What is AI compute worth? Nobody knows

Watch on YouTube

Here's a quick test. What is an hour of AI compute worth?

Not what it costs from your provider. What it's worth. If you hesitated, you're in good company. Almost nobody can answer that, including many of the people selling it.

On theCUBE at the New York Stock Exchange, our CEO Paweł Czech named the problem plainly: "We're always buying the sun. We're always paying that maximum premium. And this is because nobody knows how much is your capacity actually worth."

This page is about that last sentence: why it's true, what it costs, and what changes when prices can be seen.

A market without a price tag

Most things we buy at scale have a reference price. Electricity, oil, bandwidth and even shipping containers have a number someone publishes, dates and argues over. You might pay more or less than that number, but you know where you stand.

AI compute mostly doesn't work that way. Capacity is sold in bundles, under contracts, by a small number of very large suppliers. Units differ from one seller to the next. Prices change without much notice. And the buyer often can't see what they actually received, only what they were billed.

When buyers can't compare, they buy the safe option. In AI, that's usually the biggest supplier at the top price. That's the premium Paweł was talking about. It isn't a fee anyone chooses to charge. It's what happens when nobody can check.

Three things a real market needs

If compute is going to be priced fairly, three things have to be true.

  1. Prices you can see. Published, dated and comparable, not handed out deal by deal.
  2. Units you can compare. A number is only useful if the next supplier quotes the same unit.
  3. Delivery you can check. A price means little if you can't confirm you got what you paid for.

None of these is exotic. They're the basics of every working market. AI compute just hasn't had them yet.

Where we're starting: Model Watch

We can't fix all three alone, and we're not going to pretend we can. But we can start with the first one: putting prices where people can see them.

That's what Model Watch is for. On /research, we publish sourced data points on model prices, each with the date we read it, so anyone can check our work.

Model Watch #1 looked at list prices for paid models. Read on 22 October, four paid models list $2.00 per million input tokens, with outputs from $6 to $10.

Two things about that number matter as much as the number itself.

These are other companies' list prices, not ours. Model Watch measures the market. It isn't a pitch.

It has a date on it. Prices move. A price without a read date is a rumour. A price with one is a data point you can compare with the next post, and the one after that.

Each Model Watch post adds a point. Over time, those points add up to something buyers haven't had: a public record of what AI actually costs, and how that changes.

Output list price, USD per 1M tokens

  • Claude Sonnet 5.5Anthropic$10
  • gpt-6.1-solOpenAI$10
  • gpt-6-solOpenAI$10
  • gpt-5.6-terraOpenAI$12
  • gemini-4-argon (Standard)Google$10
  • gemini-3.1-pro-preview (prompts up to 200k)Google$12
  • grok-4.7xAI$6
Source: Starfield dated price sheet, prices read 2026-10-28. Paid models listing $2 per 1M input tokens on that date. Other companies' list prices, not ours. Each one carries the date it was read.

Why this matters for where AI runs

Pricing is a big part of why the sun keeps winning. If you can't tell what capacity is worth, you can't tell whether a smaller, closer and better-fitting option is a good deal. So the default stays the default.

Once prices can be seen and compared, the question shifts from "who's the biggest?" to "what fits this job, here, at this price?" That's the question distributed capacity is built to answer.

Where we're going

Paweł described the goal on theCUBE: "Our goal is to deliver that capacity instantly at fair market value that is determined by an actual market."

That's the direction we're building in. We deploy standardized compute close to power, connectivity and demand, then connect it through one programmable network for AI inference. Every workload can generate an auditable record of commitment, placement and delivery. That record is the third thing a market needs: proof of what was delivered, not only what was billed.

Visible prices are where it starts. Model Watch is our first step.

Follow the numbers

If you buy AI capacity and want to talk about what it's worth, use Request capacity below.

Quotes

“Nobody knows how much is your capacity actually worth.”

Related

Watch the full interview · Model Watch