The State of the AI Economy

June 25, 2026

Azeem Azhar, William Gildea, Hannah Petrovic, PhD, Nathan Warren & Marija Gavrilov

1

Exponential View

www.exponentialview.co

Independently produced by Exponential View. Built from public disclosures and Exponential View’s own models; all conclusions are our own. �© Epiiplus1 Ltd 2026

2

There is a visibility problem in the AI economy. Until now, it has been impossible to deconstruct real customer demand.�

The supply side of the AI economy is well-documented. Most semiconductor companies and hyperscalers are public �and disclose their activities in some detail. Sell-side analysts have done a great job decomposing their performance.

​

The demand side, what customers are actually paying for and if the revenues are real, has been obscure.

The largest labs are private, and even public companies bury AI revenue inside segment totals.

​

Without understanding genuine demand, it is impossible to judge the health of the AI economy that underpins $22.7 trillion of stock market valuation and has driven US GDP growth in the past six quarters.

​

We hope that this report serves as a reference source on the current state of play, free of hype and fear, while helping us all have a more informed conversation about the gravitational pull AI is exerting on the economy and the world at large.

​

Special thanks to those who kindly reviewed an early draft of this presentation and gave us feedback:�Alex Imas, Shanu Mathew, Patrick Rutherford, Jaime Sevilla and Amy Sutter.

​

– Azeem and the Exponential View team

Why we’ve done this

3

We flag, investigate and maintain our datasets using analyst research, augmented by a proprietary system that scans, crawls and synthesizes insights.

Every revenue line traced to primary filings, audited accounts, transcripts and credible reporting; plus cloud-attribution where a private firm’s revenue surfaces in a public firm’s accounts (OpenAI via Azure, Anthropic via Bedrock).

We build full per-company models specifically for GenAI financials (split out from top-line reporting), covering key drivers of revenue, profitability and cost in the P&L, cash flows and balance sheets.

A proprietary line-level revenue model: Sourced, scored, triangulated, deduplicated

Bottom-up, 1,000+ firms

›

Each line carries a rigorous confidence score before it enters any model, so weak inputs can’t inflate the number.

Confidence-scored before �it counts

›

Company financial models checked against top down

›

Spend only counted once

Revenue is counted at every layer but never summed across them: attributed by value-add so the same dollar isn’t double- or triple-counted.

Audit trail: Sample source table

$40�value-add

$30�value-add

$30�value-add

e.g. $100 app spend that sends $60 to a model provider, which spends $30 on inference hosting, is counted as $100, not $190:

Scope: Global ex-China · App, model & infrastructure revenue counted · Excludes chips, AI ad-uplift, legacy-software features and financing.

$100 rev.

$60 rev.

$30 rev.

Apps

FM labs

Hosting

We grade filed figures highest, above other primary sources, corroborated 3^rd-party estimates and single-sourced claims.

All derived numbers inherit the lowest grading from input sources.

These models are reconciled against independent proxies: silicon �(chip-maker revenue), build cost, �segment mix, industry research, �traffic and capacity.

Additional soft signals we use include unofficial sources such as:

• Public comments by executives and related parties.
• Proxy and sample metrics.
• Commentary and unverified estimates and leaks in traditional, new and social media.

Audit trail: Sample company revenue model

AI demand is more clearly validated by realized revenue than previous platform shifts. Generative AI ecosystem revenue has already surpassed $175 billion annualized (after removing double-counting from provider revenues).

​

CapEx intensity is growing well above historical large-cap technology norms to deliver the AI buildout. And third-party financing is increasingly entering the financing mix.

​

The open question is whether cheapening artificial intelligence can create enough volume and margin to service the buildout.

4

The top line

5

Contents

1 | Demand:�It’s real, big & fast

6

Revenues are driven by real external customers. �The sector is growing 3x faster than any IT wave before it.

This demand has created a compute supercycle: 10x more compute, new energy generation, larger data centers and mounting backlogs where supply cannot keep up.

$110bn trailing 12-month revenues – now at a $175bn pace

​

Source: Exponential View analysis.�Note: Global ex-China. Deduplicated app, foundation model, and infrastructure hosting revenues. Excludes chip manufacturing.

Generative AI economy revenue, deduplicated

$bn/year, Jan 2023 - Jun 2026

1 | Demand: It’s real, big & fast

7

1.6x gap�the spread reflects the growth rate

$175bn�annualized run rate

$110bn�banked trailing�12-month revenue

Real, external demand drives AI revenues

1 | Demand: It’s real, big & fast

8

Illustration of how we model and deduplicate revenues between providers

App layer

Customer

pays app license/fees, or buys tokens directly

Foundation model layer

Chip sales (CapEx from hosting layer)

Ad uplift (Google/Meta AI ad revenue)

Non-AI-native apps (Counted via token spend)

Not counted:

Revenues flow down the stack

We source, triangulate, model & audit to verify & deduplicate:

Rigorous company-by-company financial modeling to build a bottom-up deduplicated revenue model.

