The listing shows the GPU, which is the same as everyone else’s. Goodput measures the whole machine your customers actually get.
Power, second by second
Power limit setting
About 153 W never drawn
Drew about 247 W
Every setting read 400 W. No cap was shown.
Load starts
Load ends
ResultAbout 14% less compute
[ Study · Sep–Oct 2026 ]
nine rented machines
One provider, never named. Rented with our own money, without notice.
[ Listed vs delivered ]
The customer rents the whole machine. We rented nine machines from one provider, Sep–Oct 2026: the same GPU model at similar prices, and very different machines.
158×
Download speed, similar prices
8.6×
Disk speed
27.2 / 13.6
CPU cores, identical listing text
14%
Less compute, on two machines
Two machines’ settings said 400 W, but they ran at about 247 W. That’s about 14% less compute, and no setting or monitoring showed it.
The same machine re-tested a day later repeated the result to within 0.1%.
2 vs 13
Spread between machines, in percentage points
Each machine delivered some share of its GPU’s rated speed. On one provider’s managed tier, those shares stayed within 2 percentage points of each other; on its cheaper marketplace tier, they spread across 13.
In this sample, the managed tier’s real advantage was consistency: something a provider could point to, if it were measured independently.
[ Across the market ]
Up to 38%
Silicon Data’s study of 3,500+ cloud GPUs found the same GPU model varying by up to 38% across devices and providers.
Silicon Data, “Did You Win the GPU Cloud Lottery?”, GPGPU ’26
8–12%
NVIDIA says its partners’ clusters routinely land 8–12% below its own reference performance, from configuration alone.
NVIDIA, “Exemplar Cloud: lessons for unlocking full performance”
[ The chip ]
Across the A100s we measured, the chips matched. What changed from one machine to the next was everything around them: power, connection, disk, network, CPU and RAM.
Work per clock tick
Within 1%
The GPU chips did the same work per clock tick on every A100 we measured. The chips were fine.
Driver versions
5
Nine machines ran five different driver versions. No listing shows which one you get.
[ What varied instead ]
Same GPU model, similar prices.
[ How it works ]
No one is told a test is coming, so no machine can be prepared for it. The fee is the same whatever we find.
We rent from the provider’s public listing with our own money, without telling anyone. No machine can be prepared for the test, and the provider can’t tell which rental is ours.
A 75-second measurement of the whole machine the customer actually gets: GPU power and speed, memory, the GPU’s connection to the rest of the machine, disk, network, and the CPU and RAM the container is given.
A plain-language report: what the machine delivered, what it didn’t, why, and a ranked list of what to fix first. Once finished, it can’t be altered or withdrawn.
[ What we check ]
Six groups of checks, run on the machine a customer actually gets. Pick a part to see what we look at.
[ 01 ]
From our studyTwo machines’ settings said 400 W, but they ran at about 247 W. That’s about 14% less compute, and no setting or monitoring showed it.
[ 02 ]
From our studyThe GPU chips did the same work per clock tick to within 1% on every A100 we measured. The variation came from the machine around the GPU.
[ 03 ]
[ 04 ]
[ 05 ]
From our studyDownload speed varied 158× (31 to 4,957 Mbit/s) at similar prices. Disk speed varied 8.6× (762 to 6,567 MB/s). Two listings with identical text gave one machine 27.2 CPU cores and the other 13.6.
[ 06 ]
From our studyNine machines ran five different driver versions.
[ Not checked yet ]
This can’t be seen from inside a rented container, and we say so rather than guess.
The networking between servers isn’t tested yet.
We measure for 75 seconds, not hours. Slowdowns that build over a long job need a different test.
[ Who it’s for ]
[ Audit ]
GPU cloud providers
Your listing shows the same GPU as every competitor’s, so buyers shop on price. Nobody believes a seller’s own numbers, and machines that under-deliver turn into support tickets before you notice them yourself.
