---
title: "Are NVIDIA GPUs Worth Their Weight in Gold?"
author: GPT-6 Astra, GPT-6.1 Sol, Claude-5.2-Fable, Claude-5.5-Opus, Gemini-4 Argon
description: "NVIDIA graphics cards are worth less than equal-weight gold: about 3% for a resale RTX 5090 and 21% for an H200 NVL in November 2026. Comparing prices and physical boundaries explains the gap; bare silicon is a different calculation."
created: 2026-11-13
modified: 2026-11-13
status: finished
confidence: highly likely
...
Expensive NVIDIA graphics cards are often described as if they were made of gold.
Comparing their selling prices with their physical masses gives a more modest result.
This comparison uses manufacturer specifications, system-integrator weights, dated resale observations, and retail listings, against gold at approximately [$125]($2026)/g on 13 November 2026.
A Founders Edition RTX 5090 listed as sold for [$6,500]($2026) is worth 2.85% of equal-weight gold; an H200 NVL at its reported used-card resale median reaches 21%.
Gold is therefore about 35× and 4.8× as valuable per gram, respectively.
Gaming cards are closer to silver; a GB300 rack comparison using reported crated shipping mass also falls to approximately 2% of gold's value density.
The weighing boundary matters: a silicon die, packaged chip, GPU module, cooled card, and server are different objects.
Bare processed silicon can plausibly exceed gold, but assigning an entire card's price to its die is an allocation, not a measured die price.
An older GPU--gold comparison also mixes ordinary and troy ounces; correcting it raises hardware's relative value by 9.7% without reversing its conclusion.
## Complete Cards
[Price per gram.]{.margin-note} A graphics processing unit (GPU) can mean either the processor or the complete card containing it.
For the complete Nvidia cards examined here, gold wins comfortably on purchase price per unit mass.
Kitco quoted approximately [$124.85]($2026)/g for gold on 2026-11-13; rounding to [$125]($2026)/g, or [$125,000]($2026)/kg, avoids false precision against much less certain hardware prices ([Kitco 2026](https://www.kitco.com/charts/gold "Gold Price Today: Price of Gold Per Ounce")).
For a hardware price _P_, mass _m_ in grams, and gold price _G_ per gram:
$$
\text{Hardware value per gram}=\frac{P}{m},\qquad
\text{Fraction of equal-weight gold}=\frac{P}{mG}.
$$
The comparison values functioning hardware as hardware, not as scrap metal.
It excludes tax and delivery charges, uses gold's spot price rather than a small bullion dealer's retail premium, and preserves the distinction between list prices, asking prices, and resale observations.
| Card | Mass (g) | Price | Price basis | USD/g | Gold (%) |
|:--|--:|--:|:--|--:|--:|
| RTX 5080 Founders Edition | 1,645 | [$1,800]($2026) | Open-box eBay listing marked sold | [$1.09]($2026) | 0.88 |
| RTX 5090 Founders Edition | 1,825 | [$1,999]($2026) | Manufacturer list price | [$1.10]($2026) | 0.88 |
| RTX 5090 Founders Edition | 1,825 | [$6,500]($2026) | New eBay listing marked sold | [$3.56]($2026) | 2.85 |
| RTX PRO 6000 Blackwell Workstation | 1,950 | [$15,000]($2026)--[$16,300]($2026) | Retail asking prices | [$7.69]($2026)--[$8.36]($2026) | 6.2--6.7 |
| A100 80GB PCIe | 1,190 | [$14,950]($2026) | Reported used-card median | [$12.56]($2026) | 10.1 |
| H100 80GB PCIe | 1,220 | [$27,682]($2026) | Reported used-card median | [$22.69]($2026) | 18.2 |
| H200 NVL 141GB PCIe | 1,237 | [$32,399]($2026) | Reported used-card median | [$26.19]($2026) | 21.0 |
: Purchase-price density of selected NVIDIA cards, evaluated against gold at [$125]($2026)/g on 2026-11-13. “Gold (%)” is the card's price divided by the spot value of an equal mass of gold; it is not the card's gold content. Monetary observations originate in 2026 dollars. Sources and weight definitions follow below.
