Skip to main content
THE SEMICONDUCTOR HUB
Intelligence / Explainers / Where AI compute demand comes from

Where AI compute demand comes from

Demand for AI chips follows demand for AI compute. This explainer sets out who uses generative AI, for what, and how one published analysis turns those uses into an estimate of the compute needed in 2030.

Consumers and businesses

Generative AI compute is used in two broad ways: consumer applications (B2C), such as assistants and image tools offered to the public, and business uses (B2B), where companies apply models to their own work.

  • McKinsey forecast, 2024In the base case for 2030, about 70 percent of generative AI compute demand comes from consumer (B2C) applications and about 30 percent from business (B2B) uses.
    Source: McKinsey on Semiconductors, No. 9, March 2024, pp. 4, 6.

Training and inference

  • Training is heavy on data and compute; each inference run needs far less.
    Source: McKinsey on Semiconductors, No. 9, March 2024, p. 4.

A model is trained once, or a few times, but a deployed model answers requests continuously, so inference adds up across every user.

Six business archetypes

  • Business uses fall into six archetypes: coding and software development, creative content generation, customer engagement, innovation, simple concision and complex concision.
    Source: McKinsey on Semiconductors, No. 9, March 2024, pp. 4 to 5.
Coding and software development
Writing, reviewing, testing and explaining code.
Creative content generation
Drafting text, images and other media, such as marketing material.
Customer engagement
Assistants that answer, route and follow up customer questions.
Innovation
Generating and screening ideas or candidate designs in research and product development.
Simple concision
Summarising everyday documents and pulling out their main points.
Complex concision
Condensing large or specialised bodies of material, such as technical or legal records.
  • Complex concision is not expected to scale.
    Source: McKinsey on Semiconductors, No. 9, March 2024, p. 7.

How the estimate is built

  • The business estimate assumes that a company pays at most about 20 percent of the value a use creates.
    Source: McKinsey on Semiconductors, No. 9, March 2024, p. 7.
  • McKinsey forecast, 2024Generative AI needs about 25 × 10³⁰ floating-point operations (FLOPs) of compute in 2030, in the base case.
    Source: McKinsey on Semiconductors, No. 9, March 2024, p. 6.

The analysis models more than one adoption scenario; the figures on this page are its base case. They are the report's forecasts, not the Hub's.

What it means for chips

Compute demand turns into demand for accelerators, memory and servers and, further upstream, for wafers and fabs. Anatomy of an AI server sets out the parts that demand lands on.

Work in progressNot yet covered here: the wafer and fab demand that AI adds, the other scenarios in the analysis, and the electricity AI servers use. Suggest a source or a correction

Facts restated in our own words from McKinsey on Semiconductors, No. 9, March 2024, with the printed page. Forecasts are labelled as such and are the report's, not the Hub's. Where the report names a primary source, that source is the authority: check it before relying on a figure.