Comparisons 10 min read

Thunderbolt eGPU Enclosure 2026: Local AI Buying Guide for Laptops and Mini PCs

eGPU enclosures for local AI on a laptop or mini PC: Sonnet Breakaway Box 750ex, Razer Core X Chroma and AKiTiO Node Titan compared using manufacturer specs, and which mini PC to pick depending on the port (OCuLink, Thunderbolt 4, USB4).

S By Selfhostr · no sponsored content
Thunderbolt eGPU Enclosure 2026: Local AI Buying Guide for Laptops and Mini PCs
This article may contain affiliate links (no extra cost to you, it funds the site). As an Amazon Associate, Selfhostr earns from qualifying purchases. See the disclosure.

👍 What we like

  • Gives a laptop or mini PC the VRAM of a desktop card to run larger LLMs.
  • One cable for the GPU, laptop charging and, depending on the enclosure, extra USB or Ethernet ports.
  • You can swap the card as your needs change.
  • Sonnet clearly states the continuous power available to the card (375 W).

👎 What to watch

  • The Thunderbolt 3 link (PCIe 3.0 x4) slows model loading and hurts models that don't fit in VRAM.
  • Not compatible with Apple silicon Macs.
  • Bulky enclosures (3.2 kg for the Sonnet), and older models whose availability needs checking.
  • The graphics card is sold separately and must fit the enclosure's dimensions.
📑 Contents ▾

An eGPU enclosure lets you connect a desktop graphics card, and therefore its VRAM, to a laptop or a mini PC, provided the machine has a port that carries PCIe: Thunderbolt, OCuLink with a suitable dock, or USB4 if its manufacturer confirms PCIe tunneling. Among the three Thunderbolt 3 enclosures compared here, the Sonnet eGPU Breakaway Box 750ex specifies the continuous power reserved for the card (375 W), the AKiTiO Node Titan states up to 500 W for the card and takes cards up to 32 cm long, and the Razer Core X Chroma, which does not specify that power, accepts 3-slot cards. On a mini PC with an OCuLink port, such as the Minisforum UM890 Pro, an OCuLink dock provides a more direct link than Thunderbolt 3. Apple silicon Macs, for their part, do not support eGPUs.

Every specification quoted comes from the manufacturers’ pages (Sonnet, Razer, OWC/AKiTiO, Minisforum, GEEKOM, ASUS); we have not tested any of these devices.

What an eGPU brings, and what is not one

Running a large language model (LLM) locally almost always runs into video memory. A laptop or mini PC without a dedicated graphics card, or with a mobile GPU that has little VRAM, can only load small models, often heavily quantized. An eGPU enclosure is a chassis with its own power supply and a PCIe slot for a desktop card, connected to the machine by a single cable.

Two neighbouring categories are not eGPU enclosures. A Thunderbolt PCIe expansion chassis can be used to add a 10 GbE network card or a storage card, and its manufacturer sometimes explicitly excludes graphics cards: look for the word “GPU” on the spec sheet. A dock with a built-in GPU, connected over OCuLink or USB4, carries its own graphics chip and follows a different logic, since you do not choose the card.

How to choose an eGPU enclosure

1. Your machine’s port

This is the first thing to check, even before choosing an enclosure. An eGPU needs a port that carries PCIe: Thunderbolt 3, Thunderbolt 4 or Thunderbolt 5, or an OCuLink port with an OCuLink dock (see the mini PC section). A plain USB-C port is not enough. On a USB4 port without Thunderbolt certification, PCIe tunneling is not guaranteed: you have to confirm it with the manufacturer. On Windows, Razer requires an eGFX-compatible Thunderbolt 3 port and Windows 10 RS5 or later; Sonnet states that its enclosure works with Windows PCs equipped with Thunderbolt 5, 4 or 3. On Linux, the Thunderbolt device has to be authorized (for example with bolt) before the card shows up.

