Skip to content
Slow Variables

Ask the data

Answers come from the same store as the site; every number is checked against the record it cites.

Capture lens

← Switch to diffusion lens

As of 2026-09-30: compute & physical concentrating, model consistent with normal, training input faster than normal, serving & orchestration unmeasured, deployment & application consistent with normal, adopters consistent with normal, labour & consumers consistent with normal.

Who keeps the surplus, layer by layer. The compute layer's bar is sized by its filed segment operating income; a layer without a filed margin series shows a hairline, and its indicators carry the reading.

1. Compute & physical

concentrating13 indicators
semis 57.3% · cloud 42.7% as of 2026-06-30

2. Model

consistent with normal22 indicators
unmeasured

3. Training input

faster than normal4 indicators⇄ return arrow
unmeasured

4. Serving & orchestration

unmeasured
unmeasured

5. Deployment & application

consistent with normal9 indicators
unmeasured

6. Adopters

consistent with normal10 indicators
unmeasured

7. Labour & consumers

consistent with normal15 indicators
unmeasured

Bar = share of filed segment operating income across NVIDIA, AMD, AWS, Google Cloud and Microsoft Intelligent Cloud, the one layer group with segment filings. A layer without a filed or estimated margin series shows a hairline.

Margin stack by quarter · share of filed segment operating income

24-06
24-09
25-03
25-06
25-09
26-03
26-06
SemisCloudLabs and apps: no filed or estimated margin series yet, so their share is unmeasured rather than drawn.

  1. 1. Compute & physical

    concentrating13 indicators

    Semis, hyperscalers and neoclouds, data centres and power, memory/networking.

    Rent concentration in semiconductorsgrade A
    concentratinglagging85/95 confidence, good evidence, some ambiguity
    75.0%as of 2026-07-26(2 obs)
    Cloud backlog (remaining performance obligations)grade A
    concentratingleading80/95 confidence, good evidence, some ambiguity
    $638Bas of 2026-05-31
    Semiconductors' share of stack operating incomegrade A
    concentratinglagging70/95 confidence, good evidence, some ambiguity
    57.3%as of 2026-06-30(8 obs)
    Cloud segment operating marginsgrade A
    stablelagging75/95 confidence, good evidence, some ambiguity
    39.4%as of 2026-06-30(2 obs)
    Circular financing scalegrade A
    concentratingleading75/95 confidence, good evidence, some ambiguity
    $903Bas of 2026-08-17(14 obs)
    Concentration among covered chip filers (HHI)grade A
    unclearleading60/95 confidence, mixed or hard to operationalise
    0.85 indexas of 2026-06-30(2 obs)
    Concentration among covered cloud filers (HHI)grade A
    dispersingleading60/95 confidence, mixed or hard to operationalise
    0.35 indexas of 2026-03-31(3 obs)
    Non-Nvidia share of AI compute shipped (H100-equivalents)grade B
    emergingleading50/95 confidence, mixed or hard to operationalise
    26.0%as of 2026-03-31(3 obs)
    Hyperscaler capex to AI revenue (trailing year)grade A
    dispersingleading40/95 confidence, limited or vague
    2.51×as of 2026-06-30(16 obs)
    Hyperscaler server useful lives (10-K estimates)grade A
    not yet measurablelagging65/95 confidence, mixed or hard to operationalise
    5.00 yearsas of 2025-12-31
    Neocloud borrowing cost (CoreWeave facility rates)grade A
    not yet measurablecoincident60/95 confidence, mixed or hard to operationalise
    10.5%as of 2026-06-30(4 obs)
    Training power draw growth (doubling time)grade B
    emergingleading50/95 confidence, mixed or hard to operationalise
    506 daysas of 2026-08-07(213 obs)
    Hardware price-performance growth (doubling time)grade B
    consistent with normalleading40/95 confidence, limited or vague
    1282 daysas of 2025-11-06(20 obs)

  2. 2. Model

    consistent with normal22 indicators

    Frontier labs, neolabs, open-weight and Chinese labs.

