Nvidia Raised Prices a Third Time This Year. Bubbles Don't Do That.

Nvidia Raised Prices a Third Time This Year. Bubbles Don't Do That.

To investors,

In late July, Nvidia raised prices again. The company notified the board partners that assemble and sell its graphics cards — the ASUSes, MSIs, Gigabytes and Zotacs of the world — that the memory bundled with its GPUs would cost more, reportedly 20 to 30 percent more. It was reportedly the company's third pricing adjustment of 2026. Within days, AMD matched the move with increases of at least 10 percent on its own cards, and as much as 15 percent on some.

The surface reading is that this is a story about gamers paying more for video cards. It is not. This is a story about the physical price of the artificial intelligence buildout, and it is the strongest current data point the "AI bubble" crowd cannot answer.

Follow the mechanism. AI data centers need memory — enormous amounts of it — and that memory comes from the same fabs that make the DRAM in your laptop and the GDDR chips, the graphics memory soldered onto video cards, on your shelf. Memory makers have shifted 70 to 80 percent of new capacity toward high-bandwidth memory, or HBM, the stacked DRAM that plugs directly into AI accelerators, and toward server memory. Consumer memory competes for the leftovers. The leftovers now sell for 89 percent more than a year ago; some individual memory kits have nearly quadrupled. DRAM contract prices rose 80 to 90 percent in a single quarter earlier this year. When a gaming card gets more expensive, that is not a gaming story. That is an AI story bleeding into the consumer market.

This Is the Abundance-Scarcity Paradox in Real Time

The AI revolution is turning intelligence into an abundant, nearly free commodity. When a thing becomes abundant, the price of everything scarce around it rises. That is the abundance-scarcity paradox, and it is the most useful frame for the next decade of markets. The scarce things in the AI economy are not the models, which are sliding toward commodity status at astonishing speed. The scarce things are physical: memory, wafers, power and time.

Memory is the purest example. HBM costs roughly three times as much as ordinary DRAM, and a single HBM stack now represents more than half the cost of a finished data-center GPU. The GPU is the product people talk about; the memory is the product they line up for. The memory makers know it, and they are remarkably candid about it. Samsung, the world's largest memory maker, warned in January that the shortage would drive price increases across the entire electronics industry in 2026. They allocate capacity to the highest bidder, and the highest bidder is a hyperscaler — one of the cloud giants like Amazon, Microsoft or Google that builds and rents out AI compute.

Here's What a Bubble Looks Like. This Isn't It.

Every bubble narrative makes a testable prediction, and the AI-bubble narrative is no different. It says the buildout is overbuilt, capex will be cut, and the spenders will eventually discover they aren't getting their money back. In August, the New York Times ran an opinion piece warning that money flowing into and out of AI increasingly recalls 2008.

Here is what an actual bubble looks like when it bursts: pricing power breaks. Companies discount. Inventory piles up. The company at the center can no longer raise prices because demand has vanished.

Now look at the data. The company at the exact center of this buildout has raised prices three times in a single year. Its main rival matched the move within days. Its most expensive workstation and AI card, the RTX PRO 6000, now sells for roughly $13,250 — about 55 percent above its launch price in a single year. Even a flagship gaming card that launched at $1,999 has changed hands for more than twice that. That is not vanishing demand. That is pricing power.

Fairness requires one hedge. Nvidia has not officially confirmed the July figures; the 20 to 30 percent numbers come from supply-chain reporting out of Taiwan that a leading hardware outlet rates 85 percent likely to be real. The direction is certain. The precision is reported, so treat the exact percentages accordingly. No bull case depends on the decimals, and no bear case is rescued by them.

The Pricing Power Is Structural

Nvidia can keep pushing prices up for a structural reason: its customers cannot wait, and they cannot substitute. The five largest cloud and AI infrastructure spenders — Microsoft, Alphabet, Amazon, Meta and Oracle — have committed $660 billion to $690 billion of capital expenditure in 2026, nearly double the roughly $380 billion they committed in 2025. All five say their markets are supply-constrained rather than demand-constrained. Microsoft alone is sitting on an unfulfilled Azure backlog of roughly $80 billion, which it describes mostly as a function of power availability, not soft demand. TSMC, the only wafer supplier that matters at the highest end, is raising wafer prices on 3-nanometer and mature nodes at the same time. When the buyers are spending more every quarter and still cannot get enough, the seller holds the cards.

The investment question is where the scarcity premium accrues, and the answer is not where most people look. It does not sit with the model companies, which swim in abundant intelligence and brutal competition. It sits at the physical layer: memory, wafers and power, the owners of the scarce inputs, extract the toll. Nvidia is the toll-taker at the chokepoint, and every price increase it pushes through is proof its customers have no alternative. But the real bottleneck is upstream, in memory, and that bottleneck has a supply calendar. Intel CEO Lip-Bu Tan has said relief will not arrive until 2028. Micron's new HBM fabs come online in 2027, SK Hynix adds capacity through 2027 and 2028, and Samsung's newest plant starts producing in 2028. Industry researcher IDC expects the shortage to persist into 2027 at least. Between now and the new fabs, scarcity sets the price.

What Kills the Story

Being right about a trend means being explicit about what breaks it. Three things.

First, memory is a boom-and-bust industry with permanent scars. After the 2022 collapse, Samsung cut DRAM output roughly in half. When the new fabs come online between 2027 and 2030, and if AI demand pauses even for a quarter, memory prices can crater faster than they spiked. IEEE Spectrum, the engineers' trade journal, flags exactly this boom-and-bust risk.

Second, the customer base is dangerously concentrated. A handful of hyperscalers buy the bulk of AI compute, and every one of them is designing custom silicon to reduce dependence on Nvidia. Inference efficiency improves every generation. Demand for raw compute could plateau faster than anyone models.

Third, the passive capital is already wobbling. Track the iShares Semiconductor ETF (SOXX), and you see roughly $10.2 billion of net inflows this year against about $4.35 billion of net redemptions in the last month alone, per Ainvest fund-flow data. Money is trimming exposure to chips at the very moment the chipmaker is raising prices. That gap — investors selling while the company prices up — is the live version of the entire bubble-versus-buildout debate.

None of that changes the current signal. A bubble that is actually deflating shows up first in the price of memory and in the center company's ability to raise prices. Neither metric is flashing. The bears keep pointing at valuations and capex totals; the market keeps pointing at prices, and prices are the one number neither side controls, because they are what willing buyers actually pay. Nvidia has now asked buyers to pay more three times in one year. Three times, they said yes.

Watch the price of memory the way you watch the yield curve — it is the earliest honest gauge of whether the buildout's pricing power is intact. When memory prices stop climbing, that is when the story changes. Until the fabs come online, the toll collector keeps the rate on the board.

Abundance is the story. Scarcity is the toll. And the toll just went up.