AI Compute Economics

What does AI compute really cost?

Cloud list prices reveal only part of the picture. We archive GPU and accelerator configurations, Spot prices, regional catalogue signals and model coverage every hour, turning snapshots into a durable market history over time.

Archive snapshotData as of 13 Jul 2026, 08:07Tracking since 12 July 2026
298AWS Region/VM configurations in the current hardware snapshot
1,259AWS Spot price events in the latest collection window
3,526Azure VM price components in the daily catalogue snapshot
351regional Bedrock model listings in the current snapshot
AWS EC2 · current cross-section

Entry prices by Region and accelerator

For each Region we show a compact selection: the lowest-priced current Linux/shared configuration for every accelerator, memory and partition profile. On-Demand is a list price; Spot is an observed market price.

13 configurations
VM configurationAcceleratorMemory / unitvCPUOn-Demand / hourSpot median / hourvs On-Demand
g6f.largeNVIDIA12.5% L4 shareMedian across 2 AZs2.8 GiB2€0.221€0.0759−65.7 %
inf1.xlargeAWS1× InferentiaMedian across 3 AZs8 GiB4€0.2493€0.078−68.7 %
g4ad.xlargeAMD1× Radeon Pro V520Median across 2 AZs8 GiB4€0.4141€0.1204−70.9 %
g5g.xlargeNVIDIA1× T4gPrice from 1 AZ16 GiB4€0.4597€0.0975−78.8 %
g6f.2xlargeNVIDIA25% L4 shareMedian across 2 AZs5.6 GiB8€0.5197€0.2034−60.9 %
g4dn.xlargeNVIDIA1× T4Median across 3 AZs16 GiB4€0.5757€0.2221−61.4 %
g6.xlargeNVIDIA1× L4Median across 3 AZs22.4 GiB4€0.8805€0.362−58.9 %
inf2.xlargeAWS1× Inferentia2Median across 2 AZs32 GiB4€0.995€0.8546−14.1 %
g6f.4xlargeNVIDIA50% L4 shareMedian across 2 AZs11.2 GiB16€1.04€0.4948−52.4 %
g6e.xlargeNVIDIA1× L40SMedian across 3 AZs44.7 GiB4€2.04€2.04±0 %
dl2q.24xlargeQualcomm8× Qualcomm AI10015.6 GiB96€10.14
p4d.24xlargeNVIDIA8× A100Median across 3 AZs40 GiB96€24.00€24.00±0 %
p4de.24xlargeNVIDIA8× A100Median across 2 AZs80 GiB96€30.00€12.14−59.5 %

Spot is the median of the latest observed prices per Availability Zone in the collection window. A catalogue entry or price event does not prove capacity is currently available. Prices exclude tax; EUR is derived using the disclosed ECB reference rate.

The data asset grows forward

Recovering 89 days of Spot history

AWS exposes Spot history for a limited period. The collector works through the oldest safe seven-day windows across all Regions, then preserves every new observation permanently.

Recovered Region-days21 %

112 of 534 Region-days across 6 Regions archived after deduplication.

Normalised snapshots
55
Backfill observations
49,467
Oldest recovered day
14 April 2026
ECB reference rate
1 EUR = 1.143 USD
AWS catalogue coverage

Where accelerator configurations are listed

Regions vary substantially in the size of their accelerator catalogues. Bars show zonal catalogue offerings; the figures separate configurations, offerings and current price events.

Frankfurteu-central-1
VM SKUs
60
AZ offerings
149
Spot events
223
Irelandeu-west-1
VM SKUs
30
AZ offerings
87
Spot events
135
Londoneu-west-2
VM SKUs
53
AZ offerings
122
Spot events
178
Pariseu-west-3
VM SKUs
25
AZ offerings
57
Spot events
86
Stockholmeu-north-1
VM SKUs
50
AZ offerings
106
Spot events
172
N. Virginiaus-east-1
VM SKUs
80
AZ offerings
309
Spot events
465

An “AZ offering” means EC2 lists the instance as generally offered in that Availability Zone. It is not evidence of capacity that can be launched immediately.

