Cloudzat OCuLink & eGPU Intelligence
eGPU for Local AI
For local AI, VRAM capacity and software support usually matter more than gaming frame rates. An eGPU can add CUDA or other accelerator support to a mini PC or laptop, but the model, runtime and connection must be planned together.
Plan an eGPU for local AI
The calculator prioritizes model fit, runtime support and complete system cost.
Fit the model in VRAM first
When model weights and runtime overhead fit entirely in GPU memory, inference avoids much of the penalty associated with CPU offload. A slower 24GB card can therefore be more useful than a faster 12GB card for a larger model.
CUDA remains the broadest compatibility path
Many local AI applications prioritize NVIDIA CUDA. AMD and Intel support continues to improve, but the exact runtime, operating system and model format must be checked before purchase.
Connection affects loading and offload
Initial model loading, CPU offload and data-heavy workflows can expose eGPU link limits. Once a text model is resident in VRAM, generation may be less sensitive than gaming to the external connection.
Compare the complete machine
Add the host computer, GPU, dock, power supply, storage and electricity cost. A low-cost used GPU can still produce an expensive or awkward system if the host and enclosure need major upgrades.
Current eGPU accessory catalogue
Current docks, adapters, cables and power supplies with Amazon images and prices.
| Product | Type | Interface | Power | Condition | Price | |
|---|---|---|---|---|---|---|
 ADTLINKADT-Link PCIe 5.0/4.0 Oculink SFF-8611 to PCI Express x16 eGPU Adapter for GPD Win Max2 External Graphics Card (K993G-BK7-30cm) |
Egpu Dock | OCuLink x4 128Gbps class |
500W | New | $77.80 | View on Amazon |
ADTLINKADT-Link PCIe 5.0/4.0 Oculink SFF-8611 to PCI Express x16 eGPU Adapter for GPD Win Max2 External Graphics Card (K993G-BK7-50cm) |
Egpu Dock | OCuLink x4 128Gbps class |
500W | New | $79.80 | View on Amazon |
 TREBLEETTREBLEET 80Gbps eGPU Enclosure, Compatible with Thunderbolt 5 & USB4, External GPU Dock with JHL9480 Controller, Supports PD 140W Charging, Daisy Chain, DC/SFX/ATX Power Support, for Windows & Mac |
Egpu Dock | Thunderbolt 5 80Gbps class |
Separate/unknown | New | $199.99 | View on Amazon |
 RazerRazer Core X V2 External Graphics Enclosure (eGPU) |
Egpu Dock | Thunderbolt 5 80Gbps class |
Separate/unknown | New | $328.99 | View on Amazon |
RazerRazer Core X V2 - External Desktop Graphics Case for Thunderbolt Laptops - Next-Gen Thunderbolt 5 Performance - Vented Steel Housing - Supports PCIe Graphics Cards with up to 3.5 Slots | Black |
Egpu Dock | Thunderbolt 5 80Gbps class |
Separate/unknown | New | $530.61 | View on Amazon |
 PlugablePlugable Thunderbolt 5 AI eGPU Enclosure & Dock: 80Gbps, 850W PSU Included, TAA Compliant for Windows 11, NVIDIA and AMD Cards (96W Power Delivery, 2.5Gbps Ethernet, 850W PSU, USB Hub) TBT5-AI |
Egpu Dock | Thunderbolt 5 80Gbps class |
850W | New | $629.99 | View on Amazon |
 JMTJMT ADT-R29A PCIe4.0 x4 to SFF-8611/8612 OCuLink Adapter Card 64Gbps with Half and Full Height Baffle Support Chassis External |
Egpu Dock | OCuLink x4 64Gbps class |
Separate/unknown | New | $16.98 | View on Amazon |
 OwlTreePCIe 4.0 x4 64Gbps Compatible with OCuLink SFF-8612 SFF-8611 External GPU eGPU Riser Card for Mini PC |
Egpu Dock | OCuLink x4 64Gbps class |
Separate/unknown | New | $18.99 | View on Amazon |
Frequently asked questions
Can Ollama use an eGPU?
It can use a supported GPU exposed correctly by the operating system and driver stack. Verify the current Ollama platform and GPU support for the exact setup.
Is Thunderbolt fast enough for local LLM inference?
It can be usable when the model fits in VRAM. Model loading and CPU offload are more connection-sensitive than pure GPU-resident token generation.
Is a used RTX 3090 eGPU a good AI setup?
It can offer 24GB of VRAM at attractive cost, but power draw, physical size, condition, dock compatibility and total system price must be considered.