Continuous scanning and crawling across hundreds of sources to maintain and adjust the dataset.

Sourced from official filings, 1st-party disclosures, leaks, government stats, 3rd-party analysts, and proxy metrics; all sources quality-graded.

AI is scaling three times faster than any IT wave

Sources: Exponential View analysis; US Commerce; company filings; UBS; US Bureau of Labor Statistics.�Note: We use the first full year of revenues, so have started GenAI measurement from January 2023.

Realized revenue trajectory time-aligned to year zero

$bn/year, adjusted for inflation

1 | Demand: It’s real, big & fast

9

3x faster�than any prior wave

Indexing GenAI growth to past technology waves risks understating its speed when modeling:

• Demand projection
• Depreciation schedules
• CapEx paybacks

Each new $1 billion of revenue arrives faster than the last

Source: Exponential View analysis.�Note: Global ex-China. Deduplicated app, foundation model, and infrastructure hosting revenues. Excludes chip manufacturing.

Time to add $1bn additional cumulative revenue

days, log scale

1 | Demand: It’s real, big & fast

10

In 2023, the AI industry needed 180 days to add �$1 billion in cumulative revenue.

It now needs less than �two days.

90x faster

Growth has held across each adoption phase

Source: Exponential View analysis.

Revenue growth quarter-on-quarter

% change since prior quarter

1 | Demand: It’s real, big & fast

11

Agentic Coding Era

Chatbot Subscription Era

35% QoQ

equivalent to

3.2x annually

ChatGPT Team launches

Claude Enterprise launches

Claude Code released

Codex launch

OpenClaw �goes viral

This rapid demand growth is showing as a contract backlog �for hyperscalers

Sources: Exponential View analysis; company filings.

Note: Microsoft = total RPO (incl. M365/Dynamics); Amazon = total company RPO (mostly AWS); Google = revenue backlog (mostly Cloud). Oracle quarter ends one month earlier.

Combined hyperscaler backlog (remaining performance obligations)

$bn

1 | Demand: It’s real, big & fast

12

Demand has launched a compute supercycle

Sources: Exponential View analysis; World Semiconductor Trade Statistics; Japanese Semiconductor History Museum of Japan.

Note: Nominal $US

Global semiconductor market revenues

$bn/year

1 | Demand: It’s real, big & fast

13

WSTS 2026 projection:

AI has spurred an uptick in a 50-year trend of compute growth

Source: Exponential View analysis.

Note: Includes mainframes, minicomputers, PCs, servers, smartphones, IoT and AI compute. AI-server FLOPS derived from the installed base of Nvidia GPUs by generation, using FP8 from Hopper onward.

Global compute since 1971

FLOPS

1 | Demand: It’s real, big & fast

14

66% CAGR

80% CAGR

AI demand is reigniting a moribund US power sector

Sources: Exponential View analysis; US Energy Information Administration.

US electricity net generation

TWh/month

1 | Demand: It’s real, big & fast

15

2008-2024: ±0 growth

1950-2008: +6 TWh/month annual growth

2024-today:�+9 TWh/month annual growth

The size of the largest data centers has grown 50x in four years

Sources: Exponential View analysis; TOP500 / Green500; Epoch AI; OpenAI; Oracle.

Note: Rainier full build is a target.

1 | Demand: It’s real, big & fast

Power of the most powerful computers over time

MW, log scale

16

Memory and compute now take a majority of every dollar spent on the data center buildout

Sources: Exponential View analysis; Goldman Sachs; Epoch AI; Semi Analysis. �Note: Figures may not sum to 100% because of rounding.

Share of total data center build cost by component, 2021 vs 2026E

%

1 | Demand: It’s real, big & fast

17

• Each new dollar buys more silicon and less concrete: chips’ share of spend is up 50% (40%→60%).

​

• Memory is the single biggest mover, from a 2% rounding error to ~18%.

​

Leading to growing commitments for compute and energy

Sources: Exponential View analysis; Grid Strategies; Nvidia filings.

Compute & power commitments

Nvidia supply commitments ($bn, left axis) & US load-growth (GW, right axis)

1 | Demand: It’s real, big & fast

18

• Nvidia supply commitments have grown from $31bn to $95bn in the last year.

​

• The extra electricity the US grid is expected to need by 2030 has grown ~7x since 2022 (24GW→166GW), with data centers accounting for ~55% of this growth.

​

2 | Economy:�Big is still small, and early

19

Even for the highest corporate spenders, AI is a rounding error in the P&L. �It still looks early. Initiatives have focused on efficiency & cost savings, although the mix is changing. And measured revenue may understate the social gains, as consumers report benefits that don’t yet show up in the data.