The Mac case: Apple states in black and white that using an eGPU requires a Mac with an Intel processor. An Apple silicon MacBook (M1 and later) will not work with any enclosure in this comparison; on those machines, unified memory is what serves local AI. If you are still choosing your laptop, our comparison of laptops for local AI weighs that approach against RTX laptops.

2. The graphics card, and above all its VRAM

The enclosure is only a container. What matters for local AI is the card you put in it, and first of all its amount of VRAM: it determines the model size and the quantization level you can use. The most widespread tools (Ollama, llama.cpp, PyTorch) are mostly optimized for CUDA on NVIDIA. ROCm (AMD) and Intel’s software stacks are progressing, but you need to check compatibility for each tool. Our guide on how much VRAM a local LLM needs helps with sizing, and the comparison of the best GPU for local AI helps you choose the card.

3. Supported card size

Current graphics cards are often thick and long. Sonnet states full-length, double-width cards. OWC specifies for the AKiTiO Node Titan a card of at most 32 cm long, 6 cm wide and 17 cm high, heatsink included. Razer accepts cards up to 3 slots wide. A card that is 3 slots or thicker will therefore not fit in an enclosure limited to double width: always compare the dimensions of the card you want with the enclosure’s spec sheet.

4. Power supply and power reserved for the card

The total power of the supply does not tell the whole story. Sonnet specifies the continuous power available to the card: up to 375 W, from a 750 W supply, with two 8-pin (6+2) power connectors. AKiTiO states up to 500 W for the card from a 650 W supply, with dual 8-pin auxiliary power. Razer states a 700 W ATX supply without specifying the power reserved for the card. Before buying, compare the power draw (TBP) stated by the card manufacturer and its connectors with those of the enclosure.

Thunderbolt 3 is rated at 40 Gb/s (5000 MB/s according to the AKiTiO spec sheet), but the card only sees a PCIe 3.0 x4 link: the Sonnet spec sheet describes a slot that is x16 mechanical, x4 electrical, and the AKiTiO Node Titan manual says the same. For LLM inference, this limit mostly weighs on model loading and when the model spills over the VRAM, because part of the layers then runs on the machine’s processor and memory, with traffic over the cable. A model that fits entirely in VRAM is much less penalized. We do not give a performance-loss figure: it depends too much on the card, the model and the software.

6. Ports and laptop charging

These enclosures can also act as a docking station. Razer states up to 100 W of charging, four USB-A ports and a Gigabit Ethernet port. Sonnet provides 85 W to the laptop, four 5 Gb/s USB-A ports and a Gigabit RJ45 port. AKiTiO states up to 85 W of charging for laptops that accept a 15 to 85 W source. A laptop that draws more under load will still need its own charger.

The three enclosures compared

CriterionSonnet eGPU Breakaway Box 750exRazer Core X ChromaAKiTiO Node Titan
InterfaceThunderbolt 3 (40 Gb/s), Windows PCs with Thunderbolt 5, 4 or 3Thunderbolt 3Thunderbolt 3 (40 Gb/s)
Link seen by the cardPCIe 3.0 x4 electrical (x16 mechanical)limited by Thunderbolt 3 (PCIe 3.0 x4)PCIe 3.0 x4 electrical (x16 mechanical)
Power supply750 W700 W (ATX)650 W
Stated power for the cardup to 375 W continuousnot specified by Razerup to 500 W
Card sizefull length, double widthup to 3 slots wide32 x 6 x 17 cm maximum, double width
Laptop charging85 Wup to 100 Wup to 85 W
Extra ports4 x USB-A 5 Gb/s, 1 x Gigabit RJ454 x USB-A, 1 x Gigabit Ethernetnone listed (Thunderbolt 3 host port)
macOSIntel Macs only, AMD 6000 series cardsIntel Macs (macOS 10.13.4 or later), compatible AMD Radeon cardsmacOS 10.13.4 or later (Intel Macs only)

Sources: official Sonnet and Razer pages, AKiTiO/OWC spec sheet and manual.

