Compute Cost Index — $/petaFLOP-day Normalization Methodology

Gringras, David

Prices 2026-03-29 – 2026-10-19

Economics

Recorded H100 SXM rates verified on 2026-10-19 span 3.4× across 28 offerings

Including rates verified 183 days apart, the widest H100 SXM span widens to 5.4× (Lambda at $104.05 per petaFLOP-day on 2026-04-19; Azure at $565.76 on 2026-10-19).

Published on-demand rate, $ per petaFLOP-day (log scale) →

Price verified2026-03-292026-04-192026-10-19

B2003 offerings · 3 providers2.0× spread

L40S1 offering · 1 provider

MI300X17 offerings · 1 provider2.0× spread

GB2001 offering · 1 provider

H200 SXM36 offerings · 4 providers3.4× spread

H100 SXM31 offerings · 6 providers5.4× spread3.4× across 28 offerings

A100 SXM37 offerings · 6 providers3.0× spread

Hourly instance rate ÷ the instance’s dense BF16 throughput, over 24 hours. Cheaper is left.

126

Published on-demand offerings drawn, each a distinct provider, instance and region. 1 further offering could not be placed on this axis; the caption below names it with the reason.

7

GPU classes, ordered by median price per unit of peak throughput: B200 has the lowest median, A100 SXM the highest.

6

Providers, across 51 priced regions. 101 of the drawn offerings come from Azure, including its regional quotes.

3.4×

The widest spread inside a single GPU class among rates that share a verification window: H100 SXM on 2026-10-19, across 28 drawn offerings, from AWS’s rate in US East (N. Virginia) at $166.87 to Azure’s rate in South Africa West at $565.76.

5.4×

Across all recorded vintages, the widest spread inside a single GPU class: H100 SXM, from Lambda’s single-GPU rate in US (Texas) at $104.05 (verified 2026-04-19) to Azure’s 8-GPU rate in South Africa West at $565.76 (verified 2026-10-19). Those dates are 183 days apart, so this ratio reflects whatever the price did between them as well as the distance between the two sellers.

Price dispersion, by GPU class. One mark per published on-demand offering, positioned on a shared logarithmic axis of dollars per petaFLOP-day (the instance’s hourly rate divided by its dense BF16 throughput, with no sparsity uplift).

Equal distances along the axis are equal ratios.

Mark ink encodes the date the price was verified: full ink for the 2026-10-19 refresh, half-tone for older offerings, and a hairline outline with no fill for the oldest at 2026-03-29.

The short vertical rule in each row is that row’s median; marks are nudged off the centre line so a cluster stays countable.

One offering is not drawn: GCP’s H200 SXM in US Central (Iowa) (no published rate).

3 of the drawn offerings are priced as single-GPU instances and the rest as multi-GPU nodes; the denominator divides node size out, so they share an axis, but they are different products to buy.

These are published list prices. Private contract rates and customer-specific discounts are outside this dataset.

Every rate, and what building instead would cost

Corrected Sep 7, 2026Download dataset (CSV + Croissant + BibTeX)

The figure above draws the on-demand tier because that is the tier every provider publishes. The table below adds the rest (spot, one-year and three-year reserved, each cell dated with its verification). The summary and comparison normalize cloud and on-prem daily costs by dense BF16 peak capacity. The on-prem calculator separately estimates delivered-compute cost at its stated utilization and adjustable useful life. How the index is derived

Filtered prices

H100 SXM · Cloud Pricing · On-Demand

~$104.05per peak-capacity PFLOP-day

CHEAPEST ON-DEMAND

Lambda • US (Texas) • 1×-GPU node ratemanual snapshot · 2026-04-19

~$55.09per peak-capacity PFLOP-day

CHEAPEST SPOT

Azure • East US 2 • priced 2026-10-19

~$36.44per peak-capacity PFLOP-day

ON-PREM (3YR AMORT.)