Against GDP, AI revenue is still a rounding error

Sources: Exponential View analysis; St. Louis Fed.

Global AI revenues (ex. China), relative to US GDP, labor costs & corporate profits

%

2 | Economy: Big is still small, and early

20

• Still tiny: AI revenue is equivalent to 0.42% of US GDP�(vs IT sector’s 9.4%).

​

• Even a generous yardstick (corporate profits) is 32x larger than all GenAI revenues.

​

• Still early: AI revenue relative to GDP has risen 3x vs Q1 2025 (0.13%), and 10x vs Q1 2024 (0.04%).

3.0%�Profits

0.8%�Labor

0.4%�GDP

At a company level, AI spending is still relatively small:

e.g. Uber’s $1.5k per engineer barely dents the P&L

Sources: Exponential View analysis; Ramp Economics Lab (n = 70,000 US businesses), Uber filings.�Note: Top 1% / top 10% / median defined by level of AI spend compared across Ramp’s customer base. Uber figure is a max per-engineer cap, benchmarked against Ramp per-employee AI spend.

AI spend per employee

$/month, log scale, Ramp customers vs Uber cap

Uber AI spend (maxed cap) vs P&L line items

$/year (AI spend %), log scale, vs FY2025

​

2 | Economy: Big is still small, and early

21

$1.5k per-engineer cap roughly puts Uber in the top 10% of per-employee AI spend

$90m

$52bn(0.2%)

$8.7bn(1.0%)

$14bn(0.6%)

$3.4bn(2.6%)

$720m(12%)

Like previous general-purpose technologies, some gains may escape GDP measurement

2 | Economy: Big is still small, and early

22

Consumer surplus

​

Value that reaches people directly at a near-zero price. Little is sold, so GDP under-represents consumer benefit, e.g.:

• Free replacement of software & service purchases
• Learning, leisure & convenience

Producer surplus

​

Value embedded in sold goods and services. More is transacted and recorded in GDP, e.g:

• AI-enabled features that drive revenue
• Faster (valuable) releases
• Service firm margins

1980-2000: Automation

​

Programmable machine tools cut labor in every unit, raising output per worker.

Graetz & Michaels (2018)

+0.37

pp/yr GDP

+0.4

pp/yr GDP

1850-1870: Steam

​

Mechanized factories and railways, so one worker could produce and move far more to market.

Crafts (2004)

1880-1920: Electric lighting

​

Light became ~99.97% cheaper: �an hour’s wage buys ~40,000x more. Prices didn’t record this gain.

Nordhaus (1996)

≈ $0

direct GDP impact

2000-2020: Free digital goods

​

Free search, encyclopaedias and maps displaced paid services. Search alone �is worth ~$17.5k/yr/person.

Brynjolfsson et al. (2019)

≈ $0

direct GDP impact

AI impacts

Historic cases

GDP knows the price of everything but the value of nothing.�AI’s economic value exceeds measured revenue

Sources: Exponential View analysis; Stanford Digital Economy Lab

Note: Welfare value determined from responses to “Would you give up access to all AI tools like ChatGPT, Gemini, Claude, or Copilot �for one month starting tomorrow morning in exchange for [$US]?”. Revenue includes global (ex-China) consumer and enterprise spend.

Monthly GenAI revenues vs US consumer welfare

$bn/month, Jul 2024 - Mar 2026

2 | Economy: Big is still small, and early

23

Revenue: How much is spent on AI?

$3-4bn consumer surplus

(+30% on revenues)

Consumer welfare: What value do consumers place on AI?

Public companies are reporting increased impact of GenAI

Sources: Exponential View analysis; earnings calls.

2 | Economy: Big is still small, and early

Companies making claims of AI impact on earnings calls

S&P 500, Q4 2022 - Q1 2026

24

• Growing attention: Firms see AI as an opportunity to improve earnings. We tracked a 3-4x rise in mentions of AI’s impact across the S&P 500 �since 2023.

​

• Put a number to it: 50-60% of claims are now quantified, but TBD how large and meaningful these are for companies’ bottom line.�
• Still a minority: A majority of firms have not yet reported a quantified impact from AI use.

Seven in ten GenAI claims focus on cost savings or efficiency

Why initial projects prioritize efficiency, an illustrative example:

The same pure $1m impact has 4pp better profit margin from savings vs sales growth

Sources: Exponential View analysis; earnings calls. �Note: Figures may not sum to 100% because of rounding.

Claimed AI outcomes

S&P 500, Q4 2022 - Q1 2026

2 | Economy: Big is still small, and early

Note: For illustrative purposes, revenue is added at $0 cost. Adding costs-of-sales would dampen margin growth further.

