Behind the RTX 50 “One Price a Day” Phenomenon: AI Compute Squeezes GPU Capacity, ABF Substrate Lead Times of 48-56 Weeks, and the Double 11 Build Dilemma#

Zi Xing | October 29, 2026

Hello everyone, I am Zi Xing.

The Double 11 build season colliding with “one price a day” graphics cards is the most awkward combination facing anyone building a PC at the end of this year. On October 25-26, Securities Times, China Economic Net, Yicai and other media outlets once again visited Shenzhen Huaqiangbei in person: spot quotes for the RTX 5090 generally sit at RMB 51,000-55,000, with one merchant quoting RMB 53,000 tax-included for the ASUS TUF model; other reports say channel quotes have already climbed to RMB 55,000-60,000. Some merchants recall that this card was still quoted at just over RMB 40,000 last month, a rise of nearly RMB 10,000 in a little over a month. The same model can differ by one or two hundred yuan within two days, and merchants do not dare to issue written quotations casually — the market has moved so fast that even black-and-white paperwork has become unaffordable.

This site has already published two pieces on this round of the market: Chasun’s channel “financialization” analysis dissects the financial logic behind the divergence between spot prices and share prices, while my own Huaqiangbei on-the-ground market visit records the price structure at the front-line stalls. This piece deliberately avoids restating the market and instead answers two more fundamental questions: first, why prices cannot go back — taking apart the complete transmission chain, from substrate lead times to channel stock-locking, by which AI compute demand structurally squeezes upstream capacity; and second, how on earth to build a PC for Double 11 — where to spend the budget so that it is not wasted, against a backdrop of runaway GPU prices.

One disciplinary note first: all prices in this article are channel snapshot data at the time of writing (October 29, 2026); the market moves in real time, so please treat live quotes as authoritative; no figure here constitutes a forecast of future prices.

I. The Market Baseline: Prices Across All Tiers and the Timeline of This Round#

Let us first lay out the baseline needed for the decision. Using the framing of “official initial retail price → current channel price,” here are the four main tiers in one table:

A few definitional details need to be nailed down:

• The RTX 5090 has two parallel figures: Huaqiangbei spot quotes generally sit at RMB 51,000-55,000 (ASUS TUF tax-included at RMB 53,000), while reported channel quotes have reached RMB 55,000-60,000. The gap between quotations and actual transactions is itself a variable in a “one price a day” market.
• Two field-visit figures for the RTX 5060: the 21st Century Business Herald figure was previously about RMB 2,300 and has now risen to above RMB 3,600; the Yicai field-visit figure is about RMB 4,000 currently, still in the RMB 3,000 range two weeks ago, and just over RMB 2,000 a little over two months ago, for a cumulative gain close to double. Mainstream cards have not been spared either.
• The RTX 5080 has risen in steps: the manager of Jinshuoli Computer, who has worked in Huaqiangbei for 16 years, says the channel price was about RMB 12,000 half a month ago, rose RMB 4,000-5,000 in the short term, and now stands at RMB 16,000-17,000.
• The RTX 5070 platform transaction figures are not uniform: the Dewu average transaction price rose from RMB 4,899 to RMB 9,498 (about 94%), while Manmanmai records the Colorful RTX 5070 12G going from RMB 6,599 to over RMB 10,000 (a gain of over 60%) — with different models, brands and channels, the gain readings differ too, so always align definitions before comparing prices across platforms.

The flagship tier has the most scattered quotation figures, so here are the RTX 5090 quotes scattered across various reports collected into one table to avoid confusion:

Add the third-party transaction platform data on top, and the breadth of the price increases becomes complete:

On the timeline, this round is not a sudden outbreak but a three-stage acceleration. Manmanmai price-comparison data shows the increase starting in late July; accelerating markedly from September; and by late October, the market entering a “one price a day” state. Here are those three stages of acceleration as a timeline:

What characterizes this market better than “how much prices rose” is the change in quoting behavior itself. Put normal market conditions and current conditions side by side and the difference is immediately clear:

The detail of “dare not issue written quotations” deserves an extra sentence. A quotation is the foundation of a transaction: it fixes the price and gives later price comparisons and after-sales recourse something to stand on. When the same model carries a RMB 100-200 price difference overnight, issuing an order at the old price means the merchant may have to pay out of pocket to restock, so refusing written quotations becomes rational self-protection — and the cost is passed on to the buyer: without written proof, the entire chain of price comparison, negotiation and after-sales recourse loses its anchor. This is also why this article keeps stressing “treat live quotes as authoritative” — it is not boilerplate disclaiming, but the fact that quotes in this market are valid only at the moment they are given.