35% of on-demand

How $/petaFLOP-day is derived

GPU cloud rental cost index across hyperscaler and neocloud providers: hourly published rates normalized to $/petaFLOP-day. The throughput figure each rate is divided by is tabled per GPU class beneath this table; every row carries its provider, its region, its source, and the date the price shown was verified.
Provider	Instance	Region	$/hr	$/petaFLOP-day	Priced	Source
AWS	p5.48xlarge	US East (N. Virginia)	↓ $43.28 since 2026-04-19$55.04	$166.87	2026-10-19	AWS Pricing Bulk API
AWS	p5.48xlarge	Asia Pacific (Tokyo)	↓ $54.39 since 2026-04-19$68.80	$208.59	2026-10-19	AWS Pricing Bulk API
AWS	p5.48xlarge	US West (Oregon)	↓ $43.28 since 2026-04-19$55.04	$166.87	2026-10-19	AWS Pricing Bulk API
Azure	Standard_ND96isr_H100_v5	Central US	$120.93	$366.65	2026-10-19	Azure Retail Prices API
Azure	Standard_ND96isr_H100_v5	North Central US	$122.40	$371.10	2026-10-19	Azure Retail Prices API
Azure	Standard_ND96isr_H100_v5	Japan West	$156.81	$475.43	2026-10-19	Azure Retail Prices API
Azure	Standard_ND96isr_H100_v5	UK South	$122.90	$372.61	2026-10-19	Azure Retail Prices API
Azure	Standard_ND96isr_H100_v5	West Europe	$127.82	$387.52	2026-10-19	Azure Retail Prices API
Azure	Standard_ND96isr_H100_v5	Sweden Central	$127.82	$387.52	2026-10-19	Azure Retail Prices API
Azure	Standard_ND96isr_H100_v5	US Gov Arizona	$122.90	$372.61	2026-10-19	Azure Retail Prices API
Azure	Standard_ND96isr_H100_v5	West US	$127.82	$387.52	2026-10-19	Azure Retail Prices API
Azure	Standard_ND96isr_H100_v5	Korea Central	$137.15	$415.81	2026-10-19	Azure Retail Prices API
Azure	Standard_ND96isr_H100_v5	South Africa North	$173.09	$524.77	2026-10-19	Azure Retail Prices API
Azure	Standard_ND96isr_H100_v5	West US 3	$102.74	$311.48	2026-10-19	Azure Retail Prices API
Azure	Standard_ND96isr_H100_v5	Switzerland North	$145.01	$439.66	2026-10-19	Azure Retail Prices API
Azure	Standard_ND96isr_H100_v5	West US 2	↑ $4.42 since 2026-04-19$102.74	$311.48	2026-10-19	Azure Retail Prices API
Azure	Standard_ND96isr_H100_v5	Japan East	$142.56	$432.23	2026-10-19	Azure Retail Prices API
Azure	Standard_ND96isr_H100_v5	West Central US	$117.98	$357.71	2026-10-19	Azure Retail Prices API
Azure	Standard_ND96isr_H100_v5	East US 2	↑ $4.42 since 2026-04-19$102.74	$311.48	2026-10-19	Azure Retail Prices API
Azure	Standard_ND96isr_H100_v5	East US	↑ $4.42 since 2026-04-19$102.74	$311.48	2026-10-19	Azure Retail Prices API
Azure	Standard_ND96isr_H100_v5	Spain Central	$127.82	$387.52	2026-10-19	Azure Retail Prices API
Azure	Standard_ND96isr_H100_v5	South Africa West	$186.61	$565.76	2026-10-19	Azure Retail Prices API
Azure	Standard_ND96isr_H100_v5	Poland Central	$127.82	$387.52	2026-10-19	Azure Retail Prices API
Azure	Standard_ND96isr_H100_v5	Canada Central	$122.40	$371.10	2026-10-19	Azure Retail Prices API
Azure	Standard_ND96isr_H100_v5	UAE North	$140.60	$426.27	2026-10-19	Azure Retail Prices API
Azure	Standard_ND96isr_H100_v5	South Central US	$117.97	$357.66	2026-10-19	Azure Retail Prices API
Azure	Standard_ND96isr_H100_v5	Australia East	$142.60	$432.34	2026-10-19	Azure Retail Prices API
CoreWeave	HGX-H100_SXM5_80GB	US East (N. Virginia)	$49.24	$149.29	2026-03-29	CoreWeave published rate card
GCP	a3-highgpu-8g	US Central (Iowa)	↓ $9.66 since 2026-03-29$88.49	$268.29	2026-04-19	gcloud-compute.com (Google Cloud rate mirror)
Lambda	H100_SXM5_80GB-1gpu	US (Texas)	$4.29	$104.05	2026-04-19	Lambda published rate card
Oracle Cloud	OCI-H100_SXM5_80GB	Global	$80.00	$242.55	2026-10-19	Oracle Cloud price list

Unit. Every $/petaFLOP-day here divides by dense BF16 throughput. The denominator excludes the extra throughput available with 2:4 structured sparsity. For H100 SXM, the index uses 989.5 TFLOP/s; NVIDIA’s sparsity-inclusive specification is 1,979 TFLOP/s. Using that larger denominator would halve the price per petaFLOP-day.

