HUMANITYVECTOR

USA · snapshot 2026-09-07 22:35 UTC · data to 2026-07-01

Semiconductor and AI infrastructure capacity

Is the physical capacity behind AI (fabs, power, materials) expanding fast enough to meet announced demand?

No stored series for silicon-metal-production yet (metric is registered but not published); the page shows what is available.

DerivedModeled

Section 1 of 9

What changed?

Observed and derived series only. Changes are latest minus the value one, five and ten years earlier.

AI narrative Modeled

silicon-metal-production (WLD) remains unpublished. Rare earth mine production reached 390000 metric tons in 2025 [obs:263991], up 10000 metric tons or 2.6 percent from 380000 metric tons in 2024 [obs:263978]. US nonresidential construction spending was 1286376 millions of dollars in 2026-07 [obs:179179], down 16298 millions or 1.3 percent from 2025 [obs:179167] yet up 441735 millions or 52.3 percent from 2021 [obs:179119]. US electricity net generation was 354690.526 million kilowatthours in 2026-05 [obs:245396], up 10656.561 million kilowatthours or 3.1 percent from 2025 [obs:245384].

Written by grok-4.3 over the numbers in this snapshot only, validated against them (every figure and cited id had to be in the snapshot), run 72262868, prompt 1.0.0. It adds no facts.

Observed

Magnet and process inputs.

390,000 metric tons

2025

2 points, 2024 to 2025, 380,000 to 390,000 metric tons380,000390,00020242025
SpanFromChangePercentile
1 period380,000 (2024)10,000 (+2.6%)n/a
1 year380,000 (2024)10,000 (+2.6%)n/a
5 yearsseries too shortn/an/a
10 yearsseries too shortn/an/a
World · 2 observations since 2024 · obs 263991
Observed

Includes fab and data-centre construction.

1.29M millions of dollars, seasonally adjusted annual rate

2026-07-01

60 points, 2021-08-01 to 2026-07-01, 841,975 to 1.34M millions of dollars, seasonally adjusted annual rate841,9751.34M2021-08-012026-07-01
SpanFromChangePercentile
1 period1.28M (2026-06-01)1,499 (+0.1%)14th
1 year1.3M (2025-07-01)-16,298 (-1.3%)17th of 282
5 years844,641 (2021-07-01)441,735 (+52.3%)88th of 234
10 years731,328 (2016-07-01)555,048 (+75.9%)78th of 174
United States · 295 observations since 2002-01-01 · obs 179179
Observed

The power the new load must come from.

354,690.53 million kilowatthours

2026-05-01

60 points, 2021-06-01 to 2026-05-01, 301,903.18 to 446,319.53 million kilowatthours301,903.18446,319.532021-06-012026-05-01
SpanFromChangePercentile
1 period330,706.52 (2026-04-01)23,984.01 (+7.3%)63rd
1 year344,033.97 (2025-05-01)10,656.56 (+3.1%)65th of 628
5 years320,180.96 (2021-05-01)34,509.56 (+10.8%)77th of 580
10 years316,866.99 (2016-05-01)37,823.53 (+11.9%)45th of 520
United States · 641 observations since 1973-01-01 · obs 245396
  • silicon-metal-production (WLD): not published, metric is registered but not published

Section 2 of 9

How unusual is it?

The primary series against its own stored history (features_v1 rarity and surprise).

AI narrative Modeled

No primary series stored, so unusualness is not computed.

Written by grok-4.3 over the numbers in this snapshot only, validated against them (every figure and cited id had to be in the snapshot), run 72262868, prompt 1.0.0. It adds no facts.

Not computed: the primary series is not stored, so there is no history to compare against.

Section 3 of 9

Why might it be happening?

Documented mechanisms, statistical model drivers, hypotheses and unknowns are kept apart.

AI narrative Modeled

Documented mechanism: AI compute demand pulls on accelerators, memory, leading-edge fabs, and the electricity and water those fabs use via the chain ai-compute to semiconductor-fab to electricity-system [mechanism:b735fd46-d9c5-4c63-a93e-8b408410b01d]. Statistical model drivers: none stored. Hypotheses: AI compute affects economic output via productivity effects [rel:812f2c15-fcfa-420c-b874-302b1bf79647]. Unknown: no public series measures leading-edge wafer capacity; announced data-centre capacity is not the same as energised capacity.

