Clustering several Mac Studios for internal AI (AI Cluster): RDMA over Thunderbolt 5, MLX and exo

When an open AI model is too large for one machine’s memory, there are two paths: rent data-centre GPUs, or link several smaller machines into a cluster. Since macOS 26.2, Apple supports RDMA over Thunderbolt 5 between Macs, and since August 2026 it has promoted clustering directly on the Mac Studio with M5 Ultra. This article explains how a Mac Studio cluster works, how many cables and ports it needs, which software runs on it (distributed MLX, exo), what public test numbers show, the price of Namtech’s AI Cluster 2 and AI Cluster 4 packages — and, just as important, when a business does not need a cluster. All technical information is sourced and was checked on 24 Sep 2026.

What is RDMA over Thunderbolt 5?

RDMA (remote direct memory access) lets one machine read or write another machine’s memory directly, without going through the usual TCP/IP network stack, which cuts latency sharply. According to the MLX distributed documentation, starting from macOS 26.2 RDMA over Thunderbolt is available and enables low-latency communication between Macs with Thunderbolt 5; MLX provides the JACCL backend that uses it to reach communication latency an order of magnitude lower than the ring backend.

Why does latency matter for AI? When a model is split across machines, every generated token requires the machines to exchange intermediate results. If each exchange is slow, the whole cluster waits, and adding machines can even make things slower. The MLX documentation states that JACCL is necessary for techniques such as tensor parallelism — splitting each layer of the model across machines so they compute the same token together. Tech blogger Jeff Geerling, who tested a Mac Studio cluster loaned by Apple, writes in his test write-up that RDMA lowers memory access latency from about 300 microseconds to under 50 microseconds.

The ring backend still has a role. According to MLX, ring does not depend on any third-party library, so it is always available; it uses TCP sockets and is usually faster than MPI. Nodes are connected in a ring, so each one only talks to its two neighbours. Its main purpose is to use Thunderbolt rings for more bandwidth than Ethernet. In short: JACCL is the fast path once RDMA is enabled and the cabling is complete; ring is the fallback when RDMA is not enabled or the topology does not qualify.

Apple has made this direction official. In its August 2026 announcement of the Mac Studio with M5 Max and M5 Ultra, the company said users and teams can cluster multiple Mac Studios over Thunderbolt 5 and RDMA to create a large shared memory pool, and that four machines deliver up to 3x faster AI inference than one — as MacRumors reported on 26 Aug 2026. A footnote on Apple’s Mac Studio page says the test used preproduction M5 Ultra systems with a 36-core CPU, 80-core GPU and 512GB, measuring the time to summarise a 32K-token prompt with a 72-billion-parameter dense model via mlx-lm, comparing one system with a fully meshed 4-node cluster. These are the vendor’s numbers under the vendor’s conditions, not a guarantee for every workload.

How many Thunderbolt 5 cables and ports?

Every pair of machines in the cluster needs its own direct Thunderbolt 5 cable, so the cable count grows quickly: n machines need n(n−1)/2 cables and each machine uses n−1 ports. The reason is that MLX states the JACCL backend only supports fully connected topologies — there must be a Thunderbolt cable connecting every pair of Macs directly. exo sets the same condition: devices that want to join an RDMA cluster must be connected to all other devices, and the cables must support Thunderbolt 5.

There is no shortcut through a switch. Jeff Geerling writes that, as far as he can tell, Thunderbolt 5 switches do not exist, so you cannot plug several Macs into one central switch — every Mac has to be connected to every other Mac. He also notes that as of late 2025 he had only seen RDMA clusters of up to four Macs, although Apple said all five Thunderbolt 5 ports on the M3 Ultra Mac Studio are RDMA-enabled. The exo README adds a notable caveat: on a Mac Studio, the Thunderbolt 5 port next to the Ethernet port cannot be used for RDMA.

*(derived): total Thunderbolt 5 ports per Apple’s specifications (M5 Max: 4 rear; M5 Ultra: 4 rear + 2 front) minus the port next to Ethernet, which the exo README says cannot be used for RDMA on a Mac Studio. That caveat was written for the previous Mac Studio generation; Namtech will re-confirm it on real M5 machines when building a cluster.

The table makes a practical point: a 2-machine cluster has almost no cabling barrier, while a 4-machine cluster already means 6 criss-crossing cables and 3 ports per machine. Jeff Geerling also observes that Thunderbolt plugs do not hold as firmly as the QSFP ports on a DGX Spark, and he doubts the mess of cross-connected plugs would hold up for long in many environments. A cluster should therefore sit in a fixed location, on a shelf or in a tidy cabinet, rather than on a shared desk.

