The price is frozen. Everything else is the start date.
AWS tells you the earliest date a Capacity Block can start when you ask for one, then forgets. Nobody keeps the history. I do: every region that sells them, every P and Trn family, every block length, sampled up to every 15 minutes since late July, recorded.
This is the first readout, covering August 23 to September 14, the period with all 11 regions live and an error rate under 1%. Later issues are weekly.
Open means bookable within three days. Queue means obtainable, you wait. Sold out means no offer before the end of AWS's eight-week booking horizon.
1. Same block, same price, earliest start 48 days apart
The price of a Capacity Block has not moved once in 48 days, across 42 region-and-instance combinations. That isn't luck. AWS's pricing page says it plainly: "The current prices are scheduled to be updated next in October, 2026." Prices reset on a schedule. Between resets, the only thing that moves is when you can start.
And that moves a lot. For a 7-day H100 block, the median gap between the fastest and slowest region on the same day was 48 days. It never fell below 40 in three weeks, and reached 57.
| Instance | Fastest region, Sep 14 | Slowest region, Sep 14 | Median gap, 23 days |
|---|---|---|---|
| H100 (p5.48xlarge) | Ohio, today | Mumbai, 43 d | 48 d |
| H200 (p5en) | Oregon, 3 d | N. California, 42 d | 47 d |
| H200 (p5e) | São Paulo, today | London, 42 d | 45 d |
| A100 40GB (p4d) | Oregon, 12 d | N. Virginia, 53 d | 41 d |
| B200 (p6-b200) | Stockholm, today | Oregon, 28 d | 33 d |
| B300 (p6-b300) | N. Virginia, 5 d | Oregon, 44 d | 24 d |
| H100 x1 (p5.4xlarge) | Oregon, today | Tokyo, 22 d | 24 d |
| A100 80GB (p4de) | Oregon, 18 d | N. Virginia, 49 d | 15 d |
The rate card behind it, unchanged all series:
| Instance | $/accelerator-hour, US | $/accelerator-hour, non-US |
|---|---|---|
| A100 40GB (p4d) | 1.48 | not sold |
| A100 80GB (p4de) | 2.21 | not sold |
| H100 (p5.48xlarge, p5.4xlarge) | 5.19 | 4.72 |
| H200 (p5e) | 5.97 | 5.97 |
| H200 (p5en) | 6.87 | 6.24 |
| B200 (p6-b200) | 12.36 | 12.36 |
| B300 (p6-b300) | 14.04 | 14.04 |
Per hour, a 1-day block and an 84-day block cost the same. No volume tier, no duration discount.
The 9% non-US discount is paid for in waiting. H100 is $4.72 outside the US against $5.19 inside it, but US regions were open on 95% of days with no wait, non-US on 46% with a median 2.6-day wait. The cheaper side of the rate card is the slower side, which is the opposite of what a discount usually means.
2. What is actually scarce
Share of days each instance type was open, queued or sold out, 7-day blocks, all regions:
| Instance | Open | Queue | Sold out | Wait when obtainable |
|---|---|---|---|---|
| Trainium (trn1) | 0% | 0% | 100% | no offer since Aug 11 |
| A100 40GB (p4d) | 2% | 54% | 44% | 13 d |
| H200 (p5en) | 20% | 45% | 35% | 28 d |
| H200 (p5e) | 16% | 54% | 30% | 16 d |
| B300 (p6-b300) | 19% | 61% | 19% | 8 d |
| A100 80GB (p4de) | 0% | 83% | 17% | 22 d |
| H100 (p5.48xlarge) | 67% | 29% | 5% | 0 d |
| H100 x1 (p5.4xlarge) | 16% | 84% | 0% | 15 d |
| B200 (p6-b200) | 24% | 76% | 0% | 20 d |
H200 is the scarce chip right now, not B200. Both H200 flavors were sold out on roughly a third of days. B200 was sold out on no day, in any region, for the whole window. Of the 42 combinations that were sold out every single day, 22 are H200.
Asking for one GPU takes longer than asking for eight. The single-GPU H100 box was open on 16% of days with a median 15-day wait. The eight-GPU box was open on 67% of days with no wait at all. Head to head, same region, same day, the one-GPU box was the slower of the two on 92 days against 49. Same chip, same price per GPU-hour.
Longer H200 blocks vanish; longer H100 blocks don't. Share of days obtainable, by block length:
| Instance | 1 day | 7 days | 14 days | 28 days | 84 days |
|---|---|---|---|---|---|
| H100 (p5.48xlarge) | 96% | 95% | 93% | 84% | 57% |
| B200 (p6-b200) | 100% | 100% | 97% | 95% | 27% |
| H200 (p5e) | 82% | 70% | 70% | 70% | 28% |
| H200 (p5en) | 76% | 65% | 42% | 37% | 10% |
| A100 80GB (p4de) | 100% | 83% | 54% | 54% | 0% |
| A100 40GB (p4d) | 100% | 56% | 50% | 50% | 0% |
H100 is a deep pool. H200 p5en is a shallow one: the calendar has gaps you can fit a week into but not a month, so four 7-day blocks may start weeks before one 28-day block at the same price. Across everything, 84-day blocks were sold out on 73% of days: the long-commitment instrument on AWS is not a Capacity Block.
Three weeks of movement, share of days sold out for 7-day blocks:
| Instance | Aug 23–29 | Aug 30–Sep 5 | Sep 6–12 |
|---|---|---|---|
| B300 | 39% | 25% | 0% |
| H200 (p5en) | 35% | 41% | 33% |
| H200 (p5e) | 33% | 33% | 24% |
| A100 40GB | 29% | 50% | 50% |
| H100 | 3% | 11% | 1% |
| B200 | 0% | 0% | 0% |
B300 loosened hard. A100 kept tightening, and its 84-day blocks have never existed: anyone still planning around A100 blocks should move now rather than when the pool disappears. B200 has not tightened at all, which is worth saying because the obvious prediction was that it would.
