Apple evaluates on-device AI that threatens memory chip demand

Apple is in early discussions with PrismML, a Khosla Ventures-backed startup that claims it can shrink large artificial intelligence models to run directly on iPhones. The company recently compressed Alibaba's 54 GB Qwen model to under 4 GB, allowing its 27 billion parameters to operate on an iPhone 15 or newer.

PrismML achieves this by reducing how internal model data is stored, dropping requirements from 16 bits to just one to three possible values. Chief Executive Babak Hassibi said the resulting models use 10 to 15 times less memory and run six to eight times faster. However, the compression sacrifices a few percentage points of overall performance and weakens factual recall.

The discussions come as Apple launches the public beta of iOS 27, an overhaul of its Siri assistant. Running capable models locally would help Apple reduce cloud-computing costs, eliminate remote server delays, and maintain its privacy pitch for sensitive user data. "The more you can do on device, the better it is," said Carolina Milanesi, an analyst at Creative Strategies.

Memory cost pressures

For investors, the technology strikes at a critical commodities debate: whether AI efficiency gains will dampen the explosive demand for memory chips. Memory constraints are currently a major financial burden for hardware makers. Morgan Stanley estimates Apple's average dynamic random access memory cost per bit could jump roughly 190% year over year in fiscal 2027, with NAND costs rising about 180%.

Those rising component costs are expected to force Apple to raise the starting price of comparable iPhone 18 models by about $200 to protect its margins. PrismML argues its approach could allow models requiring eight GPUs to run on a single GPU, or shift workloads entirely from servers to consumer devices.

The market has historically punished any suggestion of reduced memory needs. Micron shares plunged in March after Google published a paper on cutting memory use, though the stock later recovered. Analysts note that PrismML's claims must survive real-world scaling. "The ultimate test will be millions of queries, thousands of device combinations and robust testing at scale," said Tarun Pathak of Counterpoint Research.

Furthermore, shrinking models may not actually reduce total semiconductor demand. Gil Luria, an analyst at D.A. Davidson, noted that moving AI to the edge simply shifts processors and memory from datacenters into phones. "It's not that you're not going to need the chip," Luria said. "You're still going to need the GPU, and you're still going to need the memory." He added that running AI on individual devices can be less efficient than shared datacenter infrastructure because phone chips often sit idle.