Everyone pays Virginia now.

On October 7, AWS raised Capacity Block prices about 15%. That was expected. The bigger change wasn't: AWS now bills the same price for a GPU whichever region it runs in. Why would AWS give up its cheaper regions? This series has a hint. Meanwhile, H100 got harder to book for the third week running.

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. The update: 15% on every chip

The last old rate was recorded at 15:28 UTC on October 7, and the first new one at 15:43. Every region and chip with an offer showed the new card by October 9. All 44 region-chip pairs this series tracks went up. None went down.

That is 15% on every line, within a fraction of a point. It's the third increase this year, after about 15% on H200-linked blocks in January and about 20% on July 1 (both from public reports, before this series began). Last issue used July's 20% as the example. This time it was smaller.

On last issue's example, an eight-instance, 28-day H100 block, the price goes from $223,200 to $256,800, about $33,500 more. Blocks bought before October 7 keep the old rate, even when they start after it.

2. One price in every region. Why?

Before October 7, H100 and the H200 p5en cost about 9% less outside the US: $4.72 against $5.19 for H100, and $6.24 against $6.86 for p5en. Now each chip has one price in every commercial region. For a team that had moved to Tokyo, London or São Paulo for the discount, this was a 26% increase, not 15%. The same 28-day H100 block outside the US went from $203,000 to $256,800.

Why give it up? AWS hasn't said. But the data suggests the discount had stopped doing its job. A lower price is a way to pull buyers toward spare capacity, and by late September the cheaper regions had less to spare than the US. In the last week of September, non-US H100 was sold out on 12% of days and US H100 on none. The discounted regions were already the fuller ones. Read this as one plausible reason, not AWS's.

It changes how to choose a region. The discount was a reason to accept a longer wait abroad, and that wait was real: this week the typical non-US H100 wait was 33 days, against 23 in the US. The reason to accept it is gone. The only thing that separates regions now is when you can start.

On October 12, a 7-day H100 block could start in Tokyo in 9 days and in Stockholm in 43, at the same price.

3. H100 slid for a third week

The 8-GPU H100 was open on 33% of days in the week of September 22, 17% the week after, and 7% this week. The typical wait went from 5 days to 17 to 30. It still wasn't sold out on any day. It's a queue that keeps getting longer.

The short blocks went first. São Paulo's 1-day H100, same-day the week before, was 29 days out. Sydney's 7-day went from 2 days to 36. Stockholm's 1-day went from 2 to 21.

If your job runs on one GPU, the single-GPU H100 (p5.4xlarge) was easier: open on 14% of days, with a typical wait of 11 days against 30. On days when both were offered in the same region, the eight-GPU block had the later start 43 times out of 54.

4. H200 is still scarce. Blackwell has room.

The H200 p5en was sold out on 57% of days and open on 4%. Mumbai's p5en went sold out at 1, 7 and 28 days. The other H200, p5e, held at its usual third of days sold out.

Blackwell is still the easiest to book. B200 was open on 57% of days and B300 on 50%, and neither was sold out on any day. The typical B200 wait was 2 days. N. Virginia's 1-day B300 went from 39 days out to 3.

The new prices make the trade-off sharper. B200 is now $14.21 per accelerator-hour, against $7.89 for H200 p5en and $5.97 for H100. Whether that's worth it depends on how much faster your job runs on it. What these numbers show is that if you need to start soon, Blackwell is where the open capacity is.

5. The last week before the update was the fastest

Last issue, open blocks were going about twice as fast as usual, and I said that fit buyers booking ahead of the update. The week after that, September 29 to October 5, was faster still: only 55% of open blocks were still open five days later, the lowest the series has recorded. In earlier weeks it was 75% to 80%.

This week, with the update landing on Wednesday, it was 59%. That's too close to call. If the rush was about the price, the pace should ease now that the price has moved. Next issue will show whether it did.

Sold out is still close to final: 4% of sold-out blocks had an offer the next day and 9% within three.

After the update

1. Re-run your budget at +15%, or +26% if you were buying outside the US. Blocks bought before October 7 keep the old rate.
2. Choose regions on wait time alone. Every commercial region now charges the same. Last month's cheaper region may now be the slowest.
3. For H100, check single-GPU blocks if one GPU will do. The wait was about a third as long.
4. If you need to start soon, price Blackwell. It costs more per hour but it was the most reliably open this week.
5. The rate is set when you buy. That held for this update and it holds for the next one. All three updates this year went up.

This week, October 6 to 12

Of 139 combinations: 19 open, 93 queued, 27 sold out. Open combinations fell from 35 the week before. 33 changed state. Prices changed in every region-chip pair. Trainium has not been offered in any region since August 11, nine weeks now.

Worse: São Paulo H100 1-day, same-day to 29 days. Sydney H100 7-day, 2 days to 36. Stockholm H100 1-day, 2 days to 21. Mumbai H200 (p5en) 1-, 7- and 28-day, into sold out. Oregon B300 28-day, 3 days to 56. N. Virginia A100 80GB 1-day, 6 days to 46.

Better: N. Virginia B300 1-day, 39 days to 3, and 7- and 28-day, 39 to 16. Seoul B300 7- and 28-day, sold out to open. Sydney H200 (p5e) 1-day, into open.

About the data

A personal collection of AWS Capacity Block offerings across all 11 regions that sell them, sampled up to every 15 minutes. This issue covers October 6 to 12, with week-on-week comparisons back to September 22. API errors and throttles are excluded and never counted as absent capacity. All figures are single-instance offers; larger block sizes are not covered. Price changes before July are from public reporting, not from this series.

Everything above is aggregate. The per-combination history behind it, 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: the first full weeks at the new prices.