At the parts counter, AI has deeper pockets

A person building a PC and a cloud company filling a data center are not shopping from the same shelf. One wants a pair of desktop DIMMs and perhaps a graphics card. The other is ordering racks filled with HBM, server DRAM, enterprise SSDs, accelerators, networking, and cooling. They still meet farther upstream, where a small group of memory makers, foundries, packaging lines, and component suppliers decides which orders deserve scarce capacity.

That is where the AI build-out has started reaching the consumer parts counter.

On August 25, TrendForce estimated that capital spending by major cloud providers will rise 98% in 2026 and another 50% in 2027. DRAM and NAND are expected to consume 47% of that spending this year and 68% next year.[1] Those figures are a forecast from a market-research firm, not an audited bill from every cloud company, but the direction is hard to miss: AI infrastructure is purchasing memory at a scale that can change what suppliers choose to manufacture.

For anyone without an enterprise budget, the practical issue is whether a normal PC gets cheaper from here.

I would not count on a broad drop through early 2027. Some individual parts will go on sale, weak products will be discounted, and used hardware can move differently from new stock. RAM, SSDs, and graphics cards face enough upstream pressure that a market-wide return to the cheap-memory years looks unlikely in the near term.

The squeeze is in allocation, not a shared shopping cart

AI accelerators use HBM, while desktops usually use DDR memory and graphics cards use GDDR. Treating them as interchangeable would be wrong. The link is manufacturing strategy.

IDC describes memory makers reallocating cleanroom space and capital away from conventional DRAM and NAND toward higher-margin HBM and high-capacity server products. It expects 2026 DRAM and NAND supply growth of 16% and 17%, respectively, below historical norms.[4] TrendForce estimates that HBM and RDIMMs together will take 51% of DRAM bit supply in 2026.[1] More bits going into server products leaves consumer buyers competing for what suppliers still devote to PCs and phones.

The price signals are already ugly. TrendForce says server DRAM contract prices rose a cumulative 64% in the second half of 2025 and projects an approximately 270% increase during 2026. It puts enterprise SSD price growth at 35% in the second half of 2025 and a projected cumulative 235% this year.[1] Those are enterprise prices, so they should not be copied directly onto a retail DDR5 kit or a two-terabyte consumer SSD. They do show how much more attractive the enterprise side of the factory has become.

Micron’s own results make that incentive visible. The company reported $41.46 billion in fiscal third-quarter revenue, compared with $9.30 billion a year earlier, while spending $7.1 billion on capital expenditures. It said HBM4 was already shipping in volume for a lead customer and expects HBM4E volume production in 2027.[5] Micron is also qualifying a new PC-client QLC SSD, so it has not abandoned consumers. Its most urgent investment story is plainly AI and server memory.

The spillover reaches smaller components too. TrendForce reported that Samsung Electro-Mechanics planned average price increases of 25–30% for certain consumer-grade multilayer ceramic capacitors sold to OEM and ODM customers in the fourth quarter. Taiwanese and Chinese suppliers were expected to raise some mid- and high-capacitance consumer-grade prices by 10–20%, with utilization at some suppliers approaching 90%.[2] A capacitor is not the glamorous line item in a gaming build, but thousands of quiet parts still contribute to the cost of boards, power delivery, storage, and finished systems.

What I expect at the PC aisle

Memory and storage have the clearest case for staying expensive. IDC says major PC vendors have warned customers about tougher conditions and describes 15–20% price hikes or contract resets across the industry. Its moderate downside scenario has PC average selling prices rising 4–6% in 2026; its pessimistic scenario puts the increase at 6–8%.[4] IDC is explicit that these are risk scenarios rather than a replacement for its official forecast. I find the moderate direction more persuasive than the exact percentage because vendor inventory, promotions, and regional pricing can blur the retail result.

SSDs may get relief before server memory does, but probably not in one clean drop. NAND has a history of violent oversupply cycles. High prices encourage investment, weak consumer demand builds inventory, and a market that looked tight can reverse. The problem is timing. TrendForce expects meaningful additional server DRAM and HBM supply from process migrations and new fabs mainly in the second half of 2027, with combined bit supply rising 27%.[1] Production ramps take time, and new capacity aimed at lucrative server products does not automatically become cheap client NAND or desktop DDR.

