The launch of the A20 Pro chip, which powers the new iPhone 18 and the foldable iPhone Duo model, marks a pivotal milestone in the integration of mobile artificial intelligence. Built on TSMC's cutting-edge 2-nanometer process, this architecture features a neural engine with 32 cores, doubling its computational capacity compared to previous generations. This power boost is accompanied by a significant expansion in memory bandwidth, which has increased by 50%. By reaching an estimated throughput of 115 GB/s, Apple is now rivaling the performance of its own laptop processors, such as the M4 chip.
Despite these technical strides, Random Access Memory (RAM) capacity remains the primary bottleneck of the ecosystem. Although a shift to 12 GB of RAM is anticipated, this volume imposes a strict ceiling for running complex large language models locally. On a smartphone, every request requires moving the model's entire parameter set between the memory and the processor. Consequently, even if data throughput—the "plumbing"—is optimized, the total amount of available memory restricts the size of the AI models the device can manage alongside standard system tasks.
Faced with this hardware constraint, Apple's strategy relies on aggressive software optimization. The current industry trend is shifting toward compressed models and much leaner "mixture-of-experts" architectures capable of impressive performance with fewer parameters. Adopting 8-bit floating-point calculations allows for denser processing, making the local execution of models with 7 to 8 billion parameters technically viable. This presents a notable economic paradox: the scarcity of memory components, exacerbated by global demand from data centers, limits onboard capabilities while driving up the final cost of the devices.
Beyond hardware specifications, the real challenge lies in the concrete application of this computing power. For European users, the situation is paradoxical: although the hardware is present, local artificial intelligence features remain inaccessible for the time being due to regulatory hurdles. The success of this generation will therefore depend less on the raw power of its neural cores and more on the ability of developers to offer applications that truly leverage these local resources, rather than relying on remote cloud infrastructures.