NVIDIA Admits the GPU Shortage Is ‘Beyond Its Control’: DRAM Root Cause, Record Highs Across Every Tier, and a Reshuffled Build Economy#

Zixing | October 20, 2026

Hello everyone, Zixing here.

From tallying the combined GPU and storage ledger in late September, to the October 1 DDR5 supply-pause warning, the October 11 DDR5 build decision guide, and the October 12 teardown of the RTX 5090 discontinuation rumor — I have written four hardware market pieces in a row, and every judgment rested on “channel data plus leak reporting.” In the discontinuation piece I said at the time: NVIDIA may never formally announce anything, and if you wait for official confirmation before deciding, the signal will probably never come.

On October 16, that changed. NVIDIA itself spoke up — the same day, in two cities, with two executives, delivering two statements that looked completely different yet complemented each other. In Seoul, the company admitted the shortage is “beyond its control”; in Taipei, it denied discontinuation. This piece does three things: clarifies how the official line differs from the earlier rumors, dissects the DRAM root cause, then puts the full price-tier picture across China, the US, Germany, and South Korea into one table — and finally talks about what it means for anyone building a PC around Double 11.

1. Two Cities, One Day: What NVIDIA’s Official Line Actually Said#

First, the facts of the two October 16 statements.

The Seoul session: NVIDIA held an RTX Spark launch event at COEX in Samseong-dong, Seoul. After the event, Ida Keita, vice president of developer relations, told Korean outlet THE ELEC that the consumer GPU shortage is rooted in memory (DRAM) supply and reflects a whole-semiconductor supply chain problem spanning GPUs and DRAM. His exact words:

“This is not something NVIDIA can control.”

The Taipei session: The same day, Jeff Fisher, senior vice president of personal computing, addressed the discontinuation rumors at an event in Taipei: the RTX 5090 is an important core product for the company, will not be discontinued, and will keep shipping — while conceding that supply capacity is constrained and offering no timeline for a fix.

Read together, the two statements carry more information than any leak. Here is a table comparing the official line with the earlier rumors:

The most important rows in that table are the last two: what the company officially denied was “discontinuation”; what it did not deny was “tightness.” Fisher said the 5090 will not be discontinued but never said when capacity recovers; Keita pointed the root cause straight at DRAM — a substantive shift in how NVIDIA executives talk about the consumer market. On earlier earnings calls, NVIDIA CFO Colette Kress’s line was strong demand and healthy inventory, with supply shortages challenging the gaming business into the first quarter of next year and beyond, while TSMC capacity and GDDR7 inventory remain adequate and the bottleneck sits in the material supply of board-manufacturing partners. Aligning Kress and Keita, they point the same direction: wafers aren’t short, memory chip inventory isn’t short — what’s short is the DRAM material reaching the board makers. Two phrasings of the same causal chain.

2. Why the Official Verdict and the Rumors Are Two Different Things#

In the October 12 rumor teardown I laid out a credibility grading for the rumor: the leaker’s tier, how widely media adopted it, the forgery risk of the screenshot — landing on the conclusion that “a rumor is not a fact.” Looking back today, that framework still holds, but one premise has been rewritten by the official statement — the truth of the rumor is no longer the premise in question; the shortage itself is now a fact the vendor has acknowledged.

The differences come in three layers:

1. Different source tiers. The earlier GB202 allocation rumor came from leaker tweets and a screenshot of a purported factory notice; however well it fit commercial logic, it stayed third-party. The October 16 “beyond control” came from an NVIDIA VP speaking to media after a formal event — a quotable, official source.
2. Different questions answered. The rumor answered “will the 5090 still be made”; the official denial answers that too — but Keita’s statement answers the more upstream question: “why is it short.” One is a product-line decision; the other is supply chain structure. Denying discontinuation did not ease the tightness — it nailed the cause of the tightness onto a link NVIDIA itself admits it cannot control.
3. A timeline echo. The rumor burst (around October 10), the site listing pulled, Micro Center changing status, and then the two-city statements (October 16) — less than a week apart. The speed itself shows this was no routine “no comment”: the shortage is severe enough that two VPs had to step out separately to set the record.

