Four Mini PCs Combine Into One? Lenovo’s ThinkCentre X Ultra Doesn’t Say How

2026年9月8日

I run local LLMs on a mini PC with 128GB of unified memory, the GMKtec EVO-X2 (hereafter, the EVO-X2). Large models fit; whatever does not fit in 128GB simply does not run.

On 3 September 2026, Lenovo announced a 1.6-liter mini desktop, the ThinkCentre X Ultra. It carries the same 128GB, and Lenovo says connecting four of them lets you treat them as one unit. Which port do those four machines actually connect through?

This is a survey as of September 2026.

The Lenovo ThinkCentre X Ultra’s Specifications

These are the figures from Lenovo’s press release on 3 September. For a mini PC, the configuration is on the larger side.

It is the same kind of small box as my EVO-X2 — 128GB of unified memory — but it differs in having 10GbE wired LAN and two Thunderbolt 4 ports. When I attached an external SSD to the EVO-X2, the link came up at 10Gbps (measured on 3 September 2026). The unit also has USB4 ports, but I have not tried using them to connect one machine to another.

What I Looked Into

I do not have the physical unit. I checked the following four sources to see how much could be learned about the way it connects.

• Lenovo’s press release from 3 September (primary source for specs, price, and timing)
• Lenovo’s PSREF specification database and the product page on Lenovo’s Japan site (whether it is listed)
• Reports on two machines that have already done the same thing (NVIDIA’s DGX Spark, and a mini PC built on a similar processor)
• The ports available for linking machines together on my own hardware (the only part I actually measured)

512GB Across Four Units — Is the Connection Method Public?

The press release covers this feature in a single line: “a cluster-capable configuration that connects up to four units as one." A cluster, here, means a setup where multiple computers are connected and used as if they were one machine. With four units, that works out to 128GB × 4 — 512GB of memory.

But which port is used, and how many Gbps it delivers when it is, does not appear anywhere in the materials. PSREF, Lenovo’s own specification database, has no page for this product yet, and as of 5 September 2026 the product page on Lenovo’s Japan site was not yet live either (I checked where the product page URL actually leads).

Two Kinds of Fast Port on the Unit Itself

Since this has not been disclosed, the only option is to narrow down the candidates from the ports the unit actually has.

Thunderbolt is best known as a port for displays and external SSDs, but it can also be used to connect devices directly to each other as a network. With two ports, each unit can link hands with the one on either side, connecting several units without buying a switch.

What Do the Prior Examples Connect Through?

Two machines have already done the same thing. In both cases, reports say the rated speed does not simply show up as-is.

The Case of the NVIDIA DGX Spark

The DGX Spark ships with a dedicated port for linking two units (ConnectX-7, 200GbE) built in from the start. A third-party benchmark article that connected two DGX Sparks reports that the numbers fall as you change which layer you measure.

The article explains that data held by the GPU is first copied into system memory before it is handed to the network card, and that copying step gets in the way. Even with a dedicated port on board, the speed actually usable for an AI computation came in at less than half the rated figure.

The Case of a Ryzen AI Max+ 395 Machine

For a mini PC built on the same family of processor as the X Ultra, there are reports of a user building their own cluster. What is used there is a direct USB4/Thunderbolt connection. The port is rated at 40Gbit/s, but with a Thunderbolt 3 cable, what actually comes out is written as about 9Gbit/s. Compared with the DGX Spark’s effective figure above (9–10 GB/s, roughly 80Gbit/s), that is an order of magnitude lower.

There is a limit on the number of units, too. Linux’s thunderbolt-net has no mechanism for relaying traffic, so with two ports the most you can do is line up three units in series. Bonding the two ports into one wider link is described as waiting on a kernel patch to land. The software used is llama.cpp’s RPC: you run rpc-server on one machine and point llama-server on the other at its address.

Given that the X Ultra advertises four units, how it gets past this limit looks like something we will have to wait for Lenovo to explain.

Where Does Inter-Node Speed Actually Matter?

This matters more than the connection method itself. When you split a model across multiple machines, there are two ways to divide it, and they need completely different amounts of traffic.

For local LLMs run at home, ① is what gets used in most cases. llama.cpp’s RPC also splits by layer and hands off to the next machine. Under ①, what flows between machines is a small value per token, so either 10GbE or Thunderbolt might well be enough. What is more likely to matter is the latency that occurs on every round trip.

