Two Mac Studios have more computing resources than one. Yet in Apple's AI benchmarks, adding a second machine made token generation slightly slower in one configuration.
Apple's October 15, 2026, research found that distributed AI inference performance depends on how work is divided between machines. The results raise practical questions for enterprises expanding local AI infrastructure, where additional hardware brings networking requirements and costs without guaranteeing proportional performance gains.
The tests also show that increasing model capacity and accelerating AI responses require different approaches to distributing workloads.
Why two Macs weren't always faster
In research published by Apple's machine learning team, engineers tested large language models on M5 Ultra Mac Studios using Apple's open-source MLX framework. The machines communicated over Thunderbolt 5 connections supporting remote direct memory access (RDMA).
With the 27-billion-parameter Qwen 3.8 model and an 8,192-token context, one Mac Studio generated 19.69 tokens per second. Two machines using pipeline parallelization produced 19.44 tokens per second, approximately 1.3% slower in that test.
Pipeline parallelization distributes a model's layers between machines, allowing models too large for one system to run across multiple computers. However, processing layers sequentially does not necessarily accelerate token generation.
Tensor parallelization instead divides computations within individual layers. Apple's two-Mac configuration reached 33.27 tokens per second using this method, while four Macs achieved 50.89 tokens per second.
This approach requires frequent communication between machines, making network latency an important performance consideration. It can also increase implementation complexity compared with pipeline parallelization.
Apple recorded a different result for long-context processing. At a 65,536-token context, pipeline parallelization improved initial prompt processing by up to approximately 1.8 times over one machine. That improvement concerned processing the prompt before generation, not the speed of generating subsequent tokens.
The findings add measured performance data to Apple's earlier local AI infrastructure plans, which introduced multi-Mac clustering as a way to support larger models.
What eWeek found: Four Macs increased AI generation speed by 158%
In Apple's Qwen 3.8 benchmark, four Macs using tensor parallelization achieved 2.58 times the token generation speed of a single Mac running the same model. Generation speed increased from 19.69 to 50.89 tokens per second, a 158% improvement under Apple's test conditions.
Doubling the hardware from two tensor-parallel Macs to four increased generation speed from 33.27 to 50.89 tokens per second, a gain of approximately 53%.
These calculations use Apple's published measurements for one model, an 8,192-token context, 128 generated tokens and a batch size of one. They do not establish performance under simultaneous enterprise requests or sustained production workloads.
For infrastructure architects, the results expose a procurement trade-off. Pipeline parallelization can accommodate larger models, while tensor parallelization can improve generation speed but requires more communication between machines.
Hardware demands also influence operating costs. Microsoft's local coding AI requirements illustrate how avoiding per-call inference fees can shift spending toward high-memory workstations rather than eliminate infrastructure expenses.
Networking adds another consideration. The proposed Equinix Inference Exchange takes a different approach to connecting AI compute resources, illustrating how infrastructure design shapes deployment options.
Apple's results establish measurable generation-speed gains, but not production reliability or a cost advantage over dedicated GPU infrastructure. Independent testing under concurrent workloads, alongside total-cost comparisons, would give enterprises a firmer basis for deciding how many machines to deploy.
Want to learn more AI tips, tricks, and prompting techniques? Let us teach you How to Talk to AI for free! Try our six-minute course at The Neuron Academy, our practical learning platform designed to help professionals use AI more confidently at work.
Learn a few simple ways to write better prompts and get more useful results from AI, or browse our other AI course for free for seven days. eWeek readers get free 7-day access. Check out all the lessons here →