Diffusion CapitalDiffusion Capital

Monitor

The AI economy.

Track model capabilities, usage, prices, compute economics and infrastructure spending.

Latest reading: 2 November 2026

Model capabilities

How capable the best AI models are and how fast open models catch up. Epoch AI's Capabilities Index combines many benchmarks into one score, the Remote Labor Index measures the share of real freelance projects AI completes, and SemiAnalysis's GDPx scores models on long agent tasks in 11 fields.

Epoch Capabilities Index

Best: GPT-6 Astra · OpenAI.

169.1

+15.8 vs best a year earlier

2 November 2026Source: Epoch AI

Remote Labor Index

Best: Claude Opus 5.5 · Anthropic.

21.3%

+18.8 pp vs best a year earlier

22 October 2026Source: Scale AI

Best GDPx score

Best: Claude Opus 5.5. Open: DeepSeek 4.1 Flash, 5.4%.

12.0%

1 November 2026Source: SemiAnalysis

Open-weights lag

DeepSeek V4.1 Flash 1017 matches closed models from Mar 2026.

8.0 mo

+1.4 mo vs a year ago

2 November 2026Source: Epoch AI

Epoch Capabilities Index by model and release date

232 released models; the dashed line is the best so far · latest release 26 October 2026

Best released so far
Click a lab to show only its models

Epoch Capabilities Index (ECI)

80100120140160180Apr 24Jul 24Oct 24Jan 25Apr 25Jul 25Oct 25Jan 26Apr 26Jul 26Oct 26

Dots show released models' latest scores (CC BY 4.0); the dashed line shows the best so far. The linear trend fits record-setting models and reports points/year; hover for model details.

Source: Epoch AI, Capabilities & benchmarking ↗

GDPx score by field and model

6 of the 20 models tested, each field's best in bold · 1 November 2026

LowerHigher(0.0% … 40.0%)
ClaudeOpus 5.5Gemini 4CarbonGPT-6AstraMuse WonderFullClaudeFable 5.1GPT 6.1SolCybersecurity9.316.70.035.23.70.0Quant trading27.413.120.218.328.617.1Biology20.525.65.15.120.52.6Recursive self-improvement17.619.124.915.816.522.8Materials science16.520.621.610.99.719.4Hardware design18.75.816.51.78.311.4Software engineering8.911.16.72.24.48.9Alignment8.73.46.29.21.25.1Chip design4.81.44.40.60.85.6Finance0.02.80.00.00.00.0Legal0.00.00.00.00.00.0

Shows every model that leads a field. Safety filters lower some models' scores in cybersecurity, biology and alignment.

Source: SemiAnalysis GDPx ↗

Remote Labor Index: best score so far

Share of real freelance projects completed · 22 October 2026

Projects completed, %

0%5%10%15%20%25%Oct 25Nov 25Dec 25Jan 26Feb 26Mar 26Apr 26May 26Jun 26Jul 26Aug 26Sep 26Oct 26Claude Opus 5.5 21.3%

Each logo marks a new record, dated by when Scale added the model. Experts check whether the client would accept each result.

Source: Scale AI & Center for AI Safety ↗

What markets expect

AI usage

How much AI is used. First as the companies report it: tokens processed (a token is a word fragment, the unit AI models read and write), users and revenue. Then as measured on OpenRouter, a marketplace that routes developers' requests to hundreds of models.

Google tokens per day

Across Google's products and API, as Google reports it.

107T

×6.7 vs a year ago

May 2026Source: Google

Doubao tokens per day

ByteDance's Doubao models, daily average, as ByteDance reports it.

180T

×11.0 vs a year ago

Jun 2026Source: Volcano Engine

ChatGPT weekly users

As OpenAI reports it, collected by Epoch AI.

1.2B

×1.5 vs a year ago

Sep 2026Source: Epoch AI

Anthropic annualised revenue

Latest month's revenue times 12, as reported, collected by Epoch AI.