II. Why Prices Cannot Go Back: The Transmission Chain from AI Data Centers to PC-Build Shops#

Read enough market articles and it is easy to attribute price rises to “scalper speculation.” But channel speculation cannot explain two things: first, the persistence of the gains — from late July to late October, three months without any meaningful pullback; second, the structural nature of the increases — low-end discrete graphics cards have changed little, with the rises concentrated in the high-end RTX 50 segment. The real cause has to be sought upstream: the crowding-out of capacity by AI chips has already been transmitted from the wafer layer to the packaging layer.

The key player this time is the ABF substrate (Ajinomoto Build-up Film) — the base board indispensable for high-performance chip packaging, needed by both AI chips and high-end GPUs. According to market research firm TrendForce, ABF substrate lead times have stretched to 48-56 weeks, with the longest wait for new orders at about 13 months; NVIDIA, AMD and Intel have together prepaid over USD 5 billion to Taiwanese substrate makers to lock in capacity. The trouble is this: the new capacity locked by those prepayments is absorbed by AI chips first, and the substrate resources available to gaming GPUs are being structurally squeezed.

Take this transmission chain apart and every link is backed by facts:

Why does the substrate become the bottleneck? Some mechanism background: a GPU chip is not mounted directly onto the graphics card; it must first be packaged onto a high-density substrate — and high-end chips use exactly ABF substrates. Substrate capacity is independent of foundries; expanding it requires building new production lines and ramping yields, and the current longest wait for new orders is about 13 months — a timescale that means the packaging-layer gap cannot be closed by working overtime. So the types of buyers crowding in front of substrate capacity right now become the key to understanding this market:

The scale on the demand side deserves a look of its own. Per reports, OpenAI recently signed large orders of 4-5 million graphics cards with NVIDIA and then a 2.4 million card order with AMD — the scale of these two orders is currently attested only by a single reprint source with no official confirmation, but even if only partially fulfilled, their volume is far beyond the normal demand of the consumer market. Meanwhile, there is one easily overlooked increment on the demand side: the RTX 5090 can run some quantized large-parameter language models and can also carry fairly complex AI image generation workloads, so local large model deployment and AI image generation are pushing SMEs and developers into the buying pool — the first time a gaming flagship card has simultaneously taken on the role of “compute equipment.”

Official statements on the supply side present a subtle combination: the market once circulated news that NVIDIA had discontinued the RTX 5090; NVIDIA executives have publicly denied the discontinuation and confirmed the product will continue shipping — while at the same time acknowledging that supply chain capacity constraints exist. Translated: the card is still being produced, but the rate of production cannot keep up with the demand being allocated; there is also a claim circulating in the market that “NVIDIA is directing more chip resources to professional computing products,” likewise unconfirmed officially. For PC builders, the official denial of discontinuation is good news and the official acknowledgment of capacity constraints is bad news — and the fact that both are true at once shows precisely that the supply gap is not a rumor but structural.

One more thing to watch is how rumors and the market interact. The Securities Times on-site report poses the question directly in its headline: “Have NVIDIA discontinuation rumors set off a price surge?” — the rumor has been officially denied, but its effect on channel sentiment is already visible in quotes: it is precisely the imagination of “possible discontinuation” stacked on top of the reality of “capacity constraints” that pushes channels toward panic stock-locking and hoarding, which in turn pushes quotes even higher. For unverified news of this kind, this article’s principle is to describe only the phenomenon of it “stirring the market,” and not to reason about prices from it as if it were a capacity fact.

That is the answer to “why prices cannot go back”: substrate expansion is measured in years (longest wait for new orders about 13 months), giant long-term contracts run 6-12 months per signature, and a single AI order runs to millions of cards — all three timescales are far longer than the Double 11 promotion cycle. Channel speculation can dissipate with sentiment; a misallocation of capacity cannot.

III. Double 11 Build Decisions: Spend the Budget Where Prices Are Steady#

Structure dissected, back to the decision. The conclusion first: in this round, the hardest hit is the high-end segment where the RTX 5080 and RTX 5090 sit, while CPU, storage and memory prices are overall steady or even edging down — the rational strategy for building is to avoid the GPU peak and put the money into the steady components.

