Last week the free technical preview of Whelk went live: native apps for iPhone, iPad, Mac, and Android, running against models on your own hardware, with a hub you own routing all of it. Here’s the bet behind it.
Intelligence becomes a utility you own
Two curves are crossing, and everything downstream follows from it.
Model quality is a wasting edge. Whatever is frontier today is commodity in eighteen months, open-weight shortly after, running on hardware you own shortly after that. The gap between the best rented model and the best owned one shrinks every quarter, and it only ever shrinks in one direction.
Hardware is sprinting the other way. A machine that costs less than one employee-month of a cloud AI contract now runs models that were state of the art two years ago, in an office closet, silent, on wall power. Next year that machine is cheaper and the models are better. This never reverses either.
When those curves cross for good, intelligence stops being a subscription and becomes a utility you own, like your network, like your storage. Nobody rents wifi from a vendor who keeps a copy of every packet. That era is coming for AI, and it’s closer than the incumbents can afford to admit.
The perimeter comes back
The cloud AI decade trained everyone to accept something no company would have accepted in writing: the most sensitive thinking in the building, typed into a text box that ends on someone else’s servers, under someone else’s retention policy, discoverable by whoever can reach that vendor with an order. Courts already have. The rule against it exists at every serious company, and it loses to a browser tab every single day, because the compliant tools are worse.
That trade only made sense while the intelligence lived exclusively in the cloud. It doesn’t anymore. The future looks like the past used to: a perimeter. Your prompts, your agents, your record of what was said and what it was allowed to do, inside your walls, verifiable at the network layer. Not a privacy policy. A topology. The firms that answer to privilege and regulators get there first because they have no choice, and everyone else follows because there stops being a reason not to.
Whoever owns no models wins this market. Every model release, from anyone, makes an unbundled platform better at a cost of zero, while every inference-bundling vendor prices against their own compute and meters you to protect their margin. The durable asset is not the model. It’s the surface you live in and the record it keeps, and those belong to you.
If this is useful, the next one will land in your inbox.
What has to exist
That future needs a software layer nobody had built: apps good enough to beat the best consumer software ever shipped, pointed at hardware you own instead of a vendor cloud. That’s Whelk, and the strategy is a ladder.
The individual, free forever. Native apps, one surface, no chat mode and agent mode: any conversation can be armed into a working session with live tools and approvals on your lock screen. Work is a capability, not a separate app.
The network. Whelk Hub: a service on a machine you own that gives every device one address and routes your models, your agents, and your connectors. This is where a person becomes a household and a laptop becomes a stack.
The team. Teams arrives at GA, per-seat, with a tenet that never breaks: usage is never metered, on any tier. No message caps, no token meters, no model tiers. Your hardware does the work, so there is no meter to read. Seats are the only thing anyone is ever billed for.
The firm. Whelk Server, in testing: policy, oversight, audit, and deployment on infrastructure the firm runs, for the organizations the cloud AI industry is structurally unable to serve at any price.
Apple platforms and Android today. Windows, Linux, and Terminal coming soon. GA soon.
What you can run today
The preview is live and free. Models three ways: on-device with MLX on Apple Silicon, any OpenAI-compatible endpoint (Ollama, LM Studio, vLLM, llama.cpp), or through Whelk Hub. Over Tailscale, your phone reaches the box in your closet from anywhere.
- iOS/iPadOS/macOS TestFlight: https://testflight.apple.com/join/qGguG8u2
- Android (Play testing): https://play.google.com/apps/testing/app.whelk.Whelk
- Whelk Hub (Mac):
brew tap getwhelk/whelk && brew install --cask whelk - Docs: https://getwhelk.com/docs · Discord: https://discord.gg/J4rRZ3aXf
It works and it isn’t finished, and I’d rather you have it now. Run it against what you actually have and tell me what falls over.
Point it at your models, and tell me how it does!