$65B

×13.0 vs a year ago

Jul 2026Source: Epoch AI

AI usage as the companies report it

Tokens processed per day, as the companies state them · latest report Oct 2026

Click a line to show only it

Tokens per day

100B1T10T100T1Q202420252026

Each dot links to a company statement, or China's National Data Administration; monthly and per-minute figures become daily tokens. Google includes AI in Search; the dashed China series has no published method.

OpenRouter tokens per week

On OpenRouter, a marketplace that routes requests to many models.

193.6T

×39.6 vs a year ago

Week to Nov 1Source: OpenRouter

Largest maker on OpenRouter

DeepSeek's share of the week's tokens on OpenRouter.

32.5%

+5.1 pp vs prior

Week to Nov 1Source: OpenRouter

Chinese models on OpenRouter

Tokens from Chinese makers' models; the true share may be higher.

51.4%

+1.7 pp vs prior

Week to Nov 1Source: OpenRouter

Open-weights share on Vercel

Share of Vercel AI Gateway tokens served by open-weights models.

72.1%

−1.0 pp vs previous day

1 November 2026Source: Vercel AI Gateway

Weekly tokens on OpenRouter by model maker

Weekly · latest week: 1 November 2026

DeepSeekTencentZ.ai (Zhipu)XiaomiOpenAIAnthropicOther models
050T100T150T200TTokens per weekJanMarMayJulSepNov

Weekly OpenRouter tokens by model maker: the six largest makers of the last six months, the rest under "Other models". OpenRouter names only its top nine models each week.

Source: OpenRouter ↗

Open-weights share of tokens: OpenRouter and Vercel

Weekly · data to 1 November 2026

OpenRouterVercel AI Gateway
0%20%40%60%80%100%Share of tokensJanMarMayJulSepNov

Weekly open-weights token share: OpenRouter's top models (62% of last week's tokens) versus all Vercel tokens. The samples are not directly comparable.

Source: OpenRouter ↗

PNC households paying for AI

2.2% of PNC households paid for AI in May 2026 · monthly

Approximate chart readingPublished rounded value
0.00%0.50%1.00%1.50%2.00%2.50%Share of PNC households, percent2.2%Jan 23Jul 23Jan 24Jul 24Jan 25Jul 25May 26

PNC households with identified Gen AI subscription payments, through May 2026. Approximate monthly chart readings; report published 15 June 2026.

Source: PNC ↗

Anthropic vs OpenAI revenue mix

Usage = API and apps; subs = subscriptions · Anthropic reported, OpenAI estimated

Anthropic UsageSubsOpenAI UsageSubs
0.0%25.0%50.0%75.0%100.0%Revenue share, percent202420252026

OpenAI figures are estimates: 2024 projection, 2025 assumption, Sep 2026 run-rate scenario excluding 2% other. Anthropic reported; 2026 is H1.

Source: Anthropic S-1 ↗ · Source: Tanay (2024) ↗ · Source: Model

ChatGPT use by topic

Share of consumer messages by topic, Jul 2024 to Sep 2026

WritingPractical guidanceSeeking informationTechnical helpMultimediaSelf-expressionOther
0%25%50%75%100%Share of messagesJul 24Jan 25Jul 25Jan 26Sep 26

Consumer ChatGPT plans only. Shares of messages, labelled by OpenAI's classifiers; each release restates past months. Our reading: technical help fell partly because coding moved to Codex, which the data leave out.

Source: OpenAI Signals v3.0 (CC BY 4.0) ↗

ChatGPT use by purpose

Share of consumer messages by purpose, Jul 2024 and Sep 2026

AskingDoingExpressing
0%25%50%75%100%Work, Jul 2024Work, Sep 2026Non-work, Jul 2024Non-work, Sep 2026

Consumer ChatGPT plans, labelled by OpenAI's classifiers. Asking: looking for information or advice. Doing: ChatGPT makes something, such as a draft or code. Expressing: neither, such as chat or venting. Each release restates past months.

Source: OpenAI Signals v3.0 (CC BY 4.0) ↗

What markets expect

Read our related note

Price of intelligence

How fast AI gets cheaper at a fixed level of capability and a constant token blend. Compare prices over time, today's price-capability frontier and benchmark cost per task. Platform workload observations are shown separately.