Graphics cards look sticky for a different reason. AI accelerators and gaming cards do not share the same memory configuration, but they can share advanced foundry capacity, upstream wafer economics, packaging suppliers, and a vendor’s attention. TrendForce expects shipments of NVIDIA NVL72 AI racks to grow more than 50% in 2027 and estimates Vera Rubin system prices at roughly twice those of GB300, with wafer and HBM costs also rising.[6] When data-center products support that kind of selling price, neither NVIDIA nor AMD has much incentive to use scarce leading-edge capacity to start a consumer price war.

That does not mean every GPU price rises forever. A card that misses its performance target can be discounted. Retailers eventually clear inventory. AMD, Intel, used cards, and older generations can put pressure on a particular tier. My expectation is a lumpy market: brief deals and model-specific corrections inside a price structure that remains higher than consumers remember.

CPUs are the part I would watch for better bargains. They still depend on advanced manufacturing, and AI demand can affect foundry economics, but desktop CPUs are not exposed to the memory shortage in quite the same way as a full system. Competition and unsold inventory can force price cuts even when the rest of a build stays expensive. Motherboards and power supplies may see selective promotions too, although the reported MLCC increases argue against assuming every board component becomes cheaper.

Prebuilt PCs could occasionally beat DIY pricing. IDC expects smaller assemblers and white-box builders to bear more of the shortage because large OEMs can secure supply earlier and negotiate with more leverage.[4] I dislike the idea that a market built around interchangeable parts could reward sealed configurations and purchasing power. It is still plausible that a large vendor’s bulk memory contract produces a better whole-machine price than buying each scarce component at retail.

Where the forecast can break

There are two credible paths to lower prices.

First, AI spending could slow. The current build-out assumes cloud providers keep finding customers, power, financing, and useful workloads for an extraordinary amount of hardware. If projects are delayed or returns disappoint, memory orders can be revised quickly. Semiconductor markets do not glide gently toward balance; they often overshoot and then cut.

Second, consumer resistance can force discounts even while component costs remain elevated. TrendForce’s August 31 smartphone analysis says rising memory and component costs are already making brands more cautious about inventory for new Android models, while repair and secondary-market demand remains comparatively resilient.[3] PCs can follow the same pattern: people keep working machines longer, repair them, or buy used. A vendor facing weak unit sales may sacrifice margin to move stock.

Neither route makes me comfortable predicting an imminent crash. AI rack demand is expected to remain robust through the first half of 2027, while Micron says multi-year customer agreements are making its performance more predictable.[5][6] Even when supply improves, manufacturers may use the first breathing room to rebuild inventories and margins rather than hand consumers an immediate price cut.

For a consumer, that means “wait and everything will be cheaper” is not a dependable plan. Waiting still makes sense when the current machine does the job, because not buying is the strongest leverage a buyer has. When an upgrade is actually needed, I would compare the specific component against its recent price history, include used and refurbished options, and avoid paying extra for an AI label that does not improve the work or games the machine will run.

I expect the first real relief to arrive unevenly during 2027, with storage and selected CPUs or older GPUs moving before high-capacity memory and premium graphics cards. A broad return to cheap PC hardware probably needs both new supply and cooler AI spending. Right now, only the first of those is visibly under construction.

Sources

[1] https://www.trendforce.com/presscenter/news/20260825-13198.html — Memory Prices Soar; DRAM and NAND Flash to Account for 68% of Major CSP CapEx in 2027 [2] https://www.trendforce.com/presscenter/news/20260827-13202.html — SEMCO Leads 4Q26 OEM Price Hikes [3] https://www.trendforce.com/presscenter/news/20260831-13208.html — Memory costs pressure smartphone supply chains [4] https://www.idc.com/resource-center/blog/global-memory-shortage-crisis-market-analysis-and-the-potential-impact-on-the-smartphone-and-pc-markets-in-2026 — Global Memory Shortage Crisis: PC and Smartphone Impact [5] https://investors.micron.com/news/press-release/2026/Micron-Technology-Inc--Reports-Record-Results-for-the-Third-Quarter-of-Fiscal-2026/default.aspx — Micron Fiscal Q3 2026 Results [6] https://www.trendforce.com/presscenter/news/20260828-13204.html — NVIDIA AI Rack Transition Accelerates