The full timeline in order:

• Around October 10: reports emerged that NVIDIA would stop allocating GB202 chips to AIB board partners, with flagship silicon shifting entirely to the RTX PRO professional line;
• After the rumor broke: the NVIDIA US site pulled the RTX 5090 product listing, and several related Micro Center models switched from out of stock to “no longer sold, no restock”;
• October 16 (Seoul): Ida Keita admitted the shortage’s root cause is DRAM and is “beyond control”;
• October 16 (Taipei): Jeff Fisher denied discontinuation, confirming shipments continue but capacity is constrained;
• October 20-22: GTC Berlin, where Jensen Huang speaks for three straight days on the “AI factory” theme.

From rumor to official verdict took under a week, and the official answer is heavier than the rumor itself — the rumor said “production stops”; the official line says “production continues, but the supply problem is beyond even our control.”

One more piece of background worth noting: according to the Financial Times, Chinese tech companies have bought more than 1 million gaming/workstation-grade GPUs this year to plug the AI compute gap (as relayed by Chinese-language media; the original was not checked). If that magnitude holds, “AI purchase orders sweeping up gaming cards” is no longer speculation but reported demand data — and paired with the capacity data below, the picture becomes complete.

3. The DRAM Root Cause: GDDR7 and HBM Fighting Over the Same Production Line#

Keita’s “whole-semiconductor supply chain problem” unpacks into one capacity siphon. The factual layer:

1. Shared production lines: GDDR7 memory for consumer GPUs shares advanced-node capacity with HBM for AI servers — both feed on the most advanced DRAM process nodes.
2. Vendor tilt: Samsung, SK Hynix, and Micron have already allocated 70%-90% of advanced capacity to HBM first. What remains for consumer memory is at most 30 percent.
3. Demand-side scale: Per figures relayed from IDC, AI data centers will take about 70% of global memory capacity in 2026 — versus only 20%-30% in 2022. In four years, the memory industry’s capacity allocation has flipped completely.

With capacity drained, prices moved by orders of magnitude. Here is the price data I have on hand, in one table:

Note the timing gap in the last two rows: in early April, DDR5 spot sat at $31.38; by October 16 the chip spot basis read $60.60 — spot prices climbed another step in just over half a year (the two figures come from different statistical bases; they are presented side by side here without conversion).

The most striking part of this chain is the mismatch of “ample inventory, scarce material”: per Kress’s earnings-call line, TSMC capacity and GDDR7 inventory remain adequate, and the bottleneck is the material supply reaching board-manufacturing partners. In plain language: the fabs are running and chips sit in inventory, but board makers cannot get (or cannot afford) the DRAM material they need to assemble — graphics card output is capped not by NVIDIA itself but by when the memory market loosens. That is exactly what Keita’s “beyond control” means technically: NVIDIA can decide silicon allocation; it cannot decide the production schedules of Samsung, SK Hynix, and Micron.

How does memory cost pass through to the shelf price? The Seoul event offered a blunt, real-world sample. The RTX Spark laptop starts at about 4.7 million won in Korea and reaches about 10 million won configured, and an NVIDIA official on site explained the pricing candidly: “When the product was being developed, 128GB was still cheap, so we treated it as the optimal configuration — we didn’t expect memory prices to rise this much.” — a product’s memory decision was locked in during development, and by mass production the memory price had multiplied several times over; the difference could only show up in the price tag. That is the micro-level model of this whole consumer GPU price wave: board makers’ material costs get locked in high at purchase time, and retail follows upward.

