Beyond GPUs: HBM, SSDs and Memflation Are Hitting PCs Too

Even if you secure AI accelerators (GPUs), a server rack (the frame that holds stacked servers) still stalls when memory is scarce—and that same demand is pushing up consumer PC and laptop prices.

On Oct. 12, Gartner said worldwide PC shipments in the third quarter of 2026 totaled 55.4 million units, down 22.4% from a year earlier. It was not leftover inventory. DRAM (the main memory a PC or server uses while it works) costs in the second quarter jumped about five times year over year, and PC makers passed those costs into selling prices, Gartner said. What Gartner calls memflation (sharp memory and storage price rises driven by AI demand) has reached beyond the data center.

The bottleneck is not GPUs alone. AI servers also need HBM (memory stacked next to an AI chip so data moves much faster), server DDR5 (the current generation of main memory used in servers and PCs), high-capacity enterprise SSDs (a fast storage drive made with flash memory), and MLCCs (a tiny ceramic capacitor that helps keep power and signals steady). When fabs (chip factories) prioritize those parts, consumer memory and storage from the same lines shrink and get more expensive.

Beyond the GPU: HBM, SSDs and MLCCs are stuck

Memory wall: TrendForce says AI chip compute is rising fast, but if memory bandwidth (how fast memory can move data, in GB/s) does not keep up, accelerators sit idle waiting for data. To ease that limit, the industry is adding more HBM stacks and capacity. HBM demand based on 2025 AI chip shipments grew more than 130% year over year, and TrendForce expects another rise of more than 70% in 2026.
HBM3E and HBM4: Volume use today is mostly HBM3E. HBM4 is in 2026 mass-production and qualification, where analysts say shipment schedules can slip, though we could not confirm the source of that analysis. Only a few firms—SK hynix, Micron and Samsung—make HBM at scale. Expanding capacity takes years.
Enterprise SSDs: With North American CSPs (companies like Amazon, Google and Microsoft that sell cloud services) expanding AI inference (using a finished model to produce answers) infrastructure, TrendForce expects enterprise SSD bit demand (capacity) to grow more than 80% year over year in 2026, with fourth-quarter price gains outpacing other memory products. An October 2025 report said some high-capacity SSDs from Samsung, SK hynix and Kioxia faced lead times of more than a year (TrendForce citing New Daily). That report is nearly a year old, so lead times may differ now.
MLCCs and power parts: High-capacitance, high-voltage MLCCs go into AI server power stages. TrendForce coverage in August 2026 said standard MLCC lead times were about 14–18 weeks, while some high-spec parts stretched to five to 10 months, with new capacity delayed into 2027. DigiKey data cited there put some Samsung Electro-Mechanics high-capacitance MLCCs near about 40 weeks (nearly 10 months).

The point: securing GPUs alone does not finish a rack. If memory, storage and passives slip, data-center turn-up slips too.

Memflation: data-center demand becomes consumer prices

In April 2026, Gartner forecast worldwide semiconductor revenue above $1.3 trillion, up 64%, then raised that to about $1.6 trillion on Aug. 24—up 92% from $809 billion in 2025—with 2026 memory revenue forecast at $837 billion. In its April release it estimated DRAM annual prices up 125% and NAND flash (flash memory used in SSDs that keeps data when power is off) up 234% in 2026, with meaningful relief unlikely before late 2027. Gartner analyst Rajeev Rajput said memflation could “destroy, or at least delay, non-AI demand into 2028” (Gartner press release, April 8, 2026).

TrendForce earlier raised its first-quarter 2026 contract-price outlook to about 90–95% quarter over quarter for conventional DRAM, 55–60% for NAND, and 53–58% for enterprise SSDs. In September it expected fourth-quarter conventional DRAM up another 10–15% and NAND up 15–20%, with the pace of increases possibly moderating. The consumer side is already soft. PC and smartphone makers (OEMs) cut specs, shrink entry models, or shift volume toward higher-priced configurations.

Concrete cases followed. Apple raised Mac and iPad prices on June 25. In a statement it said it had “never seen a component price increase this much, this quickly,” blaming an extraordinary surge in memory and storage demand from AI data-center buildouts, and said it could no longer shield customers. BBC and others reported a U.S. MacBook Pro with 1 TB of storage rising from $1,699 to $1,999 (list prices). In September, iPhone 18 Pro and Pro Max base prices rose $100 from the prior generation, with larger storage options up even more. The Verge reported similar roughly $100 hikes from Google, Samsung and others that month.

