GPUs and Storage ‘Only Go Up’: Why 2026 PC Build Costs Are Out of Control#

Zixing | September 27, 2026

Hello everyone, I’m Zixing.

In my last two articles I ran the numbers on full builds and on monitors, and readers kept pushing in the comments: monitor prices rose only 7%, so how have GPUs, memory, and SSDs — the “decisive big-ticket parts” — been climbing nonstop since the start of the year without even a decent pullback? This week I went back through the data I gathered and took GPUs and storage apart separately — they are not the same story, but the disease grows from the same root: AI data centers “siphoning” high-bandwidth memory and advanced DRAM capacity.

This article skips scalpers and skips “wait-for-the-drop” sentiment, focusing squarely on the price-increase mechanics and supply-side causes of the two core components: GPUs and storage. I’ll use data tables to lay out the increases, MSRP, and the quantified math on build costs.

Table of Contents#

1. GPU: The RTX 5090’s “Sealed Warehouse” Spiral to 2.1x MSRP
2. Storage: DDR5 Up Severalfold Year over Year, SSDs Riding Along
3. Supply-Side Mechanics: How AI Data Centers “Siphon” Consumer-Grade Capacity
4. Memflation: Pass-Through from Contract Prices to Retail
5. Real-World Impact on Gaming Builds and Player Budgets
6. Zixing’s Rational Judgment Framework

1. GPU: The RTX 5090’s “Sealed Warehouse” Spiral to 2.1x MSRP#

Start with the GPU side. The most extreme specimen of this market cycle is the RTX 5090, so we’ll use it as the dissection subject.

1.1 RTX 5090 Market Price vs Official Price#

The RTX 5090’s official MSRP is $1,999, but by August 2026 dealer and street prices had landed in the $3,900–4,400 range; by channel accounts, used medians sit around $4,220 and new-unit medians around $4,500. Worked out, the street price is roughly 2.1x MSRP. More notable still, in May 2026 NVIDIA raised the 5090’s wholesale price by about $300 — showing the price increases aren’t just happening at retail; the upstream vendor is actively raising prices too.

1.2 RTX 5090 Spec Refresher: Why It Became the “Price Anchor”#

The key item is “32GB GDDR7.” GDDR7 high-bandwidth memory is both the fabric of the consumer flagship and the same-sourced component of AI accelerator cards. When AI data centers absorb high-bandwidth memory in volume, the material cost of consumer flagship cards rises with the tide — this is one of the physical bases for the 5090 breaking past twice its MSRP.

1.3 The “Tighter as It Rises” Hoarding Spiral#

This GPU cycle shows an unusual pattern: the more prices rise, the tighter supply gets. The reason isn’t simple channel hoarding but two stacked structural factors:

• Upstream wholesale prices raised proactively: NVIDIA raised the 5090’s wholesale price by about $300 in May, so channel acquisition costs are already higher and retail prices naturally can’t come down;
• Crowding out toward higher-margin uses: high-bandwidth memory (GDDR7) and AI accelerator cards share a component pool; the manufacturers prioritize enterprise and AI markets with higher margins, squeezing the share left for consumer flagships.

One clarification: the specific mechanism of “manufacturers sealing the warehouse and withholding supply” is industry observation, not confirmed publicly by any manufacturer, and is not stated here as established fact. What is certain: wholesale price hikes plus tight memory supply have jointly propped up stubbornly high retail prices, forming the “tighter as it rises” spiral.

2. Storage: DDR5 Up Severalfold Year over Year, SSDs Riding Along#

If GPUs are a “single-point blowout,” storage is a “full-line rout” — memory and SSD price increases over the past six months have exceeded nearly everyone’s expectations.

2.1 US Market 32GB DDR5 Price Trajectory#

Taking the most mainstream 32GB DDR5 kit as an example: a year ago it cost $80–120 in the US market; by early June 2026 the lowest price had already crossed $374.97, and by early September typical street prices sat in the $420–520 range — a year-over-year increase of roughly 300–400%.