Keita also added a data point on RTX Spark’s positioning: the device can run AI models with more than 100 billion parameters locally, with no external cloud service and no token fees. In other words, this product line competes for exactly the demand HBM capacity serves — NVIDIA is pushing local AI on consumer devices while local AI’s threshold (large memory) is the very thing squeezing gaming GPU supply. This is not a conspiracy where vendors deliberately starve supply; it is the same AI demand projected onto both ends of the product line.

As for when the shortage eases, there is no official timeline, and upstream vendors’ remarks only go so far: Micron’s CEO recently said the memory shortage could last into 2028; other reports suggest SK Hynix’s DRAM, NAND, and HBM inventories could be exhausted by the end of 2026 — the latter is rumor-grade, take it as noise, not fact.

4. The Full Price-Tier Panorama: Four Markets in One Table#

Now for this piece’s core data section. I have pulled the China, US, German, and South Korean market snapshots into one panorama table — every number is a snapshot from its source at its own point in time, with no conversion:

Several points in this table deserve to be spelled out.

First, China’s “one price per day” and stock hoarding. One detail in the Southern Metropolis Daily monitoring: some sellers “sealed their warehouses, refusing to sell, holding out for higher prices.” The early-October basis had the 5070 Ti past 8,000 yuan; by the October 19 channel spot-check it read 10,999 yuan — two different channel bases at two different times, but set side by side they are the plainest footnote to “one price per day.” Separately, Macau customs reported in September a smuggling case at a port of entry containing two 32G 5090s — when the channel price gap gets large enough to justify smuggling, that tells you how distorted regular supply has become.

Second, the four markets show different premium shapes. Germany is “average price up plus stockouts,” with a clear statistical basis (launch €2,099 to €6,000-6,800); the US is “transaction prices scattered wildly,” a $4,300-10,000 range that says channel markups have gone fully disorderly; South Korea is “extreme multiples,” with transaction prices at five times the tax-inclusive MSRP and hot cards sold by lottery or first-come-first-served. China is “steady climb plus hoarding.” One global shortage, four shapes depending on channel structure — but all pointing the same way.

Third, both tracks of China’s dual product line got pulled up. The China-exclusive 5090D v2 (24G version) rose from about 28,000 yuan in early September to 36,000 to nearly 40,000 yuan, up to over 40% — that tier exists precisely to balance price and supply, and the market has pushed it up all the same. That shows this wave is not an isolated 5090 SKU event but a wholesale rise from flagship to sub-flagship; the 5060’s 2,700-3,100 yuan quotes are the lowest absolute numbers, and they too sit in the “across-the-board rise since August” channel.

Third, one more unconfirmed piece of bad news on the supply side. Channel sources say NVIDIA has cut production across the RTX 50 line and that Q4 consumer GPU supply will be severely short — note, this is a channel claim, not officially confirmed, and it coexists with Fisher’s “no discontinuation, shipments continue” line. The two do not directly contradict (cutting output is not the same as cutting to zero), so treat it as reference, not proof.

By the way: October 20-22, Jensen Huang speaks at GTC Berlin for three straight days on the “AI factory” theme; per relayed figures, NVIDIA’s Q2 data center revenue was $89 billion, up 117% year over year, over 92% of total revenue. On one side, a data center business taking more than nine-tenths of revenue presents the future at a conference; on the other, a gaming business taking under a tenth is being squeezed into lottery sales at retail. Which way resources tilt is not hard to understand commercially.

5. The Divergence: Build Economics Have Been Redrawn#

Looking only at GPUs and memory, you’d think the whole DIY market is rising. Widen the lens and this wave is actually a violent structural divergence:

This divergence reshapes building in a structural way: the big share of a build’s cost is now decided entirely by the GPU and memory. Budgets used to spread roughly evenly across CPU, motherboard, GPU, memory, and storage, and a price spike in any one part could be balanced by adjusting the mix; now GPU and memory are locked high at both ends while SSDs and CPUs sit in historic low-price territory — for the same total budget, GPU plus memory take an ever larger slice, and everything else has never been better value.

