Applications and end demand

Primer

What this stage is

The products that turn tokens into revenue, and therefore the demand that has to justify every megawatt, wafer and GPU below. Applications differ by an order of magnitude or more in how many tokens they consume per user per day, so this page splits them by consumption rather than by industry.

How it works

At one end, autonomous coding and research agents run for hours and burn millions of tokens per user per day; a single seat can consume more than a thousand consumer chat users. At the other, consumer assistants have the largest user counts and the lowest tokens per user, and their free tiers set the floor on what a token can be sold for. Enterprise copilots sit between: shallow per user, wide by seat count. Video and image generation are compute-heavy per output and price-sensitive. On-device models move demand away from the data centre altogether.

The question for the chain is not how many users, but how many tokens, and at what price per token those tokens are paid for. Revenue per token, times tokens, has to cover the GPU-hours, which have to cover the capital below them.

What to watch

Users, seats and paying share for the largest products.
Tokens per user per day by tier, where any platform discloses it.
Subscription and seat prices, and usage-based pricing for agents.
Where the token growth is coming from: agents versus chat.
Signs of demand outrunning supply: rate limits, waitlists, price rises.

How it connects

This is the top of the chain. Tokens demanded here, divided by tokens per GPU-hour, is the GPU-hours the world needs; that number, divided by hours in a year and multiplied by the power per GPU, is the megawatts the bottom of the chain has to deliver.

