Apple announces M5 Ultra quad-die chip and M6 alongside new Mac Mini and Mac Studio

Two new chips, one AI story

Apple introduced two processors on Tuesday: the M5 Ultra, which the company describes as its most powerful chip to date, and the M6. According to TechCrunch, the M5 Ultra is Apple's first quad-die design, fusing two dual-die M5 Max chips into a single package intended for compute-heavy AI workloads. It offers up to a 36-core CPU, an 80-core GPU and 1.2 TB/s of unified memory bandwidth, which is 50% more than the M3 Ultra delivered about 18 months earlier.

A developer analysis on dev.to adds technical context: the dies are joined by a next-generation UltraFusion interconnect carrying more than 4.4 TB/s between them while presenting to the system as a single processor, with the CPU split into 12 efficiency-oriented super cores and 24 performance cores. The post also notes that this is the first Ultra chip to get Neural Accelerators, dedicated matrix-multiply hardware inside each GPU core, and that Apple claims up to 4.3x the peak AI compute of the M3 Ultra plus up to 4x faster LLM prompt processing in LM Studio, directly targeting the long wait for first tokens on large local contexts.

The M6 goes mainstream

While the M5 Ultra is aimed at professionals doing 3D rendering, visual effects and frontier model work, TechCrunch reports that the M6 targets everyday users. Apple Silicon Engineering VP Sri Santhanam said in a press release that the chip is built on a 2 nm process with a reworked CPU complex, two additional CPU and GPU cores, and a dual 16-core Neural Engine. Apple claims nearly 30% higher peak GPU compute for AI compared with the M5, which speeds up prompt processing for on-device LLMs.

The new Mac Mini comes in M6 and M5 Pro variants. Per TechCrunch, the M6 base model costs $899 with 16GB of RAM and 256GB of storage, and Apple claims 4x the AI performance of the M4-based Mini along with doubled storage and graphics speed, moving from 10 CPU and GPU cores to 12 of each. The M5 Pro model starts at $1,699 with 24GB of RAM and 512GB of storage, with a default 15-core CPU and 16-core GPU upgradeable to 18 and 20 cores. Both include Wi-Fi 7, Bluetooth 6 and a 2.5Gb Ethernet port, and ship after September 22 with macOS 27 and Siri AI.

Mac Studio pricing hides the real story

New Mac Studio models run on the M5 Max and M5 Ultra, starting at $2,499 and $5,499 respectively, with base configurations of 36GB plus 512GB storage and 96GB plus 1TB storage, according to TechCrunch.

The dev.to analysis points out that the headline price buys a binned 30-core CPU and 64-core GPU part, while the full 36-core, 80-core chip starts at $6,799. Because memory is the main reason to buy the machine for AI work, the 256GB unified memory upgrade at $4,000 matters more, pushing a serious configuration past $10,000, and 512GB configurations will not ship until late October. The post attributes the steep memory pricing to a DRAM shortage driven by datacenter demand, noting the M3 Ultra launched at $3,999 and was later repriced to $5,299.

Two other additions stand out: Thunderbolt 5 clustering with RDMA lets memory be pooled across multiple machines, with Apple claiming a four-Studio cluster reaches 3x the inference throughput of a single box, and macOS 27 introduces Core AI, a new framework for deploying full-scale LLMs locally alongside MLX.

Why it matters

Apple's own AI model efforts have trailed competitors, with TechCrunch noting that the long-awaited Siri upgrade runs on Google's Gemini. But the company has a proven record in on-device processing, from Face ID to Touch ID, and is positioning the Mac as a private, controlled environment where developers can run and fine-tune large models entirely locally. The Mac Mini has already found an audience here, with TechCrunch reporting demand from hobbyists running local AI agents such as OpenClaw and Hermes.

The dev.to analysis frames the trade-off plainly: NVIDIA hardware wins on raw inference speed and the CUDA ecosystem, while Apple's advantage is memory capacity per dollar, with up to 512GB of fast, coherent unified memory in a quiet, low-power desktop. That profile suits memory-hungry sparse mixture-of-experts models well, and the machine is strongest where data cannot legally leave a building or where bulk workloads justify zero marginal cost. The caveat remains that local open-weight models still trail the best hosted frontier models, so buyers should not expect to cancel subscriptions without noticing a difference.