About availability: these three Thunderbolt 3 enclosures are not recent. The Razer Core X Chroma dates from 2019 and the Node Titan from 2020, and some retailers list them as sold out or end of line. Check availability and condition (new or refurbished) at the time of purchase.

Which enclosure for which use

For a documented continuous power limit: Sonnet eGPU Breakaway Box 750ex

Sonnet puts a figure on the continuous power it can deliver to the card: up to 375 W, to compare with your GPU’s power draw (TBP). For a card that needs more, the AKiTiO Node Titan states up to 500 W. The Sonnet accepts a full-length, double-width card, charges the laptop at 85 W and adds USB-A and a Gigabit Ethernet port. Sonnet lists the AMD, NVIDIA and Intel Arc cards supported on Windows. Its limits: 3-slot cards do not fit, and it is bulky (18.5 x 34.0 x 20.2 cm, 3.2 kg); Sonnet rates it for an operating temperature of 0 to 35 °C and gives a 2-year warranty on new units.

For thick cards: Razer Core X Chroma

With cards up to 3 slots wide, a 700 W ATX supply and up to 100 W of charging, it is the logical choice if your card is thick. Razer does not specify the maximum power reserved for the card: check that it matches your GPU’s power draw. It is also the oldest of the three enclosures, so check its availability.

For a long card, with a longer warranty: AKiTiO Node Titan

OWC’s Node Titan accepts a double-width card up to 32 cm long, with a 650 W supply and up to 500 W for the card, and charges the laptop at up to 85 W. OWC backs it with a 3-year warranty. The manual specifies that cards requiring a driver must use a Thunderbolt-compatible version, and points to a GPU compatibility list at AKiTiO: check it before buying the card.

On a mini PC, the port decides which type of eGPU is possible. There are three cases, illustrated by three common models.

Minisforum UM890 ProASUS NUC 14 ProGEEKOM Mini IT13
ProcessorAMD Ryzen 9 8945HS (8 cores / 16 threads), Radeon 780MMeteor Lake (Core Ultra 5 or 7) depending on versionIntel Core Ultra 7 356H (16 cores / 16 threads)
Port for an eGPU1 x OCuLink (2 x USB4, eGPU to be confirmed)2 x Thunderbolt 42 x USB4 (40 Gb/s, eGPU to be confirmed), no OCuLink listed
Storage2 x M.2 2280 PCIe 4.0M.2 2280 and M.2 2242, PCIe 4.0 x4M.2 2280 PCIe 4.0 x4, M.2 2242 SATA, 2.5-inch bay
Network2 x 2.5 GbE, Wi-Fi 6E2.5 GbE2.5 GbE, Wi-Fi 7

OCuLink carries PCIe without going through a Thunderbolt controller. The Minisforum UM890 Pro has one such port, in addition to two USB4 ports. It connects to an OCuLink dock such as the Minisforum DEG1, which Minisforum rates at PCIe 4.0 x4, one generation above the PCIe 3.0 x4 of Thunderbolt 3 enclosures, and compatible up to the RTX 4090 and RX 7900 XTX. The dock is only a mount: check on its spec sheet how the card is powered. With its two DDR5 slots up to 96GB, the UM890 Pro can also keep in system memory the part of a model that spills over the VRAM. Its 16 TOPS Ryzen AI NPU does not replace the card for an LLM.

Thunderbolt 4: ASUS NUC 14 Pro

The ASUS NUC 14 Pro has two Thunderbolt 4 ports. Sonnet explicitly states that its 750ex is compatible with Thunderbolt 4 Windows PCs; for the Razer Core X Chroma and the AKiTiO Node Titan, have ASUS and the enclosure’s manufacturer confirm eGPU support before buying. The NUC itself measures 117 x 112 x 54 mm and weighs 0.60 kg. The Intel AI Boost NPU only exists on the Core Ultra 5 and 7 versions; here again, the eGPU’s card is what runs the model.