Figures

Metric	Value	Unit	As of	Scope	Confidence	Source	Note
google_tokens_processed_per_month	3,200	trillion tokens per month across Google products and APIs (floor: 'over 3.2 quadrillion')	2026-05-19	Google (all tiers; Gemini app, Search AI features, Workspace, Cloud APIs)	high	Google, 'Google I/O 2026: Sundar Pichai's opening keynote' (The Keyword blog, 19 May 2026)	As printed: 3.2e+15 tokens per month across Google products and APIs (floor: 'over 3.2 quadrillion'). Monthly tokens processed across all Google surfaces as stated at I/O 2026, about 7x the roughly 480 trillion per month stated at I/O 2025; the largest public statement of inference volume by any provider. Arithmetic: 3.2e15 / 30 days = about 1.07e14 tokens per day. On the Q2 2026 earnings call (22 Jul 2026) Google added that its model APIs alone process 'approximately 22 billion tokens per minute', which is 22e9 x 60 x 24 x 30 = about 9.5e14 tokens per month, roughly 30 percent of the all-in figure.
gemini_app_monthly_active_users	950	million monthly active users of the Gemini app	2026-07-22	Google Gemini app (light tier: consumer chat)	high	Google, 'Alphabet earnings call Q2 2026: Sundar Pichai remarks' (The Keyword blog, 22 Jul 2026)	As printed: 9.5e+08 monthly active users of the Gemini app. Monthly active users of the standalone Gemini app as stated on the Q2 2026 earnings call, up from 400 million at I/O 2025 and 900 million at I/O 2026; the same remarks say AI Mode in Search has passed 1 billion monthly users. Monthly, not weekly, so not directly comparable with OpenAI's weekly figure.
chatgpt_weekly_active_users	900	million weekly active users of ChatGPT (floor: 'more than 900M')	2026-02-28	OpenAI ChatGPT (light tier: consumer chat)	medium	Newsweek, 'OpenAI Hits 900 Million Weekly Users, Raises $110B in Fresh Funding' (28 Feb 2026), quoting OpenAI's written statement	As printed: 9e+08 weekly active users of ChatGPT (floor: 'more than 900M'). OpenAI's own statement as quoted by a named publication when it announced a $110 billion funding round; openai.com refused automated fetching (HTTP 403), so the company page itself was not opened. Later press reports of about 1 billion weekly users (mid-2026) were not confirmed by OpenAI.
chatgpt_paid_consumer_subscribers	50	million paying consumer subscribers to ChatGPT (floor: 'more than 50 million')	2026-02-28	OpenAI ChatGPT consumer tiers (light tier)	medium	Newsweek, 'OpenAI Hits 900 Million Weekly Users, Raises $110B in Fresh Funding' (28 Feb 2026), quoting OpenAI's written statement	As printed: 5e+07 paying consumer subscribers to ChatGPT (floor: 'more than 50 million'). Paying consumer subscribers across ChatGPT's consumer plans, as stated by OpenAI in Feb 2026; arithmetic: 50 million / 900 million weekly users = about 5.6 percent paid conversion, which is the floor-setting fact for the light tier (most tokens are served free). Same page and quote as the weekly-user row.
m365_copilot_paid_seats	30	million paid Microsoft 365 Copilot seats (floor: 'over 30 million')	2026-06-30	Microsoft 365 Copilot (medium tier: enterprise copilots)	high	Microsoft Corp, Form 8-K Exhibit 99.1, FY26 Q4 earnings press release (29 Jul 2026)	As printed: 3e+07 paid Microsoft 365 Copilot seats (floor: 'over 30 million'). Paid seats of the enterprise copilot as of the quarter ended 30 June 2026; Microsoft reported 15 million (Jan 2026) and 20 million (Apr 2026) in the two prior quarters, and on the call said net seat adds more than doubled quarter over quarter. Arithmetic at the $30 list price: 30e6 x $30 x 12 = about $10.8 billion of annualized list revenue before discounts.
m365_copilot_seat_price	30	USD per user per month (list price)	2023-07-18	Microsoft 365 Copilot (medium tier)	high	Microsoft, Official Microsoft Blog, 'Furthering our AI ambitions: Announcing Bing Chat Enterprise and Microsoft 365 Copilot pricing' (18 Jul 2023)	The list price announced at launch and still the reference seat price quoted for the 30 million paid seats above; Microsoft's current pricing pages timed out during this run, so the 2026 page was not re-checked. A seat price fixed since 2023 while tokens per seat rise is the core margin question for the medium tier.
claude_code_run_rate_revenue	2.5	USD bn annualized run-rate revenue (floor: 'over $2.5 billion')	2026-02-12	Anthropic Claude Code (heavy tier: agentic coding)	high	Anthropic, 'Anthropic raises $30 billion Series G funding at $380 billion post-money valuation' (12 Feb 2026)	As printed: 2.5e+09 USD annualized run-rate revenue (floor: 'over $2.5 billion'). Company-stated run-rate revenue of the Claude Code agentic coding product about nine months after general availability; secondary reports put the figure much higher later in 2026 but no later company statement was found. Anthropic's total run-rate revenue in the same release was $14 billion, so Claude Code was about 18 percent of it.
agentforce_arr	1.5	USD bn annual recurring revenue (floor: 'exceeded $1.5 billion')	2026-07-31	Salesforce Agentforce (medium tier: enterprise agents)	high	Salesforce press release, 'Salesforce Delivers Record Second Quarter Fiscal 2027 Results' (26 Aug 2026)	As printed: 1.5e+09 USD annual recurring revenue (floor: 'exceeded $1.5 billion'). ARR of the Agentforce product alone at the end of Salesforce's fiscal Q2 2027 (quarter ended 31 Jul 2026), up from $1.2 billion one quarter earlier; the combined Agentforce and Data 360 ARR was nearly $3.9 billion. The same release counts 3.2 billion 'Agentic Work Units' delivered in the quarter, up 97 percent quarter over quarter, Salesforce's own consumption unit.
agentforce_price_per_conversation	2	USD per conversation (list price, Agentforce Service Agent)	2026-09-27	Salesforce Agentforce (medium tier)	high	Salesforce, 'Agentforce Pricing' page (accessed 27 Sep 2026)	Consumption list price per customer conversation handled by an Agentforce agent; the same page prices Flex Credits at $500 per 100,000 credits (20 credits per agent action, see conversions) and the Agentforce 1 Edition at from $550 per user per month. The clearest public example of an enterprise agent priced per unit of work rather than per seat.
servicenow_ai_acv	1	USD bn annual contract value (floor: 'crossed $1 billion')	2026-06-30	ServiceNow AI products including Now Assist (medium tier)	high	ServiceNow, Inc., Form 8-K exhibit, 'ServiceNow Reports Second Quarter 2026 Financial Results' (22 Jul 2026)	As printed: 1e+09 USD annual contract value (floor: 'crossed $1 billion'). Annual contract value of ServiceNow's AI SKUs at the end of Q2 2026; the company's stated target is $1.5 billion of AI ACV by the end of 2026. The figure covers all ServiceNow AI products, not Now Assist alone.
apple_on_device_model_parameters	3	billion parameters in Apple's on-device foundation model	2025-06-09	Apple Intelligence on-device model (edge tier)	high	Apple Machine Learning Research, 'Updates to Apple's On-Device and Server Foundation Models' (9 Jun 2025, updated 17 Jul 2025)	As printed: 3e+09 parameters in Apple's on-device foundation model. Size of the language model that runs locally on iPhone, iPad and Mac for Apple Intelligence features; requests it cannot handle go to Apple's server-based mixture-of-experts model on Private Cloud Compute. Sets the scale of what 'edge' inference means today: a few billion parameters on the device, so those tokens never reach a data-centre GPU.
openrouter_programming_share_of_tokens	50	percent of total token volume on OpenRouter classified as programming (floor: 'exceeded 50%')	2026-01	OpenRouter marketplace traffic (heavy tier: coding agents; third-party developer demand only)	medium	OpenRouter with a16z, 'State of AI: An Empirical 100 Trillion Token Study' (arXiv 2601.10088, Jan 2026)	Share of tokens routed through the OpenRouter API marketplace whose classified use is programming, up from about 11 percent in early 2025; the only public statement found of inference volume split by class of use. OpenRouter sees only developers who route through it (over 100 trillion tokens studied), not first-party traffic at OpenAI, Google or Anthropic, so it is a sample of the heavy tier, not the market.

