Apple’s artificial intelligence strategy rests not on data centers or trillion-parameter language models, but on silicon. For 15 years, the company built neural engines directly into its custom chips, betting that on-device processing would eventually define how AI works. In 2026, that bet is paying dividends as competitors scramble to catch up.
In an exclusive conversation with The Deep View ahead of WWDC 2026, Doug Brooks, senior product manager of Apple silicon, discussed how Apple’s long-term investments in neural processing, unified memory, power-efficient computing, and hardware-software integration have positioned its devices for the AI era. This interview has been edited for brevity and clarity.
The balanced architecture: CPU, GPU, Neural Engine, and unified memory
The critical breakthrough was unified memory architecture, as unlike traditional chips that shuttle data between separate CPU, GPU, and neural processing units, Apple’s design treats all three as a single pool. This fundamentally altered what on-device AI could accomplish.
According to Brooks, the tent poles of Apple silicon are performance, efficiency, and unified memory. The architecture provides CPU, GPU, unified memory, and the Neural Engine all contributing to performance across the chip, with particular importance for agentic workflows where the whole chip contributes to different parts of the task. As Brooks explained, it is not just about the GPU crunching on an LLM anymore; it is about the whole chip contributing to different parts of the task, tool-calling, and the things that are happening around those workflows.
The M4 chip, introduced in 2024 as the first Apple processor built from the ground up for AI, demonstrated the payoff. Its Neural Engine processes 38 trillion operations per second, enabling the device to run substantial foundation models entirely locally. The M5 chip, announced in May 2026, accelerated that capability further, with faster CPU and GPU performance alongside a more capable Neural Engine and unified memory that allows devices to run larger AI models on-device and execute AI inference through either the GPU’s Neural Accelerators or the dedicated Neural Engine.
Apple built dedicated acceleration into multiple layers of the chip. According to Brooks, the CPU includes neural accelerators, originally called ML accelerators, that handle AI workloads in the CPU domain, particularly low-latency tasks like speech recognition. The GPU now carries neural accelerators as well, providing a huge boost in performance because the GPU is the most scalable engine. As systems grow larger with more GPU resources, available AI capabilities expand to tackle bigger problems.
This architecture cascades across Apple’s entire product line since MacBook Pro, iPad Pro, and Apple Vision Pro all benefit from the same unified memory principle, each capable of running meaningful AI workloads without relying on cloud processing. The A17 Pro, Apple’s first 3-nanometer smartphone chip, carries a 16-core Neural Engine, twice as fast as its predecessor in the A16.
Walk into any of the frontier AI labs, and you will find wall-to-wall Macs. Developers across OpenAI, research labs, and AI startups have converged on macOS as their primary platform. According to Brooks, the Mac has long held strength in the developer community; the emergence of on-device AI crystallized that advantage. Productivity, rich development tools, and strong compile times have always been strengths on macOS. Now, so many of the tools in the industry are either Mac-only or Mac-first, making the platform the natural choice for cutting-edge AI work.
Mac minis and Mac Studios have emerged as the platform of choice for running AI agents and agentic workloads. According to Brooks, for agentic workflows, developers often want a system under their control, isolated from their primary machine, and capable of running 24 hours a day, seven days a week. A Mac mini provides exactly that combination at compelling price-performance.
The economics of local AI workloads have shifted dramatically. Brooks noted that agentic workloads consume three to ten times more tokens than traditional inference tasks. Running those models locally avoids rising costs of cloud inference and eliminates privacy and security concerns that come with sending data, code, or intellectual property to external servers.
Modern optimization and quantization techniques have made this shift viable. According to Brooks, developers are now running 70-billion-parameter and 120-billion-parameter models on laptops and writing software completely disconnected because they are running powerful models locally on Apple silicon. Brooks noted a hybrid approach is particularly interesting because agents can decide what needs to happen locally and what needs to happen in the cloud based on the workload.
For a decade since 2016, Apple was TSMC’s priority customer, a relationship that guaranteed preferential access to the most advanced manufacturing nodes, however, that relationship is changing.
Nvidia and AMD now compete directly with Apple for TSMC’s most cutting-edge capacity. AI accelerators consume substantially more wafer area per chip than smartphone processors, and as artificial intelligence workloads exploded across the industry, AI-focused manufacturers began outbidding Apple for advanced nodes. In February 2026, reports surfaced that Apple is exploring manufacturing some lower-end processors outside TSMC for the first time in a decade, a tacit acknowledgment that preferential access is no longer guaranteed.
In February 2025, Apple committed more than $500 billion in spending over the next four years and pledged to manufacture advanced silicon at TSMC’s Fab 21 in Arizona, which will employ over 2,000 workers. Apple announced plans to hire around 20,000 people, with the vast majority focused on R&D, silicon engineering, software development, and AI and machine learning. Apple is doubling down on the silicon-for-AI bet even as the competitive landscape shifts.
The biggest test of Apple’s AI strategy arrives in spring 2026 with a complete overhaul of Siri. For years, Siri lagged behind Google Assistant and OpenAI’s ChatGPT in conversational ability and task completion. Apple’s new Siri, powered by the next generation of foundation models and deep personal context integration, will handle multi-step requests that previously required cloud processing or manual app switching.
At WWDC 2026, Apple introduced its latest generation of language foundation models and released the Foundation Models framework giving third-party developers direct access to the on-device 3-billion-parameter language model that powers Apple Intelligence. This move signals Apple’s confidence in what on-device inference can accomplish; it also distributes the responsibility for innovation across the ecosystem rather than concentrating it within Apple’s own applications.
According to Brooks, developer adoption of on-device AI capabilities is accelerating across all device categories. Applications like Localy AI and Draw Things run on-device LLM and image-generation workloads on iPhone, iPad, and Mac with highly optimized models. Creative and entertainment applications are leading the charge, with tools like DaVinci Resolve adding AI capabilities that put significant power into the hands of artists and creators.
via MacRumors