2.2 Lowest Prices for US DDR5 Kits by Capacity#

From 16GB to 128GB, the lowest prices in every capacity segment are in the “4 to 5x” range. The 128GB DDR5 6400 kit soared from $329 to $1,782, showing the high-capacity segment was hit just as brutally — not just entry kits.

2.3 China Market: Memory and SSDs Both Rising#

The China numbers are just as striking. A 32GB DDR5 kit climbed from about ¥900 in the same period of 2025 to nearly ¥4,000, up about 322%; 16GB DDR5 soared from about ¥450 to ¥1,800, up about 300%; a 1TB SSD rose from ¥410 to ¥950, up about 132%. Media reports also cite mainstream dual-channel 2×16GB memory retailing around ¥1,600, versus only about ¥800 during the 2025 618 sale — “essentially doubled.”

2.4 DDR4 Is No Safe Harbor Either#

Many assumed “switching to DDR4 dodges the price hikes” — reality quickly disproved that:

The 32GB DDR4 kit went from $60–90 in October 2025 to $150–180 by January 2026. The reason: manufacturers are progressively retiring old DDR4 production lines and shifting capacity to higher-margin products, further tightening supply of the old spec. So the “retreat to DDR4” route is closed in 2026 too.

3. Supply-Side Mechanics: How AI Data Centers “Siphon” Consumer-Grade Capacity#

The symptom shows up at retail, but the real root cause is on the supply side. This section takes the mechanism apart.

3.1 AI Servers’ Enormous Storage Appetite#

A single AI server demands about 8–10x the memory chips of an ordinary server, dooming AI data centers to devour vast storage capacity. One important caution: different sources measure “capacity consumed” differently — Tencent News says AI consumes about 53% of global monthly memory chip output, while TrendForce estimates AI workloads absorb about 20% of global DRAM wafer capacity. The two figures use different bases — one is “global monthly memory output,” the other “DRAM wafer capacity” — and must be treated separately in the text, not conflated or naively averaged.

3.2 The Big Three Adjust Their Capacity Structures#

Facing AI’s fat margins, the three memory makers (Samsung, SK hynix, Micron) have proactively tilted resources toward high-margin products like HBM and DDR5:

Micron’s moves deserve the most attention here: pausing quotes, then raising everything about 20% on resumption, and announcing the end of its consumer brand Crucial. When the manufacturer is prepared to give up even its “consumer brand,” it shows they have bet capacity and brand resources on the enterprise and AI markets, and consumer supply has been substantively contracted.

4. Memflation: Pass-Through from Contract Prices to Retail#

Upstream contract price trends are the key signal for whether retail prices can still fall back. This section looks at the transmission chain.

4.1 Quarterly Trajectory of DRAM Contract Prices#

TrendForce analyst Xu Jiayuan’s judgment: in Q4 2025, DRAM contract prices rose 40%+ across applications, expected to be the steepest quarter of this cycle; in Q1 2026 contract prices are expected to rise another 15%+, after which increases gradually converge, with the rise possibly extending into the second half of 2026. Contract prices are the upstream leading indicator for retail prices — as long as contract prices are still rising, even if only at a converging rate, retail-side cost pressure will not disappear in the short term.

4.2 Institutional Definitions and Forecasts of “Memflation”#

Gartner uses the term “memflation” to describe this cycle, forecasting 80% DRAM price inflation in 2026 with shortages lasting into the second half of 2027. TrendForce’s framing is closer to the present: prices have entered a plateau — neither steepening like Q4 2025 nor clearly falling back. A plateau means no big dive is visible in the short term, and no continued loss of control either — a state of “high-level standoff.”

Note: these “back to H2 2027” timelines are institutional forecasts, not established facts; SK hynix and others have even bleaker takes like “maybe not until 2030,” so the timing of any retreat remains uncertain.