Azure Retail Prices · daily snapshot

Explicit A100 and H100 catalogue prices

Azure names the accelerator unambiguously for only some VM meters. We therefore show configurations with an explicit A100 or H100 signal and retain the full VM price as the billing unit.

RegionVM configurationSignalList price / hour
France CentralfrancecentralStandard_NC24ads_A100_v4A100€4.02
Germany West CentralgermanywestcentralStandard_NC24ads_A100_v4A100€4.18
North EuropenortheuropeStandard_NC24ads_A100_v4A100€3.86
Sweden CentralswedencentralStandard_NC24ads_A100_v4A100€4.18
Sweden CentralswedencentralStandard_ND96is_H100_v5H100€100.64
UK SouthuksouthStandard_NC24ads_A100_v4A100€4.02
UK SouthuksouthStandard_ND96is_H100_v5H100€96.77
West EuropewesteuropeStandard_NC24ads_A100_v4A100€4.18
West EuropewesteuropeStandard_ND96is_H100_v5H100€100.64

No €/GPU-hour normalisation: the public price catalogue does not provide a reliably structured accelerator count for these rows. Prices exclude tax; a catalogue listing is not a capacity commitment.

AWS Bedrock · model catalogue

Regional model coverage is part of compute economics

VM price is not the only constraint. With managed AI, the Region determines which foundation models appear in the catalogue. We preserve every regional model listing and lifecycle state.

Frankfurteu-central-1
35
active model listings
Providers
10
Not active
2
Irelandeu-west-1
51
active model listings
Providers
11
Not active
9
Londoneu-west-2
65
active model listings
Providers
14
Not active
1
Pariseu-west-3
24
active model listings
Providers
5
Not active
7
Stockholmeu-north-1
34
active model listings
Providers
12
Not active
2
N. Virginiaus-east-1
99
active model listings
Providers
17
Not active
22

A model listing indicates catalogue coverage, not guaranteed quota or runtime capacity. The collector explicitly has no model invocation permission.

Methodology & limits

From provider signal to comparable data point

Raw data stays private and auditable. The public view contains only a bounded, traceable derivation with source, observation time and a clear unit.

  1. 1
    Collect hourly

    EC2 hardware, regional offerings and Spot events arrive hourly; stable price and model catalogues plus the ECB rate are collected daily.

  2. 2
    Keep facts separate

    Catalogue listing, zonal offering, price event and actual free capacity remain different facts. The index does not blend them.

  3. 3
    Archive exact decimals

    Source price, currency, unit, effective date and checksum remain unchanged. EUR is derived separately with the dated ECB rate.

  4. 4
    Publish a bounded view

    Instead of mirroring provider catalogues, the board shows aggregates and a small derived comparison set. The raw archive stays private.

This cross-section does not compare reservations, Savings Plans or effective contract discounts. Workload performance, data transfer, storage, availability and utilisation are also excluded; a low hourly price is not automatically a low-cost workload.

Frequently asked questions

Are these prices per GPU-hour?
No. AWS and Azure bill the figures shown as the full instance or VM price per hour. Hardware count and partition are disclosed separately, but the price is not artificially normalised to one GPU unit.
Does a Spot price mean the instance is available?
No. A Spot price event is a market and catalogue signal. It does not tell you whether the instance can actually be launched at that moment or how long it will run without interruption.
Why are reservations and discounts missing?
Reservations often combine duration, upfront payment and recurring components. A single hourly figure would be misleading without utilisation assumptions, so the first cross-section stays with On-Demand and observed Spot.
Why collect the history now?
Spot price history is available only for a limited retrospective window. Every permanently archived day enables later volatility, seasonality and regional analysis that a current price list alone cannot provide.

Run your AI workload economically?

We compare cloud, EU Region, managed API and owned infrastructure against your workload, latency and privacy requirements.

Discuss a projectView services