If you have ever estimated what a model will cost in Frankfurt or Tokyo, you probably did it the
way I did: take the us-east-1 price, apply a regional uplift, move on. Every cloud cost model I
have seen treats region as a multiplier on a base rate.
I read all 1,586 of Amazon Bedrock's on-demand, in-region token meters and compared each one
against its us-east-1 equivalent. 1,424 of them are comparable, across 32 regions. They produce
123 distinct multipliers.
If region were a multiplier, there would be at most 32.
The distribution
30.3% identical to us-east-1
Q1 ×1.000
median ×1.183
Q3 ×1.200
min ×0.964
max ×1.700
Nearly a third of the meters cost exactly what Virginia costs. The top of the range is São Paulo,
where Mistral 7B output runs $0.20 → $0.34.
Four meters are cheaper outside us-east-1 — the input and output of two Nova Sonic models in
us-west-2, at ×0.967 and ×0.964. Small, but it means the uplift does not even have a floor at 1.
One region, twenty-five different multipliers
London is the clearest case. Here is eu-west-2 input, against the same model in Virginia:
×1.000 Kimi K3 $3.300 -> $3.300
×1.000 Grok 4.7 $2.200 -> $2.200
×1.300 Llama 3 8B $0.300 -> $0.390
×1.333 Mistral 7B $0.150 -> $0.200
×1.412 Nova Pro $0.800 -> $1.130
×1.500 Gemma 3 4B $0.040 -> $0.060
×1.533 Qwen3 32B $0.150 -> $0.230
×1.550 Devstral $0.400 -> $0.620
×1.600 Ministral 3B 3.0 $0.100 -> $0.160
Those are nine of London's 47 comparable input meters; across both directions it has 94 meters
and 25 distinct multipliers. Two models cost the same as Virginia. Most of the rest are half
again as expensive. "London is
55% more" is simultaneously true and false in the same table, and which one you get depends on
the model you happened to pick.
Across all 32 regions, only 9 charge a single consistent multiplier — and seven of those nine
have four meters each, which is a sample, not a policy. The two with real coverage are us-east-2
(129 meters, all ×1.000) and me-central-1 (13 meters, all ×1.000).
Put differently: 23 of 32 regions price different models differently relative to Virginia.
Azure's first axis is not the region
Azure is the useful contrast, because it is not organised the way AWS is at all.
A Global Standard deployment has one price. Not one price per region — one price, full stop;
the same number is served for every region Azure lists. Pinning that same model to a region or a
data zone is what costs money:
regional median ×1.100 range 1.100 – 1.320 10 distinct multipliers
datazone median ×1.100 range 1.100 – 1.200 6 distinct multipliers
So on Azure the question is "am I pinning this to a geography", and the answer is usually a flat
10%. On AWS the question is "which model, in which region", and there are 123 answers.
Azure is not perfectly uniform either — 18 of the 21 models with regional pricing do vary between
regions. But 10 distinct multipliers against AWS's 123, over a comparable number of meters, is a
difference in kind rather than degree.
What I would actually do about it
Don't estimate a region from a reference price. This is the practical consequence and it is
the whole post. There is no factor you can multiply by that is right for more than a slice of the
catalogue. Read the number for the model you are actually calling, in the region you are actually
calling it in.
Check the two models you run most, not the average. The median ×1.183 is nearly useless for
budgeting: your bill is a weighted sum over the specific meters you hit, and those meters range
from ×0.964 to ×1.700.
On Azure, price the deployment type first. If you are on Global Standard, region is not a cost
axis for you at all, and a migration between regions changes nothing on the invoice. That is worth
knowing before you spend a sprint on it.
What I measured, and what I got wrong first
Bedrock on-demand standard tier, in-region placement only. Batch, Flex and Priority are
separate tables and so is cross-region inference.
That exclusion is the whole methodology, and I know because I skipped it the first time. Grouping
without pinning the deployment type produced a tidy headline — "Nova 2.0 Lite costs 6.16× more in
some regions" — which was Flex in one region compared against Priority in another. Nothing was
invented; the rows were real. They were just answers to a question nobody asked. A price table
with four tiers in it will hand you a dramatic number if you let it.
A meter counts only when the same model and direction exist in both regions, so a model that is
simply absent from a region contributes nothing rather than an infinite ratio. Prices are read
from the vendor's published tables, never derived from a reference figure — which is what this
whole post argues you also should not do.
Figures as published on 2026-10-11.
The per-region tables, with a region switcher and the basis stated on the page:
aichangewatch.com/aws/pricing
Top comments (0)