4.3 One “False Alarm” of Price Cuts#

In April 2026 there was a wave of “DDR5 price cut” coverage, triggered by Google releasing an AI compression algorithm that reduced inference memory demand, briefly making markets optimistic. After investigating, PCWorld concluded it was noise, not a trend: US 7-day/30-day price movement was just 1–2%, and a kit that fell from $500 to $369 still worked out to about 4x last year’s level. The lesson: judging that “storage is about to fall” off one or two easing signals is an easy way to misread the 2026 market.

5. Real-World Impact on Gaming Builds and Player Budgets#

Only by putting GPU and storage inflation into a whole system can you grasp how deeply it hits player budgets.

5.1 Concrete Component Price Shock Examples#

Even an entry 512GB SSD has climbed to about ¥500, a single 16GB DDR5 stick touched ¥2,000, and the AMD Ryzen AI Max+ 395 mini PC doubled from about ¥10,000 to about ¥20,000. The last row stings the most: in 2026, DIY self-built systems often cost more than prebuilts of the same configuration — the old iron law that “building it yourself is cheaper” has been broken by this cycle.

5.2 GPUs Displace CPU/Motherboard as the Dominant Cost Item#

The most far-reaching structural change of this cycle is that the GPU’s dominance of system cost has been further reinforced. The GPU was always the biggest cost item in a mainstream gaming rig, but the CPU, motherboard, memory, and SSD still held a substantial share; now that GPUs have hit 2.1x MSRP and memory and SSDs have multiplied several times over, the cost center of a build has shifted even further toward GPUs and storage.

5.3 The Overall Upward Shift of Mainstream Build Tiers#

For mainstream builders, memory and SSDs — the two parts you used to “worry about least” — are now the fastest-rising components; for flagship players, the GPU premium is the biggest budget killer. The entire “how much budget buys which machine” ladder has shifted upward as a whole. ¥7,000 used to build a very decent 2K rig; the same configuration now costs visibly more.

6. Zixing’s Rational Judgment Framework#

Finally, as a hardware editor, let me distill the core logic of this “only goes up” cycle into a few actionable judgment principles:

1. Separate “must-buy” from “can wait”: if your current memory and SSD are still adequate, don’t chase prices — the market has entered the “plateau” TrendForce describes, neither steeply rising nor clearly falling in the short term, so there’s no must-buy urgency. But if you genuinely need to build now, accept that costs are rising.

2. Don’t apply “wait-for-the-drop” thinking to this cycle: in the past, waiting brought price cuts; this cycle is backed by AI data centers’ long-cycle demand for high-bandwidth memory and advanced DRAM capacity. As long as HBM and high-end DRAM remain attractive to the manufacturers, consumer supply will stay tight. Rather than watching price headlines, watch upstream capacity allocation and institutional judgments on the cycle window.

3. Understand the shared root of “GPUs and storage”: GPUs rely on GDDR7 high-bandwidth memory; storage relies on DRAM/NAND dies. Different product categories, but both are being squeezed by the same force on the supply side — AI data centers’ siphoning of advanced capacity. That’s why they rise together, “only up,” rather than fluctuating independently.

4. Face the reality that “DIY is no longer cheap”: when self-built systems cost more than prebuilts, the traditional logic of saving money by building your own is broken. If saving money is the sole goal, consider brand prebuilts or ready-made systems; if you want personalization and a performance tier, you’ll pay the premium.

5. Stay cautious with unverified information: the specific spot premiums for mainstream GPUs other than the RTX 5090 (RTX 5070/5080, AMD RX series) were not independently verified in this round of research; the specific mechanism of “manufacturers sealing the warehouse” likewise has no public confirmation. Neither should be stated as fact, let alone used as a basis for decisions.

I hope this “GPU + storage” ledger helps everyone see the real mechanics of 2026’s runaway build costs — not to stoke panic, but to read the long-term supply-side structure and then judge from your own actual needs.

— Zixing, NTGame Hardware & Tech Editor-in-Chief