    Frontier-lab revenue run-ratesgrade C
    concentratingcoincident55/95 confidence, mixed or hard to operationalise
    $40.0Bas of 2026-08-13
    Lab recoupment ratiograde C
    emerginglagging50/95 confidence, mixed or hard to operationalise
    0.22×as of 2026-08-13(12 obs)
    Price per unit of capability (cheapest METR-hour on OpenRouter)grade A
    not yet measurablecoincident40/95 confidence, limited or vague
    $0.30as of 2026-09-10(32 obs)
    Frontier labs' share of tracked venture dollarsgrade A
    stableleading50/95 confidence, mixed or hard to operationalise
    94.8%as of 2026-06-30(32 obs)
    Frontier lab valuation to run-rate multiplegrade C
    emergingleading45/95 confidence, limited or vague
    28.3×as of 2026-06-16(10 obs)
    Lab vertical integration (acquisitions and acquihires, trailing four quarters)grade A
    concentratingleading55/95 confidence, mixed or hard to operationalise
    10as of 2026-09-30(11 obs)
    Open-weights gap to the closed frontier (ECI points)grade A
    dispersingleading55/95 confidence, mixed or hard to operationalise
    5.77 indexas of 2026-08-20(3 obs)
    Enterprise multi-model usage share (Menlo)grade B
    not yet measurablecoincident45/95 confidence, limited or vague
    60.0%as of 2023-11-10
    Model-layer token concentration on OpenRouter (HHI, top table)grade B
    not yet measurablecoincident40/95 confidence, limited or vague
    0.11 indexas of 2026-09-09(19 obs)
    METR 50% time horizongrade A
    faster than normalleading80/95 confidence, good evidence, some ambiguity
    17.4 h(8.5 h–55.1 h)⚑disputedas of 2026-04-07
    METR 80% time horizongrade A
    faster than normalleading70/95 confidence, good evidence, some ambiguity
    3.1 h(1.6 h–6.6 h)as of 2026-04-07
    50%/80% horizon ratiograde A
    consistent with normalleading70/95 confidence, good evidence, some ambiguity
    6.34×as of 2026-03-05(2 obs)
    Agent-workdays per human workday in frontier research (self-reported)grade D
    emergingleading30/95 confidence, limited or vague
    3.10×as of 2026-08-15
    Human interventions on 4–8 hour agent tasks (self-reported)grade D
    emergingleading30/95 confidence, limited or vague
    50.0%as of 2026-07-31
    Safety brakes inside the labs (events, trailing year)grade D
    emergingleading40/95 confidence, limited or vague
    3as of 2026-09-01(3 obs)
    Frontier training compute growth (doubling time)grade B
    emergingleading60/95 confidence, mixed or hard to operationalise
    170 daysas of 2025-07-09(66 obs)
    Inference price at fixed capability (halving time)grade B
    emergingcoincident45/95 confidence, limited or vague
    68 daysas of 2024-12-13(6 obs)
    ARC-AGI-2 best scoregrade B
    emergingleading55/95 confidence, mixed or hard to operationalise
    95.0%as of 2026-09-03(203 obs)
    Ord half-life misfit (observed 50/80 ratio vs constant hazard)grade A
    consistent with normalleading60/95 confidence, mixed or hard to operationalise
    2.04×as of 2026-03-05(2 obs)
    Agent reliability across repeated runs (tau-bench pass^k)grade B
    emergingleading55/95 confidence, mixed or hard to operationalise
    56.2%as of 2026-05-18
    Progress against the AI 2027 quantitative predictionsgrade B
    emergingleading50/95 confidence, mixed or hard to operationalise
    75.0%as of 2026-07-31
    Internal-to-public deployment gap at the frontiergrade C
    consistent with normalleading45/95 confidence, limited or vague
    4.00 monthsas of 2026-06-30

  3. 3. Training input

    faster than normal4 indicators⇄ return arrow

    Human data, RL environments, post-training platforms, synthetic data, licensing, evals.