Price of GPT-5-level AI

Cheapest model at GPT-5's level: Claude Haiku 5.5 (ECI 150).

$0.200

÷17.2 vs a year ago

Oct 2026Source: Epoch AI · OpenRouter

Input tokens per output token

Across OpenRouter's traffic, monthly. Agents re-send large contexts.

37×

×3.0 vs a year ago

Sep 2026Source: OpenRouter

Cached share of agent tokens

OpenRouter's cached prompt tokens, share of agent tokens, roughly.

87%

Sep 2026Source: OpenRouter

Tokens per answer

Median output tokens per question, models from the past year.

27k

×1.9 vs models of the year before

Oct 2026Source: Artificial Analysis

Price of a fixed level of AI capability

Cheapest model at each fixed capability level · latest price change Oct 2026

Click a line to show only it

USD per million tokens · 3 input : 1 output · no caching

$0.01$0.1$1$10$100202420252026

Cheapest included model at each fixed ECI threshold. Fixed blend: three input tokens per output token, no caching. Annualised decline uses first and current prices.

Source: Epoch AI ↗ · Source: OpenRouter ↗

Input vs output tokens

Share of each month's tokens · latest month: September 2026

Input (prompt) tokensOutput (completion) tokens
0%20%40%60%80%100%Jan 24Jul 24Jan 25Jul 25Jan 26Jul 26

Monthly OpenRouter input (prompt) and output (completion) tokens. Agents resend long contexts at each step, increasing input's share.

Source: OpenRouter ↗

What caching saves

Caching cuts this agent workload's bill by 75.3% · Sep 2026 example

Cached inputFresh inputOutput

USD per million output tokens

With caching$51.40
Same workload, uncached$208.00
$0$50$100$150$250

Sep 2026 example: 87% cached input, 12% fresh input, 1% output, at Sonnet 5 prices (7 October 2026). Costs per million output tokens, excluding cache-write fees.

Source: OpenRouter ↗

Compute economics

Ornn tracks GPU rental prices. SemiAnalysis's InferenceX models serving revenue and surplus at assumed token prices, costs and paying demand. Use the assumptions panel to compare models and scenarios.

H100 rental price

Ornn index from executed spot rentals, USD per GPU-hour.

$3.02

+22.8% vs 4 weeks ago

1 November 2026Source: Ornn

B200 rental price

Ornn index from executed spot rentals, USD per GPU-hour.

$9.44

+23.9% vs 4 weeks ago

1 November 2026Source: Ornn

Modelled revenue, GB300-hour

Kimi K3 at the source's tariffs, per capacity-hour, best setup.

$42.07

×3.1 since 24 August 2026

2 November 2026Source: InferenceX

Modelled surplus, GW-year

1 GW of GB300s serving Kimi K3 for a year, InferenceX assumptions.

$63.5B

×4.5 since 24 August 2026

2 November 2026Source: InferenceX

GPU rental prices by chip

Daily · latest: 1 November 2026

B200H200H100 SXMRTX PRO 6000A100
$0.00$2.00$4.00$6.00$8.00$10.00US dollars per GPU-hourAugSepOctNov

Ornn's daily 4 PM ET settlement index of executed spot rentals, in USD per GPU-hour. The public series covers the last three months.

Source: Ornn ↗

GPU model age and accounting lives

Years since each GPU model launched against the useful lives builders book · rent, last 4 weeks

Useful life the builders assume
5–6 years
AmazonNebius
Meta
MicrosoftAlphabetOracleCoreWeave
A100
2020
$1.83/h list
+8% since late 2025
H100
2022
$3.60/h list
+20% since late 2025
H200
2024
$4.86/h list
+39% since late 2025
B200
2024
$7.69/h list
+37% since late 2025
0123457

Years since model introduction

Dashed lines: accounting useful lives from filings. Bars: years since model introduction. Prices: provider-balanced advertised medians; listing availability is unverified.

