Qualcomm’s NPU Meets a Near-60% Memory Bill

On July 8, Omdia reported that memory costs had reached nearly 60% of first-quarter materials costs for smartphones priced below $400. That is the budget into which Qualcomm is selling faster mobile NPUs.

The accelerator score ends at the edge of the chip

In 2015, Qualcomm presented cognitive compute as a Snapdragon 820 feature: the chip would supply the capability, and handset makers would turn it into products. Qualcomm’s pitch made chip performance the measure of progress.

TOPS still measures accelerator throughput. Geekbench AI broadens the comparison by testing CPUs, GPUs, and NPUs across Android, iOS, and desktop systems. Neither metric tells a product team whether model weights and context fit in usable memory, whether the memory bus can feed the accelerator efficiently, or whether the enclosure can dissipate the resulting heat.

An NPU benchmark therefore cannot by itself settle whether an AI phone can ship profitably. The product team must balance usable memory, bandwidth, energy, thermal headroom, cloud routing, and total materials cost.

Memory constrains both model fit and handset cost

An on-device language model occupies memory before it produces a token. Its weights and context consume capacity, and generation repeatedly moves data through a narrow power budget. A faster accelerator cannot process data that the memory subsystem fails to deliver.

SK hynix raised LPDDR5T transfer speed to 9.6Gbps from LPDDR5X’s 8.5Gbps, a 13% gain. The faster interface feeds compute more quickly, but it does not create additional capacity or reduce the cost of securing it.

On April 18, Nikkei Asia projected that global DRAM supply would satisfy only 60% of demand through 2027 as memory makers focused on AI. The same report projected that memory would rise from 20% to roughly 40% of a low-end smartphone’s manufacturing cost by mid-2026.

Omdia’s July 8 report found Q1 memory costs already near 60% of materials for smartphones priced below $400. The April and July figures are not a direct time series: Nikkei used a “low-end” category and manufacturing cost, while Omdia specified a price cutoff and materials cost. They are non-comparable signals of pressure, not evidence that Omdia’s segment exceeded Nikkei’s forecast.

Omdia also expects global shipments of smartphones below $400 to fall more than 22% year over year in 2026.

Ming-Chi Kuo expects data centers to tighten the market further. On June 29, he estimated that 15%–20% of the memory capacity assigned to consumer electronics in 2026 would shift to data centers in 2027. A phone team can shrink or quantize a model, but it still needs an LPDDR contract for every unit it plans to sell.

Qualcomm prices one line of the phone’s bill

Qualcomm sells a SoC inside a bill of materials that another company must finance. The handset maker still buys memory, display, battery, cooling, cameras, storage, and radio components, then tries to recover the combined cost from the customer.

On July 24, Bloomberg reported that Qualcomm had told customers in a letter that it planned double-digit percentage price increases after exhausting its ability to absorb higher supplier costs. The letter concerned Qualcomm’s own costs; it did not identify handset DRAM as their cause. It did show how separate suppliers pass cost pressure into the same finished device.

A handset maker facing both increases must choose among a higher retail price, less memory, cuts elsewhere in the device, or a lower margin. Qualcomm can set the price of its component, but it cannot make that tradeoff for the customer.

Apple can trade software against hardware

In 2024, Apple Intelligence paired an approximately 3-billion-parameter on-device model with larger models running on Apple silicon servers. Apple coordinates its model family, operating-system behavior, silicon, privacy policy, and server routing under one product authority. It can optimize the model, change the software path, or route an eligible request to Private Cloud Compute.

A PrismML demonstration reportedly ran a 27-billion-parameter Qwen 3.6 model on an iPhone 17 Pro. That result demonstrates technical fit on one device, not the procurement required for a product line. A phone maker must reproduce the configuration across every unit it intends to ship.

Apple’s integration widens its options, but it does not create DRAM. SK hynix, Samsung Electronics, and Micron still allocate physical supply, so an optimized design can work in a demonstration and fail as a product if its maker cannot secure enough memory at an acceptable cost.

Qualcomm is moving beyond one handset component

On June 25, Qualcomm raised its 2029 non-handset revenue target from $22 billion to $40 billion and set a target of $15 billion in data-center chip sales. Its High Bandwidth Compute architecture makes memory bandwidth central to that data-center strategy.

The company also agreed to acquire Modular for nearly $4 billion, adding a chip-software platform and proprietary programming language. Modular reaches above the silicon into the layer that maps models onto hardware and manages deployment constraints.

The non-handset target would reduce Qualcomm’s dependence on phone economics, while Modular would let it sell more of the inference stack. Neither move lowers an Android maker’s LPDDR bill, but each extends Qualcomm beyond one component inside a device it cannot price.

In 2015, Qualcomm’s drawing ended at Snapdragon 820. In 2026, nearly $4 billion for Modular and a $40 billion non-handset target push that boundary outward. But for a smartphone below $400, Omdia’s near-60% Q1 memory share leaves the immediate choice with the handset maker: cut memory, raise the price, accept less margin, or ship fewer AI phones.