Why AI inference needs a benchmark.

Compute and inference are two links in one value chain. A trusted reference price for inference is the natural place to start.

Carry Research8 min read

The cost of AI is often discussed in terms of chips. GPUs are a vivid symbol of the AI economy, and the price of compute matters enormously to the data centres, clouds and investors who own and rent that capacity. Alongside them, almost every company building with AI buys inference: answers, code and analysis, delivered as tokens and billed by the million.

One value chain, two exposures

Compute sits at the top of the chain. It is financed over years and rented by the hour, and its owners and buyers carry the risk that its price moves. Inference is what that compute produces, and its cost reaches every company putting AI into a product or a workflow.

The AI value chain: compute produces inference, which every application consumes. Carry builds markets for compute and inference, with a flagship benchmark for inference. Compute Inference Applications GPUs and data centres Tokens, answers, code Every company using AI Owned and rented by the hour Bought by every company using AI WHERE CARRY BUILDS MARKETSFLAGSHIP BENCHMARK
Carry builds markets across compute and inference, anchored by a benchmark for the price of inference.

Both exposures are worth managing. A data centre pricing a multi-year lease, a finance team budgeting for AI and an investor studying margins all benefit from trusted, consistent prices. Inference is where a reference price can add the most today, which is why Carry started there.

Measuring price at constant capability

Inference has a remarkable property: models become more capable and more affordable at the same time. A good benchmark captures both, by holding capability constant and measuring what that level of intelligence costs.

That is what makes the index useful for hedging. When it moves, it tells participants that the cost of a given level of intelligence has moved, which is exactly the exposure they want to manage.

A benchmark should follow the price of intelligence itself.

What a reference price needs

  • Verifiable inputs. Prices anyone can check, such as the rates vendors publish themselves.
  • Consistent rules. A methodology that is versioned, governed and kept on record.
  • Independent oversight. A governing board that oversees the methodology and its use.
  • A single flagship. One benchmark the whole market can meet at, so liquidity gathers in one place.

From benchmark to market

Grain, metals and oil each came to rely on a trusted reference price, with a deep market built around it. Intelligence is on that path now. Carry Indices publishes the reference price for inference; Carry Exchange is building the market for compute and inference around it, so participants on both sides can manage their exposure.

This article is for information only and is not investment advice.

Keep reading

All insights
Research10 min read

Measuring price at constant capability

Models get cheaper and better at the same time. A useful benchmark has to separate the two.

ReadEssay7 min read

What grain, metals and oil teach us about pricing intelligence

Every great commodity eventually needed a trusted price. The pattern is remarkably consistent.

ReadMarket note5 min read

Designing a hedge that fits

How closely a hedge tracks what you actually pay shapes how well it protects you.

Read