25

$11m revenue • $7m cost

$4m profit

-$1m costs

Cost reduction: 25%

Time savings: 23%

Throughput increase: 22%

Quality improvement: 18%

Conversion improvement: 7%

Revenue gain: 6%

TODAY

30% net margin

$10m revenue • $7m cost

$3m profit

COST SAVINGS

40% net margin

$10m revenue • $6m cost

$4m profit

SALES GROWTH

36% net margin

+$1m revenue

As in prior waves, early adopters are outgrowing their peers

Sources: Exponential View analysis;�Bessen, Goos, Salomons & Van den Berge (2020):�“What Happens to Workers at Firms that Automate?”.

Sources: Exponential View analysis; Ramp Economics Lab �(n = 70,000 US businesses).�Note: High intensity = Top 25% AI spenders by share of revenue.

The historic case study:

Firm-level employment with/without automation

% change 2000-2016

The AI economy today:

Revenue growth with high vs no AI usage

% change since Nov 2022

​

2 | Economy: Big is still small, and early

26

Δ37%

Δ92%

3 | CapEx:�The largest buildout in tech history is paying back (for now)

27

Hyperscalers & neoclouds have committed to $2 trillion of cumulative CapEx to 2026, putting pressure on growing revenues to pay back, especially as more is funded by external capital. These economics set the tone for data center and token production finances.

Hyperscaler and neocloud CapEx reaches $2T cumulatively through 2026E

​

3 | CapEx: The largest buildout in tech is paying back (for now)

Sources: Exponential View analysis; company filings.

Note: 2026 is based on guidance values. Oracle uses a 5/12:7/12 split based on FY2026 reported and FY2027 guidance values.

Hyperscaler and neocloud CapEx

$bn, PP&E + leases

28

Total CapEx ≠ AI CapEx.

Announced numbers include pre-planned CapEx for existing cloud & SaaS businesses, metaverse (Meta), and logistics (Amazon)

AI-linked CapEx adds $535bn above the pre-AI trend by 2026E

Sources: Exponential View analysis; Hyperscaler earnings & guidance; US telecom & carrier guidance; The International Energy Agency.

3 | CapEx: The largest buildout in tech is paying back (for now)

Annual CapEx per industry

$bn

29

$535bn above pre-AI trend

Forecasts have chased the CapEx curve higher

3 | CapEx: The largest buildout in tech is paying back (for now)

Sources: Exponential View analysis; Barclays, Citi, Goldman Sachs, JP Morgan, Morgan Stanley, New Street Research, SemiAnalysis, UBS; company filings.

Note: Forecasters use slightly different baskets (the Big Five hyperscalers vs broader AI infrastructure). Actuals here correspond to the cash CapEx (purchases of property and equipment) for Microsoft, Alphabet, Amazon, Meta, Oracle, CoreWeave, and Nebius.

Hyperscaler / AI infra CapEx forecasts by analyst and forecast date

$/year

30

The marginal AI-infra dollar is increasingly externally financed

Sources: Exponential View analysis; company filings.�Note: Debt is net of repayments (not gross issuance) and includes all debt instruments (bonds, commercial paper, etc.).

3 | CapEx: The largest buildout in tech is paying back (for now)

Hyperscaler and neocloud CapEx by funding source

$, 2020-2026E

CapEx by funding source

%, 2020-2026E total

31

Hyperscaler CapEx primarily cash, but risk to wider economy from their market cap weight �vs rest of market

Paying cash keeps a bad bet inside the firm, only denting profits

External funding moves risk outside firms, �as third-parties expect repayment

Neocloud CapEx primarily debt-funded

The 2026E depreciation charge approaches $111 billion

​

3 | CapEx: The largest buildout in tech is paying back (for now)

Source: Exponential View analysis.�Note: IT equipment is depreciated over 6 years, and buildings over 14 years. Required revenue values exclude OpEx.�Headroom is the portion of revenue beyond that required to meet the depreciation expense.

32

Revenues cover the ongoing expense, not yet the cumulative bill

3 | CapEx: The largest buildout in tech is paying back (for now)

Sources: Exponential View analysis; company filings.

Note: Meta contributes to industry CapEx but initiatives are focused on ad uplift, so not recognized as pure GenAI revenue, or currently have minimal direct monetization (e.g. Meta AI assistant, Muse Spark).

Quarterly AI revenues & CapEx depreciation

$bn/quarter, hyperscalers & neoclouds only

Cumulative AI revenues & CapEx depreciation

$bn, hyperscalers & neoclouds only

33

Still ~half-covered: cumulative revenue has nearly covered cumulative depreciation, but still has to cover the expected headroom

​

Q4 2025: Quarterly revenues first exceed CapEx depreciation

AI infra revenue now just clears today’s depreciation hurdle

3 | CapEx: The largest buildout in tech is paying back (for now)

Source: Exponential View analysis.