Written by grok-4.3 over the numbers in this snapshot only, validated against them (every figure and cited id had to be in the snapshot), run 72262868, prompt 1.0.0. It adds no facts.

Reviewed

Documented mechanisms

Statistical

Model-supported drivers

Modeled

No active forecast for this topic records driver contributions (the champion may be a baseline with no drivers).

Candidate or hypothesized relationships

Hypotheses

Not settled by stored data

Unknowns

  • No public series measures leading-edge wafer capacity; the site stores upstream materials and corporate pulses instead.
  • Announced data-centre capacity is not the same as energised capacity; the gap between the two is not stored.

Section 4 of 9

What has not changed?

Fundamentals judged against the 10th to 90th percentile band of their own year-over-year changes (or of their levels when the history is short).

AI narrative Modeled

R&D expenditure as share of GDP was 2.59737836465431 percent of GDP in 2023 [obs:102615], inside the 10th to 90th percentile band of past changes from -0.035262 to 0.075445 percent of GDP and inside the level band 1.961711 to 2.476088 percent of GDP. Researchers per million people was 1651.5851690472 in 2020 [obs:118861], outside the level band 1137.958671 to 1587.701882.

Written by grok-4.3 over the numbers in this snapshot only, validated against them (every figure and cited id had to be in the snapshot), run 72262868, prompt 1.0.0. It adds no facts.

SeriesLatestBasisNormal band (p10 to p90)Historical rangeReading
R&D expenditure as share of GDP Observed
R&D as a share of world GDP.
2.6 percent of GDP
2023, obs 102615
year-over-year change 0.0262 (+1.0%)
26 past changes
-0.0353 to 0.07541.92 (2007) to 2.6 (2023)
since 1996; latest at the 100th percentile of levels
within its normal band
Researchers per million people Observed
Researchers per million people.
1,651.59 researchers per million people
2020, obs 118861
level
6 observations
1,137.96 to 1,587.71,080.91 (2000) to 1,651.59 (2020)
since 2000; latest at the 100th percentile of levels
outside its normal band
  • oecd-industrial-production (WLD): no observations, no current observations for WLD

Section 5 of 9

What absorbs the shock?

Each buffer names the series or graph entity that evidences it, or says it is not measured.

AI narrative Modeled

Imports, spare capacity, inventory and substitution are not measured. Diversification shows rare earth production top-3 country concentration at 89.7 percent of world production in 2025 [obs:409075]. Financing shows US real private fixed investment at 4551.12 billions of chained 2017 dollars in 2026-04 [obs:181431].

Written by grok-4.3 over the numbers in this snapshot only, validated against them (every figure and cited id had to be in the snapshot), run 72262868, prompt 1.0.0. It adds no facts.

Imports

Not measured

No stored series evidences this buffer.

rare-earth-us-net-import-reliance is referenced but not available (see gaps). US net import reliance for rare earths.

Derived

89.7 percent of world production · 2025 · yoy -1.1% · obs 409075

Three-country share of rare-earth output.

Observed

4,551.12 billions of chained 2017 dollars, seasonally adjusted annual rate · 2026-04-01 · yoy +3.9% · obs 181431

Real private fixed investment carries the build-out.

Spare capacity

Not measured

No stored series evidences this buffer.

Fab utilisation is not published as a public series.

Inventory

Not measured

No stored series evidences this buffer.

Accelerator and memory inventories are not ingested.

Substitution

Not measured

No stored series evidences this buffer.

Substitution between accelerator generations is not measured here.

  • rare-earth-us-net-import-reliance (USA): not published, metric is registered but not published

Section 6 of 9

What response is underway?

Capacity pipeline, corporate pulses, capital spending and production series, with their year-over-year change.