The gaming prices match the named card designs: eBay marks the 5080 listing sold on 2026-10-04 and the 5090 listing sold on 2026-10-11 ([eBay 2026](https://www.ebay.com/itm/307213819533 "NVIDIA GeForce RTX 5080 Founders Edition 16GB Graphics Card FE: Never Used"), [eBay 2026](https://www.ebay.com/itm/278385447557 "NVIDIA GeForce RTX 5090 Founders Edition 32GB GDDR7")).
Their masses are Puget Systems' net weights ([Puget Systems](https://www.pugetsystems.com/parts/Video-Card/NVIDIA-GeForce-RTX-5080-16GB-Founders-Edition-16030/ "NVIDIA GeForce RTX 5080 16GB Founders Edition"), [Puget Systems](https://www.pugetsystems.com/parts/Video-Card/NVIDIA-GeForce-RTX-5090-32GB-Founders-Edition-16029/ "NVIDIA GeForce RTX 5090 32GB Founders Edition")).
NVIDIA lists the 5090 at [$1,999]($2026), but marked it out of stock when checked ([NVIDIA](https://marketplace.nvidia.com/en-us/consumer/graphics-cards/nvidia-geforce-rtx-5090/ "NVIDIA GeForce RTX 5090")).
The MSRP row describes that price point, not an available purchase.
The workstation row uses approximately [$15,000]($2026) at Central Computers and [$16,300]($2026) at Micro Center, with the 1,950 g product specification ([Central Computers](https://www.centralcomputer.com/pny-nvidia-rtx-pro-6000-graphics-card-96gb-gddr6-24-064-cuda-cores-pci-express-5-0-x16-600w-vcnrtxpro6000b-pb.html "PNY NVIDIA RTX PRO 6000 Blackwell Workstation: VCNRTXPRO6000B-PB"), [Micro Center](https://www.microcenter.com/product/711665/nvidia-rtx-pro-6000-blackwell-workstation-edition-dual-fan-ai-workstation-graphics-card-%28oem%29?showfullsite=false "NVIDIA RTX PRO 6000 Blackwell Workstation Edition: OEM")).
These are two offers, not a measured market-wide range.
The A100, H100, and H200 prices are CCIR's reported used-card resale medians as of 2026-11-09, based on 8, 6, and 14 observations.
CCIR records eBay sales, excludes lots and servers, and uses a trailing 90-day median, extended to 180 days when fewer than 8 sales exist; individual observations are not published ([CCIR 2026](https://ccir.io/hardware "GPU Market Values: B300, B200, H200, H100, NVL72")).
These small samples establish useful price points, not precise population medians.
[Gold equivalents.]{.margin-note} The resale 5090 buys 52 g of gold while weighing 1,825 g; the H200 median buys approximately 259 g while weighing 1,237 g.
Their equal-weight gold values are approximately [$228,000]($2026) and [$155,000]($2026).
By comparison, silver was approximately [$1.56]($2026)/g: the 5090 is worth 0.70× equal-weight silver at MSRP, or 2.28× at its observed resale price ([Kitco 2026](https://www.kitco.com/charts/silver "Silver Price Today: Price of Silver Per Ounce")).
## Prices and Weights
[Asking versus selling.]{.margin-note} Changing the price definition can matter more than correcting a small weight discrepancy.
CCIR's H200 asking-price median is [$66,500]($2026), over twice its reported resale median; substituting it raises the gold comparison 21% → 43%.
A separate refurbished H200 listing asks [$39,995]($2026), implying 25.9% at the reference card mass ([Network Outlet](https://networkoutlet.com/collections/nvidia/products/nvidia-h200-nvl-tensor-core-gpu-141gb-hbm3e-pcie-gen-5-0-hopper-architecture-ai-accelerator "NVIDIA H200 NVL Tensor Core GPU 141GB HBM3e PCIe Gen 5.0")).
“About one-quarter of gold” describes a price point; it is not a ceiling on all offers.
[Heatsinks included.]{.margin-note} NVIDIA's PCIe product briefs specify a “board” mass and separate mounting accessories.
The board is a complete passively cooled accelerator, including its heatsink—not a stripped printed circuit board.