Source: SEC EDGAR ↗ · Source: GetDeploying (CC BY 4.0) ↗

Assumptions

Model · prices and source assumptions
Model
Token price$3.00 in · $15.00 out per M tokens$0.30 cached input per M tokensMoonshot's list price
Assumed lab revenue share30%InferenceX scenario assumption
SemiAnalysis's assumptions · adjustable
Billable capacity sold60%
Chip cost
GB300: $2.31 per hour, InferenceX's figure

Speed target: 60 tokens/s per user. Cache and power assumptions follow InferenceX; annual figures use utility GW.

Same chip, more revenue per hour

Kimi K3: revenue per GPU-hour at list prices, first estimate and today · weekly

First sampledToday
GB300
$42.07 · ×3.1 since 24 August 2026
MI355X
$16.04 · ×7.8 since 17 August 2026
GB200
$12.92 · ×7.4 since 17 August 2026
TPU v7
$10.61 · unchanged since 19 October 2026
B300
$9.41 · unchanged since 17 August 2026
B200
$9.03 · ×1.5 since 24 August 2026
$0$10$20$30$50

Modelled revenue per sold GPU-hour for Kimi K3, from the first sample to today. Each date uses the best eligible setup.

Source: SemiAnalysis InferenceX ↗

Modelled revenue and costs per utility GW-year

Kimi K3: one gigawatt-year in each chip's best setup · InferenceX scenario

Serving surplusAssumed lab shareCostRevenue
GB300Dynamo vLLM
$63.5B surplus · $104.3B revenue
TPU v7vLLM
$24.9B surplus · $46.2B revenue
MI355XvLLM Disagg
$21.9B surplus · $40.3B revenue
GB200Dynamo vLLM
$16.7B surplus · $36.3B revenue
B200Dynamo vLLM
$10.6B surplus · $27.7B revenue
B300vLLM
$7.8B surplus · $26.0B revenue
$0B$50B$100B$150B

InferenceX scenario for Kimi K3 using the assumptions above, with revenue, costs, assumed lab share and surplus for each chip's best setup.

Source: SemiAnalysis InferenceX ↗

What markets expect

Capex

Reported company investment, guidance and future commitments, alongside operating cash flow and financing raised. Company figures cover different businesses, fiscal periods and lease definitions. Figures come from filings and earnings calls; sums and ratios are our arithmetic.

Capex, 4 quarters to Q3 2026

Six largest AI builders, from Q4 2025; quarters all six reported.

$741B

+22.6% vs four quarters to Q2 2026

Q3 2026Source: SEC EDGAR

Company capex guidance

10 companies; mixed fiscal years and lease definitions.

$997B

+$35B vs first guidance

29 October 2026Source: SEC EDGAR

Off-balance-sheet commitments

$1.35T of leases not yet started, $1.93T of purchase commitments

$3.28T

×4.4 Jun 2025 to Jun 2026

24 October 2026Source: SEC EDGAR

Capex as share of cash flow

Six largest AI builders. Above 100%, they use cash, bonds or shares.

119%

+14 pp vs prior

Q3 2026Source: SEC EDGAR

Capex per quarter

Six companies, calendar quarters · latest: Q3 2026

AmazonAlphabetMicrosoftMetaOracleSpaceX
$0B$50B$100B$150B$200B$250BCash capex, US dollarsQ1 2025Q2 2025Q3 2025Q4 2025Q1 2026Q2 2026Q3 2026

Calendar-quarter cash capex. Incomplete quarters show how many of the six companies have reported.

Source: SEC EDGAR ↗

Historic US investment and the AI build-out

Peak annual spending relative to one year's US GDP · AI: DWS scenario, Sep 2026

Historic US annual peaksUS AI, 2026Other years, 2024–29

Spending relative to one year's US nominal GDP, percent

0%5%10%15%18715.4%Railroads19240.9%Electrification19440.4%Manhattan Project19640.4%Interstates19660.8%NASA (Apollo)20004.4%IT & software20141.0%Shale≈$650B2026≈2.0%AI 2026, US≈$3.85T2029≈10.3%AI total, 2024–29 US

DWS's Sep 2026 illustrative US scenario includes chips: $650B in 2026, $3.85T gross in 2024–2029 including 2026. The total uses 2029 GDP, extrapolated at 4.5% nominal growth.

