Your Browser Became an AI Computer and Nobody Announced It

There was no launch event for this. Over the last two years the browser quietly stopped being a document viewer and became a runtime capable of executing real machine learning models on your own hardware. The shift is structural rather than novelty — it changes who pays for compute, and who holds your data.

What made it possible

• ✦WebAssembly let compiled C and C++ libraries — ffmpeg, Tesseract, whole codecs — run at near-native speed in a tab.
• ✦WebGPU exposed the graphics card to JavaScript, turning inference from a slideshow into something interactive.
• ✦Small models got genuinely good. A few hundred megabytes now does what needed a server rack in 2020.

The practical threshold in 2026 sits around 3 billion parameters. Below that, client-side inference is viable on ordinary consumer hardware. Above it, browser memory limits bite hard and a server is still the right answer.

The economics quietly flipped

That third row explains a lot about the web you use. Free tiers on upload-based tools have minute caps and watermarks because every conversion costs the operator money. When the compute is yours, that pressure disappears — which is why local-first tools can credibly promise unlimited free use without it being a bait-and-switch.

What this actually runs today

• ✦Speech recognition with Whisper — transcription and video captioning with no upload.
• ✦Image segmentation for background removal that never sees a server.
• ✦Super-resolution models that upscale and clean up photos on your GPU.
• ✦OCR that turns screenshots into selectable text in seconds.

Where it still falls down

It would be dishonest to pitch this as strictly better. Local inference is bounded by whatever machine you happen to be using, so a five-year-old laptop will be slower than a rented GPU — sometimes much slower. First use pays a real download cost while the model arrives. And frontier-scale models simply do not fit; anything needing hundreds of billions of parameters belongs on a server and will for the foreseeable future.

The sensible posture is local-first rather than local-only: do the work on-device when the model is small enough to do it well, and reach for the network deliberately when a task genuinely needs more than your hardware has. That is the framing most serious writeups of the trend have landed on, and it matches what we see building these tools.

Every AI tool on this site follows that rule. Not as a marketing position — because for the jobs these tools do, it is simply the better engineering trade.