The 21st Century Business Herald field visit gives a clear category divergence: storage and CPU prices are overall steady with very little volatility; memory prices edged down from September to October; low-end discrete graphics card prices have changed little; and price rises are concentrated in the high-end RTX 50 segment. This divergent structure is itself a decision map:

This table hides an important piece of circumstantial evidence: if this round of increases were a broad chip shortage or pure financial speculation, prices should rise across the board with every category climbing together; the actual divergence is precisely “high-end discrete cards up, low-end discrete cards steady, CPU and storage unmoved, memory still edging down.” Price pressure lands precisely on the high-end RTX 50 segment — exactly the segment whose substrate resources are being crowded out by AI chips first. In other words, the category divergence is itself the fingerprint that “structural capacity crowding-out” has left on this round of the market, and the most reliable basis for PC builders making budget decisions.

A few concrete recommendations, following this map:

First, deliberately drop half a tier when choosing a tier. At close to three to four times the official price, what the 5090 buys you back is no longer gaming performance but anxiety; the 5080’s 2x premium is equally far from its value center. If your resolution target was 2K anyway, the real transaction price of the 5070 tier (about a 94% gain) has already started overdrawing the budget of the next tier — what is truly least affected is a configuration that keeps expectations at the 5060 or below, or that simply bridges with a low-end card.

Second, exploit the category divergence in budget allocation. Since CPU and storage prices are steady and memory is still edging down, the parts that are “least worth skimping on in a whole machine” at this moment are precisely these: the marginal cost of fitting out large-capacity memory and fast storage is lowest right now, while the GPU is precisely the part you should least be chasing at a high. Getting every component in the case other than the graphics card sorted first, and leaving an upgrade slot for the GPU, is the most stable build path under this round of the market.

Third, compare prices more between prebuilts and discrete parts. The discrete parts market is directly exposed to channel spot volatility — in a market where nobody dares issue a quotation, discrete-part price comparison has almost no anchor; branded prebuilts follow manufacturer pricing systems, and their adjustment pace usually lags the spot channel, but there is no unified data supporting the specific magnitude or pace of pass-through — so the conclusion can only be operational: ask on both paths, take the current actual selling price as authoritative, and pick whichever is closer to the official guide price; do not decide on the old impression that “prebuilts are more expensive” or “discrete parts are cheaper.”

Fourth, the risk exposure of used and mining cards is widening. A channel cutoff of new stock pushes up the used market, a mechanical result of supply and demand; more noteworthy is the deterioration of condition in circulation — when the price gap on new cards is large enough, the incentive for refurbished and mining cards to slip into circulation rises accordingly, and this round also brings in purchasing demand from local large model deployment and AI image generation, the kind that runs at full load 7x24. In scenarios where used cards have no official warranty chain to check, a RMB 100-200 intraday price difference means the card-inspection window itself is depreciating. If you consider used cards during Double 11, be sure to confirm the source of the condition and the warranty documentation; better to spend more on channel stock that is still under warranty than to take in “leftover stock” of unclear origin.

Fifth, waiting or buying now — judge by layering the demand. Here is an executable framework:

The AI deployment line deserves a separate word of caution. In this round, the reason SMEs and developers buy the RTX 5090 is that it can run some quantized large-parameter language models and carry fairly complex AI image generation workloads — the use case itself holds up, but the balance of the decision has shifted: when this card is being contested on three layers at once by cloud-giant long-term contracts, AI mega-orders and local deployment demand, and the current price is about 3.4-3.9x the official price, “buying one flagship card yourself” needs to be re-compared against cloud options, and the accounting has to be done by asking “whether this card’s use case really runs at full load every day.” For lightweight use that runs inference only occasionally, buying a flagship card at a market peak for peak demand is the easiest purchase to regret.

Finally, turn the judgment framework into an executable checklist to run through in order before ordering on Double 11:

1. First fix your resolution and tier expectations: if targeting 4K high settings, face the high-premium reality of the 5080/5090 and consider dropping a tier or waiting; if targeting 2K or below, look first at the 5060 tier and below, which is least affected.
2. Check the live transaction price, not the listed page price: with one price a day, the same model can differ by RMB 100-200 within two days; the reference value of a listed price is only valid in the moment.
3. Non-GPU components first: CPU and storage prices are overall steady and memory is edging down, so these can be fitted out as planned or even at a higher spec, at no loss to the market.
4. Leave an upgrade slot for the graphics card: reserve PSU wattage and case space for a future card swap, and add a high-end card once supply loosens, avoiding a second replacement at the peak.
5. Get quotes on both prebuilt and discrete-part paths: on both, use the deviation between the current actual selling price and the official guide price as the yardstick; do not decide on the old impression that “prebuilts are more expensive” or “discrete parts are cheaper.”
6. Three checks on used channels: check the source of the condition, check the warranty documentation, check whether it has carried full-load compute duty; if any one is in doubt, walk away.
7. Keep the transaction records: in a market where not even merchants dare issue written quotations, order snapshots, chat records and invoices are the only anchors for after-sales recourse — archive them immediately after ordering.

