Nine thousand dollars. That is what a GeForce RTX 5090 now costs at the top of
the market, and no, it does not come with a car to soften the blow.
Read that back.
Even the floor is silly. The cheapest one a third-party seller will part with
sits at $6,699, a 235 percent markup on the $1,999 launch price Nvidia was
quoting eighteen months ago. The US median has settled nearer $4,700.
Whichever number you pick, they all say the same thing: the halo gaming card
of 2026 costs more than most people’s whole rig, chair and monitor thrown in.
The part that stings is that this is not Nvidia gouging for sport, and it is
not scalpers, though the scalpers are having the time of their lives. The
culprit is memory. The 5090 carries 32GB of GDDR7, and GDDR7 has become one of
the most supply-starved parts in the entire industry, because Samsung, SK
Hynix and Micron have all swung their factory capacity toward the HBM chips
that AI data centres are begging for, which means every wafer that turns into
a data-centre memory chip is a wafer that does not turn into the memory
soldered onto your graphics card.
We are losing that fight.
The numbers are bleak. The cost of 16GB of GDDR7 reportedly climbed from
around $65 to $80 per card in mid-2025 to more than $200 by year’s end, and
back in May Nvidia warned its partners to brace for roughly a $300 jump in
what it costs to build a 5090. In March the street price was a “mere” $3,500
to $4,000. It has close to doubled since, while we all stared at our wishlists
and gave up.
Where the ceiling is, nobody knows, because the AI memory gold rush shows zero
sign of cooling. One knock-on we did not see coming:
AMD has reportedly overtaken Nvidia in GPU sales, mostly because a Radeon you can afford beats a GeForce you cannot.
If you were circling a 5090, put the mouse down. Our
2026 GPU value guide
has the cards that still make sense. This one does not. (Figures via
Wccftech
and
TechTimes.)