Beyond cables, the documentation lists a few operating conditions. According to MLX, even though TCP/IP is not used with Thunderbolt RDMA, the Thunderbolt Bridge must still be disabled and an isolated local network set up for each Thunderbolt connection; the mlx.distributed_config tool can discover the cabling, check for a valid mesh and generate the configuration commands. According to exo, RDMA ports may fail to discover each other on different macOS versions, so every machine must run exactly the same macOS version, down to beta numbers.

Enabling RDMA: on-site only

This is a big difference from ordinary servers: RDMA over Thunderbolt cannot be enabled remotely. The MLX documentation says that until the feature matures, enabling it cannot be done remotely even with sudo; you have to boot into macOS Recovery, open Terminal from the Utilities menu, run rdma_ctl enable and reboot. You can then verify with ibv_devices. For a 4-machine cluster that means four hands-on procedures, one per machine — a technician has to be there.

Software: distributed MLX and exo

There are two main ways to run a model on a Mac cluster today: use the distributed layer built into MLX directly, or use exo — an open-source layer built on MLX that handles device discovery and model splitting for you. Both share the same foundation: the exo README states that exo uses MLX as its inference backend and MLX distributed for communication between machines.

Distributed MLX suits teams that want fine-grained control: list the machines in a JSON hostfile, choose a backend (ring, JACCL, MPI and others) and launch with mlx.launch. The MLX documentation covers both data parallelism and tensor parallelism. It is flexible, but the operator needs to understand networking and distributed configuration.

exo aims for a simpler experience. According to its README, exo automatically discovers devices running exo with no manual configuration, picks how to split the model based on a real-time view of the cluster topology (each device’s resources and each link’s latency and bandwidth), supports tensor parallelism with up to 1.8x speed-up on 2 devices and 3.2x on 4, and is compatible with common APIs including OpenAI Chat Completions, Claude Messages, OpenAI Responses and Ollama. exo also says it shipped day-0 support for RDMA over Thunderbolt 5, cutting latency between devices by 99%, and its macOS app requires macOS Tahoe 26.2 or later. Jeff Geerling notes that exo 1.0 launched alongside his test (18 Dec 2025), with RDMA for Thunderbolt 5 Mac clusters as its headline feature, released under the Apache 2.0 licence.

One detail shows Apple itself has acknowledged exo: on Apple’s Mac Studio page (checked 24 Sep 2026), the AI apps section invites users to use exo to cluster multiple AI workstations.

What does a third-party 4-machine test show?

The most detailed public test available is Jeff Geerling’s, on 4 Mac Studio M3 Ultra machines — an older chip generation than M5 Ultra — with 1.5 TB of unified memory in total, linked over RDMA/Thunderbolt 5, with hardware loaned by Apple. The figures below are third-party numbers measured on the hardware and software of late 2025; they show a trend, not the speed of an M5 Ultra cluster.

According to Jeff Geerling’s article, exo speeds up as nodes are added and reaches 32 tokens per second with Qwen3 235B on the full cluster; with Kimi K2 Thinking — a 1-trillion-parameter model with about 32 billion parameters active at a time — the cluster still reaches around 30 tokens per second. He also tested DeepSeek V3.1 with 671 billion parameters. Notably, llama.cpp spreads model layers across nodes over RPC, so its performance drops as nodes are added, whereas exo — at the time the only tool he used that supported RDMA — gets faster with more machines.

The other side is documented too. He used prerelease software, many bugs were fixed during testing, and he found RDMA over Thunderbolt still new: when it works it works great, and when it does not, the whole cluster has to be shut down and rebooted. Running HPL over Thunderbolt (TCP, not RDMA) made machines crash and reboot. Managing a Mac cluster has its own friction, such as system upgrades that cannot be run over SSH and require clicking through the UI. He concludes that the bigger question is whether you need a full cluster at all, because a single Mac Studio is already very capable.

Namtech’s AI Cluster 2 and AI Cluster 4 packages

Namtech packages two cluster configurations on the same base machine: the Mac Studio M5 Ultra 30-core CPU/64-core GPU, 256GB. AI Cluster 2 has 2 machines with 512GB of unified memory in total; AI Cluster 4 has 4 machines with 1TB in total. The hardware prices below are Namtech’s selling prices on 24 Sep 2026, VAT included; software, deployment and maintenance are quoted separately for each project’s scope.