IV. When Does This Round End: What Can and Cannot Be Determined#

The worst thing a decision article can do is pretend to know the answer. In this round, the boundary between what can and cannot be determined is clear.

What can be determined: the pace of capacity recovery is dictated by substrates, and the longest wait for new ABF substrate orders is about 13 months; giant long-term contracts run 6-12 months per signature; the digestion cycle of the AI mega-orders (reportedly 4-5 million cards and 2.4 million cards) is likewise measured in years. In other words, even if upstream went all-out on expansion today, the loosening on the supply side cannot come quickly on any timescale — the merchants’ judgment that “the probability of staying high in the short term is relatively large” is self-consistent with the capacity data.

What cannot be determined: first, the graphics card subsidy and inventory policies of the various e-commerce platforms for Double 11 — there is no reliable information on this round, and any claim that “platforms will release stock to damp prices” is speculation; second, the specific timetable for price declines, on which neither media nor merchants have a definitive conclusion; third, the rebalancing point between AI demand and gaming supply, which depends on the fulfillment pace of the mega-orders and the progress of substrate expansion, both of which currently lack public quantitative measures.

The above judgments are collected into a comparison table for quick reference at any time:

One more reminder about the “wait for the drop” mindset. In his earlier channel “financialization” analysis, Chasun dissected this: the expectation of “waiting for the last edition” turns hardware into a quasi-financial asset, and once the positive feedback loop of “the longer you wait the more expensive it gets, the more expensive it gets the longer you wait” forms, prices dance away from supply-demand fundamentals on their own. This round of the graphics card market is replicating exactly that logic — and there is only one way to break the loop: bring the purchase decision back to real demand, judging by the tiered framework in section III of this article, rather than ordering on the gambler’s intuition of “will it be cheaper if I wait one more week.”

So this article does not forecast prices. The only thing worth writing into the conclusion is a reminder: in a market where not even a written quotation can be issued, the validity of any current quote is measured in hours — if you decide to act, take the live market at the moment of transaction as authoritative; if you decide to wait, treat it as a decision measured in quarters, not in days.

In Closing#

From the stalls of Huaqiangbei to the production lines of Taiwanese substrate makers, this round of price increases is really a complete industrial transfer line: AI compute demand has changed the rules by which upstream capacity is allocated, and gamers are simply the link with the weakest bargaining power on that allocation chain. After several years of writing hardware market coverage, this is the first time I have given the advice “do not buy the flagship” during a build season — not because the performance is not worth it, but because the current price has nothing to do with performance anymore. Double 11 promo posters go up every year, but this year the real bargains may be hidden in CPU and storage prices. Good luck with your builds, and do not argue with the market.

One editorial aside: after finishing this piece, I looked over that Dewu price-trend table again — the RTX 5090 average transaction price went from RMB 24,099 a year ago to RMB 59,929 today; in the space of a year, the pricing logic of a gaming card has been rewritten completely. The market teaches more than reviews do — the next time someone says “the waiting camp never loses,” I will send them that TrendForce figure of 48-56 week lead times.

* All prices in this article are channel data at the time of writing; the live market is authoritative.

Further Reading#

• DDR5 Up Another 10-15% Within a Week, Huaqiangbei Graphics Cards Double in Ten Days: The Channel “Financialization” Inflection Point Under the Spot-Price/Share-Price Divergence — the divergence logic of the two curves: spot prices and share prices
• Huaqiangbei Graphics Cards “One Price a Day”: The RTX 5090 Channel Price of RMB 55,000-60,000 and the Double 11 Build Dilemma — the front-line stall price structure and a breakdown of the gains
• NVIDIA Admits the GPU Shortage Is “Beyond Control”: DRAM Root Cause, All-Time Highs Across Every Price Band, and the Reconstruction of Build Economics — official statements and a breakdown of how the supply gap is framed