Apple unveils 2nm M6 and quad‑die M5 Ultra chips for desktop AI

Apple announced a pair of new silicon offerings on August 25, 2026 that target the growing desktop AI market. The M6, built on a 2‑nanometer process, powers a refreshed Mac mini, while the M5 Ultra, a quad‑die design that merges two M5 Max chips, equips the new Mac Studio. Both devices are positioned as on‑premise alternatives to cloud inference, allowing developers to run large language models and image generators locally.

The M6 chip features a 12‑core CPU that Apple describes as the world’s fastest core, a 12‑core GPU, and a Neural Accelerator embedded in each GPU core. It also includes a dual 16‑core Neural Engine that delivers roughly twice the peak compute of the previous generation. Memory support tops out at 32 GB of unified RAM with a bandwidth of 170 GB/s, a modest 10 percent increase over the M5. Apple claims the M6 offers up to 1.2× faster multithreaded CPU performance and about a 30 percent boost in peak GPU compute for AI tasks.

The M5 Ultra represents Apple’s most ambitious desktop silicon to date. It combines two dual‑die M5 Max chips via an updated UltraFusion interconnect, creating a single quad‑die processor. This architecture yields up to a 36‑core CPU, an 80‑core GPU, and a unified memory bandwidth of 1.2 TB/s, which is 50 percent higher than the earlier M3 Ultra. Configurations can be ordered with up to 512 GB of unified memory, enabling the chip to host large language models with hundreds of billions of parameters entirely on the device. Each GPU core also carries a Neural Accelerator, and the chip houses a 32‑core Neural Engine and a Media Engine with four ProRes encode/decode blocks and hardware‑accelerated AV1 decoding.

Apple’s UltraFusion technology is central to the M5 Ultra’s performance. The interconnect links the four dies with more than 4.4 TB/s of bandwidth and a connection density over six times that of the prior generation. This high‑speed fabric allows the four silicon pieces to share a single memory pool without the latency penalties typical of multi‑chip systems, effectively turning the quad‑die package into one logical processor from the operating system’s perspective.

These chips arrive in two price points. The Mac mini with M6 starts at $899 and is available for pre‑order, with shipments slated for September 22, 2026. An M5 Pro‑based Mac mini option offers up to 64 GB of memory and higher bandwidth for users needing more capacity. The Mac Studio equipped with M5 Ultra begins at $2,499 for the M5 Max configuration, with the ultra variant configurable up to 512 GB of memory. Apple also introduced a leasing program called Apple Upgrade, priced at $48.99 per month over 36 months for the Mac Studio.

Apple frames the new hardware as a solution for local inference, reducing reliance on per‑token cloud pricing. The company lists applications such as LM Studio Bionic, Ollama, and Draw Things as examples of software that can leverage the increased compute and memory. All performance claims stem from Apple’s internal testing against prior‑generation devices using selected industry‑standard benchmarks. The 512 GB memory ceiling is an optional configuration rather than the base model, and the new Siri AI features require macOS 27, currently available through the Apple Beta Software Program.

The launch underscores Apple’s continued partnership with TSMC, which supplies the 2‑nm process, though the foundry is not explicitly named. Apple also highlights that the new Mac mini is manufactured with 100 percent renewable electricity across its supply chain, aligning the hardware rollout with the company’s sustainability goals. Analysts will watch how the on‑device AI proposition competes with cloud‑centric offerings from other silicon vendors and whether developers adopt the high‑memory desktop for fine‑tuning large models.

Looking ahead, Apple’s roadmap suggests further scaling of its Fusion architecture and potential expansions of unified memory limits. Observers should monitor developer adoption rates, benchmark results from independent labs, and the impact on Apple’s broader AI services strategy as macOS 27 rolls out later this year.