Headroom after quarterly CapEx depreciation

% = (Revenue – Depreciation) ÷ Revenue

34

All GenAI revenues�32%

19%�Hyperscaler &�neocloud revenues only

• GenAI revenues now cover the quarterly depreciation of AI infrastructure. Q1 26 headroom reached 19% for hyperscaler/neocloud revenues �and 32% across all GenAI revenues.

​

• Coverage remains thin. Depreciation absorbs roughly 81% of hyperscaler/neocloud GenAI revenue and 68% of total GenAI revenue before additional costs.

​

• The next test is incremental coverage. As committed AI capex enters service, the depreciation base will rise. Revenue growth, utilization and pricing must continue to compound or headroom will compress again.

Q4 2025: Quarterly revenues first exceed CapEx depreciation

Rental rates suggest demand is absorbing existing supply

3 | CapEx: The largest buildout in tech is paying back (for now)

Sources: Exponential View analysis; SemiAnalysis.

H100 1-year rental contract price

$/hour/GPU. H100 is the most liquid, most-traded GPU in the merchant market: a useful signal.

35

$3.05: Launch-era scarcity premium

$1.70: Peak overbuild fears as Blackwell ships

$2.40: Surging inference demand

Data center economics set the hurdle for token pricing

Sources: Exponential View analysis; Epoch AI; SemiAnalysis.��Note: Illustrative model. 1GW of IT capacity (6.7k GB200 NVL72 systems, 480k GPUs), cost of ownership per Epoch AI (May 2026), annualized incl. cost of capital, 6-year IT life. Token output from SemiAnalysis InferenceX: FP4, 8k-in/1k-out, 50 tokens/sec/user, 65% utilization (Apr 2026). Token output range reflects +10-25% throughput uplift from speculative decoding. Open-weight models incur no model-licensing fee. Closed-weight column adds a 25% licensing fee of Kimi’s $1.29 blended price.

​

3 | CapEx: The largest buildout in tech is paying back (for now)

36

REQUIRED CUSTOMER VALUE FOR 25% ROI

● Kimi K2.5 (1T) ● Kimi-class model under closed licensing (illustrative)

DIVIDE $7.9bn/yr BY TOKEN OUTPUT:

÷ TOKENS PER GW / YEAR

75-85 quadrillion

= COST FOR INFERENCE PROVIDER

$0.10

$0.42

REQUIRED PRICE FOR 50-75% GROSS MARGIN:

per 1m tokens

$0.20 – $0.40

$0.84 – $1.68

$0.25 – $0.50

$1.05– $2.10

$7.9bn annual cost to own and operate 1 GW of AI capacity

Capital costs $7.0bn/year · 89%

65%

20%

Servers $4.6bn

480,000 GPUs in 6,700 systems · 6-year life

Facility $1.4bn

Building shell, power & cooling · 14-year life

DC network $1.0bn

Switching & interconnect fabric · 6-year life

Land + utility $33m

Land at cost of capital · utility 14-year life

OpEx $900m/year · 11%

66%

Energy $594m

Electricity to run the fleet · 66% of OpEx

Other OpEx $308m

Staff, maintenance & overhead

per 1m tokens

per 1m tokens

$0.32 licensing fee (~25% of the blended selling price of $1.29)

Gross rental yields suggest useful lives extend past six years

3 | CapEx: The largest buildout in tech is paying back (for now)

Sources: Exponential View analysis; Silicon Data.�Note: Yield = (On-demand rate x 50% utilization x 8760 hours) ÷ original list price.

GPU yield at 50% utilization

%, excludes OpEx

37

Older GPUs�earn yields long beyond their six-year depreciation life

Years since release:

1yr

6yr

7yr

8yr

9yr

Depreciation charge (6-year): 17%

Newer GPUs earn yields well above depreciation charge

3yr

4yr

After 6 years, chip CapEx fully depreciated

Longer GPU useful life stretches headroom

3 | CapEx: The largest buildout in tech is paying back (for now)

Sources: Exponential View analysis; Meta.�Note: Buildings depreciated over 14 years as constant.

Range of headroom after CapEx per chip depreciation schedule

Q1 2026, 3-9-year schedules, % = (Revenue – Depreciation) ÷ Revenue

38

If chip life is shorter, revenues don’t repay CapEx (this would require initial H100 purchases becoming obsolete today)

Mark Zuckerberg

Meta Q3 2025 earnings call

​

“... the kind of very worst case would be that we effectively have just prebuilt for a couple of years, in which case, of course, there would be some loss and depreciation, but we’d grow into that and use it over time.”

Overbuild can be a bet on longer chip depreciation

At standard 6-year chip life, Q1 2026 infrastructure revenue headroom is 19%

Extending useful chip life boosts margins: Using chips for 8 years (as old as T4s) raises infrastructure headroom to 36%

4 | Tokens:�The unit of value for the AI economy?

39

Token volumes are growing 14x annually, propelled by agentic workloads and highly elastic demand. Token-based pricing has made this especially pertinent, but it also represents an opportunity for the industry to attribute and evaluate the output from token consumption.