    Public text data exhaustion year (Epoch projection)grade B
    emergingleading40/95 confidence, limited or vague
    2028 yearas of 2024-06-06
    Continual-learning ladder (highest production rung)grade D
    emergingleading45/95 confidence, limited or vague
    3.00 rungas of 2026-05-27(2 obs)
    Active RL-environment vendors (directory count)grade B
    emergingleading35/95 confidence, limited or vague
    24as of 2026-07-15
    Expert-data market run-rate (Mercor)grade C
    faster than normalleading70/95 confidence, good evidence, some ambiguity
    $2.0Bas of 2026-06-30

  4. 4. Serving & orchestration

    unmeasured

    Inference serving, routing and gateways, agent infrastructure.

    No published indicator for this layer yet.

  5. 5. Deployment & application

    consistent with normal9 indicators

    Deployment services, coding agents, vertical and consumer apps, rollups.

    Application layer's share of enterprise generative-AI spend (Menlo)grade B
    emergingcoincident55/95 confidence, mixed or hard to operationalise
    51.4%as of 2025-12-31(2 obs)
    Public cloud software valuation multiples (Clouded Judgement)grade B
    emergingleading50/95 confidence, mixed or hard to operationalise
    4.40×as of 2026-09-04
    Vertical versus horizontal application fundinggrade A
    dispersingleading45/95 confidence, limited or vague
    2.30×as of 2026-03-31(11 obs)
    Startups' share of enterprise AI application revenuegrade B
    emergingcoincident50/95 confidence, mixed or hard to operationalise
    63.0%as of 2025-12-31
    Developer productivity uplift (METR RCTs)grade A
    consistent with normalleading70/95 confidence, good evidence, some ambiguity
    -4.0%⚑disputedas of 2026-02-24
    Enterprise pilots with measurable P&L impactgrade C
    consistent with normalcoincident55/95 confidence, mixed or hard to operationalise
    5.0%⚑disputedas of 2025-06-30
    FDA-authorised devices built on a large language modelgrade A
    consistent with normalleading60/95 confidence, mixed or hard to operationalise
    1as of 2025-12-23(3 obs)
    Frontier models completing expert legal tasks end to end (Harvey LAB)grade D
    emergingleading40/95 confidence, limited or vague
    7.1%as of 2026-05-26
    Waymo paid rides per weekgrade C
    consistent with normalcoincident65/95 confidence, mixed or hard to operationalise
    500,000 rides per weekas of 2026-03-27

  6. 6. Adopters

    consistent with normal10 indicators

    Enterprises and individuals using the products; indicator-only layer.

    Share of work hours assisted by generative AIgrade B
    consistent with normalcoincident80/95 confidence, good evidence, some ambiguity
    6.3%as of 2026-06-30
    US firms using AI (Census BTOS)grade A
    consistent with normalcoincident70/95 confidence, good evidence, some ambiguity
    22.4%(21.7%–23.1%)as of 2026-08-09
    US businesses paying for AI (Ramp AI Index)grade B
    emergingleading60/95 confidence, mixed or hard to operationalise
    56.1%as of 2026-08-31
    FDA-authorised AI-enabled devices, growthgrade A
    consistent with normalcoincident70/95 confidence, good evidence, some ambiguity
    26.7%as of 2026-06-30(2 obs)
    US state AI bills introduced (legislative year)grade B
    emerginglagging55/95 confidence, mixed or hard to operationalise
    1035as of 2025-12-31
    Augmentation vs automation share (Anthropic Economic Index)grade B
    consistent with normalcoincident60/95 confidence, mixed or hard to operationalise
    51.4%as of 2026-05-31
    Customer-support productivity uplift from a generative-AI assistantgrade B
    consistent with normalcoincident75/95 confidence, good evidence, some ambiguity
    14.0%as of 2023-04-30
    Executives reporting no AI impact on their own firmgrade C
    consistent with normallagging70/95 confidence, good evidence, some ambiguity
    90.0%as of 2026-02-28
    Work-related share of ChatGPT consumer messagesgrade B
    emergingcoincident65/95 confidence, mixed or hard to operationalise
    27.0%as of 2025-06-30
    Usage as a share of theoretical exposuregrade B
    consistent with normalcoincident60/95 confidence, mixed or hard to operationalise
    0.49×as of 2026-06-30(2 obs)

  7. 7. Labour & consumers

    consistent with normal15 indicators

    Wages, employment and consumer surplus — the welfare residual; indicator-only layer.