Source: DWS US scenario (Sep 2026) ↗ · Source: IMF WEO (Oct 2026) ↗

Investment and the timing of productivity gains

US nonfarm business · productivity growth 1.6% → 3.0% per year

CAPEX share of GDPProductivity period starts 1996
1.0%2.0%3.0%4.0%5.0%1990199119921993199419951996199719981999200020012002200320042005

Annual CAPEX/GDP and the start of the productivity period. Growth rates are sector output per hour; they do not isolate the technology's causal effect.

Source: BEA/FRED ↗ · Source: BLS / FRED ↗ · Source: Goldman Sachs ↗

Promised beyond capex

Capex plus spending committed but not yet started · quarterly filings

CapexFuture leasesGPU lease letterPurchases
Alphabet
$1.21T
Meta
$915B
Microsoft
$729B
Amazon
$630B
Oracle
$398B
SpaceX
$134B
00.51$1.5T

Latest four quarters of capex plus off-balance-sheet leases not yet started, non-cancellable purchases and SpaceX's GPU lease letter of about $40B.

Source: SEC EDGAR ↗

Capex versus operating cash flow

Six builders' capex against their operating cash flow · quarterly filings

capex = cash flowcapex > cash flowbubble = bonds + shares raised
$0B$0B$50B$50B$100B$100B$150B$150B$200B$200B$250B$250BOperating cash flow, last four quartersCapex, last four quartersAmazonAlphabetMicrosoftMetaOracleSpaceX

OCF may include customer advances. Bubbles show company-wide bond principal/carrying values and equity proceeds since 2025.

Source: SEC EDGAR ↗

New AI capacity in 2027

Capacity expected to come online worldwide in 2027 alone.

+50 GW

19 October 2026Source: SemiAnalysis

AI revenue, annualised

End-customer AI spending, counted once, outside China.

$229B

×3.5 vs a year ago

Aug 2026Source: Exponential View

AI capex, 2024 to 2029

Chips, data centres and energy worldwide; about $5T funded by debt.

~$11T

25 August 2026Source: SemiAnalysis

Capex and published AI revenue scenarios

Annual capex guidance against a published return scenario and today's AI revenue

US dollars a year

Company capex guidance
JPMorgan revenue scenario
AI revenue today
00.20.40.6$1T

Company guidance uses mixed fiscal years and lease definitions. JPMorgan and Goldman figures have different return-model scopes; EV measures end-customer revenue outside China.

Source: SEC EDGAR ↗ · Source: JPMorgan ↗ · Source: JPMorgan range (24/7 Wall St.) ↗ · Source: Goldman Sachs (Bloomberg) ↗ · Source: Exponential View ↗

AI revenue growth scenarios

Annualised AI revenue, then fixed growth rates · latest Aug 2026

AI revenue, annualised3.5× a year2× a year1.5× a year

US dollars a year

Use the left and right arrow keys to inspect months. Press Escape to close the readout.

$0B$500B$1T$1.5T2026202720282029JPMorgan, 10% returnGoldman, break-even

EV's annualised revenue extended at fixed growth rates. Markers show $650B, JPMorgan's scenario with a different investment scope. Crossing it does not establish a return.

Source: JPMorgan ↗ · Source: JPMorgan range (24/7 Wall St.) ↗ · Source: Goldman Sachs (Bloomberg) ↗ · Source: Exponential View ↗

What markets expect

Public opinion and politics

What Americans think of AI and the build-out, and how politics responds ahead of the midterms on 3 November. Most oppose a data centre in their community, and support for safety rules spans both parties. AI money is now flowing into campaigns, and 34 local data-centre measures are on the ballot.

Oppose a local AI data centre

US adults, September 2026; 21% support one.

72%

+7 pp vs March

30 September 2026Source: Quinnipiac

Back AI safety standards

Even if it slows AI; 9% oppose. US adults, September 2026.

86%

30 September 2026Source: Quinnipiac

Spent by AI-focused PACs

Federal spending by the nine PACs below; one filing conflicts.

$80.0–80.2M

30 October 2026Source: FEC

Data-centre votes, 3 November

20 in Ohio. All 4 earlier 2026 votes went against them.