The storage split deserves special attention: DRAM is surging while SSDs (the NAND camp) are falling. The cause is again capacity allocation — this shortage centers on DRAM (system memory and VRAM), while NAND demand displacement is nowhere near as extreme. For builders, this is one of the few “tailwinds” in the current budget: if you need big storage, now is actually a relatively good time.

The PC market data confirms the pressure from the demand side: per IDC as relayed, global PC shipments were 62.7 million units in Q3, down 20% year over year (relayed basis, single source). Rising system costs suppressing upgrade demand — that pass-through lines up with the memory price timeline. By the way, the console market is going through the same cost repricing — the used PS5 Pro creeping toward 10,000 yuan is a market I covered in the era of 10,000-yuan consoles; that’s a different battlefield, so I won’t repeat it here.

6. The Secondhand Signal: New Cards in Chaos, Old Cards Anchored#

When new-card channel prices get pushed to absurd levels, the used market’s reference frame changes: a used card’s price anchor is last generation’s performance and real transaction values, not scarcity premiums. Per channel intelligence (single source), the used RTX 4090 has recently been seen as the relatively rational option for users who genuinely need the upgrade — listed prices swing a lot, so I won’t quote specific numbers here, only the structural logic:

1. Relative value improves. When new flagship channel prices hit five times MSRP (the South Korean basis), the previous generation flagship’s used price naturally sits lower in relative terms — it’s not that the 4090 got cheaper; the new cards got more absurd.
2. But absolute levels are high too. In a rising market there is no “bargain window” in used; listings rise with new cards as a matter of course. Counting on used prices falling against the trend has no data behind it right now.
3. The buyer-beware nature hasn’t changed. In a high-price environment, secondhand risk exposure (no warranty, opaque condition and provenance) only grows — the premium buys immediate certainty, and whether that’s worth paying is a personal call.

In one sentence: the used route today is “relatively rational,” not “cheap.” Treat it as a backup path around channel chaos, not an arbitrage opportunity. Concretely, three tests:

1. Your old card still holds up — stay put, watch the supply-side signals, and use the observation checklist above;
2. You truly need flagship performance and accept no-warranty risk — a used 4090 is a viable backup, but verify condition and provenance yourself before the deal; market analysis cannot hedge this risk;
3. First build, no card-inspection experience — skip used; a mainstream new-card tier is the safer landing spot.

7. Double 11 Build Advice: Pick Your Budget Tier#

Finally, the practical layer. For this year’s Double 11 build decisions, I split the budget into four tiers:

Mainstream tier (RTX 5060, quoted at 2,700-3,100 yuan) — if you need it, buy. This is the lowest-priced tier in the lineup and the least affected by this shortage — both the official verdict and the four-market data point to the flagship end as the epicenter; mid-range cards are not at the storm’s center. Per the decision matrix in the October 12 piece, the “mid-range, buy as usual” judgment still holds under the new official line. One caveat: “across-the-board rises” are a fact and quotes can move up at any time; if you genuinely need the card, don’t deliberately wait for Double 11 — in a shortage like this, promo discounts and price hikes coexist as a norm, so don’t expect Double 11 to beat the channel premium back down.

High-end tier (5070 Ti past 8,000 yuan, 5080 above 11,000 yuan) — waiting is justified, but wait for the right thing. The CFO’s line says the supply shortage will challenge the gaming business into the first quarter of next year and beyond, and Micron’s CEO even points at 2028 — in a window like that, “waiting for Double 11 sales” is waiting for the wrong thing; the signal worth waiting for is capacity change on the supply chain side. If your current card works, keep using it; if you must buy, prefer established channels with stock on hand and proper invoices, and steer clear of scalper-inflated sources.