Choose ② and the picture changes — the port’s speed shows up directly in the result. Which approach the X Ultra’s four-unit setup assumes has not been disclosed at this point.

Which Port Can I Actually Use on My Own Hardware?

Before getting to four units, I checked whether I could even connect two of my own machines. Here is what I found.

I do not have 10GbE on hand. Connecting two machines over 1Gbps wired LAN is one-tenth of the X Ultra’s 10GbE, and a few dozenths of the DGX Spark’s dedicated port. Split ①’s way, it looks likely to work; under ②, latency would likely run longer than the computation itself.

The word cluster sounds like it is about the number of units, but what comes first is the port that connects them. That only became clear once I looked at my own hardware.

What Actually Increases at 512GB?

What a 512GB configuration across four units reliably increases is the size of model that fits. A model that did not fit in 128GB now does.

Speed is a different matter — it is not guaranteed to go up. Under split ①, while one token is being processed, the machines not currently handling it wait their turn. Adding more units just lengthens the chain each token has to travel. Using it to serve several requests at once is where more units pay off; for one person using it alone at a desk, that benefit looks unlikely to show up.

On cost, this means assembling four units starting at €3,100 each. On top of that, a 10GbE switch and cabling get added. For a comparable scale of use, there is also the option of a Mac Studio with 512GB of unified memory in a single unit, which needs no inter-machine connection at all. The difference looks like this: the X Ultra is something you grow by adding units, while the Mac Studio is complete as one unit from the start.

Who It Suits, Who It Doesn’t

What This Article Does Not Confirm

• I have not touched an actual ThinkCentre X Ultra unit. It ships in November
• The actual procedure for connecting four units is unknown, since Lenovo has not published it. Whether it’s 10GbE or Thunderbolt 4 is a guess based on the ports on the unit
• I do not have 10GbE on hand, and I have not used the EVO-X2’s USB4 ports for machine-to-machine links, so I could not test what tok/s would result from actually connecting machines

In Summary: The Port That Ties Four Units Together Is Still Unknown

• The Lenovo ThinkCentre X Ultra was announced with 128GB of unified memory in a 1.6-liter case, and Lenovo says up to four units can be treated as one (announced 3 September 2026; ships November; from €3,100)
• Which port is used to connect them has not been disclosed. The fast ports on the unit are one 10GbE and two Thunderbolt 4, and those two are the only candidates
• In a prior example, even the DGX Spark, which carries a dedicated 200GbE port, has a measured report showing the traffic actually usable for an AI computation falls to about 10GB/s
• A cluster built from a similar mini PC, over a direct USB4/Thunderbolt link, gets an effective 9Gbps; one report also notes it can’t relay traffic and tops out at three units
• That said, if a model is split by layer, the amount crossing between machines is small, and latency per round trip appears to matter more than bandwidth
• What I have actually verified is 10Gbps on the EVO-X2 port used for an external SSD, and 1Gbps wired LAN on the desktop PC [measured]. Before talking about four units, I have never connected even two

The idea that adding units lets you fit more is easy to follow, but what grows is capacity, and speed is decided by something else entirely. If the first ceiling you hit running a large model at home is memory capacity, and the next one is the bandwidth connecting your machines together, then choosing hardware for a local LLM may not be something you can settle by asking “how many GB" alone.

For a look at hardware that can carry 512GB in a single unit, here is a related article.

Sources

• Lenovo press release, “Lenovo Advances Hybrid AI for Business with New Think-Family Devices, Displays, and Security Solutions" (3 September 2026; source for specs and price)
  https://news.lenovo.com/pressroom/press-releases/hybrid-ai-for-business-devices-displays-solutions/
• Multimodal Flow, “Clustering Two NVIDIA DGX Spark Systems: 200GbE Looks Fast, but NCCL Tells the Real Story" (source for the two-DGX-Spark measurements)
  https://multimodalflow.net/en/blog/dgx-spark-dual-node-nccl-rdma/
• Strix Halo Wiki, “Clustering" (source for the effective USB4/Thunderbolt figures and the unit-count limit)
  https://strixhalo.wiki/AI/Clustering

Hardware Featured in This Article

This is the 128GB mini PC I own, used here as the point of comparison.