OpenClaw on a GPU server: why the GPU matters for your AI agent

OpenClaw is the open-source AI assistant everyone is installing this year: one Gateway that you run on your own machine, which connects your chat apps to an AI agent able to browse the web, run commands, handle files and media, and use skills. Many people buy a Mac mini just to keep it running. A server does the same job, always on, and away from your personal computer. This article explains what a GPU changes for OpenClaw on a server, using the measurements of our AI agent browser test.

In short

OpenClaw drives a real browser (a managed Chromium profile: tabs, clicks, screenshots, PDFs). Without a GPU, that browser draws pages with the processor: in our test, a 1080p WebGL page ran at 13 images/s with 3.9 of 4 vCPU busy. With the Intel UHD 770: about 260 images/s with 0.6 vCPU.
Websites see a real graphics card: Intel UHD Graphics 770 instead of SwiftShader or no WebGL at all, which is a typical signature of an automated browser.
The CPU stays free for the agent itself: the Gateway, its tools and your skills.
Media and local models can use the GPU too: video conversion with Quick Sync, image and text models with OpenVINO.

We did not benchmark OpenClaw itself: the figures below come from our test of Chrome driven by an AI agent on the same hardware (4 vCPU, Intel UHD 770, Debian 13).

1. What OpenClaw needs on a server

According to the official documentation:

Need	What OpenClaw uses
Runtime	Node 24 or 26, one Gateway process (Linux, macOS, Windows/WSL2, Docker)
Model	a provider API key, or local models (an Ollama server, through its native API)
Browser	a managed browser profile (Chrome, Chromium, Brave or Edge) that the agent controls: tabs, clicks, typing, snapshots, screenshots, PDFs
Access	the Web Control UI listens on `127.0.0.1`: reach it through an SSH tunnel or Tailscale

The browser is the heavy part. Every page the agent opens is drawn, scrolled and screenshotted, often in a loop. That is where the GPU comes in.

2. The browser: 20× faster with the GPU

On a server without a GPU, Chrome falls back to SwiftShader, a graphics card emulated by the processor. Same 1080p WebGL page, headless Chrome, 4 vCPU server:

Bar chart: headless Chrome on a 1080p WebGL page, 13 images per second with 3.9 busy vCPU without GPU, about 260 images per second with 0.6 vCPU on the Intel UHD 770 — Without a GPU, the browser alone saturates the server. With the Intel UHD 770 it renders 20 times faster and leaves the CPU almost idle (measurements from our AI agent browser test).

For an agent, the second line matters as much as the first: when the browser takes 3.9 of 4 vCPU, the Gateway, the model calls and your skills wait. A heavy page (a map, a dashboard, a video) is enough to slow the whole assistant down.

3. What websites see

Sites that fight bots read the graphics card reported by the browser (WEBGL_debug_renderer_info):

Browser setup	Graphics card reported
Server without GPU, default headless	`SwiftShader` (emulated)
Server without GPU, virtual display	no WebGL at all
Server with the Intel UHD 770	`Intel UHD Graphics 770`

SwiftShader or a missing WebGL is a classic sign of an automated browser. A real GPU does not make an agent invisible, and you must respect each site’s terms, but it removes one obvious reason to be blocked or shown endless challenges. The details and the Chrome flags we used are in our AI agent browser article.

4. Media and local models

OpenClaw handles media (images, audio, video) sent from your chat apps. On our servers the same Intel GPU also has:

a video engine (Quick Sync / VA-API) to convert or shrink videos without loading the CPU: see our ffmpeg Quick Sync commands;
OpenVINO support for vision, OCR and small language models on the GPU: see AI inference on an integrated GPU.

For the main model, a hosted API (Anthropic, OpenAI…) remains the most capable choice; a small local model served by Ollama works for simple tasks and keeps your data on your server.

5. Why a server rather than your own computer

Isolation: OpenClaw can run commands and read files. On a dedicated server it only sees what you put there, not your personal machine.
Always on: your assistant answers at night and while your laptop is closed.
Private access: keep the Web Control UI on 127.0.0.1 and reach it through an SSH tunnel or Tailscale, with no port open on the Internet. Restrict who can talk to the agent with OpenClaw’s allowlists.

6. Mac mini or Linux GPU server?

	Linux GPU server (iGPU Start / Power)	Dedicated Mac mini (M4)
Price	from €14.99 / €29.99 per month excl. VAT	from €95 per month
Browser on the GPU	yes, Intel UHD 770	yes, Apple GPU
Apple-only features (iMessage, Apple apps)	no	yes
Best for	an always-on agent that browses, converts media, runs small local models	agents that need macOS

Both are available from us: choose the Mac mini only if your agent needs macOS.

Which plan?

iGPU Start (2 vCPU, 8 GB) is enough for an agent that browses now and then. If it browses all day, handles video or runs a local model next to the Gateway, take iGPU Power (4 vCPU, 14 GB, twice the GPU share of Start): our browser figures were measured on 4 vCPU.

Want it ready to use? Ask by email when you order (contact@linuxgpu.eu): we can deliver the server with OpenClaw’s prerequisites installed and Chrome set up to render on the GPU. You then connect your own model account and chat apps.

Try it

See the plans: iGPU servers from €14.99 per month excl. VAT, dedicated Mac mini from €95 per month.