Issue 02 / GPU Market ReportJuly 2026

Rubin ships, and the used market moves

Nvidia's next generation reaches production this year, pulling last-generation Hopper and Ampere into the secondary channel. Here is what held, what softened, and where the fakes are hiding.

By Ryan Hammett, ClustermeltJuly 2026Prices as of July 1, 2026
The market this month

New silicon, and what it does to the used market

Nvidia's Rubin generation, the successor to Blackwell, is now in full production, with cloud and partner availability in the second half of 2026 across AWS, Google Cloud, Microsoft, OCI and the specialist GPU clouds, while Blackwell Ultra is already shipping. On an annual release cadence, each new generation does two things to the used market at once: it pulls a wave of retired Hopper and Ampere hardware into the secondary channel as operators refresh for density and power, and it resets what buyers will pay for last-generation silicon. That repricing is underway now. It is not uniform, and the exceptions are where the value, and the risk, sit.

H2 2026

Nvidia Rubin reaches cloud and partners

/ Nvidia

~85%

what refurbished H100 still holds of new-unit pricing

/ secondary-market trackers

$160M

in Nvidia GPUs relabelled and smuggled, charged this year

/ Tom's Hardware

Trends that matter

Where the market is moving

The signals worth acting on this month, and what each one means if you are buying or selling.

01 /

Ampere is the generation on sale

The A100 and the wider Ampere line are the most available hardware in the secondary channel and have been depreciating steadily through 2026 as fleets rotate into Blackwell; independent secondary-market trackers put used A100 down a further 10 to 15 percent over the year. What stands out in our own read is how deep the discount already runs: across the Ampere cards we price, secondary units trade 31 to 60 percent below new.

Ampere secondary price, discount to new

A40

48GB

-60%

A100 PCIe

80GB

-57%

RTX 3090

24GB

-40%

RTX A6000

48GB

-31%

How far each card's secondary price sits below new in this month's read. Shown for the Ampere cards whose new-side reading is solid enough to stand behind.

What it means

If you are buying Ampere for cost-efficient inference, time is on your side. If you are holding it to sell, the liquid window is now, not next quarter.

02 /

Hopper is holding a surprisingly high floor

Even as Blackwell ramps, refurbished H100 has held in the mid-80 percent range of new-unit pricing, a far narrower discount than the 30-to-50-percent haircut buyers expect from 'refurbished,' because Hopper inference demand has not let up. Rental economics say the same from the other side: H100 cloud time has fallen to roughly a $2-per-hour floor at the cheapest providers, the flip side of more supply, not a demand collapse.

What it means

Do not wait for a Hopper fire sale. The softening is a slow grind, and the sub-floor 'bargains' are usually the counterfeit trap, not a deal.

03 /

Inference-class cards are decoupling

Not every last-generation card is falling. Inference-favorite accelerators like the L40S have held value far better than training-class Ampere, resisting the deep discounts hitting the A100 and H100. The RTX 4090 is a related story: with 40-series production ended, used stock stays firm on scarcity rather than merit.

What it means

Where you can afford to wait for a better price, and where you cannot, now splits by workload, not just by generation.

04 /

More supply means more risk

A rising-supply market is a rising-fraud market. US authorities charged a roughly $160 million scheme that smuggled Nvidia H100 and H200 boards to China by relabelling them under a fictional brand to dodge export controls. As channel volume grows, so does the share of misrepresented and gray-origin units mixed into it.

What it means

This is where a broker earns its fee: provenance and authentication, not just a number.

Authentication watch

The number is only worth what its provenance is

The judgment a listing aggregator cannot copy, and the reason a broker exists in a used market.

A rising-supply market is a rising-fraud market, and price alone is not provenance. This month's pass again set aside the cheap Hopper tier that no discount can vouch for: sub-floor listings, China-sourced 'new' claims on export-controlled parts, and non-standard conversions that do not match a real catalog SKU. Every number in this report comes from supply we would actually put in front of a buyer, which is why our Hopper read sits well above the bargains you can find elsewhere. That is the whole job: the cheapest listing and the best deal are rarely the same line item, and telling them apart is what a broker is for.

Red flags on used GPU hardware

  • An H100 priced well under the going floor. The discount is the tell, not the deal.
  • A 'new' export-controlled card sourced through China. Check the paperwork, not the price.
  • Specs that do not match the SKU: disabled interconnect lanes, cut-down memory, non-standard conversions.
  • No serial, no warranty record, no test report. Provenance you cannot check is provenance you do not have.
What to watch next

The next ninety days

  1. 01

    Ampere keeps sliding into Q3. With Rubin in production and Blackwell still shipping, the retirement wave repricing the A100, A40 and A30 has further to run.

  2. 02

    Hopper grinds down, not off a cliff. As long as refurbished H100 holds the mid-80s percent of new and demand stays real, expect gradual softening, not a crash.

  3. 03

    The inference premium widens. If cards like the L40S keep resisting discounts while training-class Ampere falls, the value gap between them is the trade to watch.

  4. 04

    Authentication risk climbs with volume. More supply, a large gray-market premium, and tighter export enforcement mean more misrepresented units in the channel, not fewer.

Sources

Price-direction figures for individual SKUs are drawn from independent secondary-market listing trackers and are directional. Every price in the Clustermelt read below is computed from our own observed market data.

The price read

The numbers

Current secondary prices for every SKU we broker, computed from our own observed market data rather than list prices. Free to read, and free to check us on.

The Clustermelt price read

Secondary-market prices for the SKUs we broker, used-grade and trustworthy supply only. We exclude the sub-floor listings that price alone cannot vouch for, so these are numbers you could actually transact against, not the lowest number on the internet.

Hoppersecondary / used

H100 PCIe

80GB

$31,995$28,000 to $34,995

H100 NVL

94GB

$29,999$29,999 to $34,995
Adasecondary / used

L40S

48GB

$6,790$6,790 to $7,000

RTX 6000 Ada

48GB

$5,700$4,750 to $7,359

L4

24GB

$3,120$3,000 to $3,500

RTX 4090

24GB

$2,400$2,299 to $2,630
Amperesecondary / used

A100 PCIe

80GB

$12,550$6,866 to $18,500

A100 PCIe

40GB

$5,750$5,500 to $7,900

RTX A6000

48GB

$4,000$3,500 to $8,500

A40

48GB

$4,699$4,500 to $5,136

A30

24GB

$1,800$1,640 to $2,250

RTX 3090

24GB

$1,050$800 to $1,200

Informational only. Observed secondary-market estimates, not quotes, offers, or advice. Do not rely on these figures as the sole basis for a purchase.

Price your config

These are ranges. Yours is a number.

The table above is the market. What you actually pay depends on quantity, condition, warranty and lead time. Send us the spec and we will price it against vetted supply, with provenance checked before it ever reaches you.

Monthly

Get this every month

No live requirement right now? We publish a new read each month. One email, no other use.

* required

By submitting, you agree to our Terms of Service and Privacy Policy.

Get started

Need a number you can trust?

Tell us what you need or what you have. We handle the rest, from first conversation to final payment.