Conversions

From → To	Factor	Unit	As of	Confidence	Source	Note
Applications and end demand → Tokens	13	USD of API tokens per developer per active day, Claude Code enterprise average	2026-09-27	high	Anthropic, Claude Code documentation, 'Manage costs effectively' (accessed 27 Sep 2026)	Anthropic's own average for enterprise deployments billed per token, with 90 percent of users under $30 per active day and $150-250 per developer per month. Illustrative tokens (model mix is not published): at Sonnet 5 list prices of $2 per million input and $10 per million output (claude.com/pricing), $13 buys 1.3 million output tokens or 6.5 million uncached input tokens, so a typical coding seat is in the low millions of tokens per active day, the heavy-tier band in consumption_tiers.csv.
Applications and end demand → Tokens	0.1	USD per Agentforce agent action (list, via Flex Credits)	2026-09-27	high	Salesforce, 'Agentforce Pricing' page (accessed 27 Sep 2026)	Arithmetic: Flex Credits cost '$500 USD/Per 100k Credits' = $0.005 per credit; an Agentforce action consumes 20 credits, so 20 x $0.005 = $0.10 per action (voice actions are 30 credits = $0.15). Together with $2 per conversation this is the public price of a unit of enterprise agent work, which the 3.2 billion Agentic Work Units per quarter can be multiplied against.
Applications and end demand → Tokens	0.23	USD per second of generated video, Runway Gen-4.5 on the Standard plan	2026-09-27	high	Runway, 'Pricing' page (accessed 27 Sep 2026)	Arithmetic: Standard plan $12 per month for 625 credits = $0.0192 per credit; Gen-4.5 burns 12 credits per second, so 12 x $0.0192 = $0.2304 per second (about 52 seconds of video per month). On the Max plan ($76 for 9,500 credits, $0.008 per credit) the same second costs $0.096. This is what a consumer pays; the compute cost behind it is not disclosed.
Applications and end demand → Tokens	0.4	USD per second of generated video with audio, Google Veo 3.1 Standard (720p and 1080p), API list price	2026-09-27	high	Google AI for Developers, 'Gemini Developer API Pricing' page, Veo 3.1 section (accessed 27 Sep 2026)	Per-second API price for the standard Veo 3.1 video model; the same table lists Veo 3.1 Fast at $0.10 per second (720p) and $0.60 per second for 4K standard output. Arithmetic: an 8-second clip costs 8 x $0.40 = $3.20 at standard quality, versus about 8 x $0.2304 = $1.84 on Runway's Standard plan. The quote is the table cell for the standard 720p/1080p rate; the page's row label is 'Veo 3.1 Standard video with audio price (default)'.
Applications and end demand → Tokens	1,000,000,000,000	tokens per year that define a 'trillion-token' Azure Foundry enterprise customer	2026-07-29	medium	Microsoft FY26 Q4 earnings call, 29 Jul 2026, Satya Nadella remarks (transcript hosted by StockAnalysis)	A threshold Microsoft uses to count its largest API customers, not an average; arithmetic: 1e12 / 365 = about 2.7 billion tokens per day per such customer, or roughly 2,700 heavy-tier developers at 1 million tokens a day each. On the FY26 Q3 call Microsoft put the count at over 300 customers on track for the year. Third-party transcript, so quote accuracy depends on the transcriber.
Applications and end demand → Power generation and fuel	0.34	Wh per average ChatGPT query	2025-06	medium	Sam Altman (OpenAI CEO) personal blog, 'The Gentle Singularity' (June 2025; the page shows no date)	The only OpenAI-attributed energy-per-query figure; no methodology is published and 'average query' is undefined, so treat as an order of magnitude. Arithmetic: 0.34 Wh x 1,000,000 queries = 340 kWh per million queries; at 900 million weekly users making one query a day each, 0.34 Wh x 900e6 = 306 MWh per day, about 12.8 MW of continuous load.
Applications and end demand → Power generation and fuel	0.24	Wh per median Gemini Apps text prompt	2025-08-21	medium	Google Cloud blog, 'Measuring the environmental impact of AI inference' (21 Aug 2025)	Google's own full-stack measurement (accelerators, host CPU and DRAM, idle capacity and data-centre overhead) for the median text prompt in the Gemini app in May 2025; a median, not a mean, so heavy prompts, image and video generation and agentic use are not represented. Arithmetic: 0.24 Wh x 1,000,000 prompts = 240 kWh per million prompts.