Is the GenAI economy a token economy?�Sort of.

40

“The input is electrons, the output is tokens. In the middle is Nvidia.”

– Jensen Huang

“Tokens, the fundamental units of data our models process…”

​

– Sundar Pichai

Global token volumes exceed 30Q/month, growing 14x YoY

4 | Tokens: The unit of value for the AI economy?

Sources: Exponential View analysis. �Note: Global, inc. China

Inference tokens processed

Quadrillion tokens per month (left axis), growth rate multiple (right axis)

41

Year-on-Year change

Total (API + subscription + internal)

API-only

The transition from chat to agents is multiplying token use

Sources: Exponential View analysis; OpenRouter; Bai, Huang, Wang, Sun, Mihalcea, Brynjolfsson and Pentland 2026.

4 | Tokens: The unit of value for the AI economy?

​

Agent coordination density

% tool use per prompt, OpenRouter

Token consumption per task

Average tokens consumed per task, log scale

42

Cheaper tokens and better models amplify demand

Sources: Exponential View analysis; Epoch AI.

4 | Tokens: The unit of value for the AI economy?

​

43

• More capable models able to cover a wider range of economically useful tasks, increasing value and use.

​

• Token volume rising as reasoning models spend more tokens “thinking”.

​

• Price declines encourage more use and make previously uneconomical applications viable.

Tokens processed per output token:

12 → 36

Blended price per million tokens:

$17 → $2

Epoch Capabilities Index:

112 → 158

Token demand appears elastic: As prices fall, usage grows faster

4 | Tokens: The unit of value for the AI economy?

Sources: Exponential View analysis; Google; OpenAI; ByteDance.�Note: Time-series, not cross-sectional: price and usage both trend with time, so β may overstate pure price-elasticity.

Google price elasticity (avg. price vs volume)

$/million tokens vs trillion tokens/month, log-log scale, 2023-2026

44

Across providers, magnitude of elasticity ≈ 1.2-1.8:�every 10% price cut → 12–18% more tokens → total token spend still rises

Sam Altman OpenAI: “Three Observations”, 2025

“The cost to use a given level of AI falls about 10x every 12 months, and lower prices lead to much more use”

​

Price −90% | Volume ↑

Sundar Pichai Google: I/O 2025

“…we were processing 9.7 trillion tokens a month. Now, over 480 trillion — 50x more.”

​

Price −97% | Volume 50x

Tan Dai Volcengine / ByteDance, 2025

“Doubao’s daily token usage exceeded 50 trillion this month, up from �4 trillion in Dec 2024.”

​

Price −50% | Volume 12x

trend β = -1.7

R² = 0.93

Increasing usage

Decreasing prices

Token-based pricing is AI’s ‘pay-per-click’ moment

4 | Tokens: The unit of value for the AI economy?

​

Sources: Exponential View analysis; IAB/PwC Internet Advertising Revenue Report.

Evolution of AI pricing models

Annual digital ad revenue

$bn/year, log scale, 1996-2024

45

Attribution with CPC enabled a sustainable and�profitable market. Spend could be linked to ROI.

Untracked banner ads had a limited market size and crashed with the dot-com bust

2002: Pay-per-click (Google AdWords)

1

FREE

Unrecognized in budgets

2

SUBSCRIPTION

Seat-based pricing without tracing to specific value

3

TOKEN-BASED PRICING

Metered usage enables (and requires) attribution to projects

$500m

Each generation lifts token output per gigawatt of capacity

Source: Exponential View analysis. Global ex-China.

4 | Tokens: The unit of value for the AI economy?

​

Tokens produced per gigawatt of data center capacity

Trillion tokens per GW per month

46

• Every gigawatt buys more token output each month.�
• This enables a supercycle, even as physical constraints put pressure on the AI economy.�
• Driven by:
1. Labs achieving higher efficiency through smaller models and better serving.
2. Hardware gains, although slower moving (e.g. Hopper → Blackwell → Rubin).
3. Workload mix shifting to inference vs training.

This efficiency is increasing monetization per GW of capacity while revenues per token fall

Source: Exponential View analysis. Global ex-China. Annualized revenue.

4 | Tokens: The unit of value for the AI economy?

​

Revenue generated from tokens & data center capacity

$m/TTok (left axis), $bn/GW (right axis)

47

• Revenue per trillion tokens has fallen since its 2023 peak, mirroring price declines.

​

• Efficiency gains drive lower token prices, which are more than offset by higher demand.

​

• Industry-wide revenue per GW of data center capacity passed $7bn/GW.

Tokens are AI’s billing metric, but not yet a unit of value

4 | Tokens: The unit of value for the AI economy?

