Issue 03 / GPU Market ReportAugust 2026

The H200 shows up used

Nvidia's newest Hopper card is already trading secondhand, SXM modules are pricing below their PCIe twins, and for the first time supplier asks stand behind our dealer read.

By Ryan Hammett, ClustermeltAugust 2026Prices as of August 1, 2026
The market this month

New silicon, and what it does to the used market

Nvidia's Vera Rubin platform went from ramping into full production, with shipments promised for the fall, to systems standing up at OpenAI, CoreWeave, Google Cloud, Microsoft Azure and Meta inside two months. A generation arriving early compresses the retirement schedule of the one before it, and that is what our channel read caught this month. The most striking result is not a price cut. It is that the H200, Nvidia's newest Hopper part, now trades secondhand at all, and gives up only single digits to a new unit. Underneath it, the picture splits by form factor rather than by generation: retired fleets come out of racks as whole servers, so SXM modules are reaching the secondary channel faster than the PCIe cards buyers actually know how to install, and they are pricing accordingly.

8%

all the used H200 NVL gives up to a new one in this month's read

/ Clustermelt

+40%

rise in H100 one-year rental pricing, October 2025 to March 2026

/ SemiAnalysis

50

servers with allegedly forged export paperwork in July's Taiwan case

/ AFP

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 /

The H200 is already a used card

We now read the H200 NVL 141GB across both new and secondary supply. The used central lands at $32,900 in a $26,875 to $34,799 band, against a new central of $35,798. That is roughly eight percent off for Nvidia's newest Hopper part, at a moment when the Ampere cards two generations back are giving up more than half. No trade publication has reported H200 reaching the resale channel; this is what our own supplier and dealer sourcing turned up, which is the whole reason we publish a channel read rather than a news summary.

Discount to new, newest silicon to oldest

H200 NVL

141GB

-8%

L40S

48GB

-24%

RTX 6000 Ada

48GB

-30%

RTX 4090

24GB

-43%

A100 PCIe

80GB

-57%

A40

48GB

-60%

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

What it means

If you assumed the H200 was a new-only part, it is not, and the discount is thin enough that supply, not price, is the thing to negotiate for. If you are holding H200, there is a real bid today.

02 /

SXM trades below PCIe, and that is not a defect

This is the first issue to price an SXM part. The A100 80GB SXM4 reads $9,900 secondhand in a notably tight $9,900 to $10,000 band, while the same silicon in a PCIe card reads $12,550. The module is the faster configuration on paper, with NVLink and a higher power envelope, but it cannot go into a standard server: it needs a matching HGX baseboard. Retiring fleets come out as whole nodes, so SXM supply arrives ahead of the demand that can absorb it, and the price reflects a narrower buyer pool rather than worse hardware.

What it means

If you already run HGX baseboards, SXM is the cheapest route to A100 80GB right now. If you do not, the PCIe premium is buying deployability, not performance, and it is worth paying.

03 /

Rental prices went up, which is why used Hopper has not cracked

The widely repeated expectation that Hopper would fall off a cliff has now been wrong for three quarters. Independent analysis puts H100 one-year rental contracts up almost 40 percent, from a low of $1.70 per GPU-hour in October 2025 to $2.35 by March 2026, with the conclusion that rental pricing is more likely to keep rising than falling, and providers reporting no Hopper capacity coming off contract at all. Our own read agrees from the hardware side: refurbished H100 PCIe still asks $34,995 and has not moved.

What it means

Waiting for a Hopper fire sale has been a losing trade all year. If you need the capacity, renting is not the cheaper hedge it was, and the sub-floor listings are still the counterfeit trap rather than the bargain.

04 /

Supplier asks now stand behind part of the read

Until this month every figure we published rested on public dealer and marketplace listings alone. It now rests on two independent kinds of source on the H200 NVL secondary read, both A100 PCIe reads that clear our bar, and both L40S reads: what dealers list publicly, and what suppliers we have actually spoken to are asking. Two independent kinds of source agreeing is a stronger read than either alone, and where only one stands behind a number we treat it as a starting point rather than a settled price.

What it means

A price that a listing and a supplier ask agree on is much closer to a price you can transact at. A price only one of them supports is a starting point for a conversation.

05 /

The paperwork is the product

On July 24 Taiwanese prosecutors raided and detained seven people, including an Nvidia employee and two Super Micro staff, over roughly 50 Super Micro servers carrying advanced Nvidia chips. The alleged method was forged business documents, with some units clearing Taiwan customs and reaching China via Japan. Nvidia's response noted that diverted products get no service, support or updates. Ten days earlier, US Commerce moved in the other direction, reclassifying the UAE so that AI chips and servers can ship there license-free.

What it means

The hardware in these cases was genuine; the documents were not. As the export map shifts in both directions at once, provenance you can actually check is worth more than a certificate you cannot.

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.

This month's enforcement news landed on servers, not cards, which is where our own supply increasingly comes from too. In the July 24 Taiwan case the hardware was genuine Nvidia silicon in genuine Super Micro chassis; what was allegedly forged was the paperwork that moved roughly 50 of those servers to China via Japan. That is the failure mode a price check cannot catch, and it is the reason we ask for provenance on a lot before we ask anything else. Our pass again set aside the cheap Hopper tier no discount can vouch for: sub-floor listings, China-sourced claims of new stock on export-controlled parts, and non-standard SXM-to-PCIe conversions that do not match a real catalog SKU. We also excluded a whole-rack lot bundling two different card SKUs with switches and utility servers, because there is no honest way to derive a per-card price from a single bundled number, and inventing one would be worse than publishing nothing.

Red flags on used GPU hardware

  • An H100 or H200 priced well under the going floor. The discount is the tell, not the deal.
  • A server whose only proof of origin is its own export paperwork. July's case ran on forged documents, not forged hardware.
  • An SXM module offered as a drop-in upgrade. Without the matching HGX baseboard it is not one.
  • 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

    Rubin arriving early compresses the Hopper retirement curve. Nvidia's own May release promised fall shipments; by late July systems were standing up at named customers. Watch for the first real secondary volume in H100 SXM5 and H200.

  2. 02

    The H200 discount widens from here. Eight percent off new is what a thin, early secondary supply looks like, not what a settled one looks like.

  3. 03

    SXM keeps outrunning PCIe on the supply side, because retired fleets are servers, not cards. Expect the SXM-to-PCIe discount to persist rather than close.

  4. 04

    More of the read earns higher confidence as supplier asks accumulate. That migration, not a price move, is the thing to watch in the next few issues.

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

H200 NVL

141GB

$32,900$26,875 to $34,799

H100 PCIe

80GB

$32,290$28,000 to $34,995

H100 NVL

94GB

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

L40S

48GB

$6,790$6,750 to $8,750

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$9,951 to $18,500

A100 SXM4

80GB

$9,900$9,900 to $10,000

A100 PCIe

40GB

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

A40

48GB

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

RTX A6000

48GB

$3,950$3,800 to $4,800

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.