## 2.1    Sector Classification

## Source details

**Canonical URL:** [2.1    Sector Classification](https://www.imf.org/-/media/files/publications/wp/2025/english/wpiea2025081-print-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2025/english/wpiea2025081-print-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2025/english/wpiea2025081-print-pdf.pdf.json)

---

### Scope and primary NAICS alignment
- AI production occurs primarily within NAICS sectors 518/519 (data processing, internet publishing, and other information services) and 5415 (computer systems design and related services), though it is not exclusive to these sectors.
- The AI production ecosystem comprises four distinct firm types:
  - (i) specialized AI research labs (e.g., OpenAI, Anthropic);
  - (ii) pure data center operators (e.g., Equinix);
  - (iii) cloud services and infrastructure providers; and
  - (iv) vertically integrated technology companies (e.g., Microsoft, Google) that span research, deployment, and integration with existing products like Google Search, Gmail, and Microsoft Office.
- These firms’ core activities predominantly align with NAICS 518/519 and 5415, making these sectors central to measuring AI-related economic activity.

### Data center and firm NAICS specifics
- Data centers are most often categorized as NAICS 518210 (Data Processing, Hosting, and Related Services), which includes activities such as application hosting, cloud storage services, computer data storage services, or computing platform infrastructure provision.
- Examples of NAICS codes for large firms:
  - Equinix: categorized under NAICS 518210 (as a data center company).
  - META: NAICS code 519290.
  - Alphabet: operates under 519, 518 and 541511.
  - Microsoft: has 518 and 541511 as one of its NAICS codes.
  - IBM: codes include 5415 (54151 and 541512).

### Delineation from broader ICT and ‘tech’ classifications
- The definition used here represents a narrower scope than broader classifications like the Information and Communications Technology (ICT) sector, which spans both manufacturing (computers, electronics) and services (telecommunications, software, IT services).
- The scenario simulations in the IMF-ENV model are built around a TFP shock to the ICT sector, as the latter constitutes the smallest plausible proxy for the AI sector that can be lifted from the GTAPv11 database.
- The NAICS 518/519 and 5415 classification differs from, but overlaps with, the commonly used ’tech’ category, which typically includes:
  - hardware manufacturers (for example, Apple),
  - digital platform and service providers (Microsoft, Google, Meta, Alibaba),
  - semiconductor firms (Nvidia, TSMC, ASML).
- Hardware manufacturers and semiconductor firms are excluded from the AI-producing sector definition used here.

### Caveats and classification challenges
- Certain activities of AI companies are classified under traditional sectors (e.g., Equinix as a data center company is also a lessor of real estate under NAICS 531110); such non-AI activity codes are excluded to avoid capturing non-AI activities.
- AI production is increasingly embedded across activities due to hybrid business models (e.g., Tesla investing in autonomous vehicles), complicating a clean one-to-one correspondence between AI-producing sectors and NAICS codes.
- The chosen NAICS-based classification is narrower than ICT and therefore may understate broader economic linkages where AI activity is embedded in other sectors.

*Italic: Source: wpiea2025081-print-pdf - 2.1    Sector Classification*

### 3.3    Scenarios

#### Scenario design and calibration
- Baseline scenario: does not account for AI growth trends; energy and emissions trends calibrated solely based on policies implemented until the year 2024.
- AI sector input cost shares (global average, GTAPv11, 2017): labor and capital approximately 30 percent; about a quarter of intermediate inputs from other compute services; roughly 5 percent from manufacturing; energy (mainly electricity) about 1 percent of input costs in 2017.
- Recent trend: electricity cost share in AI platforms and service companies increased from 0.8 percent in 2019 to 1.5 percent in 2023.
- U.S. assumption: IT sector electricity intensity rises to 4 percent by 2030 (up from 1 percent in 2017). For other countries, input shares are kept identical to 2017 values.
- Two AI scenarios (both include the same AI-driven TFP shock in the IT sector, applied to the Value Added bundle within the nested-CES production function and directed at production that incorporates new capital):
  - AI under current energy policies: AI shock with no changes in the electricity generation mix relative to the baseline.
  - AI under alternative energy policies: same AI shock but with additional supply-side measures (feed-in tariffs aligned with regional long-term strategies following NDC-LTS) to increase renewables’ share.