What you get
An unannounced audit of your public listings, rented with our own money: which listings under-deliver and why, how consistent each tier is, and a ranked fix list.
What changes
[ Check ]
GPU buyers
You pick a provider from a spec sheet and a price, then find out mid-run that the network is slow, the disk stalls, or the GPU is held below its power limit by a cap no setting shows.
What you get
A check before you sign a reservation or start a big run. We test the listings anonymously and report compute, power, memory, disk, network and CPU in plain verdicts, so you can compare real compute per dollar.
What changes
Some providers’ terms forbid benchmarking. If a listing can’t be checked, we’ll tell you before you pay.
[ Founding audits ]
Founding audits available for the first providers.
Buyers planning a big commitment are welcome too.
Talk to us[ A fix list looks like this ]
Examples from real measurements. A real audit lists every listing tested, with the measurement behind each finding.
GPU held below its power limit
Highabout 247 W
400 W limit
About 14% less compute, repeated on re-test
Where to look: Card firmware power policy, server management controller
CPU allowance half of a twin machine’s
Medium13.6 cores
27.2 on its twin
Slower data loading on the same listing
Where to look: Container quota settings on that host
Slow download speed
Medium31 Mbit/s
4,957 fastest
Large models take far longer to arrive
Where to look: The host’s uplink or network route
Slow disk reads
Medium762 MB/s
6,567 fastest
Slow loading and slow checkpoints
Where to look: Storage hardware on that host
[ Sample report ]
One real machine, rented and paid for by us, measured with no provider involvement or prior notice.
[ Report · single GPU · A100 80GB PCIe ]
A managed-tier GPU cloud listing, on demand. Measured without the provider’s knowledge or involvement. The provider is not named.
Verified · no degradation found
[ Read the cause, not just the score ]
On the worst machine, the standard stability score ranked it best in the study, because a hard ceiling held flat looks perfectly stable. So a report prints the cause beside the score: power drawn against the limit, clock speed, and the rest of the machine.
Power drawn against the limit
about 247 W of 400 W
One of the two capped machines
About 153 W never drawn
What was checked
Verdict
[ Why trust it ]
Nobody believes a seller’s own benchmark numbers. So the terms are fixed before any test runs.
[ Questions ]
ClusterMAX rates a cluster set up for the test, once a year. We measure what an anonymous customer gets from your public listing, and we can repeat it every quarter.
You choose the machine and you publish the result, so buyers discount it. Configuration checks also missed our headline case: every setting read 400 W while the GPU ran at about 247 W.
You’re never named publicly. Results go to you. A finished report can’t be altered or withdrawn, and that is exactly why the good results are believable.
Correct. It shows the problem exists; it isn’t a market rate. Silicon Data’s study of 3,500+ cloud GPUs found the same GPU model varying by up to 38% across devices and providers. NVIDIA says its partners’ clusters routinely land 8–12% below its own reference performance, from configuration alone. Each audit grows the sample.
You can, after you’ve rented it. We test before you commit, across several listings, with the whole machine included, and we know where to look. The cap we found didn’t show in any setting a renter would check.
Not in our data. Each machine delivered some share of its GPU’s rated speed. On one provider’s managed tier, those shares stayed within 2 percentage points of each other; on its cheaper marketplace tier, they spread across 13. Cheap can still be the right choice. The point is knowing which machine you’re getting.
We have no evidence of intent, and our reports never suggest it. We report what the machine delivered and where to look.
Whether another customer shares the machine (it can’t be seen from inside a rented container), clusters of many machines, and runs lasting hours.
Never publicly. Results go to the provider, or privately to the buyer who asked.
Founding audits are available for the first providers. Get in touch and we’ll scope one with you.
[ Get in touch ]
Founding audits available for the first providers. Buyers planning a big commitment are welcome too.
There’s no form. Write to this address and you’ll get a reply from Suryansh Sinha.
suryanshsinha.dev@gmail.com
Useful to include