Tesla cut memory capacity on its next AI chips instead of only raising prices. On Oct. 1, Elon Musk said on X that AI5 memory would drop to half the plan—72 GB of LPDDR5—and AI6 by a third to 144 GB of LPDDR6, to secure volume for Optimus (humanoid robot) production and cut cost, while holding bandwidth steady so performance impact would be negligible. The next day he nudged AI5 up to 96 GB, saying, as Wccftech reported, that 72 GB would have tied the chip to the lowest-spec LPDDR5 parts available. That is still below the original 144 GB plan.

October’s Gartner numbers show the outcome. Third-quarter PC shipments of 55.4 million units match early-2023 levels, driven not by excess inventory but by higher component and selling prices and fewer low-cost devices. Businesses, Gartner said, focused spending on targeted buys such as workstations and AI PCs rather than broad PC refreshes.

IDC’s preliminary figures differ: it reported third-quarter shipments down 20.1% to 62.7 million units, and said buying ahead of price increases also played a part (per Newsbytes; original not checked).

U.S. plants help later, not this quarter

There is a push to widen supply inside the United States. In August 2026, SK hynix broke ground on an advanced HBM packaging plant in West Lafayette, Indiana. Planned investment is more than $4 billion, with a cleanroom targeted for October 2028 and next-generation HBM mass production in the second half of 2029. Wafers made in Korea would be packaged and tested in the U.S. for U.S. customers. Micron has also outlined U.S. DRAM fabs and future HBM packaging backed by CHIPS Act support.

Those projects do not fill today’s missing parts. HBM wafer production and advanced packaging still lean heavily on Asia, and U.S. ramp dates are mostly 2027–2029. Until then, HBM, DDR5 and enterprise SSD demand from North American CSPs is often locked first in long-term supply agreements. Diversifying U.S. infrastructure investment matters, but the lead times and prices developers and buyers feel are likely to stay elevated for a while.

What we could not confirm, or should treat with care

Treating this bottleneck as a U.S.-only data-center crisis overstates it. North American CSPs are a major demand axis, but price and lead-time spillover is global across consumers and OEMs.
“Some high-capacity SSDs with lead times over a year” rests on a single New Daily report via TrendForce, not an official figure for every product or vendor.
MLCC lead times and stockouts summarize securities and local-media reports via TrendForce; we did not read manufacturers’ official lead-time tables.
Gartner’s “five times” refers to second-quarter 2026 PC DRAM costs versus a year earlier, not every memory product or every quarter’s contract price.
Tesla’s cut is not a reduction of in-car infotainment RAM; it is a lower planned capacity on next-generation AI chips (AI5 and AI6) for Optimus and Full Self-Driving. Figures follow Musk’s X posts and coverage of them.
We could not open the Gartner press releases directly (April 8, Aug. 24, Oct. 12) in the environment used to write this post; we confirmed their contents through search results. Gartner’s forecasts change—April’s was raised in August—so check the latest release, and October’s PC shipment figures are preliminary.

What could change for developers

What follows is our estimate and advice based on the reports.

Laptop, Mac and phone quotes: Where Apple has raised list prices, discounts, education pricing and used markets tend to move with them. For team gear, prioritize available capacity and lead time over a locked long-term unit price, and compare base configs plus self-upgrades.
Do not watch only cloud GPUs: Behind instance waitlists sit storage IOPS (reads and writes per second), local NVMe (a fast SSD interface) and memory-bandwidth caps. When costing training or inference, measure storage and network bottlenecks too.
Local LLMs (large language models, the kind of AI behind chatbots like ChatGPT) and home labs: RAM, VRAM (memory dedicated to a graphics card) and NVMe prices are tied to AI server demand. Size needed memory and bandwidth before loading a big model, settle on the configuration you need first, then buy when stock is available, with some schedule buffer.
Lead times and alternates: If you build servers, power or boards yourself, pre-qualify alternate high-spec MLCC parts and power-IC part numbers and second sources. Plan around the reported lead times of five to 10 months for some high-spec parts.

What to watch

Next-quarter PC and smartphone shipments and ASPs from Gartner and IDC
Memory contract prices (fourth quarter through early 2027) and enterprise SSD bit supply from TrendForce and peers
Actual milestones at SK hynix Indiana and Micron’s U.S. fabs
How much HBM4 qualification and shipments ease HBM3E demand
Tesla AI5 volume-production timing (we could not confirm a source for it) and the memory capacity that ships