​

48

INTERNET

Pageviews

Sessions & clicks

Advertisers pay on attention and conversion:

CPM → CPC → CPA

MOBILE INTERNET

Megabytes

Active users

Apps valued on engagement:

DAU / MAU, retention

GEN AI

Tokens

“Intelligence”?

The value-producing unit is still undefined

ELECTRICITY

Lamps

Kilowatt-hours

Edison’s first customers paid per light bulb installed. Metering came later.

Quality-adjusted tokens come closest to a usable unit of value

4 | Tokens: The unit of value for the AI economy?

​

49

Output tokens

Input tokens

Reasoning tokens

Input & reasoning tokens are a cost of production: exclude them

More consumption here often buys more capability, but it’s not the user-facing output

Capability

×

Scale for quality of output

Output from “better” models is more valuable to users

=

Intelligence

Quality-adjusted �output tokens

Volume of useful output

Users draw on intelligence to complete tasks of increasing complexity

Token volume

Token value

Regardless of the measure you pick, the trend is on the up

Sources: Exponential View analysis; Epoch AI; METR; Artificial Analysis.�Note: METR tasks primarily consist of software engineering, machine learning, and cybersecurity tasks.

4 | Tokens: The unit of value for the AI economy?

Epoch Capabilities Index

Score: GPT-5 = 150,�Claude 3.5 Sonnet = 130

METR Task Horizon

Human task duration with 50% model success rate, log scale

Artificial Analysis Intelligence Index

Score: 0-100

50

Quality-adjusted output kept pace with raw volume growth

Sources: Exponential View analysis; OpenRouter; Epoch AI; METR; Artificial Analysis.�Note: ECI, METR & AAII indexed to average Jan 2025 scores for comparative purposes.

4 | Tokens: The unit of value for the AI economy?

​

Tokens: Total, output & quality-adjusted

Trillion tokens per month, Jan 2025 vs Apr 2026

51

• Raw output tokens grew more slowly than total tokens (30x vs 39x): Increasing volumes spent on input and reasoning.

​

• Wide spread in score improvement (33x-59x): �Different measures and different scoring scales mean we can only draw directional conclusions.

​

• Quality-adjusted output tokens seem to be growing: Output volumes and capabilities both higher than January 2025.

Total tokens

Output�tokens

+39x

+30x

+33x

+59x

+35x

Total tokens

Output tokens

ECI-adjusted

METR-adjusted

AAII-adjusted

5 | Stack:�Where the value is captured

52

Revenue is concentrated, but apps and models are gaining share. Labs retain pricing only while they hold the frontier; the economic value of yesterday’s frontier diminishes quickly into open weights. Where margin accrues across the stack will depend on technical advancements, which cut across competitive dynamics.

The stack turns capital and energy into cognitive work

5 | Stack: Where the value is captured

​

53

Revenue is concentrated today, but the mix is shifting

Sources: Exponential View analysis; company filings. �Note: Revenues not subject to deduplication adjustment.

5 | Stack: Where the value is captured

Annualized GenAI revenue

$million, Q1 2024

Annualized GenAI revenue

$million, Q1 2026

54

Value is moving up the stack, towards apps and models

5 | Stack: Where the value is captured

​

Source: Exponential View analysis.�Note: Global ex-China. Excludes chips (which are capitalized as CapEx by the hosting layer). �Figures may not sum to 100% because of rounding.

Quarterly GenAI revenues, deduplicated by layer to reduce double-counting

$bn per quarter

55

2.95x in one year

7%

11%

82%

4%

9%

87%

3%

8%

89%

Pricing power follows competitive pressure, not stack position

Source: Exponential View analysis. �Note: Hosting & Apps are defined by share of revenues. Foundation Models are defined by share of tokens (due to open-weight competition). Chips are defined by share of compute (H100-eq) due to vertical integration of dedicated hyperscaler chips. Herfindahl–Hirschman index is a measure of market concentration.

5 | Stack: Where the value is captured

​

Market share of leading companies & Herfindahl–Hirschman Index (HHI)

% per layer (left axis) & HHI (right axis)

56

• Upstream suppliers price tokens to capture all the margin available, absent competition downstream.

​

• Nvidia is the largest chip provider, but vertical integration by AWS & Google into custom silicon may reduce its prominence.

​

• FM revenues are concentrated with OpenAI & Anthropic, but open-weight models offer low-cost competition for quality tokens.

Largest provider

Others

HHI

Frontier labs can defend premium pricing, for now

Sources: Exponential View analysis; Epoch AI.

5 | Stack: Where the value is captured

​

Epoch Capabilities Index

Score at release, OpenAI/Anthropic/open-weight

Output price

$/million output tokens

57

OpenAI

Anthropic

Others

Labs must outrun open-weight commoditization �to hold margin

5 | Stack: Where the value is captured

​

58

Lightweight open-weight

Summarize an article

Leading open-weight

Draft working code

Closed-weight frontier

Multi-day agentic work

Last year’s frontier is commoditizing fast

Sources: Exponential View analysis; Epoch AI.