34

21 September 2026Source: Ballotpedia

Where the parties differ on AI

Share agreeing, by party · Pew, Gallup and Quinnipiac polls, 2026

DemocratsRepublicans

Share agreeing, percent

0%25%50%75%100%More concerned thanexcited about AI+7AI will mean fewerjobs in 20 years+7No confidence governmentcan regulate AI+13Strongly oppose alocal AI data centre+17Data centres do more harmthan good (Pennsylvania)+40

Party gaps in percentage points; Pennsylvania samples likely voters. Pew Research Center bears no responsibility for the analyses or interpretations of the data presented here.

Source: Pew Research Center ↗ · Source: Gallup ↗ · Source: Quinnipiac ↗

Concern about AI, by country

Adults more concerned than excited about AI · 37 countries, spring 2026

United StatesMedian of 37 countries 37%

More concerned than excited, percent

0%20%40%60%Australia52%United StatesBrazilCanadaGreeceItalyPolandPalestineUKNetherlandsColombiaSwedenChileFranceHungarySpainMexicoIndonesiaPeruArgentinaPakistanKenyaGermanySouth AfricaMalaysiaPhilippinesJapanTürkiyeThailandGhanaSingaporeNigeriaIndiaIsraelSouth KoreaBangladeshSri Lanka

Adults more concerned than excited about AI: 37 countries, spring 2026. Pew Research Center bears no responsibility for the analyses or interpretations of the data presented here.

Source: Pew Research Center ↗

AI PAC spending on federal races

Independent spending by AI-linked PACs in 2026 federal races · FEC filings to Oct 30, 2026

Pro-industry · $37.1–37.3MPro-safety · $42.9M

Independent spending, US dollars

051015$20MThink Big$18.6MAmerican Mission$16.1–16.3MBuilding the Future$2.5MJobs and Democracy PAC$18.5MDefending Our Values PAC$7.9MTomorrow Together$6.1MYou Can Push Back$4.3MGuardrails Alliance$3.6MDream NYC$2.5M

Federal candidate spending in 2026, checked 30 October 2026. Range: one conflicting $0.2M filing. Logos identify backers or personal donor affiliations; details on hover.

Source: FEC ↗ · Source: Transformer ↗

AI and data-centre policy

As of Oct 28, 2026

Loading measures…

Selected federal and state measures. Open an entry for its scope, dates and source.

Funding and coverage

Money received by each of the 13 identified federal AI-policy PACs, at its latest reporting date. Parent groups pass money to spending PACs, so these receipts cannot be added together. The card and bars count candidate spending once. Funding reports can lag spending notices.

Funding and coverage
PACMoney receivedReported to
Think Big$25.0M14 October 2026
American Mission$24.0M14 October 2026
Building the Future$0.0M30 September 2026
Jobs and Democracy PAC$20.7M30 September 2026
Defending Our Values PAC$8.4M30 September 2026
Tomorrow Together$2.5M14 October 2026
Guardrails Alliance$3.2M14 October 2026
Dream NYC$2.8M31 August 2026
You Can Push Back$4.5M14 October 2026
Leading the Future$76.0M14 October 2026
Public First$4.9M30 September 2026
Innovation Council$0.0M30 April 2026
Race for the Future$0.0M14 October 2026

The range reflects an American Mission notice for $200,000, filed on 11 October and absent from its later regular report. No amendment or withdrawal was identified. The lower figure follows the regular reports plus later spending; the upper figure includes this notice too.

Related groups below have different mandates or reporting systems. State races, ordinary company PACs and policy advocacy are outside the bars. Donors behind nonprofit intermediaries are not fully disclosed, so there is no verified total for all US AI political funding.

Anthropic says its $40M to Public First Action is restricted to policy and public education. The Anthropic logos above refer to individual donors, including Dario Amodei and Jan Leike.

What markets expect

Figures are shown as published by the named source and are revised by their publishers. This page is for information only and is not investment advice or a recommendation to buy or sell any security. This dashboard is maintained and updated with the help of AI. Errors can occur; we apologise for any.