#### Modeling of medium-term constraints and sensitivity checks
- Sensitivity checks on:
  1. Growth potential of renewables between 2025-2030 (Current/Alternative policies with smaller renewables scale-up).
  2. Investments in transmission and distribution (Current/Alternative policies with no additional investments in T&D).
- Constraint implementation in IMF-ENV:
  - Caps annual growth rates of solar PV and wind generation to be equal to or below the average growth rates seen in the last five years.
  - Fixes the sectoral investment pathway of the T&D sector to the baseline pathway to model absence of additional T&D investments.
- Power generation from hydropower and nuclear technologies is capped at baseline generation levels in all scenarios.

#### Results — Electricity supply and generation mix (by 2030)
- Change in total electricity supply by 2030, AI scenario under current energy policies, relative to baseline:
  - U.S.: increase by 8 (525 TWh) percent.
  - Europe: increase by 3 (145 TWh) percent.
  - China: increase by 2 (237 TWh) percent.
- Under AI + alternative energy policies: total electricity supply increase is identical to current policies, but composition shifts toward renewables.
  - In the U.S., Europe, and China, solar and wind generation offsets about 58, 35 and 166 TWh of generation from other sources, respectively.

#### Results — Electricity prices (by 2030)
- Under AI + current energy policies, electricity price increases relative to baseline:
  - U.S.: 0.9 percent.
  - Europe: 0.45 percent.
  - China: 0.35 percent.
- Price increases are smaller under alternative energy policies due to feed-in tariffs lowering generation cost of solar and wind.
- If two medium-term constraints bind (slower renewables expansion and no additional T&D investments), price increases under current policies could escalate to:
  - U.S.: 8.6 percent.
  - Europe: 3.6 percent.
  - China: 5.3 percent.
- Modeling indicates grid (T&D) capacity is a more critical factor contributing to price increases than renewables scale-up alone.
- Without further T&D investments, electricity for AI expansion may need to be reallocated from other economic activities, adversely affecting energy-intensive manufacturing:
  - Example effect in the U.S.: annual growth in these sectors’ value added would experience an average reduction of 0.3 percent point compared to baseline, negatively impacting annual GDP growth by 0.1 percent point.

#### Results — Emissions and GDP impacts (through 2030)
- Under AI + current energy policies, 2030 increase in GHG emissions relative to baseline:
  - U.S.: 5.5 percent.
  - Europe: 3.7 percent.
  - China: 1.2 percent.
  - Global average increase: 1.2 percent.
- Cumulative global GHG emissions increase between 2025 and 2030 due to IT sector expansion:
  - Under current energy policies: 1.7 Gt (comparable to Italy’s energy-related GHG emissions over a 5-year period).
  - Under alternative energy policies (modest power-sector decarbonization via feed-in tariffs): cumulative increase limited to 1.3 Gt by 2030, which is 24 percent less global emissions than under current energy policies.
- GDP effects:
  - Under current energy policies, the AI TFP shock raises the average annual growth rate of global GDP by 0.5 percentage point between 2025 and 2030.
  - Under alternative energy policies, GDP benefits are slightly diminished due to feed-in tariff fiscal costs.
  - Fiscal cost of feed-in tariffs ranges from 0.3 to 0.6 percent of GDP across various countries; in simulations, these costs are financed through increased lump-sum taxes, causing a slight reduction in household consumption.
- Social cost of emissions (median SCC estimate preserved from source):
  - Median social cost of carbon (SCC): $39 per ton.
  - Additional social cost of 1.3 to 1.7 Gt of carbon-equivalent emissions: about $50.7 to $66.3 billion, or 1.3 to 1.7 percent of the AI-driven increase in real world GDP between 2025-2030.

#### Uncertainty and interaction with technological change
- Demand for compute and electricity from AI service producers is highly uncertain:
  - Algorithmic and model efficiency improvements (e.g., emergence of more efficient, open-source models like DeepSeek) can reduce compute costs and electricity demand.
  - Lower compute costs can also stimulate AI use; development of more energy-intensive reasoning models increases electricity demand.
- This uncertainty risks delaying critical energy investments, potentially producing underinvestment and higher energy prices.

#### Policy implications and recommendations
- Energy policies should prioritize stimulating the supply side to support AI growth sustainably.
- Implementing policies that incentivize renewables (examples modeled: feed-in tariffs) can:
  - Enhance electricity supply responsiveness.
  - Mitigate price surges.
  - Reduce emission impacts from AI-driven electricity demand.
- Without sufficient expansion of renewables and T&D investments, AI-driven electricity demand can produce substantial price pressures (up to 9 percent in the U.S. in adverse cases) and notable increases in GHG emissions.
- Balancing AI expansion and climate objectives requires supply-side measures alongside investments in transmission and distribution infrastructure to avoid constraining economic activity and exacerbating emissions accumulation.

