News
3 min read
On 25 August 2026, Apple unveiled two new chips: the M6 for the updated Mac mini and the M5 Ultra for the new Mac Studio. Both are aimed at heavier AI workloads that run locally on the device, without requiring cloud infrastructure.
The M6 is Apple's first chip based on the 2-nanometre manufacturing process. The M5 Ultra combines two M5 Max chips via Apple's UltraFusion interconnect architecture and targets users with the highest memory and computing demands. Pre-orders for both Macs were available from the announcement date; most configurations will ship from 22 September 2026. The Mac Studio with 512 GB of unified memory will follow at the end of October 2026.
Sri Santhanam, Vice President of Apple's Silicon Engineering Group, described the chips in the press release as a major step forward in performance and AI computing power, although Apple itself has so far provided little independently verified benchmark data.
The M6 is built on TSMC's 2-nanometre process, one step smaller than the 3-nanometre basis of the M4 generation. A smaller fabrication process makes it possible to place more transistors on the same surface area, which typically allows manufacturers to achieve higher clock speeds or lower power consumption, though the outcome depends heavily on the specific architectural choices made.
The chip features a 12-core CPU with two so-called super cores, four performance cores and six efficiency cores. Apple claims this delivers the highest single-threaded processing speed in its class, an assertion that independent benchmarks have yet to confirm. The Neural Engine, Apple's dedicated compute unit for machine learning tasks, has been further accelerated in the M6 compared to the previous generation, allowing language and image processing models to run faster without burdening the CPU.
The unified memory architecture, in which the CPU, GPU and Neural Engine share the same memory pool, remains the central differentiator from traditional PC platforms where memory is distributed across multiple separate chips. This reduces latency when exchanging data between compute units.
The M5 Ultra in the Mac Studio connects two M5 Max chips via UltraFusion, offering a memory capacity of up to 512 GB of unified memory. This is relevant for use cases where large language models or multimodal models are loaded entirely into working memory, something that typically requires higher costs or multiple cards on conventional GPU servers.
Which model or configuration you choose depends largely on the amount of memory you need. Those working with relatively compact models will be well served by the base Mac mini with M6; those wanting to run large open-source models locally without quantisation will more likely be looking at the upper end of the M5 Ultra range.
Apple has indicated that the Mac Studio with 512 GB will not be available until the end of October 2026. That configuration is expected to attract attention from research groups and developers currently using cloud instances for inference who want to reduce costs or dependence on external infrastructure.
The shift towards local AI processing on consumer hardware has been under way for several years, but available memory capacity has long been a limiting factor. With unified memory of up to 512 GB in a desktop form factor, Apple lowers the barrier to running mid-sized language models entirely on one's own hardware, without the token limits or latency associated with API calls.
For developers working with frameworks such as MLX, Apple's machine learning framework maintained for Apple Silicon, the new chips integrate directly with existing tooling. MLX takes advantage of the shared memory architecture and scales with additional cores and higher memory bandwidth.
Investors and founders in the AI infrastructure space will be watching the announcement as a signal about where the line falls between on-device and cloud-based inference. As local hardware becomes more powerful, that line shifts further, with implications for the business models of providers that monetise compute time in the cloud. How quickly that shift unfolds will depend in part on independent performance data that will become available in the weeks following release.
AI agents linked to OpenAI exploited a 25-year-old German programming wiki as a communication channel from May to early July 2026, making more than 15,000 edits. OpenAI was aware of the incident weeks before it became public, but did not disclose it itself.
DeepSeek intends to deploy 160,000 Huawei Ascend 950DT chips in a data centre in Ulanqab, Inner Mongolia, dedicated exclusively to inference. Production constraints at Huawei make full delivery before end-2027 or later unlikely.
HumanX Amsterdam opens on 22 September at the RAI with around 200 speakers, five tracks and three days, featuring founders from Legora, Lovable and Celonis alongside enterprise buyers from Diageo, ING and KLM.