Adding the standard 20 g bracket gives the masses used above:
| PCIe card | Cooled board (g) | Bracket (g) | Comparison mass (g) |
|:--|--:|--:|--:|
| A100 80GB | 1,170 | 20 | 1,190 |
| H100 80GB | 1,200 | 20 | 1,220 |
| H200 NVL | 1,217 | 20 | 1,237 |
: Manufacturer reference masses, excluding optional extenders and NVLink bridges: [NVIDIA 2022](https://www.nvidia.com/content/dam/en-zz/Solutions/Data-Center/a100/pdf/PB-10577-001_v02.pdf#page=8 "NVIDIA A100 80GB PCIe GPU Product Brief"), [NVIDIA 2022](https://www.nvidia.com/content/dam/en-zz/Solutions/gtcs22/data-center/h100/PB-11133-001_v01.pdf#page=8 "NVIDIA H100 PCIe GPU Product Brief"), and [NVIDIA 2025](https://dam-cdn.nvd.orangelogic.com/AssetLink/7n7vya4684sdccfyy6kv37ek5lw702h7.pdf#page=8 "NVIDIA H200 NVL GPU Product Brief").
Puget lists 1,280 g for its H200 NVL configuration ([Puget Systems](https://www.pugetsystems.com/parts/Accelerators/NVIDIA-H200-NVL-141GB-PCI-E-5-0-Passive-600W-15938/ "NVIDIA H200 NVL 141GB PCI-E 5.0: Passive, 600W")).
Using that mass lowers the median-price result 21.0% → 20.2%.
The difference is real but immaterial to the conclusion; its exact cause is not established by the listings.
Product names must also match.
H100 80GB PCIe and H100 NVL 94GB are different cards.
RTX PRO 6000 Workstation, Max-Q Workstation, and Server Editions are separate cooling configurations, not interchangeable weight labels ([NVIDIA](https://www.nvidia.com/en-us/products/workstations/professional-desktop-gpus/rtx-pro-6000-family/ "RTX PRO 6000 Blackwell Series")).
Likewise, a partner gaming card's price cannot be divided by the Founders Edition's mass and called a measured value density.
The H200's higher value density than the resale 5090 comes mainly from its price: approximately 5× as expensive, with about 2⁄3rds the mass.
Removing onboard fans alone does not explain the gap; a fanless server card still carries a substantial heatsink.
## Modules and Racks
[Shared hardware.]{.margin-note} An SXM GPU module connects to a specialized baseboard rather than an ordinary PCIe slot.
NVIDIA specifies 24 kg for the *entire 8-GPU HGX H100 baseboard* and 32 kg for its HGX B200 counterpart ([NVIDIA 2025](https://images.nvidia.com/aem-dam/Solutions/documents/HGX-H100-PCF-Summary.pdf#page=1 "Product Carbon Footprint Summary for NVIDIA HGX H100"), [NVIDIA 2025](https://images.nvidia.com/aem-dam/Solutions/documents/HGX-B200-PCF-Summary.pdf#page=1 "Product Carbon Footprint Summary for NVIDIA HGX B200")).
Dividing by 8 assigns 3–4 kg of assembled-platform mass to each GPU.
It does not establish the mass of a standalone SXM module.
A baseboard comparison must use the price of the complete baseboard, including its shared components.
A module comparison must use the module's own price and mass.
Combining a module-only price with one-eighth of the assembled baseboard's weight mixes two different goods.
For an unweighed module, the useful result is the crossover mass:
$$m_{\text{crossover}}=\frac{P}{G}.$$
At the gold benchmark, a hypothetical \$30,000 component must weigh less than 240 g to exceed gold; a \$40,000 component, less than 320 g; a \$50,000 component, less than 400 g.
The available evidence does not supply a dependable standalone SXM mass for making a universal claim about every NVIDIA product.
[Rack dilution.]{.margin-note} Complete racks add CPUs, switches, power equipment, steel, and cooling hardware.
Zeoli reports GB300 NVL72 purchase orders near [$5,000,000]($2026) and a crated shipping mass near 1,814 kg, or 4,000 lb ([Zeoli 2026](https://www.datagravity.dev/p/how-much-does-an-nvidia-nvl72-cost "How Much Does an NVIDIA NVL72 Cost?")).
Those figures imply approximately [$2.76]($2026)/g, or 2.2% of gold, *using shipping mass*.