*Source: IMF Working Paper content unit "3.3    Scenarios" from the provided PDF chapter.*

### 2.1    Sector Classification

### 2.1    Sector Classification

### Scope and primary NAICS alignment
- AI production occurs primarily within NAICS sectors 518/519 (data processing, internet publishing, and other information services) and 5415 (computer systems design and related services), though it is not exclusive to these sectors.
- The AI production ecosystem comprises four distinct firm types:
  - (i) specialized AI research labs (e.g., OpenAI, Anthropic);
  - (ii) pure data center operators (e.g., Equinix);
  - (iii) cloud services and infrastructure providers; and
  - (iv) vertically integrated technology companies (e.g., Microsoft, Google) that span research, deployment, and integration with existing products like Google Search, Gmail, and Microsoft Office.
- These firms’ core activities predominantly align with NAICS 518/519 and 5415, making these sectors central to measuring AI-related economic activity.

### Data center and firm NAICS specifics
- Data centers are most often categorized as NAICS 518210 (Data Processing, Hosting, and Related Services), which includes activities such as application hosting, cloud storage services, computer data storage services, or computing platform infrastructure provision.
- Examples of NAICS codes for large firms:
  - Equinix: categorized under NAICS 518210 (as a data center company).
  - META: NAICS code 519290.
  - Alphabet: operates under 519, 518 and 541511.
  - Microsoft: has 518 and 541511 as one of its NAICS codes.
  - IBM: codes include 5415 (54151 and 541512).

### Delineation from broader ICT and ‘tech’ classifications
- The definition used here represents a narrower scope than broader classifications like the Information and Communications Technology (ICT) sector, which spans both manufacturing (computers, electronics) and services (telecommunications, software, IT services).
- The scenario simulations in the IMF-ENV model are built around a TFP shock to the ICT sector, as the latter constitutes the smallest plausible proxy for the AI sector that can be lifted from the GTAPv11 database.
- The NAICS 518/519 and 5415 classification differs from, but overlaps with, the commonly used ’tech’ category, which typically includes:
  - hardware manufacturers (for example, Apple),
  - digital platform and service providers (Microsoft, Google, Meta, Alibaba),
  - semiconductor firms (Nvidia, TSMC, ASML).
- Hardware manufacturers and semiconductor firms are excluded from the AI-producing sector definition used here.

### Caveats and classification challenges
- Certain activities of AI companies are classified under traditional sectors (e.g., Equinix as a data center company is also a lessor of real estate under NAICS 531110); such non-AI activity codes are excluded to avoid capturing non-AI activities.
- AI production is increasingly embedded across activities due to hybrid business models (e.g., Tesla investing in autonomous vehicles), complicating a clean one-to-one correspondence between AI-producing sectors and NAICS codes.
- The chosen NAICS-based classification is narrower than ICT and therefore may understate broader economic linkages where AI activity is embedded in other sectors.

*Italic: Source: wpiea2025081-print-pdf - 2.1    Sector Classification*

### 3.3    Scenarios

### 3.3    Scenarios

### Scenario design and calibration
- Baseline scenario: does not account for AI growth trends; energy and emissions trends calibrated solely based on policies implemented until the year 2024.
- AI sector input cost shares (global average, GTAPv11, 2017): labor and capital approximately 30 percent; about a quarter of intermediate inputs from other compute services; roughly 5 percent from manufacturing; energy (mainly electricity) about 1 percent of input costs in 2017.
- Recent trend: electricity cost share in AI platforms and service companies increased from 0.8 percent in 2019 to 1.5 percent in 2023.
- U.S. assumption: IT sector electricity intensity rises to 4 percent by 2030 (up from 1 percent in 2017). For other countries, input shares are kept identical to 2017 values.
- Two AI scenarios (both include the same AI-driven TFP shock in the IT sector, applied to the Value Added bundle within the nested-CES production function and directed at production that incorporates new capital):
  - AI under current energy policies: AI shock with no changes in the electricity generation mix relative to the baseline.
  - AI under alternative energy policies: same AI shock but with additional supply-side measures (feed-in tariffs aligned with regional long-term strategies following NDC-LTS) to increase renewables’ share.