The model-size column is an estimate for choosing a configuration, using the same formula as our article on the real cost of running an LLM on-premise: usable memory ≈ 75% of RAM, a 4-bit model ≈ 0.5 bytes per parameter, with 20% reserved for context — applied to the cluster’s total RAM. It is a theoretical capacity ceiling that does not subtract the overhead of splitting a model across machines, and it is not a speed measurement.

*Namtech estimate, not a benchmark: usable memory ≈ RAM × 0.75 (the share macOS gives the GPU by default) per machine, a 4-bit quantised model ≈ 0.5 bytes per parameter, 20% reserved for context; distribution overhead not included. Corresponding list prices (before discount): 192,000,000₫ · 324,000,000₫ · 648,000,000₫ · 1,296,000,000₫. The machines are sold and VAT-invoiced by Decorp; Namtech advises on the configuration and installs the software.

The most notable jump in the table is not from a 2-machine to a 4-machine cluster, but from the 96GB machine to the 256GB machine: capacity nearly triples while it remains one machine, with no cables, no RDMA to enable and no macOS versions to keep in sync. So if you are weighing a cluster of two 96GB machines, compare it first with one 256GB machine — see both configurations side by side on the Mac comparison page.

On power, Namtech’s cost article cites Apple’s figure: a Mac Studio M5 Ultra draws up to about 385W for the whole machine. It follows that a 4-machine cluster running flat out has a ceiling of roughly 1.54 kW (derived: 4 × 385W); plan the power supply and a UPS for the whole cluster when choosing its location. Apple has also announced that a 512GB M5 Ultra will launch in late October 2026, with no price yet; once available, one 512GB machine may replace a 2-machine 256GB cluster for many needs.

When do you NOT need a cluster?

Most businesses do not need a cluster. A single Mac Studio M5 Ultra with 96GB — estimated at up to about 115 billion parameters at 4-bit — already runs open models such as Qwen, SEA-LION or Gemma for internal assistants, document Q&A and RAG, through MLX, Ollama or LM Studio. The 256GB machine extends that to about 307 billion parameters, still in one box. A cluster only pays off when the model you must use is larger than that.

Weigh three limitations before deciding. First, cabling: a fully connected cluster’s cable count grows quadratically — 4 machines already need 6 cables, with no switch to consolidate them. Second, on-site setup: RDMA can only be enabled in Recovery on each machine, and every machine must run the same macOS version, so every reinstall or upgrade needs a technician on site and a coordinated update plan. Third, distribution overhead: the machines exchange data constantly, so 4 machines are not 4x faster — exo reports up to 3.2x and Apple states up to 3x in its own testing.

Remember that what a Mac cluster solves is memory capacity — loading very large models on-site. If your real need is serving many concurrent users, batch processing or training models, revisit the Apple Silicon vs GPU analysis in our guide to choosing internal AI hardware before choosing a cluster.

How does Namtech build a Mac Studio cluster?

Namtech follows a measure first, commit later principle: for every customer considering an AI Cluster, we identify the model to run and the expected load, build and measure on that model before confirming the configuration, and do not promise any speed figure before measuring. If the results show one 256GB machine is enough, we will recommend one machine.

When a cluster is warranted, the work covers: planning the location and Thunderbolt 5 cabling for a fully connected topology, enabling RDMA on-site on each machine, aligning macOS versions, configuring the network for each link, installing distributed MLX or exo, and exposing an OpenAI-compatible API so internal applications can use the cluster as a single AI server. The machines are sold and VAT-invoiced by Decorp; Namtech advises on the configuration, installs and operates it. See every configuration and price on the Mac Studio page.

Clustering Mac Studios over RDMA/Thunderbolt 5 lets you load models larger than one machine’s memory right in your office, but it costs you extra cables, on-site setup and distribution overhead — so choose a cluster only when the model truly needs it, and measure before you buy.

Frequently asked questions (FAQ)

Note: This article compiles public sources checked on 24 Sep 2026. The M3 Ultra cluster speed figures are third-party numbers (Jeff Geerling, Dec 2025) on an older chip generation and do not represent an M5 Ultra cluster; the “3x” figure is Apple’s own test under its conditions. Model sizes are estimates based on memory capacity, not benchmarks. Machine prices are Namtech selling prices including VAT as of 24 Sep 2026 and change with Apple’s list prices. For reference only; not technical or investment advice.