The price didn't move. B200 finally did.

Last issue set the baseline: across 42 region-and-instance combinations the price of a Capacity Block had not moved once in 48 days, and the only thing that moved was when you could start. This is the first weekly on top of that, September 15 to 21, and it is a good argument for watching weekly. The price still didn't move. The map underneath it moved a lot.

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. B200 stopped being instant

Last issue said B200 had not tightened at all, and that this was worth saying because the obvious prediction was that it would. This week it did.

In N. Virginia, the earliest start for a 7-day B200 block went from 4 days to 39 inside the week, the largest single move of any combination I track. The 28-day B200 block in the same region went from a queue to sold out. Stockholm B200, open with an immediate start last week, moved to a three-to-four-day queue across every block length. Oregon's 28-day B200 sold out too.

Keep it in proportion: B200 was still the least sold-out Blackwell, with no fully sold-out 7-day block in any region on any day this week. The story isn't that B200 ran out. It's that the one pool you could always walk into started making you wait and pulled its longer blocks first — the exact tightening the last issue said hadn't happened. The obvious prediction was just early.

2. If you default to N. Virginia, this was your week to notice

N. Virginia is the region most teams reach for first, and it was the most turbulent room on the board — open on 7% of days, against Oregon and N. California one time zone away at 27% and 50%, both sold out on none.

Every one of the week's largest tightenings was in N. Virginia: the 7-day H100 block there went from a 2-day start to 26, the B200 move above, the 28-day H100 from 19 days to 29. It is not that N. Virginia is uniformly scarce — its A100 actually reopened this week (section 4). It is that it moves the most, in both directions, and a team that only ever asks there is riding all of that volatility for no reason the price can justify.

The region gap that framed the last issue held. For a 7-day H100 on September 21, Oregon started today while London was 37 days out; the median gap across the week was 40 days. Same block, same price, still worth about six weeks depending only on which room you ask in.

3. H100 got easier — almost everywhere you wouldn't think to ask

The chip everyone worries about was the easy one this week. The 8-GPU H100 block was sold out on no day in any region, open on 59% of them, at a 2-day median start. Inside the US it was open on 86% of days at a start of under a day.

The scarcity that remains is locational, not chip-wide, and it is quietly stubborn where it lands: Stockholm's H100 slid from an immediate start to a 13-day queue this week, and it is still true that asking for one H100 is slower than asking for eight — the single-GPU box was open on 12% of days at a 14-day wait while the eight-GPU box sat at 59% and two days, same chip, same price per GPU-hour. If H100 is what you need and you can move regions, this is a buyer's market. If you're pinned to one region, it may not be yours.

4. A100 came back from the dead — the short blocks, anyway

Last issue said A100 was on its way out and told anyone planning around it to move now. One week later the pool refilled. The 80GB A100 (p4de) was sold out on zero days this week, obtainable every day as a roughly 32-day queue, and N. Virginia's A100, sold out a week earlier, reopened across both the 7-day and 28-day blocks. The 40GB A100's share of sold-out days fell by more than half from the baseline.

The exit is real but slower than it looked — with one exact edge that hasn't moved: the 84-day A100 block still does not exist. 80GB A100 was obtainable on 0% of days at that length, 40GB on 7%. So the correction is narrow. If you want a short or medium A100 block, the door that looked like it was closing is open again. If you want a long one, it was never open and still isn't, and that is the horizon on which A100 genuinely leaves.

The one number that still hasn't moved

Zero prices changed this week. The only wobble in the whole set was 40GB A100 in Oregon reading $1.47 one day and $1.48 the next, which is rounding, not a move. The rate card is exactly the last issue's: H100 at $5.19 per accelerator-hour in the US and $4.72 outside it, B200 at $12.36 everywhere, on down the ladder.

That matters this week for one reason: AWS's own pricing page says the current prices are scheduled to be updated next in October, 2026. We are a week or two from the first price change this series has ever recorded. Everything above — B200 rationing, A100 refilling, N. Virginia swinging 25 days on an H100 — happened with the price nailed to the floor. When it finally lifts, that will be the first time in the life of this record that both levers moved at once. Next issue may be the one that catches it.

This week, September 15 to 21

Of 143 tracked combinations: 33 open, 82 queued, 28 sold out. 25 changed state. Prices changed in none.

Worse: N. Virginia B200 7-day, 4 days out to 39. N. Virginia H100 7-day, 2 days to 26. Stockholm H100 (all lengths), immediate to a 13-day queue. Oregon B300 7-day, deeper into a 44-day queue.

Better: N. Virginia B300 7-day, a 6-day queue to a 2-day start. N. Virginia A100 (both flavors), sold out back to an obtainable queue. São Paulo single-GPU H100, a 6-day queue to an immediate start. Oregon 40GB A100 1-day, a 7-day queue to a 3-day start.

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 15 to 21. 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 the per-combination daily status and earliest start date for every region, instance and block length, 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 22 to 28.