Before the price moves.
This is the last full week before AWS's scheduled October price update, and the price still hasn't moved: same rate card on every probe through 06:48 UTC on October 1. Underneath it, this was the most lopsided week the series has recorded. The older chips had their worst week. The newest had their best.
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. What the October update can change, and what it can't
AWS's pricing page says the current prices are scheduled to be updated in October. It doesn't say which way. The two updates already made this year both went up, by public reports: about 15% on H200-linked blocks in January, and about 20% across P4de through P6-B300 on July 1. The rate card this series has recorded since July is exactly the July 1 card, untouched.
The rule that matters is AWS's own: a block is charged at "the prevailing rate at the time of purchase," even if it starts after the price changes. You pay up front when you book, and the price doesn't change after that. Blocks can start up to eight weeks out, so anything you will need between now and late November can still be bought at today's rate.
Two things make that lock worth more than it looks. There is no duration discount, so the money at stake grows in step with block length. An eight-instance, 28-day H100 block costs $223,200 at today's rate, and a repeat of July's 20% would add about $45,000 to it. And the lock only covers blocks that are actually for sale. For H200 (p5en) this week there were few: none open, 84-day blocks offered on 0% of days, 28-day blocks on 18%. B200 had its 84-day block on offer 69% of days and H100 49%. The October question is mostly about whether there is anything to buy at the old price.
2. H200 had its worst week on record
The H200 p5en was sold out on 66% of days, the highest of the five weeks since all regions went live. Before this week it had ranged from 29% to 50%. It was not open on any day in any region.
Oregon, the fastest region for this chip in issue 1, saw its 7-day earliest start go from 2 days to 57, the largest single move of the week. N. California went sold out at 1, 7 and 28 days. Ohio's 7-day block sold out. The other H200, p5e, held at its usual third of days sold out, but Oregon's 28-day p5e went from 12 days out to 38.
This is the chip where depth is thinnest. Longer blocks disappear first, and this week the 84-day H200 p5en was offered on no day at all.
3. H100: the deep pool got shallow
H100 has been the dependable one: open on 59% to 78% of days in every week so far. This week it was open on 33%, its lowest. The median wait went from 2 days to 5, and it was sold out on 7% of days, up from none.
The region you ask in mattered more than ever. The median gap between the fastest and slowest region for a 7-day H100 block was 53 days, level with the widest week in the series. On September 28, São Paulo could start one the same day; London was 52 days out. London's 7-day and 28-day H100 blocks both went to sold out. Tokyo, N. California and Ohio all went from open to a queue. São Paulo was the exception, going from a 6-day queue to same-day.
The non-US discount kept its catch. H100 is $4.72 outside the US against $5.19 inside it. This week US regions were open on 46% of days and never sold out. Non-US regions were open on 24% of days, sold out on 12%, and the typical wait was longer.
4. Blackwell went the other way
Last issue's headline was that B200 had stopped being instant. A week later it was instant again. B200 was open on 63% of days, its best week on record. The median wait went from 15 days to same-day. In N. Virginia, the 7-day B200 that had blown out to 39 days came back to under one. Oregon, Stockholm and Ohio all opened up too.
B300 swung the same way: open on 48% of days after 0% the week before, with Oregon's 7-day block going from 44 days to same-day. N. Virginia's B300 went the opposite direction, 2 days to 43, which is a reminder that none of this moves as one market.
Two readings to take from it. The newest chips are now the easiest to get: B200 costs $12.36 per accelerator-hour against $6.87 for H200 p5en, and this week it was the one with room. And one week is not a forecast. The same N. Virginia B200 block moved by more than 30 days, in opposite directions, in consecutive weeks. A start date from last week is not a quote.
5. Open blocks went twice as fast
Of the blocks that were open on a given day this week, 84% were still open a day later, 67% three days later and 58% five days later. In every earlier week, between three in four and four in five survived five days. This week four in ten were gone. Sold out still didn't come back: 2% of sold-out blocks had any offer the next day, 6% within three.
Faster turnover in the last week before a scheduled price update is what you would expect if buyers were bringing bookings forward. I can't see who books, so read this as consistent with that rather than proof of it. Either way, the practical window to decide on an open block has shrunk from about five days to about two.
Before the update
- Anything you'll need before late November, price it now. The purchase date sets the rate, not the start date.
- Lock long blocks first. With no duration discount, the dollars at stake grow with length, and long blocks are what disappear first.
- If it's H200, look today. For p5en there may be nothing to lock at any price. B200 and H100 have room. H200 mostly doesn't.
- Don't sit on an open block. This week four in ten were gone within five days.
- Tell finance. The whole fee is charged in the month you book, not the months you run.
This week, September 22 to 28
Of 147 combinations: 40 open, 72 queued, 35 sold out, the most sold out of any week this series has reported. 40 changed state. Prices changed in none. Trainium has still not been offered in any region since August 11, seven weeks now.
Worse: Oregon H200 (p5en) 7-day, 2 days out to 57. N. Virginia H200 (p5en) 1-day, 20 days to 56. N. Virginia B300 7-day, 2 days to 43. London H100 7- and 28-day, into sold out. N. California H200 (p5en) 1-, 7- and 28-day, into sold out.
Better: N. Virginia B200 7-day, 39 days to under one. Oregon B300 7-day, 44 days to same-day. Ohio B200 28-day, 28 days to 3. São Paulo H100, a 6-day queue to same-day.
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. This issue covers September 22 to 28; the price check runs through 06:48 UTC on October 1. Week-on-week comparisons use the five weeks since August 25. API errors and throttles are excluded and never counted as absent capacity. All figures are single-instance offers; larger block sizes are not covered. Earlier price changes are from public reporting, not from this series, which began after July 1.
Everything above is aggregate. The per-combination history behind it, and reopen alerts, are held back.
Need one instance, one region, one block length, before the price changes? 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 29 to October 5, which may be the first with the October rates.