5 | Stack: Where the value is captured

​

Price per capability frontier

Blended price $, grouped by performance at GPQA Diamond (PhD-level science)

59

Labs earn a time-limited premium on the current frontier

Among self-selecting OpenRouter users, token share is moving to open-weight

Sources: Exponential View analysis; OpenRouter.�Note: While OpenRouter is not a cross-section of the market, its data shows the behavior of self-selecting “model-routing” users.

5 | Stack: Where the value is captured

​

Weekly OpenRouter token share

% per model author

60

72%

33%

Google + OpenAI + Anthropic:

Under pricing pressure, labs push into apps and infrastructure

5 | Stack: Where the value is captured

​

Sources: Exponential View analysis; TechCrunch; OpenAI; Anthropic.

61

Hosting

Apps

Foundation�models

Consumer�surplus

Today

Building vertical apps: Law

Claude for Legal

Codex for Legal

Compress every layer, and consumers capture the surplus

5 | Stack: Where the value is captured

​

62

Today

Scenario 2: General-purpose models

Scenario 3: Reduced compute needs

Scenario 4: All 3 combined

Scenario 1: Open-weight models catch up

Tokens are cheaper for apps to consume & �hosting to serve

Foundation model labs build application layer�and their own infrastructure

Frontier model labs absorb integration work; generic "wrapper" pricing power is squeezed

Models and apps cost less to run, growing the applicable market

Cost reductions grow the applicable market

Significant consumer surplus as benefits delivered beyond price

Some local hosting, with chips distributed�to edge devices

Apps defend with proprietary data, domain-specific workflows & own models

Labs unable to charge license fee for �frontier models

Frontier models can replace AI application/integration layer

Models become smaller, and compute �more efficient

The perfect storm of efficiency improvements

AI demand is more revenue-validated than �any prior platform shift.

The investment case comes down to whether falling prices can move enough token volume to earn a return on CapEx.

63

64

What we count in, and what we exclude

• We count revenue at every layer of the stack: apps (subscriptions and AI-first software), foundation-model APIs, and AI cloud and compute sold as discrete services. Each layer represents real spend by a paying customer. A company that operates across layers (e.g. foundation model providers with customer-facing apps) has its revenue split across them. We count global ex-China revenues, and exclude chips and hardware (a cost to the compute layer, not a customer payment), AI features in legacy software, advertising uplift from AI (primarily Alphabet and Meta). We also exclude CapEx and financing from measures of revenue.
• CapEx is counted as the AI-attributable portion of the seven stack-builders’ infrastructure spend (hyperscalers and neoclouds), including both cash PP&E �and leases.
• Tokens are counted as every token processed, input and output, across all major providers and surfaces.

How we deduplicate, and why

• Revenue is counted at each layer but never summed across them: $100 of app spend that sends $60 to a model provider which in turn spends $30 on cloud hosting for inference is attributed (in line with added value) $40/$30/$30 to sum to the same $100 without double-/triple-counting, which would otherwise result in an erroneous $190 figure.
• CapEx is counted once, on the balance sheet of the entity that actually owns the asset. Compute that is jointly leased, or rented by a foundation model provider from a hyperscaler, sits with the owner-operator: not also with the renter.
• For tokens: when inference is served by a foundry running another company’s foundation model, those tokens are attributed once, to the model actually run, so a model offered through foundries isn’t double-counted.

Sources

All figures are built bottom-up from primary and specialist sources, and triangulated against top-level estimates and proxies. Revenue and CapEx are grounded on company filings (SEC 10-K, 10-Q and 8-K) and executive disclosures, cross-checked through cloud attribution (e.g. Azure and Bedrock) where a private company’s revenue appears in a public firm’s accounts. Our systems daily scan and crawl available sources to create, maintain and improve the breadth and depth of our data, with source attribution and confidence grading. This high-quality dataset is used to build up full financial models (including P&L) for major companies, and driver-based models for smaller companies.

CapEx’s AI share is carved per company and reconciled against silicon (chip providers’ revenue), build cost, segment composition, and sell-side research. �Token volumes are reconciled from executive statements, third-party sampling and analysis, traffic volumes, and triangulated against revenues and available compute capacity.

Methodology: How we count revenues, CapEx, tokens

Authors

Marija Gavrilov

Azeem Azhar

Hannah Petrovic, PhD

Nathan Warren

William Gildea

Managing Director

Founder

​

​

​

Senior Researcher

​

Senior Researcher

​

Product Manager

​

​

We welcome feedback and contributions �at aieconomy@exponentialview.co

For advisory requests and institutional inquiries, please contact helen@exponentialview.co

Follow our analysis on

Exponential View

66

(www.exponentialview.co)

Subscribe to Exponential View to receive our research in your inbox each week