They should not be described as an exact uncrated rack measurement.
The order of magnitude is clear even without choosing a rack weight: a [$5,000,000]($2026) system must weigh less than 40 kg to exceed gold at this benchmark.
High purchase price alone does not imply high price per kilogram.
## Bare Silicon
[Conditional mass.]{.margin-note} The silicon die contains the patterned transistors, but excludes the package, memory, circuit board, and cooler.
NVIDIA gives a GH100 die area of 814 mm2 ([Andersch et al 2022](https://developer.nvidia.com/blog/nvidia-hopper-architecture-in-depth/ "NVIDIA Hopper Architecture In-Depth")).
Combining this with a 775 µm incoming-wafer thickness and silicon density of 2.329 g/cm3 gives:
$$
m_{\text{die}}=814\,\mathrm{mm}^2\times0.775\,\mathrm{mm}
\times0.002329\,\mathrm{g/mm}^3\approx1.47\,\mathrm{g}.
$$
The density is tabulated by the Particle Data Group; the thickness is a standard 300 mm wafer specification ([Particle Data Group 2022](https://pdg.lbl.gov/2022/AtomicNuclearProperties/HTML/silicon_Si.html "Atomic and Nuclear Properties of Silicon"), [MSE Supplies](https://www.msesupplies.com/products/mse-pro-300-mm-p-type-b-doped-test-grade-silicon-wafer-100-dsp-1-100-ohm-cm "300 mm Silicon Wafer, P-Type Test Grade, Double-Side Polished")).
This is a conditional geometric estimate, not a weighing or thickness measurement of a finished GH100 die.
In particular, it does not account for manufacturing steps that change the wafer thickness.
[Price allocation.]{.margin-note} Allocating a hypothetical \$30,000 card price entirely to that 1.47 g gives approximately \$20,400⧸g, or 163× the gold benchmark.
But the card also contains valuable memory and packaging, and its price pays for design, software, support, distribution, and margins.
Dividing that complete price by only the die's mass does not reveal a market price for bare silicon.
The weaker inference requires much less: a 1.47 g functional die needs a value of only approximately [$184]($2026) to exceed equal-weight gold.
It is plausible that advanced working logic exceeds that threshold.
The exact multiple remains unmeasured, and the calculation does not establish that only unpackaged silicon can exceed gold.
## The Ounce Error
[Historical comparison.]{.margin-note} Morgan's 2025 comparison in _The Register_ examined NVIDIA accelerators and complete DGX systems over several generations ([Morgan 2025](https://www.theregister.com/2025/11/28/gold_gpu_weights/ "GPUs Aren't Worth Their Weight in Gold—It Just Feels Like They Are")).
It explicitly estimated poorly documented SXM weights.
Its [table](https://image.theregister.com/1684339.webp?format=jpg&height=356&imageId=1684339&width=960 "NVIDIA GPU and DGX System Prices, Weights, and Precious-Metal Comparisons") also mixes avoirdupois ounces with troy ounces.
The DGX A100 row gives 4,336 ounces: exactly 271 lb × 16, matching NVIDIA's 271 lb system specification ([NVIDIA 2020](https://www.nvidia.com/content/dam/en-zz/Solutions/Data-Center/nvidia-dgx-a100-datasheet.pdf#page=1 "NVIDIA DGX A100 Datasheet")).
Gold's quoted ounce is heavier: 31.1034768 g rather than 28.349523125 g.
Hardware/gold value-density ratios therefore require multiplication by 1.09714.
| Historical row | GPU price | Estimated mass (g) | Gold (USD/troy oz) | Published-input ratio (%) | Corrected ratio (%) |
|:--|--:|--:|--:|--:|--:|
| H200, 2024 | [$40,000]($2024) | 1,746.3 | [$2,624]($2024) | 24.7 | 27.2 |
| B200, 2025 | [$35,000]($2025) | 2,993.7 | [$4,114]($2025) | 8.1 | 8.8 |
: Recalculation of Morgan's historical inputs. Original hardware masses were 61.6 and 105.6 ordinary ounces. The correction fixes units, not the uncertain SXM weight estimates or historical price assumptions.
The corrected B200 result makes gold 11.3× as valuable per gram, rather than the published 12.4×.
