The AI memory supercycle comes for the $200 phone

In late June, Xiaomi’s POCO brand did something phone vendors almost never do: it repriced handsets that were already on shelves. In the Philippines, the POCO X8 Pro went from PHP 18,999 to PHP 25,999, a jump of roughly 37 percent, while the X8 Pro Max climbed PHP 11,000. GizGuide, which documented the new prices on June 25, noted Xiaomi had already warned that the memory shortage would squeeze profitability and that raising prices might be the only way to cope.

Mid-cycle repricing is the visible symptom. The disease is upstream, and this week it got its clearest diagnosis yet.

Research firm Omdia now expects global shipments of smartphones priced under $400 to fall more than 22 percent in 2026, dragging the whole market down 12 percent year over year, as Tom’s Hardware reported Tuesday from the firm’s July analysis. The stranger number sits underneath the forecast: between the third quarter of 2025 and the first quarter of 2026, memory’s share of manufacturing cost nearly doubled for cheap phones. For sub-$400 devices, DRAM and NAND now approach 60 percent of the bill of materials. The much-quoted “up to 64 percent” is real but applies at the extreme end of Omdia’s model — ultra-budget phones retailing under $99, where there is the least everything-else to dilute the memory line items.

Phones above $400 are forecast to grow 5.7 percent this year. The cheap end takes the hit because it has no margin to hide in.

Manufacturers, Omdia principal analyst Zaker Li says, are under pressure to “significantly raise retail prices on low-end models just to maintain microscopic profit margins.”

Follow the wafer

None of this started in phones. Samsung, SK hynix, and Micron have been steering wafer capacity toward high-bandwidth memory for AI data centers, and the commodity DRAM and NAND that budget handsets depend on is what got starved. The sellers’ side of that trade is now posting numbers that read like typos.

On July 7, Samsung guided second-quarter operating profit to roughly 89.4 trillion won — about $58 billion — against 4.68 trillion won a year earlier, a nineteen-fold increase. Tom’s Hardware reports that the chip division alone is tracking toward some 300 trillion won in operating profit for 2026, which would exceed everything the semiconductor business has earned since Samsung entered it four decades ago, combined. The same reporting puts the quarter’s operating profit ahead of Nvidia’s $53.5 billion, and DRAM and NAND operating margins at 40 to 50 percent in the first half.

Further down the chain it gets stranger. Shenzhen’s Longsys, the module maker that owns Lexar, told investors to expect first-half net profit of 9.2 billion to 11 billion yuan — against 14.8 million yuan a year earlier, per the South China Morning Post. At the top of the range that is an increase of more than 74,000 percent, on a company whose business is buying wafers and packaging them into the storage and memory that goes into ordinary devices.

A supercycle like that is supposed to summon its own correction. Supply is in fact responding, and in the same week the Omdia forecast landed. Look at where it’s pointed.

The cavalry is riding somewhere else

On July 3, Kioxia announced sample shipments of its tenth-generation BiCS flash, built with SanDisk: 332 layers, a 1-terabit triple-level-cell die, 59 percent better bit density than the generation before it, a 4.8Gb/s interface. Tom’s Hardware notes it is the densest 3D NAND yet at over 29 Gb/mm², edging out Samsung’s 400-plus-layer V10 class. More bits per wafer is exactly the medicine a NAND shortage calls for. But read Kioxia’s targeting: enterprise and data-center SSDs for AI workloads. The older 218-layer BiCS9 is what serves client devices.

Meanwhile at JEDEC, the memory standards body is nearing completion of a standard called SPHBM4 — Standard Package HBM — that takes the same DRAM dies as HBM4 and, via a new base die that serializes the interface 4:1, drives them over 512 data signals instead of 2,048. The relaxed pin count means the stacks can sit on ordinary organic substrates rather than silicon, cutting out the scarce, expensive advanced-packaging step that constrains how much HBM the industry can ship. It is a genuinely clever piece of engineering economics.

It is also aimed entirely at AI accelerators. Our read: SPHBM4 makes HBM-class memory cheaper to package and easier to scale, which if anything strengthens the case for keeping wafer allocation right where it is. Nothing in the standard touches LPDDR, the mobile DRAM that budget phones actually buy, and Kioxia’s densest NAND is spoken for by data centers before a $150 handset sees a gigabyte of it. The supply response is real; it is simply queued in priority order, and handset buyers are at the back. Relief for hyperscalers is being engineered on a faster clock than relief for anyone holding a cracked Redmi.

One thing we looked for and could not find: a documented street-price increase on a named phone in Omdia’s ultra-budget, sub-$99 tier — the segment where memory hits 64 percent of build cost. The mid-tier POCO hikes above are the closest paper trail so far. If those cheapest models move, or quietly vanish from price lists instead, that will be the tell.

Omdia doesn’t expect meaningful supply relief for phone memory until mid-2027 at the earliest, with normalization possibly stretching into 2028. In the meantime, one number to hold on to, from Tom’s Hardware’s reporting on Samsung’s quarter: a single 12GB LPDDR5X module now runs about $145 on the open market. That is one component — and more than the full retail price of every phone in Omdia’s cheapest tier.