Apple is in preliminary discussions with Silicon Valley startup PrismML, which claims to compress large AI models for local execution on iPhones. If validated, this technology would bolster Apple's privacy stance and provide crucial support for its long-anticipated Siri upgrade.

PrismML's CEO stated that Apple has begun evaluating its technology. Talks are at a very early stage, with the outcome uncertain, but are reportedly "progressing well."

This development follows the public release of the iOS 27 beta, the first widespread test of a significantly revamped Siri aimed at competing with AI assistants from OpenAI and Anthropic.

For Apple, keeping more AI processing on-device promises lower response latency, reduced cloud computing costs, and offline functionality for certain features—aligning closely with its core privacy positioning.

PrismML's public release allows users and investors to test its claims outside a lab setting. As a Counterpoint analyst summarized, a hybrid cloud and on-device AI approach can deliver a more complete, efficient, and privacy-focused experience, handling complex tasks in the cloud while processing sensitive, latency-critical, and privacy-related tasks locally.

Compression Breakthrough: 54GB Model Shrinks Under 4GB

PrismML, incubated from a Caltech research team and backed by Khosla Ventures, publicly released a compressed version of Alibaba's open-source Qwen model. It reduced the original ~54GB model to under 4GB, enabling the full 27-billion-parameter model to run directly on an iPhone 15 or newer.

PrismML's CEO explained to CNBC that the compression is achieved by drastically simplifying how model information is stored internally, reducing each numerical value from 16 bits to just 1-3 possible values, thereby significantly cutting the memory required for storage and operation.

The company claims the compressed model reduces memory usage by 10-15x, speeds up responses by 6-8x, and cuts energy consumption by 3-6x. The CEO acknowledged trade-offs: overall model performance typically drops by a few percentage points, with factual memory degrading before reasoning, math, and coding capabilities.

PrismML has released two free compressed versions compatible with iPhones, MacBooks, and PCs with Nvidia chips. The company stated Google's open-source Gemma model is its next compression target, followed by larger frontier models that currently require data center hardware.

In March, the company closed a $16.25 million seed round led by Khosla Ventures. Caltech holds the underlying patents and has granted PrismML an exclusive license.

Apple's On-Device AI Strategy: Driven by Privacy and Efficiency

Apple already runs some AI features locally on devices, including translation, some summarization, and features tied closely to personal information, with more complex requests routed to its private cloud infrastructure or external models.

An industry analyst suggested Apple's goal is likely to keep the vast majority of daily Siri interactions on-device, sending only the most demanding tasks to the cloud. Local processing means lower latency, stronger privacy, and potentially lower licensing and cloud computing costs.

A lead analyst noted that smaller models could allow Apple to move more demanding features to the iPhone locally, including computational photography, video generation, and health and fitness tools relying on sensitive health and medication data. "The more you can do on-device, the better," she said.

Apple holds a potential advantage in advancing on-device AI through its chip and software co-design, granting it more precise control over how AI runs on its devices.

Analyst Caution: Scaling Validation is the Key Hurdle

Despite impressive technical claims, analysts generally caution that PrismML's assertions need validation beyond controlled demos.

A research director highlighted that performance on long prompts, battery consumption during multitasking, and reliability across millions of requests and thousands of device combinations will be critical tests. The ultimate test will be at scale across millions of queries and device combinations.

Another research lead pointed out that power consumption may be the biggest unknown. A model powerful and frequently used enough to run agent tasks, even in the background, could significantly drain a phone's battery despite lower memory usage.

Impact on Chip Demand: A Shift, Not a Reduction

PrismML's announcement touches on the ongoing debate about whether AI efficiency gains will ultimately reduce chip demand.

Morgan Stanley estimates Apple's per-bit DRAM costs could rise about 190% year-over-year in FY2027, with NAND storage costs up about 180%. The firm expects Apple to raise the starting price of an iPhone 18 equivalent by about $200 to protect margins.

PrismML states its compression could reduce a cloud model requiring 8 GPUs to just 1, and shift server-dependent models to phones and laptops. This may reduce the memory or compute needed for a single AI task but does not necessarily lead to an overall decline in chip demand.

An analyst from D.A. Davidson noted that model slimming doesn't eliminate the need for processors or memory; it simply shifts more chips from data centers to end devices like phones. "You still need GPUs, you still need memory," he said.

He added that running AI on a single device might be less efficient than using shared data center infrastructure, as phone chips are idle most of the time. Furthermore, efficiency breakthroughs often spur more usage, not reduced spending.

The market has been sensitive to signals that AI requires less memory than expected. In March, when Google published its TurboQuant paper on cutting memory use without harming performance, Micron Technology's stock fell sharply before later recovering.