Is an eGPU worth it for AI on a used laptop?

Yes if you work mostly at a desk and want more than 16 GB VRAM. A used enclosure (about £150–£220) plus a used desktop RTX 3090 with 24 GB turns any Thunderbolt laptop into a 24 GB CUDA machine for roughly the price of a 16 GB laptop.

How much performance does Thunderbolt bandwidth cost you?

LLM inference loses almost nothing, about 0–10%, because model weights load into VRAM once. Image generation costs around 5–15% and streaming-heavy training up to 30%. Thunderbolt 3/4 gives roughly PCIe 3.0 ×4, about 3 GB/s usable.

Which laptops can use an eGPU?

You need Thunderbolt 3, Thunderbolt 4 or USB4 with PCIe tunnelling. Most business laptops we review qualify; most AMD gaming laptops of this era do not — check the exact port specification before buying.

Should I put an AMD or NVIDIA card in the enclosure?

NVIDIA only. CUDA is the ecosystem default, and AMD eGPU support for ROCm on Windows is effectively a science project. Buy a used desktop RTX card and test it like the ex-miner many of them are.

Is an eGPU setup portable?

No — the enclosure is a roughly 5 kg desk fixture. It is a desk setup, not a mobile one. If you need 16 GB or more VRAM on the move, a 16 GB laptop is the right tool instead.

eGPU for AI: Adding Desktop CUDA Power to a Used Laptop (2026)

There’s a workaround to the laptop VRAM ceiling that this site keeps bumping into: plug a used desktop GPU into your laptop over Thunderbolt. Done right, a £300 office ThinkPad plus a used enclosure and a used RTX 3090 gives you 24 GB of CUDA VRAM for less than the price of a 16 GB laptop. This guide covers when that maths works, what the bandwidth bottleneck really costs you, and the traps.

The value proposition in numbers

Approximate UK used prices (July 2026 — verify current listings before buying):

Part	Typical used price
Thunderbolt enclosure (Razer Core X / Sonnet Breakaway 550)	£130–£220
Desktop RTX 3060 12 GB	£180–£250
Desktop RTX 3090 24 GB	£550–£720
Laptop + Core X + 3090 total (with a ~£300 used ThinkPad)	~£1,000–£1,250

Compare: a ThinkPad P15 Gen 2 at £900–£1,300 tops out at 16 GB VRAM. The eGPU route lands in the same money with 50% more VRAM and roughly 2× the raw compute — the desktop 3090 outclasses every mobile GPU we’ve reviewed. That’s the pitch. Now the fine print.

What Thunderbolt bandwidth actually costs you

Thunderbolt 3/4 gives the GPU roughly PCIe 3.0 ×4 — about 3 GB/s usable, versus 15+ GB/s for a desktop slot. What that means per workload:

LLM inference: nearly free. Model weights load once into VRAM; token generation barely touches the link. Expect ~0–10% below desktop performance. A 3090 over Thunderbolt still generates 70B-class Q4 models (offloaded) and runs 13B/32B Q4 entirely in VRAM — beyond any laptop.
Image generation: small tax. SDXL/FLUX round-trips are modest; budget ~5–15% off desktop speed. Still several times faster than a mobile 3080.
Training/fine-tuning: workload-dependent. Compute-heavy QLoRA on data already in VRAM is fine; pipelines that stream large batches from disk/RAM feel the link. Budget 10–30% off.
The one real trap: display loopback. If you render on the eGPU but display on the laptop’s internal screen, frames travel back over the same cable. For AI work this is irrelevant (no frames to ship) — but plug your monitor into the eGPU if you also game.

What your laptop needs

Non-negotiable: Thunderbolt 3, Thunderbolt 4 or USB4 with PCIe tunnelling. From the machines we’ve reviewed:

✅ Dell Latitude 5540 — TB4; the classic cheap host
✅ ThinkPad T14 Gen 3 (AMD) — USB4; verified eGPU-capable, an excellent host
✅ ThinkPad X1 Extreme Gen 4, XPS 15 9520 — TB4; premium hosts whose weak internal GPUs an eGPU neatly bypasses
⚠️ Gaming AMD machines (Legion 5 Gen 6/Gen 7) — generally no TB/USB4 on these generations; check the exact port spec before assuming

Also sensible: 32 GB RAM for big-model offloading, and NVMe space for model files. CPU matters less than you’d think for inference hosting.

Stick to NVIDIA cards in the enclosure: CUDA is the ecosystem default, and AMD eGPU support for ROCm on Windows is effectively a science project (see CUDA vs ROCm).

Setup notes (Windows)

The happy path is short: connect enclosure → Windows detects → install standard NVIDIA drivers → nvidia-smi shows the card → point Ollama/ComfyUI at it. Things that go wrong: BIOS Thunderbolt security set to block PCIe devices (fix in BIOS), hybrid-graphics confusion (set apps to prefer the NVIDIA GPU), and hot-unplug during a job (don’t — treat it as docked). On Linux, eGPU hotplug is grumpier; boot with it connected. WSL2 sees the card fine once Windows does.

Buying used enclosures: the PSU is the wear item — a Core X’s 650 W supply comfortably feeds a 3090’s 350 W. Check the Thunderbolt cable is the original (cheap replacement cables cause silent link downgrades). Used 3090s deserve the same scrutiny as any used GPU: demand load-test screenshots, and run a VRAM test on arrival — 24 GB of GDDR6X ran hot in mining rigs.

Who should and shouldn’t do this

Do it if: you already own a Thunderbolt laptop, you work mostly at a desk, and your ambitions (24 GB VRAM, 32B models, serious SDXL/FLUX throughput, 13B fine-tuning) exceed what any used laptop offers. It’s also the only sensible AI upgrade path for premium ultrabook owners.

Skip it if: you need your compute on the move (the enclosure is a 5 kg desk fixture), your budget stops at one purchase (a Legion 5 Gen 7 is simpler at £700–£950), or your laptop lacks Thunderbolt — the adapter workarounds (M.2/OCuLink) work but cost you the laptop’s bottom cover and your warranty appetite.

Summary

A used enclosure (~~£150–£220) plus a used desktop RTX 3090 (~~£550–£720) turns any Thunderbolt laptop into a 24 GB CUDA machine — more VRAM than any laptop we review, at 16 GB-laptop money.
LLM inference loses almost nothing to Thunderbolt bandwidth; image generation ~5–15%; streaming-heavy training up to ~30%.
Your laptop must have TB3/TB4/USB4 with PCIe tunnelling — most business machines we review qualify, most AMD gaming laptops of this era don’t.
Use NVIDIA cards only, plug your monitor into the eGPU when displaying, and test used 3090s like the ex-miners many of them are.
It’s a desk setup, not a portable one — if you need 16 GB+ on the move, that’s what the P15 Gen 2 and Legion 7 are for.