No, RAM is not worth more than gold by weight — the true numbers are strange enough

A claim is circulating that RAM is now worth more than gold by weight, and that AI is the reason.

Half of that is right. Let us do the second half properly, because it takes about a minute.

The arithmetic

Gold on 25 August 2026 sits at $4,675.96 an ounce — $150.34 a gram.

A mainstream 32GB DDR5 kit currently runs $399 to $479, with the wider market between $380 and $589. Call it $400.

Now weigh it. A two-module kit with heatspreaders comes to roughly 80–100 grams, which puts the kit at about $4 to $5 per gram — around thirty times cheaper than gold.

And the conclusion does not depend on getting the weight right. Suppose the entire kit somehow weighed only 20 grams, which is less than one bare module. That is still $20 a gram against gold's $150. The claim fails by a wide margin at any plausible weight, which is the useful kind of check: it survives being wrong about the inputs.

The only way to make the comparison work is to weigh the silicon alone and price the finished product — to divide the cost of a thing you can buy by the mass of a thing you cannot. That is not a price comparison. It is a sentence that sounds like one.

What is actually true

The real figures did not need the embellishment:

A 32GB DDR5 kit that sold near $100 a year ago now lists around $400. A 16GB kit that was under $100 is now upward of $240.

That is a consumer product quadrupling in a year, which almost nothing outside a genuine shortage ever does.

Why it is happening

Samsung, SK Hynix and Micron control more than 95% of global DRAM output. All three are moving wafer capacity away from consumer memory and toward high-bandwidth memory for AI accelerators.

AI is expected to consume 20% of total DRAM production in 2026.

The shortage is therefore not a disruption. Nothing broke, no fab burned down, no ship ran aground. Three companies made the same rational decision at the same time — sell the same silicon into the market paying more — and consumer memory is what is left over.

That is a more interesting story than the gold line, because it is a structural fact about who gets served when a general-purpose input becomes scarce. It is the same shape as Broadcom pointing AI at Spring and finding 200-plus vulnerabilities in a year: the resource goes where the return is, and everyone else adjusts.

When it ends

Not soon. Leadership at Synopsys has put the shortage as lasting at least through 2027, and some forecasts push full resolution out toward 2030. New fab capacity is the constraint, and fabs take years.

What to do if you are buying

• Buy the RAM you need now, not the RAM you might want later. The usual advice — overspecify memory because it is cheap — is currently backwards.
• Check second-hand and older standards. DDR4 pricing has risen too, but the delta is different, and an older platform may be the cheaper total build.
• If you are specifying laptops or phones for the next cycle, expect the memory tier to cost more than it used to. Manufacturers are absorbing some of this and passing on the rest.
• Do not wait for a crash. The capacity decisions driving this were made for good commercial reasons and are not about to reverse.

Why the claim spread anyway

Because "more expensive than gold" is a sentence people can repeat, and "DRAM contract prices rose 170% in 2025 as producers reallocated wafers to HBM" is not.

The check took a minute and a search for the gold price. That is the whole cost of not repeating it.

What is not established

• The exact weight of a given kit. We used a range; the conclusion holds across it and well beyond it.
• Where the claim originated, or whether its author was comparing bare die.
• Whether the 400% estimate lands. It is a projection to end-2026, not a measurement.
• How much of the increase reaches retail prices of finished devices, as opposed to being absorbed by manufacturers.