Methodology · snapshot 2026-09-14 · next check 2026-09-21

Our prices are calculated, not decided

One formula produces all 80 on-demand prices, and it runs on a public input. Pick any card and watch the rule execute: market median, times 0.70, rounded down. Nothing else sets a price on this site.0 violations across all 80 models, checked as this page loaded. Highest ratio 0.699961 against a 0.70 ceiling.See the audit

Audit all 80 models Try the pricing API

Run the rule snapshot 2026-09-14

GPU model

Public market median $2.0414
× 0.70 $1.4290
Rounded down to $0.001 $1.428

On-demand, per GPU-hour $1.428 −30%

Interruptible × 0.50$0.714
Reserved × 0.65$0.928

Every one of the 80 models on the sheet is produced by these three lines. H100 SXM page

The rule, step by step

Four stages, no discretion at any of them. The figures below follow the card selected above (H100 SXM).

1 Public market median $2.0414 The middle on-demand rate for this exact model across a large public marketplace of independent GPU hosts. Dated 2026-09-14, re-read weekly.
2 Multiply by 0.70 $1.4290 A fixed ratio, identical for every model in the catalogue and for every account. Nothing here is negotiated.
3 Round down ~~$1.4290~~ $1.428 To the tenth of a cent, always downward. Rounding can only push the discount past 30%, never under it.
4 Your fixed price $1.428 Locked at deploy for the life of the instance. Interruptible is $0.714 (× 0.50) and reserved $0.928 (× 0.65).

the whole pricing engine

on_demand     = floor(market_median × 0.70 × 1000) / 1000     # rounded DOWN to $0.001
interruptible = on_demand × 0.50                              # flat, no bidding
reserved      = on_demand × 0.65                              # 3-month term

The audit, run on the whole catalogue

The rule is a ceiling: on-demand ÷ market median must never exceed 0.70. Rather than assert that, this page divides every price by its median as it renders, and prints the worst result it finds.

0 violations across the 80 models with a published median, checked just now

0.699961highest ratio observed
ceiling 0.700000 · Q RTX 8000
30.2%average discount
the rule only guarantees 30%
0models without a median
priced from comparable listings instead

GPU	Market median	× 0.70	Rounded down	Ratio	Real discount	Verdict
B200	$7.7507	$5.4255	$5.425	0.699937	−30%	≤ 0.70
H200	$3.9879	$2.7915	$2.791	0.699867	−30%	≤ 0.70
H100 SXM	$2.0414	$1.4290	$1.428	0.699520	−30%	≤ 0.70
A100 SXM4	$0.8007	$0.5605	$0.560	0.699388	−30%	≤ 0.70
L40S	$0.7356	$0.5149	$0.514	0.698749	−30%	≤ 0.70
RTX 5090	$0.6281	$0.4397	$0.439	0.698933	−30%	≤ 0.70
RTX 4090	$0.4676	$0.3273	$0.327	0.699316	−30%	≤ 0.70
RTX 3090	$0.1547	$0.1083	$0.108	0.698125	−30%	≤ 0.70

Eight cards shown; the same check runs on all 80 at every render. Reproduce it on your own machine with the command in verify it yourself, or download the whole table from the GPU price index.

The input: what "market median" means

The reference is the median on-demand hourly rate charged for that exact GPU model across a large public marketplace of independent GPU hosts — the rate at which half the listed supply is cheaper and half is dearer.

Per model, not per family. An H100 SXM is priced from H100 SXM listings, an H100 PCIe from H100 PCIe listings. The two land at different prices because their markets do.
Median, not minimum. The cheapest listing on a marketplace is usually an unverified host with no uptime commitment; it is a lottery ticket, not a price level.
Dated. The snapshot in force is 2026-09-14, printed on every page that quotes a price and returned in the API as market_snapshot.
Thin markets are marked. Where a model has too little liquidity for a meaningful median, the sheet shows a dash instead of inventing one. Every model in the catalogue has a published median today.

Why the method is public at all

A GPU price that moves hourly cannot be planned against, and a price behind a sales call cannot be compared at all. Publishing the formula turns the price into something a buyer can audit before signing up, and re-audit afterwards.

It also constrains us: the rule is arithmetic on a public input, so drifting above median × 0.70 would be visible to anyone running the command below, on any day, without asking permission.

What moves, and what cannot

The rule is a ratio, so the sheet follows the market in both directions. What you were quoted, however, does not move.

Moves, weekly

The market reference. Re-checked every week. Snapshot in force 2026-09-14, next check 2026-09-21.
Sheet prices, in both directions. A rising median raises the price, a falling one lowers it. Medians have mostly fallen, so most checks cut prices.
New deployments only. A new sheet price applies to instances created after it, never to those already running.

Never moves

A running instance. The rate shown at deploy is the rate until you stop it.
A reserved term. Locked for the full three months at the rate agreed on day one.
The multipliers. 0.70, 0.50 and 0.65 are the product, not a promotion.
One sheet for everyone. No free tier, no promotional credit, no negotiated rate — which is what makes the rule checkable at all.
Multi-GPU. An 8× machine is exactly 8 × the per-GPU price, with no premium.

Storage and bandwidth sit outside the rule entirely: flat published rates of $0.08/GB per month and $0.01/GB, listed in full on the pricing page.

