The 70% Markup Is the Business Model Now

NVIDIA's consumer cards sell, on average, roughly 70% above their listed MSRP today, against 28% for AMD and 21% for Intel. The markup itself is not the part that matters. What matters is that only one of those three companies has the volume and the flagship halo to make a premium that large stick across an entire product stack.

That distinction is worth holding onto, because the usual explanation — memory is expensive, therefore cards are expensive — only accounts for the direction of travel, not the magnitude. Memory got expensive for everyone. NVIDIA's customers paid far more than proportionally for it, and the mechanism behind that is not a shortage.

The Flagship Sets the Ceiling

GamersNexus' year-long retail price tracking puts NVIDIA's average at about 70% over MSRP, up from 39% in February. Its flagship is where the number comes from. The average in-stock price of an RTX 5090 reached $4,880, roughly $2,900 — or 144% — above a $2,000 MSRP, after climbing $900 in six months.

Strip out the usual complaints about scalpers and one structural fact remains: the 5090 is the only consumer card on the market with more than 16GB of VRAM, and AMD has declined to field anything above the 9070 XT. There is no competing product at the top, so the price there stops responding to the market and starts responding to the seller's patience.

The bracket data shows the same thing from a different angle. Cards with 16GB or more went from 44% over MSRP in February to 67% now. Cards at 12GB or less went from 23% to 34%. The premium is tracking memory capacity, not silicon, and the gap between the two groups has widened at every snapshot since launch. The lone exception proves the point: the 8GB RX 9060 XT is the only current card whose average price fell, by $15, because demand for it is thin enough to beat supply constraints.

Memory Explains the Slope, Not the Height

There is a genuine cost story underneath this. GDDR7 sits on NVIDIA's cards while AMD and Intel remain largely on GDDR6, so the NVIDIA stack absorbed more of the shortage by design. A 3GB GDDR7 module has been reported at $60 to $70 against $20 for the standard 2GB part, and NVIDIA reportedly raised GDDR6 kit prices to board partners as well. Because it sources the memory and ships it bundled with the GPU die, raising the bundle moves every partner's cost floor at once.

But the arithmetic gets uncomfortable past that point. AIB unit shipments rose about 28% year over year from the first quarter of 2025 to the first quarter of 2026, while sales rose nearly 76%. Worked out, the average selling price of a completed card went from roughly $668 to $918. Jon Peddie Research's reading was that suppliers appear to be using well-publicized memory costs to raise card prices by more than memory actually costs them.

That is not evidence of collusion. It is evidence of opportunity. A crisis everyone can point to is the most useful thing that can happen to a channel that wants to expand margin, and the industry has spent a year pointing at it.

The Reset Mechanism Never Arrived

Normally a Super refresh resets street pricing. This cycle it did not come. The launch moved repeatedly, and at least one board partner has reportedly had cards in hand while the release stayed parked over the cost of those 3GB modules. Without a new SKU there is no moment where prices re-anchor, and NVIDIA has left its MSRPs untouched rather than restate them at what the hardware actually sells for.

The one product NVIDIA sells directly removes the partner alibi entirely. RTX PRO 6000 went from $8,565 at launch to $16,000 on NVIDIA's own marketplace. Nobody else set that number. If a 5090 sustains $4,000 to $5,000 for another year, expecting an RTX 6090 at a $2,000 sticker is hopeful rather than reasonable — and the more likely outcome is that the 90-class card drifts fully into workstation pricing.

Downstream, the channel is planning around scarcity. PC Partner, which owns Zotac, Manli and Inno3D, told investors that VGA shipments fell substantially in the first half of its fiscal year and that availability gets worse in the second, with entry-level cards facing the deepest shortages. MSI's chairman reportedly expects consumer shipments to fall 10% to 20% while unit prices rise. Intel, meanwhile, accounts for just over 2% of in-stock listings and three or fewer models in several tiers — a third competitor that cannot reach the shelf is not a check on anything.

What I Think

The conclusion is not that GPUs got expensive. It is that the number printed on the box stopped being a price and became a comparison figure that exists for marketing decks. Once a company learns it can hold a flagship at more than twice its sticker for eighteen months without losing the sale, the next launch does not get cheaper. It gets priced at what the market proved it would pay.

So: if you have working hardware, keep it. A 3080 or a 4070 Ti that still runs your games is a better asset now than it was in 2023, and the used market has already figured that out. If you genuinely need a build, buy the tier you need and stop waiting for a correction that requires either an AI capex collapse or a memory supply miracle. Neither arrives on a schedule you can plan a build around. And when the next 90-class card shows up looking like a workstation part with a gaming logo, remember this year — it was priced that way because it sold.