Dense BF16 throughput per GPU for each class priced on this page — the denominator every $/petaFLOP-day figure here is divided by, with the source of each figure noted on its row.
GPU class	Dense BF16 TFLOP/s	Offerings
H100 SXMEpoch AI ML Hardware dataset, dense BF16 Tensor Core.	989.5	31
A100 SXMEpoch AI ML Hardware dataset, dense BF16 Tensor Core.	312	37
H200 SXMThe H100 SXM figure, from the Epoch AI ML Hardware dataset: the two parts are compute-identical, and the H200’s advantage is HBM3e memory bandwidth, which this denominator does not measure.	989.5	37
MI300XAMD datasheet dense FP16/BF16 (2,614.9 with sparsity). AMD’s sparsity support differs from NVIDIA’s 2:4, so the dense figure is the only like-for-like basis across the two vendors.	1,307.4	17
B200Epoch AI ML Hardware dataset, dense BF16 Tensor Core.	2,250	3
GB200NVIDIA Blackwell spec sheet, per GPU — a GB200 superchip carries two — cross-checked against Epoch AI ML Hardware.	2,500	1
L40SNVIDIA quotes 362.05 dense and 733 with sparsity for this part; the dense figure is used, as everywhere else on this page.	362.05	1

Dense BF16 TFLOP/s per GPU, with no 2:4 structured sparsity applied to any class. Where a figure comes from is on its row.

Refresh cadence. AWS, Azure and Oracle rows are scheduled for daily API refreshes. GCP, CoreWeave and Lambda rows come from manual snapshots of their pricing pages. The Priced column records when each rate was checked.

Coverage. 7 GPU classes are priced here: A100 SXM (37 offerings from 6 providers), H200 SXM (37 offerings from 5 providers, 1 without a published rate), H100 SXM (31 offerings from 6 providers), MI300X (17 offerings from 1 provider), B200 (3 offerings from 3 providers), GB200 (1 offering from 1 provider) and L40S (1 offering from 1 provider). Lambda rows price single-GPU (1×) instances and are labeled as such. The per-GPU rate on Lambda's 8-GPU training nodes is a separate listing and arrives with the next manual snapshot refresh (these rows verified 2026-04-19).

Methodology

Compute unit = 1 petaFLOP-day of BF16 training compute = 8.64 × 10¹⁹ FLOP. Cloud rates, summary figures and comparison bars divide daily cost by peak dense BF16 capacity. The on-prem calculator divides by throughput after utilisation to estimate delivered-compute cost. At its 40% MFU assumption, one H100 SXM delivers ~0.40 petaFLOP-days a day (989.5 TFLOP/s × 0.40 = 395.8 TFLOP/s = 0.396 PFLOP/s).

Cloud pricing: $/petaFLOP-day = hourly instance price × 24 / (num_gpus × bf16_tflops / 1000). The 7 classes priced here share their divisors with the Capacity estimator.

On-premise: $/petaFLOP-day = (hardware amortization + lifetime power cost) / (daily PFLOP output × useful life in days). Includes GPU, networking ($5K/GPU), and facility share ($3K/GPU) amortized over the useful life set on the On-Premise tab — 3 years at present.

Cloud prices verified 2026-03-29 – 2026-10-19Estimation bounds = parameter sensitivity ranges

~$104.05per peak-capacity PFLOP-day

CHEAPEST ON-DEMAND

Lambda • US (Texas)manual snapshot · 2026-04-19

~$55.09per peak-capacity PFLOP-day

CHEAPEST SPOT

Azure • East US 2

~$36.44per peak-capacity PFLOP-day

ON-PREM (3YR AMORT.)

35% of on-demand

Rate cards quote dollars per hour for machines holding different accelerators in different quantities; dividing by the throughput inside each machine puts them on one axis.

The index over time

The $/petaFLOP-day index

Index base date 2026-07-20·91 daily points, 2026-07-20 → 2026-10-19. No point on 2026-08-06: its capture has no on-demand H100 SXM price from the daily-refreshed providers.