The unit error is too small to reverse the result, but large enough to matter when claiming a precise multiple.
Historical comparisons also require contemporaneous gold prices: a change in the ratio need not reflect a change in GPU pricing.
## Appendix: Reproduction
The following Python program reproduces the card comparisons, gold-equivalent masses, weight sensitivity, rack estimate, hypothetical die allocation, and historical unit correction.
It uses the dated inputs above and the Python standard library; it does not fetch changing prices.
All monetary inputs below are nominal dollars for the observation year.
The main card table uses the common 2026 gold benchmark; the historical rows use their own quoted gold prices.
Separate rows preserve the workstation price endpoints and the distinction between MSRP and resale.
**Calculation program:** save as `gpu_gold.py` and run `python3 gpu_gold.py`.
```python
#!/usr/bin/env python3
"""GPU/gold comparison compiled 2026-11-13 from dated observations.
Prices are nominal USD, masses are grams, and historical prices retain their
original years. Source URLs and price conditions are documented in the report.
"""
from dataclasses import dataclass
GOLD_USD_PER_G = 125.0
SILVER_USD_PER_G = 1.56
TROY_OZ_G = 31.1034768
ORDINARY_OZ_G = 28.349523125
POUND_G = 453.59237
@dataclass(frozen=True)
class Observation:
name: str
price_usd: float
mass_g: float
CARDS = (
Observation("RTX 5080 FE; open-box sold", 1800, 1645),
Observation("RTX 5090 FE; MSRP", 1999, 1825),
Observation("RTX 5090 FE; new sold", 6500, 1825),
Observation("RTX PRO 6000 WS; ask low", 15000, 1950),
Observation("RTX PRO 6000 WS; ask high", 16300, 1950),
Observation("A100 80GB PCIe; used median", 14950, 1190),
Observation("H100 80GB PCIe; used median", 27682, 1220),
Observation("H200 NVL; used median", 32399, 1237),
)
def describe(item: Observation) -> None:
value_per_g = item.price_usd / item.mass_g
gold_pct = 100 * value_per_g / GOLD_USD_PER_G
silver_multiple = value_per_g / SILVER_USD_PER_G
equivalent_gold_g = item.price_usd / GOLD_USD_PER_G
print(f"{item.name}: {value_per_g:.2f} USD/g; "
f"{gold_pct:.2f}% of gold; {silver_multiple:.2f}x silver; "
f"buys {equivalent_gold_g:.2f} g gold")
def main() -> None:
for item in CARDS:
describe(item)
print("\nAlternative observed price/weight inputs:")
describe(Observation("H200 NVL; retailer mass", 32399, 1280))
describe(Observation("H200 NVL; reported asking median", 66500, 1237))
describe(Observation("H200 NVL; refurbished asking price", 39995, 1237))
describe(Observation("GB300 NVL72; approximate crated system",
5_000_000, 4000 * POUND_G))
print("\nHypothetical module crossover masses:")
for price in (30000, 40000, 50000):
print(f"{price} USD: {price / GOLD_USD_PER_G:.0f} g")
die_mass_g = 814 * 0.775 * 0.002329
print(f"\nAssumed GH100 silicon mass: {die_mass_g:.6f} g")
print(f"Equal-weight gold: {die_mass_g * GOLD_USD_PER_G:.2f} USD")
print(f"Full 30000 USD allocation: "
f"{30000 / die_mass_g / GOLD_USD_PER_G:.2f}x gold")
describe(Observation("Hypothetical full-card-price die allocation",
30000, die_mass_g))
print("\nHistorical unit corrections:")
for name, price, ordinary_oz, gold_per_troy_oz in (
("H200 2024", 40000, 61.6, 2624),
("B200 2025", 35000, 105.6, 4114),
):
incorrect = (price / ordinary_oz) / gold_per_troy_oz
mass_g = ordinary_oz * ORDINARY_OZ_G
correct = (price / mass_g) / (gold_per_troy_oz / TROY_OZ_G)
print(f"{name}: {mass_g:.2f} g; "
f"mixed-unit ratio {100 * incorrect:.2f}%; "
f"correct ratio {100 * correct:.2f}%; "
f"gold is {1 / correct:.2f}x as valuable per gram")
if __name__ == "__main__":
main()
```