Players

Company	Ticker	Role	Listed	Note
Microsoft	NASDAQ:MSFT	enterprise copilot seats (Microsoft 365 Copilot) and Azure Foundry token platform (medium tier)	yes	Over 30 million paid Microsoft 365 Copilot seats at a $30 list price as of June 2026, the largest per-seat AI deployment anywhere; on the same call Microsoft said the number of Foundry customers at a 1-trillion-token annualized run rate grew 4x year over year.
OpenAI	private	consumer chat (ChatGPT, light tier) and coding agent (Codex, heavy tier)	no	OpenAI's Feb 2026 statement claimed more than 900 million weekly active users and more than 50 million paying consumer subscribers, the largest light-tier user base; openai.com blocks automated fetching so its own pages were not opened.
Alphabet (Google)	NASDAQ:GOOGL	consumer assistant (Gemini app), AI search (AI Mode, AI Overviews) and API tokens (light and medium tiers)	yes	950 million Gemini app monthly active users (Jul 2026) and over 3.2 quadrillion tokens processed per month across its products (May 2026); the one company that publishes an all-in token count.
Anthropic	private	agentic coding (Claude Code) and enterprise API (heavy tier)	no	Claude Code run-rate revenue passed $2.5 billion in Feb 2026 inside a $14 billion company run-rate; its Max plan for heavy individual users is priced 'From $100 Per month' with a choice of '5x or 20x more usage than Pro' (claude.com/pricing), and its docs put enterprise usage at about $13 of tokens per developer per active day.
Anysphere (Cursor)	private	AI code editor and coding agent (heavy tier)	no	Cursor's own blog (13 Nov 2025) said 'We've also crossed $1B in annualized revenue' alongside a $2.3 billion Series D at a $29.3 billion post-money valuation; third-party reports (Dealroom, citing a person familiar with the matter) put annualized revenue above $4 billion by June 2026. Pricing page (Sep 2026): Individual $20 per month, Teams $40 per user per month.
Cognition (Devin)	private	autonomous software-engineering agent (heavy tier)	no	Company blog, 25 Sep 2026: 'Today, Cognition crossed $1B in annualized revenue run rate.' Devin runs long autonomous tasks, the purest example of the heavy tier's many-tokens-per-user pattern.
Salesforce	NYSE:CRM	enterprise agents priced per conversation and per action (Agentforce, medium tier)	yes	Agentforce ARR exceeded $1.5 billion at 31 Jul 2026 (up over 240 percent year over year), priced at $2 per conversation or 20 Flex Credits per action; 3.2 billion Agentic Work Units delivered in the quarter.
ServiceNow	NYSE:NOW	enterprise workflow copilots and agents (Now Assist, medium tier)	yes	ServiceNow AI crossed $1 billion of annual contract value in Q2 2026 against a $1.5 billion year-end 2026 target; sold as add-on SKUs on top of existing per-seat workflow subscriptions.
Perplexity	private	AI search / answer engine (medium tier: AI search)	no	No company page with current figures could be opened in this run; third-party trackers cite roughly $750 million of annualized revenue by August 2026 and a February 2026 switch to subscription-only pricing, both unverified here.
Runway	private	text-to-video generation (high tier)	no	Sells video by the credit: Gen-4.5 burns 12 credits per second, on plans from $12 per month (625 credits) to $76 per month (9,500 credits). Midjourney, the other large private image-generation player, sells by Fast GPU hours on plans a secondary source lists at $10 to $120 per month; its official price page refused automated fetching.
Apple	NASDAQ:AAPL	on-device inference at scale (Apple Intelligence, edge tier)	yes	Runs an approximately 3-billion-parameter language model locally on iPhone, iPad and Mac, routing only overflow requests to its own Private Cloud Compute servers, so most Apple Intelligence tokens never touch a data-centre GPU.
Tesla	NASDAQ:TSLA	in-car inference on its own silicon (FSD, edge tier)	yes	Tesla's Q1 2026 update (8-K exhibit) reports 1.28 million active FSD subscriptions and says 'In April, we completed the final chip design of our next-generation AI5 inference processor'; Tesla publishes no per-car compute (TOPS) figure in its filings.