What the rule promises, and what it does not

Stating the limits is part of the method. A rule that claimed everything would be worth nothing.

It does promise

Never more than 70% of the market median for any model, on any day, for any account.
A price locked at deploy and never revised under a running instance.
The same sheet for everyone, published in full and served as JSON without an account.
A dated input, so any claim on this site can be checked against the day it was made.

It does not promise

The lowest price anywhere. Unverified marketplace hosts, bid-based capacity and promotional tiers can undercut the sheet on specific cards, especially consumer GPUs. The comparison pages name those cases instead of hiding them.
A ceiling in absolute dollars. If the median for a card rises, the fixed ratio raises its price too.
Like-for-like hardware. A marketplace listing at the same model name may sit in a different chassis, host, cooling and network. Every PowerGPU machine is a dedicated GPU in a Tier-III facility.
Anything about storage or bandwidth. Those are flat rates, set by us and published on the pricing page.

Verify it yourself

A promise that cannot be tested is marketing. Each of these takes under a minute and needs no account.

1

Check the ratio on every model

Fetch the sheet and divide. If any model returns a ratio above 0.70, the rule is broken and the claim on this site is false. As this page rendered, the worst was 0.699961.

curl -s https://powergpu.ai/v1/gpus/pricing
2

Compare against a named provider

Take any provider on the alternatives pages, open their own pricing page, and check the quoted figure against the date printed beside it. Where they are cheaper, our page says so.

https://powergpu.ai/gpu-price-index.csv
3

Price a real month

Put your own hours, storage and egress into the cost calculator, then price the same month on the provider you use today. The rule only matters if it survives your workload.

https://powergpu.ai/calculator

the audit, in one pipe

# every model, the ratio the rule promises to keep at or under 0.70
curl -s https://powergpu.ai/v1/gpus/pricing \
  | jq -r '.data[] | select(.market_median != null)
      | [.slug, .market_median, .price_per_gpu_hour.on_demand,
         (.price_per_gpu_hour.on_demand / .market_median)] | @tsv' \
  | awk '$4 > 0.70 { print "RULE BROKEN:", $0; bad++ } END { print (bad ? bad : 0), "violations" }'

Pricing methodology FAQ

The billing mechanics themselves — per-second accounting, credits, refunds — are documented in the billing docs.

How exactly is a PowerGPU price calculated?

On-demand price = the public GPU marketplace median hourly rate for that exact GPU model × 0.70, rounded down to the tenth of a cent. Interruptible = on-demand × 0.50. Reserved (3-month term) = on-demand × 0.65. For the H100 SXM: a median of $2.04 gives $1.428 on-demand, $0.714 interruptible and $0.928 reserved. Rounding down means the real discount is always slightly better than 30%.

What is "the market median" and why use it instead of the lowest price?

It is the middle on-demand rate charged for that GPU model across a large public marketplace of GPU hosts — the price at which half the supply is cheaper and half dearer. The minimum listing is not a usable reference: it usually belongs to an unverified host with no uptime guarantee, and it changes hour to hour. The median describes what that card actually costs in the market, so a fixed discount against it is a meaningful promise rather than a race against the least reliable machine on the internet.

How often do prices change?

The market reference is re-checked weekly. The snapshot currently in force is dated 2026-09-14 and the next check is due 2026-09-21. A change to the sheet only affects new deployments: the rate shown when an instance is created is locked for the life of that instance, and a reserved rate is locked for its whole term.

Can PowerGPU prices go up?

Yes, if the market median for a model rises — the rule is a fixed ratio, not a fixed number. The commitment is the ratio: never more than 70% of the median. In practice GPU medians have trended down, so most weekly checks lower prices. Running instances are never re-priced in either direction.

What does the rule not promise?

Three things. It does not promise the lowest price on the internet: an unverified marketplace host or a promotional tier can be cheaper on a given card, and the comparison pages say so when it happens. It does not promise identical hardware to a marketplace listing at the same nominal model. And it does not apply to storage or bandwidth, which are flat published rates rather than market-derived ones.

How can I verify the 30% claim myself?

Fetch https://powergpu.ai/v1/gpus/pricing without a key. Each entry carries market_median and price_per_gpu_hour.on_demand. Divide one by the other for every model in the response: the ratio never exceeds 0.70. As this page rendered, the highest ratio across the 80 models with a published median was 0.699961, with 0 violations.

Why do some models show no market median?

When a model has too little liquidity on the public marketplace, a median would be noise rather than a reference, so the field is left empty and the sheet shows a dash, and the price is derived from the closest comparable listings under the same 30% rule. Today every one of the 80 models in the catalogue has a published median.

How are competitor prices on the comparison pages collected?

By reading each provider's own public pricing page and quoting the on-demand list price per GPU-hour, before taxes, with the configuration noted and the date of reading printed on the page. They are never estimated, never adjusted, and never rounded in PowerGPU's favour. When a provider is cheaper on a card, the table shows the difference in that provider's favour.

Our prices are calculated, not decided

The rule, step by step

The audit, run on the whole catalogue

The input: what "market median" means

What moves, and what cannot

What the rule promises, and what it does not

Verify it yourself

Check the ratio on every model

Compare against a named provider

Price a real month

Pricing methodology FAQ

Deploy your first GPU in under a minute