AI Took the Memory. China Is Taking the Customers.

In early 2025, the two memory chips inside a $99 video doorbell cost the maker about $2.50. Today they cost $6.40, an increase the company never chose, on a product that clears maybe $15 of hardware profit per unit. Apple, the most protected buyer in electronics, watched the 12GB of DRAM in a flagship phone go from $27 to $68 on contract while the open market asked $145. In June, Tim Cook called price increases "unavoidable" and repriced fourteen products in a single day. In July, Apple opened talks with the Chinese memory makers it walked away from in 2022.

Every electronics maker took the hit. What differs is how many ways each one has to respond.

The same chip now sells at two prices

What a buyer pays for memory now depends mostly on who the buyer is. Datacenters absorb roughly 70% of global output, and the three manufacturers who control the market (Samsung, SK Hynix, and Micron) have moved wafer capacity to the high-bandwidth memory that AI accelerators use. HBM capacity for all of 2026 sold out early in the year. Gartner projects DRAM prices up 130% for the year. Mobile DRAM contracts rose 78 to 89% in the second quarter alone, NAND rose 70 to 75%, and flagship phones are quietly shipping with 12GB instead of 16GB. Phison's CEO predicts the shortage "will shut down many consumer electronics companies in 2026."

The spread between the front and the back of the queue is the number to watch. The same 12GB LPDDR5X package costs an Apple-sized buyer $68 under a long-term agreement and everyone else about $145 on the open market. The 8GB eMMC storage chip in a video doorbell runs about $3.50 on an incumbent supply contract and roughly $20 on the spot market. Same parts, 2.1x and 5.7x apart.

Only one shield held, and it worked by accident

We modeled three representative products on the Muir platform (a $99 video doorbell, a mid-range automotive infotainment unit, and a flagship smartphone), generating full teardown-level bills of materials and pricing every component at H1 2025 and H1 2026 contract conditions. Non-memory components are held constant across both bases, so the changes below come from memory alone.

The doorbell's memory bill rose from $2.50 to $6.40, taking memory from 8.6% to 19.5% of its factory cost (components plus final assembly; freight, duties, and channel margin sit on top). The smartphone's rose from $48 to $101, and that uses the conservative contract-lagged basis rather than the open-market print. The automotive unit rose 37%, from $16.25 to $22.25. Carmakers didn't negotiate better. Their annual contracts and AEC-Q100 qualification requirements simply reset more slowly, and the same lock means qualifying a new supplier takes 18 to 24 months. They can't be hit quickly, and they can't move quickly.

Contract leverage delayed the shock for the biggest OEMs. It did not stop it, which is why Apple raised prices anyway. Apple has four ways to respond: absorb the cost, reprice, cut specs, or add a supplier. It used two of them in a single month this summer, raising prices across its lineup and opening the CXMT and YMTC talks. Reporting on the next iPhone generation puts memory content on a path from roughly $50 toward $200 per device, and our contract-lagged model sits on that trajectory.

A doorbell maker has one way to respond. It can't absorb the cost, since $3.90 is about a quarter of its unit profit. Repricing is nearly as hard: factory-cost increases reach the shelf at roughly 2.5 to 3x, and $10 to 12 breaks the $99 price the whole smart-home category is built on. Specs are already at the floor for a product that records video. What's left is the supplier.

Re-sourcing the doorbell's memory to CXMT DRAM and YMTC-based eMMC, at the 36 to 40% discount Chinese suppliers currently offer, recovers $2.40 of the $3.90 increase. Across ten million units a year, that's roughly $24 million. The re-sourced cost still sits above 2025 levels, so the switch softens the shock rather than undoing it. But for this tier the alternative is losing money on every unit. That is why it is happening, whatever anyone thinks about supply-chain geography.

The shift to Chinese memory is structural, not cyclical

CXMT now holds 8% of global DRAM, fourth in the world and up five points in a year. YMTC holds 13% of NAND. Corsair, HP, and Dell have adopted CXMT-based DDR5, Lenovo ships YMTC storage, both companies came off the Pentagon's restricted-supplier list, and Apple's talks settle the argument. The buyer with the most alternatives decided Chinese memory was worth the risk. The buyers with fewer alternatives got there sooner.

Three things follow. Expect repricing everywhere: prices are still guided up 13 to 18% quarter over quarter into Q3, at that pace the doorbell's memory passes 20% of factory cost by year-end, and the thinner the margin, the sooner the price moves. Automotive's calm is borrowed time: the 2026 spot reality arrives in 2027 contract resets, and with an 18 to 24 month qualification wall, the tier that looks safest in our data has the least room to respond when its turn comes. And the design wins will stick. Qualification and firmware validation are sunk costs, so parts designed in during a shortage stay designed in after it. Japanese DRAM makers entered the market the same way in the 1980s, from below, on price and availability, and stayed. When this shortage ends, the incumbents will want these customers back. The customers won't be where they left them.

The trade has a cost side. Our models already show China-origin semiconductors carrying US import-duty rates several times those of Korean or Taiwanese parts, export-control policy keeps moving, and supplier control raises questions that don't show up on a purchase order. For many buyers the switch is still the only option on the table. It should at least be priced with open eyes.

How we built this

Muir's model decomposes a product from a plain-language description into a full teardown (component hierarchy, specifications, masses, manufacturing locations) and prices every line item against current market data. For this report we described three products in a paragraph each. The model generated their bills of materials, priced them under H1 2025 and H1 2026 contract conditions, and produced the incumbent-versus-CXMT/YMTC sourcing scenarios. Every price carries its derivation and source citation, and memory pricing was validated against published contract data from TrendForce, SigmaIntell, and industry reporting.

See what your product's memory exposure looks like: Book a Demo