Apple is building a Mac chip with enough memory to run the largest AI models entirely on your own hardware, a shift that could challenge NVIDIA's dominance in AI computing. According to a Bloomberg report, the company is designing an M7 Ultra processor with 1.5 terabytes of unified memory, roughly double what was planned for the M5 Ultra generation. This aggressive pivot toward AI-focused silicon represents an unprecedented acceleration in Apple's chip development cycle, with the M7 Ultra expected to arrive in 2028.

The memory capacity matters because running large language models (LLMs), which are AI systems trained on vast amounts of text to generate human-like responses, requires enormous amounts of computing memory just to load the model weights. Today, even high-end consumer Macs max out at 128 gigabytes of unified memory, a shared pool that both the processor and graphics chip can access. With 1.5 terabytes, Apple would enable users to run frontier-level AI models that currently require expensive cloud subscriptions or dedicated server hardware.

This development matters to a growing community of local AI enthusiasts who use tools like LM Studio and Ollama to run AI models on their own computers. These users value privacy, offline capability, and avoiding cloud computing costs. The Mac Mini has already become the default always-on server for running local AI agents, partly because it offers a balance of affordability and performance. An M7 Ultra with 1.5 terabytes would dramatically expand what's possible on consumer hardware.

What's Driving Apple's AI Hardware Overhaul?

Apple is skipping the Pro, Max, and Ultra variants of the M6 generation entirely to accelerate its timeline to the M7 family. The base M7 is expected in the first half of 2027, with Pro and Max models arriving by the end of that year, and the Ultra variant following in 2028. Engineers have already begun developing an AI server chip based on the M7 Ultra architecture for a 2029 launch, signaling that Apple views this technology as foundational to its future computing strategy.

The company is also preparing 1.4-nanometer M8 chips, code-named Soko, alongside other high-end Mac processors under the name Cardinal. This level of investment suggests Apple is betting heavily that on-device AI will become as essential as cloud computing once was. The shift accelerates a trend already visible in the local AI community, where users increasingly prefer running capable models on their own hardware rather than relying on external APIs.

How Does This Compare to Current Local AI Setups?

Today, running capable AI models locally on a Mac relies almost entirely on the graphics processing unit (GPU), not Apple's Neural Engine, a dedicated chip designed for low-power background tasks. When users run LLMs using tools like LM Studio, Ollama, or Apple's MLX framework, the GPU cranks to 100% utilization while the Neural Engine sits idle. The Neural Engine excels at fixed-function tasks like facial recognition or background blurring, but it lacks the flexible architecture needed for the complex mathematical operations required by large language models.

Apple's success in attracting AI developers isn't due to the Neural Engine or a canceled self-driving car project, as some have suggested. Instead, it stems from building an efficient, high-capacity unified memory pool that allows the GPU to access enormous amounts of data without performance penalties. If the M7 Ultra actually ships with 1.5 terabytes of memory, Apple would leapfrog both NVIDIA and AMD's current development platforms for AI workloads.

What Models Can Run on Current Apple Silicon?

The local AI community has already found success running smaller, specialized models on current-generation Macs. Mistral AI released the Ministral 3 family in December 2025, offering three model sizes: 3 billion, 8 billion, and 14 billion parameters. All three support vision input, allowing them to process both images and text, and all are released under the Apache 2.0 license, permitting commercial use without royalty payments.

For Mac Mini M4 owners, the 8 billion and 14 billion parameter models are particularly significant because they run comfortably on consumer-grade unified memory while delivering performance comparable to much larger models on published benchmarks. This combination of real capability paired with modest hardware requirements is exactly what local AI enthusiasts have been requesting since on-device language models first arrived in 2024.

Steps to Getting Started With Local AI on Mac

• Choose a Local AI Framework: Popular options include LM Studio, Ollama, and Apple's MLX framework, each offering different levels of ease-of-use and customization for running models on your Mac.
• Select an Appropriate Model: Start with smaller models like Ministral 3 8B or 14B, which run efficiently on current Mac hardware while delivering strong performance on knowledge and reasoning tasks.
• Monitor GPU Utilization: Watch your Mac's GPU usage when running models locally; you should see the GPU reach high utilization while the Neural Engine remains largely idle, confirming the GPU is handling the AI math.
• Consider Memory Requirements: Ensure your Mac has sufficient unified memory for your chosen model; the 8B and 14B Ministral variants work well on standard consumer Mac configurations.

The gap between today's local AI capabilities and what's coming is substantial. Current Macs can run smaller, specialized models efficiently, but the M7 Ultra's rumored 1.5 terabytes of memory would enable running the largest frontier models that currently require cloud infrastructure or expensive server hardware. For users who value privacy, offline capability, and avoiding recurring cloud computing costs, this represents a fundamental shift in what's possible on consumer hardware.

Of course, 2028 is still two years away, and the AI landscape could shift dramatically in that time. NVIDIA, AMD, and other chip makers will likely reveal new technologies before Apple's M7 Ultra arrives. However, if Apple delivers on these specifications, the company would fundamentally reshape the economics of local AI, making it practical for far more users to run powerful models entirely on their own machines.