USB4: GEEKOM Mini IT13

The GEEKOM Mini IT13 has no OCuLink port listed on its spec sheet, but two 40 Gb/s USB4 ports. As with any USB4 port, have GEEKOM confirm eGPU support (PCIe tunneling) before buying an enclosure. Its appeal lies elsewhere: up to 64GB of DDR5, a PCIe 4.0 x4 M.2 2280 slot, a SATA M.2 2242 slot and a 2.5-inch bay make it a versatile home server.

If local AI has to run around the clock, a tower or a server with the card inside often remains simpler than a laptop or mini PC connected to an eGPU: our tutorial on building a local AI server with Ollama covers that option, and the comparison of mini PCs for local AI, with or without an NPU shows what a mini PC can do on its own.

Key facts

Thunderbolt 3 on all three enclosures
Enclosure interface
750 W (Sonnet), 700 W (Razer), 650 W (AKiTiO)
Power supply
Double-width (Sonnet, AKiTiO), 3 slots (Razer)
Max card size
Not supported (Apple)
Apple silicon Macs

Our picks

Affiliate links · same price for you
Stated continuous power
📦

Sonnet eGPU Breakaway Box 750ex

  • 750 W supply, 375 W continuous for the card
  • Full-length, double-width cards
  • 85 W laptop charging
View on Amazon

Amazon · price and availability on the listing

For 3-slot cards
📦

Razer Core X Chroma

  • 700 W ATX supply
  • Cards up to 3 slots wide
  • Up to 100 W charging
View on Amazon

Amazon · price and availability on the listing

Cards up to 32 cm
📦

AKiTiO Node Titan

  • 650 W supply, up to 500 W for the card
  • Card up to 32 cm long and 6 cm wide
  • 3-year OWC warranty
View on Amazon

Amazon · price and availability on the listing

Mini PC with OCuLink
Minisforum UM890 Pro

Minisforum UM890 Pro

  • Ryzen 9 8945HS, 8 cores / 16 threads
  • OCuLink port and 2 USB4 ports
  • Up to 96GB of DDR5
View on Amazon

Amazon · price and availability on the listing

USB4 mini PC, storage
📦

GEEKOM Mini IT13

  • Intel Core Ultra 7 356H
  • 2 USB4 ports at 40 Gb/s
  • PCIe 4.0 M.2, SATA M.2 and 2.5-inch bay
View on Amazon

Amazon · price and availability on the listing

Thunderbolt 4 mini PC
📦

ASUS NUC 14 Pro

  • 2 Thunderbolt 4 ports
  • 2 PCIe 4.0 M.2 slots
  • 0.60 kg
View on Amazon

Amazon · price and availability on the listing

See all our recommended gear in the shop.

Tags: egputhunderboltlocal aihomelablaptopmini pc

Related

Comparisons

Best Laptop for Local AI and Development in 2026

MacBook Pro 16-inch M5, ASUS ROG Zephyrus G16 or Razer Blade 16: which laptop should you pick to run LLMs locally in 2026? VRAM, unified memory, configurations to aim for, and the options for a developer laptop.

Read
Comparisons

Mini-PCs for Local AI in 2026: With or Without an NPU?

Do you need an NPU for local AI? Three mini-PCs compared on their spec sheets: the Beelink Mini S12 Pro (Intel N100) and EQ14 (Intel N150), with no NPU, and the Minisforum UM890 Pro, whose Ryzen 9 8945HS includes an NPU, with up to 96 GB of DDR5 and an OCuLink port.

Read
Comparisons

Best GPU for local AI in 2026: RTX 3090, RTX 4090, RTX 4060 Ti 16 GB, RTX A4000 or RX 7900 XTX?

Which graphics card for a local LLM, Stable Diffusion or GPU passthrough on Proxmox? VRAM, power draw, size and software stack of the RTX 3090, 4090, 4060 Ti 16 GB, A4000 and RX 7900 XTX, from the manufacturers' spec sheets.

Read