Flagship tier (5090D v2 at about 36,000 to nearly 40,000 yuan, 5090 32G at about 54,000 yuan) — either accept scarcity pricing or walk away. Prices in this tier now reflect “scarcity pricing,” not “value pricing”: five times MSRP in South Korea, hoarded warehouses in China, even a smuggling case. Three concrete suggestions: first, don’t touch gray-market channels of unknown origin (smuggled cards seized by customs carry no warranty or provenance guarantees); second, if you accept the lottery/first-come-first-served rules, participate through regular channels and never leverage up to pay a channel premium; third, if the 5090 is not a genuine need for you, the optimal decision this generation may simply be “don’t buy” — the vendor itself says supply is beyond its control, and the urgency of paying scarcity pricing should not be manufactured by a promo calendar.

The used route — a backup, not a default. As above, the used 4090 is a relatively rational option per channel intelligence, but in a high-price environment the used market rises with everything else, and the risk is yours. It suits veterans who know exactly what they’re buying, not first-time builders.

I won’t rehash memory here — the “buy/wait/swap” matrix in the DDR5 build decision guide still applies; SSDs and CPUs sit in historic low-price territory (channel monitoring basis), so those two budget anxieties are currently the smallest, and you can spec them as needed.

Watchlist: What Signals a Supply Turning Point#

More important than “buy or not” is “what to watch.” Since there is no official timeline, here is a list of publicly observable signals — adjust strategy if they change, hold the judgments above if they don’t:

Every item in this table is a publicly verifiable factual state that needs no insider knowledge — watching state changes beats staring at coupon pages.

8. Risk Notes and Data Freshness#

As always, let me lay out the data’s condition before closing:

• Channel claims are not officially confirmed: “RTX 50 line production cuts, severely short Q4 consumer supply” is a channel claim; NVIDIA’s official line is only “capacity constrained, no timeline,” with no confirmation of cuts.
• Single-source data: the ZhiTai SSD and CPU historic lows, the October 19 channel spot-check prices (5090D 32GB at 49,999 yuan, 5070 Ti at 10,999 yuan), the used 4090 trend, IDC’s Q3 PC shipments of 62.7 million units, the $60.60 DDR5 chip spot — all are single-source relays, not cross-verified.
• Relayed figures: the Financial Times item on “Chinese tech companies buying over 1 million gaming/workstation-grade GPUs” is a Chinese-language relay, original unchecked; GTC Berlin and NVIDIA’s Q2 data center revenue of $89 billion, up 117% and over 92% of revenue, are likewise relayed figures.
• Rumor-grade: SK Hynix inventories running out by the end of 2026 is rumor-grade and not cited as fact.
• Basis timing differences: the 5070 Ti appears in this piece both as “past 8,000 yuan” (early October) and “10,999 yuan” (October 19 spot-check) — different channels and time points, presented side by side with no adjudication.
• Price freshness: all prices are snapshots from early-to-mid October and highly volatile; check live channel quotes the day you order.

Closing Thoughts#

From the late-September ledger to this piece, five articles in a row, and my biggest takeaway is that this shortage has finally been upgraded from a “channel phenomenon” to a “vendor verdict.” “This is not something NVIDIA can control” — spoken by a company that draws over ninety percent of its revenue from data centers, that line weighs more than any channel leak. It means the consumer GPU supply problem is not a bump that a promo cycle can smooth over, but the result of an industry-level reallocation of resources: memory lines ceded to HBM, board materials held hostage by the memory market, and AI purchase orders still expanding (Berlin’s “AI factory” theme says it outright).

For builders, my stance hasn’t changed — it has just sharpened: don’t fight a structural shortage with “wait”; adapt your budget structure to it — spec up on the historic-low parts (SSDs, CPUs) as needed, hand the GPU decision to real need rather than a promo calendar, and squeeze one more month out of your old card. Every extra month an old card holds on is an option you keep for your future self — the option of not buying at the very top.

I’m Zixing — see you in the next hands-on piece.