MarketCatalyst
← Back to Resources
MARKETCATALYST

Apple Is Building Harbor, a Mac App That Clusters Macs Into a Private AI Server

Harbor is in private preview with no release date or pricing. Apple has shown four clustered Mac Studios running AI inference about 3x faster than one.

10 Nov 2026 · 15:06 ET · 4 min read

Apple ($AAPL) is developing Harbor, a Mac app that turns Apple silicon computers into private AI servers, according to details posted by leaker Aaron Perris and reported by 9to5Mac. The app lets users run open-weight AI models on their own Mac, or link several Macs over Thunderbolt 5 to pool their memory, without sending data to the cloud. Harbor is in private preview under Apple's Claris brand, and Apple has not confirmed it.

TB5
Thunderbolt 5 links pool memory across Macs
~3x
Apple-reported speedup, four Mac Studios vs. one
Preview
Private preview; no release date or pricing

What Harbor does

Harbor featureDetail
PurposeTurns a Mac into a private AI inference server; data stays on the device
ModelsBrowse, download and load open-weight models in the app; screenshots show gpt-oss-20b coming from Hugging Face
Local APIA status panel shows server state and an API address to copy, which suggests other apps can use the models
ClusteringTwo or more Thunderbolt 5 Macs pool unified memory; a single Mac also works
SetupHarbor detects connected Macs automatically, with no command line or manual setup
StatusPrivate preview under the Claris brand; no release date or pricing announced

How the clustering works

  • Pooled memory. Connected Macs combine their unified memory into one inference resource. That allows models larger than a single machine can hold and more simultaneous requests.
  • Example screenshot. One screenshot shows a MacBook Pro with an M5 Max and 48 GB, plus two Mac minis with M4 Pro chips and 16 GB and 24 GB, according to iSpazio. Together that is 88 GB of memory.
  • RDMA setting. The screenshots show an RDMA option under Privacy and Security settings. RDMA lets one Mac read another's memory directly over Thunderbolt 5.

The groundwork Apple has laid

Apple added Thunderbolt 5 RDMA support in macOS 26.2, and its MLX framework already supports distributed inference. In August, Apple said four clustered Mac Studios deliver about 3x faster AI inference than one, MacRumors reported. Apple's October research post gave results on Mac Studios with M5 Ultra chips:

Model (Apple, 8,192-token context)One Mac StudioFour Mac StudiosSpeedup
Mistral Medium 3.5 (128B)7.70 tokens/sec23.37 tokens/sec3.0x
Qwen 3.8 (27B)19.69 tokens/sec50.89 tokens/sec2.6x
Muse Spark Open (816B)15.74 tokens/sec*36.73 tokens/sec2.3x

*Muse Spark Open does not fit on one machine, so the baseline is two M5 Ultras. Speedups are calculated from Apple's reported tokens per second and vary by workload.

How Apple traded

Nov. 10 (3:01 p.m. ET)PriceChange
Apple ($AAPL)$348.12-0.73%

What to watch

  • Public launch. New Mobile Life reported a public release is expected in the coming weeks to months, but Apple has not announced a date, pricing or final hardware requirements.
  • Hardware support. Whether Harbor works across older Macs with Thunderbolt 5 or targets newer Mac Studio and Mac mini models.
  • Local AI competition. How Apple's on-device approach compares with cloud offerings such as Google's Gemini 4 Carbon.

This article is informational only and is not investment advice.

Readers also read

MarketCatalyst is not a registered investment advisor and does not manage client assets. Content on this platform is provided for informational and educational purposes only. It is not investment advice, and MarketCatalyst is not a stock-picking or trade-alert service. Trading stocks and options involves risk, including the possible loss of principal. Consider your own goals, time horizon, and risk tolerance, and consult a qualified financial advisor before making any investment decision.