Cheapest on-demand $/petaFLOP-day for H100 SXM (8-GPU node), across the daily-refreshed providers (AWS, Azure, Oracle). GCP, CoreWeave and Lambda prices come from manual snapshots; those dated quotes appear separately below and are excluded from this daily series.

Pool composition. This is a cheapest-across-the-pool statistic, so its level moves when the pool changes even if no provider changed price. 7 stretches were drawn from fewer than the full 3 providers: 2026-07-21 → 2026-08-08, Azure only; 2026-08-13 → 2026-08-13, AWS + Oracle (OCI) only; 2026-08-20 → 2026-08-20, AWS + Oracle (OCI) only; 2026-09-07 → 2026-09-07, AWS + Oracle (OCI) only; 2026-09-12 → 2026-09-12, AWS + Azure only; 2026-09-26 → 2026-09-28, Azure + Oracle (OCI) only; 2026-10-04 → 2026-10-07, Azure + Oracle (OCI) only. From 2026-07-21 to 2026-08-08, AWS and Oracle quotes were missing from the captures. The minimum for those dates therefore covers Azure alone. The shaded bands on the plot mark them.

Each vertical rule marks a date the daily-refresh provider pool changed width, and the line is the cheapest quote in whatever pool that date had, so a step across a rule is a different set of providers being compared.

Current quotes by vintage

How old each live on-demand quote for H100 SXM (8-GPU node) is right now, oldest first; a manual-snapshot quote can be materially older than a daily-refreshed one. A provider whose regions share one verification date folds to one line; every region’s own figure remains in the rate table.

Provider	Region	$/pFLOP-day	Verified	Cadence
CoreWeave	US East (N. Virginia)	$149.29	2026-03-29	manual snapshot
GCP	US Central (Iowa)	$268.29	2026-04-19	manual snapshot
Lambda	US (Texas)	$104.05	2026-04-19	manual snapshot
AWS	3 regions	$166.87–$208.59	2026-10-19	daily
Azure	24 regions	$311.48–$565.76	2026-10-19	daily
Oracle (OCI)	Global	$242.55	2026-10-19	daily

Where the throughput figures come from

Divisors come from Epoch AI’s ML Hardware database and from AMD’s and NVIDIA’s published specifications, named per class in the throughput table. The full derivation, with its assumptions.

12 runs have a published training cost · latest Grok 4, Jul 2025 · 8 later recorded runs

Published training-cost figures

Published training-cost figures: 11 Epoch AI estimates and 1 lab statement, covering 12 frontier runs. The latest included release is Grok 4 (Jul 2025).

Runs with a published training cost, newest first, with the figure, who published it and its basis; runs without a published figure are omitted. Only Epoch AI estimates and lab disclosures with a public source URL are included.
Run	Released	Published cost	Basis
Grok 4 · xAI	Jul 9, 2025	$388M	Epoch AI estimate · 2023 US dollars
Llama 4 Behemoth (preview) · Meta Platforms, Inc.	Apr 5, 2025	$45M	Epoch AI estimate · 2023 US dollars
GPT-4.5 · OpenAI	Feb 27, 2025	$366M	Epoch AI estimate · 2023 US dollars
Grok 3 · xAI	Feb 17, 2025	$218M	Epoch AI estimate · 2023 US dollars
DeepSeek-V3 · DeepSeek	Dec 26, 2024	$5.6M	Lab statement · assumes $2 per H800 GPU-hour; excludes prior research and ablation experiments
Grok-2 · xAI	Aug 13, 2024	$32M	Epoch AI estimate · 2023 US dollars
Llama 3.1-405B · Meta Platforms, Inc.	Jul 23, 2024	$53M	Epoch AI estimate · 2023 US dollars
Claude 3.5 Sonnet · Anthropic	Jun 20, 2024	$26M	Epoch AI estimate · 2023 US dollars
Nemotron-4 340B · Nvidia	Jun 14, 2024	$21M	Epoch AI estimate · 2023 US dollars
Gemini 1.0 Ultra · Google DeepMind	Dec 6, 2023	$31M	Epoch AI estimate · 2023 US dollars
Inflection-2 · Inflection AI	Nov 22, 2023	$13M	Epoch AI estimate · 2023 US dollars
GPT-4 (Mar 2023) · OpenAI	Mar 15, 2023	$37M	Epoch AI estimate · 2023 US dollars

Source: Epoch AI, CC BY 4.0.