What is WebGPU — why AI models finally run right inside your browser

Here's a funny thing. The graphics card in your laptop sits mostly idle while you browse. The browser could barely reach it — and only ever to draw.
WebGPU changed that. A page can now ask the graphics card to compute. Not render an explosion — multiply matrices. Which is exactly why neural networks suddenly work inside a tab, with no server, no key, and no "please send us your data".
What WebGPU is, in plain words
WebGPU is a browser API that lets JavaScript hand work to the graphics card.
There used to be a wall between your code and the hardware. The GPU is driven by the operating system through its own native interface: Metal on macOS, Direct3D 12 on Windows, Vulkan on Linux. Three different, incompatible languages. WebGPU is the adapter: you write one piece of code, and the browser translates it for whatever system the page opened on.
WebGPU only works in a secure context — over HTTPS or on localhost. On a plain http site, navigator.gpu simply won't exist. That's not a bug, it's a security rule.
WebGL could draw, WebGPU can compute
WebGL showed up around 2011 and genuinely changed the web: 3D games, maps, visualisations right on a page. But at heart it was a port of OpenGL ES 2.0 — a graphics library.
Anything that wasn't "colour this pixel" took a hack. Want to multiply a pile of numbers? Pack them into a texture, pretend you're drawing a picture, then read the answer back out of the pixels. It worked. It wasn't pretty.
WebGPU was designed from scratch for modern GPUs, and general-purpose computation is a first-class feature in it, not a side effect of drawing. There's also a boring but important detail: OpenGL isn't getting new features — its development stopped. WebGPU is, because it grows alongside Metal, Vulkan and Direct3D.
Why this moved AI into the browser
Here's where it gets interesting. Inference is mostly matrix multiplication: an enormous number of identical simple steps. Exactly what a graphics card exists for — it wins by having thousands of identical hands, not by being clever.
While browsers could only compute on the CPU — through WebAssembly — models on the web stayed a toy. Small, slow, "great, eight seconds to decide that's a cat".
The Transformers.js team, moving the library onto WebGPU, reported speedups of up to 100x over WASM. That's not "a third faster". That's the gap between unusable and running live on a video feed.
What became possible inside a tab:
- translation and speech recognition, with the audio never leaving the device;
- semantic search — embeddings computed on the user's machine;
- small language models answering without a single network request;
- processing camera frames in real time.
And notice the side effect. There is physically nowhere for the data to leak: voice, photos and text stay on the device. That's not a privacy promise, that's architecture.
Where it works today, and where it doesn't
Time to be honest: WebGPU isn't everywhere, and that's its main weakness.
- Chrome and Edge — since version 113 (2023). The smoothest support.
- Chrome on Android — since 121.
- Safari — since 26, including iPhone and iPad.
- Firefox — since 141, but on Windows only. Off by default on macOS and Linux, absent on Android.
- Chrome on Linux — only since 144, and only on Intel Gen12 GPUs or newer.
Global coverage per caniuse is around 87%. That's a lot. But "around 87" means roughly one in eight visitors arrives without WebGPU. So there's exactly one working move — check, then fall back:
if (!navigator.gpu) {
// no WebGPU at all — compute on the CPU
}
const adapter = await navigator.gpu.requestAdapter();
if (!adapter) {
// the API exists, but the browser handed over no GPU
}
Two checks, not one: navigator.gpu can exist while requestAdapter() still returns null — the browser decides whether to give you an adapter. Skip the second step and your page breaks precisely for the people whose setup "looks modern". Full breakdown in why WebGPU isn't available.
What this means for you
The practical takeaway is simple. If you're building an AI feature, you now have a third option beyond "cloud API" and "a local model on your own machine": a model on the user's machine, inside an ordinary tab. They install nothing, no token bill accrues, no data travels.
The price is that you don't control the hardware. Someone shows up on a gaming PC, someone on an ageing office laptop, someone in Firefox on a Mac where WebGPU is off. So a CPU fallback isn't politeness — it's the condition for the feature working at all.
Do I need to install WebGPU?
No. It's part of the browser itself — no plugins or drivers to install. You need a recent browser, a supported combination of operating system and graphics card, and a site served over HTTPS.
Is it just WebGL 3?
No, and the difference is fundamental. WebGL descends from OpenGL and is built around drawing. WebGPU was written from scratch for modern GPU interfaces, and computation in it is equal to graphics. The only thing they share is the idea of giving a page access to the hardware.
Do I need a powerful gaming machine?
No. WebGPU runs on integrated graphics — the kind that shares a chip with the processor. A larger model will be noticeably slower there, but "compute on the GPU" is available on a normal laptop with no dedicated card.
The AI Architect 2027 course: step by step from computer basics to your own app and first customers. 543 short lessons on your phone.