AI narrative Modeled

Corporate pulse semiconductor capacity capex growth was 5.53 percent yoy in 2026-07 [obs:404028]. Corporate pulse ai infrastructure capex growth was 70.14 percent yoy in 2026-07 [obs:404026]. US future capacity pipeline 24 months was 109333.015 MW in 2026-07, up 14.5 percent yoy [obs:409101]. US real private fixed investment was 4551.12 billions of chained 2017 dollars in 2026-04, up 3.9 percent yoy [obs:181431].

Written by grok-4.3 over the numbers in this snapshot only, validated against them (every figure and cited id had to be in the snapshot), run 72262868, prompt 1.0.0. It adds no facts.

GroupSeriesLatestYear earlierChange
Corporate pulseCorporate pulse: semiconductor capacity capex growth Derived5.53 percent change yoy (2026-07-03, obs 404028)n/an/a
Corporate pulseCorporate pulse: ai infrastructure capex growth Derived70.14 percent change yoy (2026-07-03, obs 404026)n/an/a
Capacity pipelineUS future capacity pipeline, 24 months Modeled109,333.02 MW (2026-07-01, obs 409101)95,452.66 (2025-07-01)13,880.36 (+14.5%)
Capital spendingUS real private fixed investment Observed4,551.12 billions of chained 2017 dollars, seasonally adjusted annual rate (2026-04-01, obs 181431)4,380.47 (2025-04-01)170.65 (+3.9%)

The US pipeline and response measures are on the capacity page; bottleneck assessments are on bottlenecks.

Section 7 of 9

Scenarios

Base is the champion model's point estimate; upside and downside are the bounds of its 80 percent interval. No probability beyond the interval's own is assigned.

AI narrative Modeled

No forecast is published.

Written by grok-4.3 over the numbers in this snapshot only, validated against them (every figure and cited id had to be in the snapshot), run 72262868, prompt 1.0.0. It adds no facts.

No forecast published for this topic yet.

Section 8 of 9

What to watch

Each indicator is a stored series with a numeric threshold; the reading says which side of it the latest value sits on.

AI narrative Modeled

Corporate pulse semiconductor capacity capex growth at 5.53 percent yoy against threshold 0, rising confirms easing [obs:404028]. US future capacity pipeline 24 months at 109333.015 MW, year-over-year 14.5416 percent against threshold 0, rising confirms easing [obs:409101]. Rare earth production top-3 country concentration at 89.7 percent against threshold 80, falling confirms easing [obs:409075].

Written by grok-4.3 over the numbers in this snapshot only, validated against them (every figure and cited id had to be in the snapshot), run 72262868, prompt 1.0.0. It adds no facts.

Derived
Level: 5.53 percent change yoy (2026-07-03)Threshold: 0 percent change yoy

On the easing side of the threshold · rising confirms easing

A positive pulse means the curated semiconductor companies report capacity growing.

obs 404028 · United States
Modeled
Year over year: +14.5% (2026-07-01)Threshold: 0%

On the easing side of the threshold · rising confirms easing

Power for new load has to come from the generation pipeline.

obs 409101, 409089 · United States
Derived
Level: 89.7 percent of world production (2025)Threshold: 80 percent of world production

On the worsening side of the threshold · falling confirms easing

Concentration above 80 percent (the series is in percent) leaves magnet supply exposed to one policy decision.

obs 409075 · World

Section 9 of 9

Forecast track record

Backtests are rolling-origin, out of sample; live evaluations compare published forecasts with the realised values.

AI narrative Modeled

No forecast keys for this topic, so there is no track record.

Written by grok-4.3 over the numbers in this snapshot only, validated against them (every figure and cited id had to be in the snapshot), run 72262868, prompt 1.0.0. It adds no facts.

No forecast keys are attached to this topic, so there is no track record to show.

The full scorecard is at /forecasts/scorecard.

Provenance

Snapshot
3 · 2026-09-07 22:35 UTC · navigator_view_v1+navigator_topics_v1
Gaps stated
3
AI narrative
grok-4.3, run 72262868-a4f5-4800-8025-0b1e3ac7fda7, prompt 1.0.0, 16 ids cited
Earlier snapshots
none

Every value on this page is a stored observation (obs ids shown), a forecast record, a mechanism or a relationship row, or a percentile computed from those the same way the relevance engine computes its features. Series the config references but the site does not store are listed as gaps under their section, never filled in.