Demand: Who uses AI and for what: users and usage, company adoption, closed and open models, AI revenue and pricing.

How it works

Demand is where AI gets used: people asking assistants questions, businesses building it into their work, and developers calling models through APIs. That use is metered in tokens, and metered use is what turns into revenue and, in time, into new capacity in the other three layers.

OpenAI said ChatGPT had more than 900 million weekly active users in February 2026¹, and Alphabet said in October 2026 that the Gemini app had reached one billion users². Among US employer businesses, 24.6% reported using AI in the two weeks to October 18, 2026, up from 17.3% in late November 2025³.

Share of US businesses using AI rose from 17.3% to 24.6%. Employer businesses that used AI in any business function in the previous two weeks, every two-week reading Data in the table below. — Download image Sized for X and LinkedIn, with source and address included.

Survey period	Share using AI
Nov 17 to 30, 2025	17.3%
Dec 1 to 14, 2025	17.2%
Dec 15 to 28, 2025	17.8%
Dec 29, 2025 to Jan 11, 2026	17.7%
Jan 12 to 25, 2026	18.2%
Jan 26 to Feb 8, 2026	18.9%
Feb 9 to 22, 2026	17.5%
Feb 23 to Mar 8, 2026	18.2%
Mar 9 to 22, 2026	18.9%
Mar 23 to Apr 5, 2026	19.1%
Apr 6 to 19, 2026	19.8%
Apr 20 to May 3, 2026	19.8%
May 4 to 17, 2026	19.5%
May 18 to 31, 2026	20.6%
Jun 1 to 14, 2026	20.6%
Jun 15 to 28, 2026	20.6%
Jun 29 to Jul 12, 2026	21.7%
Jul 13 to 26, 2026	21.5%
Jul 27 to Aug 9, 2026	21.8%
Aug 10 to 23, 2026	22.4%
Aug 24 to Sep 6, 2026	23.2%
Sep 7 to 20, 2026	23.8%
Sep 21 to Oct 4, 2026	23.8%
Oct 5 to 18, 2026	24.6%

How use becomes infrastructure

Request. A person or a software agent sends a prompt, document or task to an AI app or API.
Inference. The provider runs a model on accelerators to produce the answer, metered in tokens. Google's model APIs processed nearly 32 billion tokens a minute in July to September 2026, more than four times a year earlier².
Billing. Tokens are billed as usage fees or covered by subscriptions. Microsoft said its AI business passed an annual revenue run rate of $37 billion in the quarter to March 2026, up 123%⁷.
Contracts. Growing use leads to multi-year capacity contracts, which show up as cloud backlog. Amazon reported an AWS backlog of $578 billion and Bedrock token volume up nearly 17 times from a year earlier⁸.
Capital spending. Cloud providers raise capital budgets to deliver that backlog.
Capacity. The spending buys chips, servers, data centers and power, which serve more use.

Google's model APIs went from 7 billion tokens a minute to nearly 32 billion. Tokens processed per minute through Google's first-party model APIs, by quarter, as stated by Alphabet Data in the table below. — Download image Sized for X and LinkedIn, with source and address included.

Quarter	Tokens per minute, as stated
Jul to Sep 2025	7 billion
Oct to Dec 2025	over 10 billion
Jan to Mar 2026	16 billion
Apr to Jun 2026	approximately 22 billion
Jul to Sep 2026	nearly 32 billion

Terms

Token: The unit of text, code, image or audio a model reads or writes; usage is metered and priced in tokens.
Inference: Running a trained model to answer a request.
Training: Building a model by adjusting its parameters on large datasets; the most compute-intensive step.
Weekly active users: People who used a product at least once in a seven-day period.
Run-rate revenue: Current revenue annualized; each company defines it differently, and it isn't audited revenue.
Open-weight model: A model whose trained weights are published so others can download and run it under its license.

How it connects

Drives Compute: Contracted use sets how much capacity cloud providers build.
Drives Energy: More capacity means more data center load on the grid.
Depends on: Model quality and price. Epoch AI estimates the cost of reaching a fixed benchmark score has fallen about 47% a quarter, or 13 times a year¹⁰.
Shaped by Chips: Accelerator supply limits how much capacity can serve new use.

Who's involved

Companies named most often in this layer's facts.

Adtran1

Early signals to watch

Public records that show changes in this layer before they're announced.

Record	What it shows	How early
Census business survey	Share of businesses using AI, and expecting to within six months	Every two weeks, about four days after collection ends³
Token and usage disclosures	Tokens processed and user counts from major providers	Quarterly, before use shows up fully in revenue²
Cloud backlog	Contracted future spending with cloud providers	Quarters to years ahead of revenue and capacity⁸
Public model usage data	Which models developers use, open and closed	Weekly; open-weight models passed half of OpenRouter tokens in April 2026⁹
SEC filings	Revenue, customer counts and spending plans	As filed, between press updates
User and worker surveys	How often people use AI	Months before company revenue reflects it

Latest facts

Demand Financial disclosure Nov 3, 2026

Adtran hyperscale revenue rises more than 200% year over year on data center interconnect demand in Q3 2026

Data center interconnect with hyperscalers drives growth. Ent/Gov/Cloud revenue grew 44% year over year to 28% of revenue, and grew 20% excluding hyperscalers.

Adtran Source: ADTRAN Holdings, Inc. 8-K, filed Nov 3, 2026

Sources

Reference figures checked Nov 3, 2026.