AI Memory Demand Drives Up PC Costs as Shipments Fall and Vendors Extend Older Platforms

On Oct. 9, Gigabyte announced that all of its B760 and H610 motherboards would support new Intel LGA1700 processors expected to launch in early 2027 through a BIOS update, including DDR4 boards, according to Ifanr. The LGA1700 platform arrived in 2021 with Intel's 12th-generation Core processors and has supported both DDR4 and DDR5 motherboards. The announcement did not disclose the new processors' models, architecture or performance. The move allows existing users to keep their boards and memory and wait for a processor upgrade.

The price pressure comes from the AI data center build-out. Ifanr reported that demand for HBM and server DRAM has led storage manufacturers to direct more capacity toward higher-margin data center products, limiting room for consumer memory supply growth. Memory and SSD quotes have risen for more than a year and have moved from procurement lists into finished PC costs. TrendForce estimated that for a mainstream notebook with a suggested retail price of $900, the CPU, DRAM and SSD accounted for about 68% of material costs in the third quarter of 2026.

PC makers have raised prices several times in recent quarters, but Omdia said those increases have not fully covered the cost gap, Ifanr reported. Consumers facing higher prices have opted to keep older computers for longer. The third-quarter shipment decline also reflects a high base from 2025, when the end of Windows 10 support spurred replacement demand, and a pull-forward of orders in the first half of 2026. Brands and sales channels stocked up before further memory price increases, moving some later-quarter orders earlier. In the third quarter, high prices deterred end buyers while channels worked through inventory, weakening orders to manufacturers. The back-to-school season did not produce its usual peak. IDC analyst Jitesh Ubrani said inventory pressure may push channels toward promotions, but prices are unlikely to return to levels seen a year earlier.

Vendors face a choice among raising prices further, absorbing some costs, lowering specifications or extending mature platforms. The return of DDR4 is one result. Ifanr cited The Verge, whose reporter found a 32GB Corsair Vengeance DDR5 kit priced at $620, while the corresponding DDR4 kit cost $260. The $360 difference can shape a build decision. DDR5 offers higher bandwidth and better energy efficiency, and DDR4 has not escaped price increases, but users who already own DDR4 memory can save on an upgrade. Gigabyte's compatibility plan offers that option. On the AMD side, Gigabyte announced a new batch of AM4 motherboards in August, alongside the reintroduction of the Ryzen 7 5800X3D, adding another DDR4 option. New platforms once pushed bundled upgrades of processor, motherboard and memory; preserving existing hardware has become part of product competitiveness.

Software developers are also trying to extend the life of existing PCs. Ifanr reported that an open-source project called Strata has gained about 19,500 stars on GitHub. It aims to run a 125-billion-parameter model, Qwen3.8-Flash-Next, on an ordinary gaming computer. In the project's test, a PC with an RTX 5070 graphics card, 12GB of VRAM, a Ryzen 5 7600 processor and 64GB of system memory generated 94 tokens per second under the Q2_0 quantization setting. That is roughly 10 times faster than a typical person reads. The model uses a Mixture-of-Experts architecture, so only some expert modules are used for each token. Strata keeps frequently used experts in VRAM, stores all expert weights in system memory, uses the GPU for core computation, has the CPU handle experts not cached in VRAM and stores large lookup tables on the SSD. Quantization shrinks the model, but harsher compression reduces answer quality. The Q2_0 version reached 94 tokens per second; the milder IQ3_S version reached 53 tokens per second on the same test hardware. Speculative decoding can add about 1.6 to 1.8 times speedup depending on whether predictions are accepted, but it does not eliminate accuracy loss from low-bit quantization. Strata supports chat, coding and optional image input, and can connect to applications and coding assistants through a compatible API. Inference runs locally. A 12GB VRAM card is only part of the requirement; system memory also shapes the experience. A 64GB machine can hold multiple 2-bit and 3-bit versions listed by the project, while a 32GB configuration usually can run only a trimmed Coder version that removes some experts and performs worse on Chinese and other non-coding tasks. The first launch can take one to three minutes as tens of gigabytes of data are loaded into memory. Long text must be pre-processed, and default requests are queued. Startup wait, memory requirements and answer quality leave Strata short of everyday ease, but it shows that better allocation of computing and storage tasks can let the same hardware run models that were previously difficult to fit. How much a computer can do depends not only on the hardware purchased but also on how well software uses it.

The contrast with cloud AI prices is sharp. Ifanr cited an a16z chart using Goldman Sachs data showing that the LLM price index fell over about three years by as much as the personal computer price index fell over about 15 years. Intelligent services are becoming cheap digital fast-moving goods, while the memory, chips and computing devices that carry them remain constrained by capacity and supply and demand. The technology industry is pushing more powerful AI into ordinary hands while ordinary users watch every dollar spent on a memory stick. The result is a period in which intelligence is getting cheaper, but the hardware needed to deliver it is getting more expensive.