One address change. Your OpenAI code stays.
The chat is already built. You don’t want a new library or a week of rewriting.
You get Point the client at api.tezgrid.com/v1. Keys, spend limits, and the dashboard are already there.
The computer is already in the room
A gaming PC after midnight. A Mac on the desk. An office machine after six. Someone already paid for it. TezGrid lets an app send a normal chat request to that machine. The owner keeps 99.8%. The new cost is mostly electricity.
Why not just buy the computers? They are already bought, in other people’s houses and offices. TezGrid keeps 0.2%. Owning the fleet would be the expensive way.
Why not only Macs? A Windows GPU, a Linux office box, and a Mac can all take the job. NVIDIA, AMD, Intel, and Apple.
Can the owner read the prompt? The link is encrypted. After it is decrypted on their machine, a person with admin access can. If that is not acceptable, don’t send it.
Chat in, answer out, streaming included. Point it at TezGrid and it works. Want a computer near your users, or a stricter privacy level? Add one header. Otherwise you change nothing else.
TezGrid picks by speed, city, trust, and price. A machine that only looks online is skipped. If an answer cuts off, the next working computer finishes it. Your app still talks to one address.
The model runs on their Mac, their Windows or Linux GPU, or their processor. Every request shows the split: 99.8% to them, 0.2% to TezGrid. Same numbers on both dashboards.
Who it is for
You build apps and you’re tired of the cloud bill. Or you own a computer that does nothing for half the day. Either way, this is the upside.
A support bot in Pune, Lagos, or Jakarta feels slow when the model sits on another continent. The delay is the distance.
You get The same request. If a working computer is closer, that one answers, and the reply names the region.
Customer chats, source code, internal plans. You decide how trusted the computer has to be.
You get Every reply says the trust level that served it. Refuse anything weaker. Read the limits before you send secrets.
Sorting mail, tagging tickets, a first draft. A 1.5B or 7B model is enough. A giant GPU is not.
You get Desktop GPUs, Apple chips, and spare processors already run these. You pay for the small model, not a cloud rack.
The chat sits inside their product. They will ask how the bill stops. You should have that answer in one sentence.
You get One API key per client, a limit you can show them, and no second charge if the same request is retried.
The internet blips. The program restarts. Your user should not be left with half a sentence.
You get TezGrid moves the job to the next working computer. Your app still talks to one address.
Gaming PC after midnight. Mac on the desk. Office machine after work. A student laptop with a real GPU counts. NVIDIA, AMD, Intel, and Apple.
You get Install, sign in, load a model, go online. You keep 99.8% of every job. More people asking nearby means more work.
Leave the router alone. Don’t paste a wallet into a terminal. Sign in with the email you already have.
You get The computer connects out over an encrypted link. The dashboard shows each job and your 99.8%. Bank payout is next. The record of every job is already there.
A founder or a client will ask why it cost this much. You shouldn’t need a diagram of five companies.
You get 99.8% to the person who ran the model. 0.2% to TezGrid. No chip maker, then a cloud, then a reseller, each taking a cut of the same request.
Price
A normal AI API is a chip company, then a cloud, then another company selling tokens. Each one adds a cut. TezGrid sends the request to a computer someone already bought. The extra cost is mostly electricity. That’s why the price can be this low, and why the owner can still keep 99.8%.
That $0.005 figure is a planning price from hardware estimates, not the bill you would get today. During the early release, every model is charged one flat rate.
Compared
That is the first question. Salad is mostly gaming PCs and a points balance. Darkbloom is a high-end Mac. GPU rentals charge for the empty hour. TezGrid is the computer you already have, paid per request.
Salad
OpenAI-style gateway, or you ship a Docker job. The PC owner earns Salad Balance and later redeems PayPal, gift cards, or games. Their published token rates start around $0.06 per million input tokens.
TezGrid. A gaming PC, a Mac, or an office machine can take the job. The owner keeps 99.8% of that request, on the receipt. No balance to redeem, and no container to ship. One web address.
Darkbloom
Apple Silicon only. Their listed models start around 36 GB of memory. A 27B model is about $0.10 per million input tokens. The 100% owner share is a public-alpha promo on that higher price.
TezGrid. A normal gaming PC, office machine, or Mac can take the job. Same one-address API. The owner keeps 99.8% of a much smaller bill. Planning price for a small model is about $0.005 per million tokens.
Vast, RunPod, crypto GPU markets
You bring a container and pay for the GPU while it sits there, including the minutes nobody is sending a chat. Hosts are often datacenters, and some markets pay in crypto.
TezGrid. You pay for the request, not for an empty hour. The person who already owns the computer keeps 99.8%. No wallet, no server to babysit.
Salad and Darkbloom numbers are from their public pages. The TezGrid small-model price is a planning estimate, not today’s invoice. Bank and UPI payout is not live yet.
Before you send real data
These computers belong to other people. The connection is encrypted. The person who owns the machine can still read a request after it is decrypted there. If that is not acceptable, don’t send it.
The request is encrypted between TezGrid and the computer that runs the model. That computer does not open a public port.
The model has to run on someone’s machine. After it is decrypted there, the person with admin access on that computer can read it. If that is not acceptable for your data, do not send it.
Every reply says what trust level served it. You can require a minimum level. The strictest level only uses computers whose owners say they run confidential hardware.
We do not yet check AMD, Intel, or NVIDIA confidential-computing claims with a hardware measurement. Until we do, that level is a preference, not a proof.
A computer in a place where nobody is asking for AI will earn little. The dashboard shows real requests. It does not invent work.
Each request shows the split. Sending that money to a bank or UPI is not ready yet. Early wallet credit is given by invitation, not by a card payment.
Don’t wait for the crowd
Builders get cheaper AI without rewriting the app. Computer owners get paid for hardware that is already sitting there. Join the first wave and you’re in before your friends are still reading about it.