Trainium stopped. Through early August it was among the easier things to get: trn1 in N. Virginia and Oregon was offered with no wait at all, trn2 in Ohio at a three-day wait. The last offer of any kind, in any region, was August 11, and every probe in the 34 days since has come back empty. An empty answer is the API saying there is no capacity rather than refusing the question, since a missing quota returns an error and errors are excluded here. I watch five Trainium pools against far more GPU pools, so read this as those five rather than as Trainium everywhere. It matters because AWS documents that Trn2 needs a Capacity Block to launch at all.
3. Open does not stay open. Sold out does not reopen.
Take a block that is open today and check back:
| Days later | Still open |
|---|---|
| 1 | 93% |
| 3 | 85% |
| 5 | 79% |
| 7 | 76% |
| 10 | 73% |
| 14 | 75% |
Nearly all the risk lands in the first five days, and then it flattens. What survives the first week tends to keep surviving. So an open block is really a five-day option rather than a standing offer: deliberate past that and roughly one in five has gone to someone else, and the only decision left is where to look next.
Going the other way is worse. A block sold out today is still sold out 82% of the time a week later and 78% of the time two weeks later. Of the 42 outages that ended inside the window, the median lasted eight days, the upper quartile eleven, the longest twenty-one. Waiting for a sold-out combination to reopen is a plan with a coin-flip attached to a week or more of delay.
Two more numbers on the same theme. The start date itself is unstable: 26% of quotes moved by a week or more within seven days, 8% by two weeks or more. And 121 of the 215 combinations tracked were never open on a single day in three weeks — never bookable inside three days, not once.
4. Where you ask is the whole decision
Share of days each region was open, queued or sold out, 7-day blocks, all instance types:
| Region | Open | Queue | Sold out |
|---|---|---|---|
| sa-east-1 (São Paulo) | 57% | 43% | 0% |
| us-west-1 (N. California) | 48% | 37% | 15% |
| us-west-2 (Oregon) | 42% | 58% | 0% |
| ap-southeast-2 (Sydney) | 38% | 62% | 0% |
| ap-northeast-2 (Seoul) | 30% | 70% | 0% |
| ap-northeast-1 (Tokyo) | 29% | 38% | 33% |
| us-east-2 (Ohio) | 23% | 56% | 21% |
| us-east-1 (N. Virginia) | 19% | 46% | 36% |
| eu-north-1 (Stockholm) | 17% | 38% | 45% |
| ap-south-1 (Mumbai) | 15% | 74% | 11% |
| eu-west-2 (London) | 1% | 78% | 20% |
N. Virginia, the region most teams default to, was sold out on 36% of days, second only to Stockholm. Oregon, one time zone away, was sold out on none.
There is no AWS-wide GPU shortage. For the same instance type on the same day, two regions agreed on status 30% of the time for both H200 flavors, which is chance. These are forty-odd separate markets that happen to share a rate card, and scarcity in one says nothing about the next.
Most pools sit in a single availability zone. A Capacity Block pool is not region-wide; it is a specific cluster inside one zone, so when that zone fills there is nowhere else in the region for the request to go. 27 of the 42 region-and-instance pools served every offer from one zone. Pools spanning more than one zone were open on 41% of days with a median 4-day wait; single-zone pools were open on 22% with a 14-day wait.
The consolation: on 28% of all observed days, somewhere could start a block immediately. The problem is rarely that there is no capacity. It is that there is no capacity where you asked.
What to do with this
- Decide region flexibility first. It is worth up to 48 days on the same block at the same price. If data gravity pins you to one region, that is your most expensive constraint, and you should know you are paying it.
- Book inside five days. One in four open blocks is gone within a week, and almost all of that loss happens early.
- Don't wait on sold out. Four in five are still sold out a week later, and the typical outage that did end lasted eight days.
- Treat H200 as a queue and B200 as available, ask for the bigger box if you can use it, and check shorter blocks before longer ones.
- Check Trainium and A100 before planning around either. Trainium blocks were routine until August 11 and have not been offered since. A100 is on its way out.
This week, September 8 to 14
Of 141 combinations: 36 open, 75 queued, 30 sold out. 21 changed state. Prices changed in none.
Worse: N. Virginia B300 28-day, 3 days out to 54. N. Virginia A100 80GB 7-day, 31 days to 52. N. Virginia H200 1-day, start tomorrow to 21 days. Ohio H200 (p5e) 7 and 28-day, into sold out.
Better: N. Virginia B200 7-day, 41 days out to 4. Mumbai H100 1-day, sold out to a 44-day queue. London H200 (p5e), sold out to a 45-day queue. Tokyo H100 single-GPU, 4 days to 2.
About the data
My own read-only sampling of AWS Capacity Block offerings across all 11 regions that sell them, up to every 15 minutes since late July 2026. Statistics cover August 23 to September 14, the period with all regions live and errors under 1%. API errors and throttles are excluded and never counted as absent capacity. All figures are single-instance offers; larger block sizes are not covered. Combinations observed on fewer than seven days are excluded from share-of-days statistics.
Everything above is aggregate. The underlying record is 215 region, instance and block-length combinations with a daily status and earliest start date, plus the sampling behind it. That breakdown, the per-combination history and reopen alerts are held back.
Need one instance, one region, one block length? Pick the three in the box below, the way you would in the console, and it comes to me. I'll reply with that combination's history since July and where it stands today. (Or email williambishop.cs@gmail.com with the three.) If your team plans training runs against Capacity Blocks, I'd like to compare notes on how you do it.
Next issue: week of September 15 to 21.