consistent with not wanting to shock every notebook that boots a single GPU. Either way, the builder action is the same: identify which SKU you are on.

Does this change what I pay for Claude, GPT, or Bedrock tokens?

Not directly. Frontier APIs have their own price lists. Indirectly, reserved NVIDIA hours are an input to anyone training or serving on AWS GPUs. A 15% Capacity Block bump is one more reason a lab or inference host revisits their margins. Track your vendor's token sheet, not this page, for API spend. Track this page if you are the one who files the AWS reservation.

Should I switch clouds this week because of 15%?

Probably not as a two-day migration. Porting a multi-node training job is a quarter, not a Thursday. What you can do this week is: buy the planned AWS block at the current rate, or get a competing reserved/rental quote (including the Nebius-style GPU clouds) for the same GPU-hours and interconnect assumptions. Comparison shopping is justified. A panic lift-and-shift is not.

Honest limitations

• AWS can still edit the page. These numbers are as published on the Capacity Blocks pricing URL on October 1, 2026. If the banner or tables change before October 7, the banner wins.
• "Across all available AWS Regions" is AWS's wording. We flag the P5/P5en regional flattening as the important reading of that sentence. If AWS keeps regional discounts in the tables after October 7, use the tables.
• We do not have a first-party archive of every 2026 quarterly Capacity Blocks update, so we do not certify "fourth consecutive" or "first A100 hike of the year" as AWS claims.
• Perf/$ versus a new rack is out of scope. Rubin-class throughput stories do not reprice an H100 block you need in November.
• No customer-specific discounts. EDP, private pricing, and credits can dominate list. This post is list Capacity Blocks, not your contract.
• Availability is separate from price. A cheaper purchase-time rate does not create idle B300s in the region you need.

What this means for what you build or pay

ML platform teams with a dated GPU reservation on the calendar: this is a purchase-by-October-6 problem. The CapEx-like hit is the up-front reservation fee, not a surprise On-Demand spike. Lock the rate, then keep the training recipe unchanged.

Teams that only use On-Demand or Savings Plans for accelerators: no price change in this notice. Still split the SKU in docs so the next person does not "save money" by switching a guaranteed job onto On-Demand and losing the cluster.

Teams on reserved A100 (P4d/P4de): you are in the October 7 list. Re-quote. Do not assume "legacy GPU" means "legacy price."

Teams that can run Trainium: current Trn1/Trn2 Capacity Block tables are the remainder. Measure the job. Do not port CUDA for a blog post.

Everyone else renting NVIDIA hours in 2026: put this next to the Nebius 20% rental hike. Two vendors, two products, one direction: reserved and rented GPU hours are not getting cheaper this autumn.

Related on explainx.ai

• Nebius raises GPU rental rates 20% (Sept 2026)
• RAM prices and AI demand: local inference cost
• AI chip architectures: GPU vs TPU vs Trainium
• NVIDIA Rubin NVL72 throughput-per-cost
• Stanford memory prices: DRAM, HBM, NAND
• Optimising costs with generative AI
• Closed-source AI vs local open-source alternatives

Primary source: Amazon EC2 Capacity Blocks for ML pricing (rates quoted as published October 1, 2026). Product mechanics: Capacity Blocks for ML and the EC2 Capacity Blocks user guide.

Capacity Block list prices and the October 7, 2026 effective rates are accurate as of October 1, 2026 against AWS's public pricing page. Confirm the live tables before you purchase.