Why We're Building TokenRelay

It's insane what we've all quietly agreed to pay for.

Every lab serving inference bills the same way: per token. The frontier labs selling their own models, the hyperscalers serving open ones — same meter, different logo. Doesn't matter if the answer was brilliant or a hallucinated dead end — the meter ran, so you pay. Tokens aren't intelligence, yet we pay for them whether or not we get the right answer. We ran the industry logic and it doesn't add up.

Think about it. We benchmark these models against humans — replacing humans, accelerating humans, human-level intelligence. But nobody pays a human per word. Nobody pays a consultant for the draft they threw away. You pay for what gets delivered.

The claims keep escalating too — AGI, superintelligence, agents that run for days. And the meter? Still counting syllables. Even if you take the cheap seats, open models on a hyperscaler, you're not escaping per-token pricing. You're just paying less per token for the same trap.

And it goes deeper than the meter. Look at the hardware underneath. Memory prices have been ripping, and the future production of the components that matter most — the HBM stacks and advanced packaging every accelerator needs — is already spoken for. Pre-booked. Blocked out. By the same handful of players selling you the tokens. When a few companies can corner the physical stuff intelligence runs on, they don't just set chip prices — they get to coordinate the price of intelligence itself. Manufacture the scarcity upstream, then meter it to you downstream, per token. Coordination is a hell of a business model.

And the slogans keep getting grander to match. "The era of abundance." "The future is for everyone." Great — at what cost? Nothing in this world is free, and abundance that arrives metered by the token isn't abundance, it's a subscription to someone else's margin. If the future really is for everyone, it can't show up billed by the syllable.

So we're building it the other way.

Work with a finish line? You tell us what done looks like, you see the price up front, you buy the completion. Exploratory work — investigate, figure out what the problem even is? You buy a bounded block of capable time. Either way, the retries, the reasoning loops, the verbose wrong answers are our cost of production. Not your line item.

Here's the vision, in two lines:

• A market where intelligence is bought like work — by the task and by the hour, not by the syllable.
• Where "what did it cost" finally means "what did it get done."

And the mission, same deal:

• Price the work, not the words — per task where there's a finish line, per block of capable time where there isn't.
• Keep tokens where they belong: on our side of the invoice, as infrastructure — never as the product.

That's the whole premise. If you've ever stared at a token invoice and wondered what you actually bought… welcome. 🤝

The rest of this blog is us working it out in public.

— The founder