    Consumer surplus from generative AI (willingness to accept)grade C
    dispersinglagging65/95 confidence, mixed or hard to operationalise
    $172Bas of 2026-03-31
    US labour productivity, year on yeargrade A
    consistent with normallagging80/95 confidence, good evidence, some ambiguity
    2.2%as of 2026-06-30
    US total factor productivity, private nonfarm businessgrade A
    consistent with normallagging75/95 confidence, good evidence, some ambiguity
    0.8%as of 2025-12-31(2 obs)
    Employment in AI-exposed occupations, year on yeargrade C
    consistent with normalcoincident80/95 confidence, good evidence, some ambiguity
    -0.2%as of 2026-06-30
    Entry-level employment shortfall in AI-exposed occupationsgrade B
    emergingleading70/95 confidence, good evidence, some ambiguity
    19.0%as of 2026-06-30
    Exposed vs unexposed occupation employment gap (Revelio)grade B
    emergingcoincident65/95 confidence, mixed or hard to operationalise
    -6.0%as of 2026-07-31
    California unemployment claims from AI-exposed occupationsgrade A
    consistent with normalleading70/95 confidence, good evidence, some ambiguity
    1.0%as of 2026-07-31
    Recent-graduate unemployment rate (NY Fed)grade A
    consistent with normallagging75/95 confidence, good evidence, some ambiguity
    5.6%as of 2026-06-30
    Cross-tracker concordance (labour)grade B
    consistent with normalcoincident70/95 confidence, good evidence, some ambiguity
    0as of 2026-07-31(4 obs)
    US labour share of income, year on yeargrade A
    emerginglagging65/95 confidence, mixed or hard to operationalise
    -3.4%as of 2026-06-30(2 obs)
    Workers highly exposed to AI with low adaptive capacitygrade B
    emergingleading45/95 confidence, limited or vague
    3,300,000 workersas of 2026-01-31
    Occupational-mix dissimilarity since ChatGPT (Yale Budget Lab)grade B
    consistent with normallagging65/95 confidence, mixed or hard to operationalise
    5.26 ppas of 2026-03-01
    Chatbot effect on earnings and hours in exposed occupations (Denmark)grade B
    consistent with normallagging80/95 confidence, good evidence, some ambiguity
    0.0%as of 2024-12-31
    Freelance writing earnings after ChatGPT (Upwork)grade B
    consistent with normalleading70/95 confidence, good evidence, some ambiguity
    -5.2%as of 2023-12-01
    Wage premium for AI skills (PwC AI Jobs Barometer)grade C
    emergingcoincident50/95 confidence, mixed or hard to operationalise
    62.0%as of 2025-12-31

Last three status changes

  1. 2026-09-10enterprise_multi_homing_shareunmeasured to not yet measurableconf — → 45 · evaluate

    First reading. Menlo's 2023 enterprise survey put multi-model use at 60% (restated January 2024); later waves state the pattern without a share and the 2025 mid-year update gives an 11% vendor-switch rate instead. One numeric point, so not yet measurable.

    obs:5a3caabe
  2. 2026-09-10neocloud_debt_termsunmeasured to not yet measurableconf — → 60 · evaluate

    First reading. CoreWeave's Q2 2026 10-Q states effective rates of 15% (DDTL 1.0), 11% (DDTL 2.0), 9% (DDTL 5.0) and 7% (non-recourse DDTL 4.0), a simple average of 10.5%, on $35.6B of total indebtedness; the August DDTL 5.5 facility prices at Term SOFR plus 5.5%. One quarter-end, so the direction rule cannot run: not yet measurable.

  3. 2026-09-10depreciation_useful_life_changesunmeasured to not yet measurableconf — → 65 · evaluate

    First reading. Amazon's FY2025 10-K states a subset of servers and networking equipment moved from six to five years from 1 January 2025; Meta lengthened to 5.5 years, Alphabet holds six and Microsoft states two to six. One annual point per filer, so the direction rule cannot run yet: not yet measurable.

    obs:81efc38c