### Modeling of medium-term constraints and sensitivity checks
- Sensitivity checks on:
  1. Growth potential of renewables between 2025-2030 (Current/Alternative policies with smaller renewables scale-up).
  2. Investments in transmission and distribution (Current/Alternative policies with no additional investments in T&D).
- Constraint implementation in IMF-ENV:
  - Caps annual growth rates of solar PV and wind generation to be equal to or below the average growth rates seen in the last five years.
  - Fixes the sectoral investment pathway of the T&D sector to the baseline pathway to model absence of additional T&D investments.
- Power generation from hydropower and nuclear technologies is capped at baseline generation levels in all scenarios.

### Results — Electricity supply and generation mix (by 2030)
- Change in total electricity supply by 2030, AI scenario under current energy policies, relative to baseline:
  - U.S.: increase by 8 (525 TWh) percent.
  - Europe: increase by 3 (145 TWh) percent.
  - China: increase by 2 (237 TWh) percent.
- Under AI + alternative energy policies: total electricity supply increase is identical to current policies, but composition shifts toward renewables.
  - In the U.S., Europe, and China, solar and wind generation offsets about 58, 35 and 166 TWh of generation from other sources, respectively.

### Results — Electricity prices (by 2030)
- Under AI + current energy policies, electricity price increases relative to baseline:
  - U.S.: 0.9 percent.
  - Europe: 0.45 percent.
  - China: 0.35 percent.
- Price increases are smaller under alternative energy policies due to feed-in tariffs lowering generation cost of solar and wind.
- If two medium-term constraints bind (slower renewables expansion and no additional T&D investments), price increases under current policies could escalate to:
  - U.S.: 8.6 percent.
  - Europe: 3.6 percent.
  - China: 5.3 percent.
- Modeling indicates grid (T&D) capacity is a more critical factor contributing to price increases than renewables scale-up alone.
- Without further T&D investments, electricity for AI expansion may need to be reallocated from other economic activities, adversely affecting energy-intensive manufacturing:
  - Example effect in the U.S.: annual growth in these sectors’ value added would experience an average reduction of 0.3 percent point compared to baseline, negatively impacting annual GDP growth by 0.1 percent point.

### Results — Emissions and GDP impacts (through 2030)
- Under AI + current energy policies, 2030 increase in GHG emissions relative to baseline:
  - U.S.: 5.5 percent.
  - Europe: 3.7 percent.
  - China: 1.2 percent.
  - Global average increase: 1.2 percent.
- Cumulative global GHG emissions increase between 2025 and 2030 due to IT sector expansion:
  - Under current energy policies: 1.7 Gt (comparable to Italy’s energy-related GHG emissions over a 5-year period).
  - Under alternative energy policies (modest power-sector decarbonization via feed-in tariffs): cumulative increase limited to 1.3 Gt by 2030, which is 24 percent less global emissions than under current energy policies.
- GDP effects:
  - Under current energy policies, the AI TFP shock raises the average annual growth rate of global GDP by 0.5 percentage point between 2025 and 2030.
  - Under alternative energy policies, GDP benefits are slightly diminished due to feed-in tariff fiscal costs.
  - Fiscal cost of feed-in tariffs ranges from 0.3 to 0.6 percent of GDP across various countries; in simulations, these costs are financed through increased lump-sum taxes, causing a slight reduction in household consumption.
- Social cost of emissions (median SCC estimate preserved from source):
  - Median social cost of carbon (SCC): $39 per ton.
  - Additional social cost of 1.3 to 1.7 Gt of carbon-equivalent emissions: about $50.7 to $66.3 billion, or 1.3 to 1.7 percent of the AI-driven increase in real world GDP between 2025-2030.

### Uncertainty and interaction with technological change
- Demand for compute and electricity from AI service producers is highly uncertain:
  - Algorithmic and model efficiency improvements (e.g., emergence of more efficient, open-source models like DeepSeek) can reduce compute costs and electricity demand.
  - Lower compute costs can also stimulate AI use; development of more energy-intensive reasoning models increases electricity demand.
- This uncertainty risks delaying critical energy investments, potentially producing underinvestment and higher energy prices.

### Policy implications and recommendations
- Energy policies should prioritize stimulating the supply side to support AI growth sustainably.
- Implementing policies that incentivize renewables (examples modeled: feed-in tariffs) can:
  - Enhance electricity supply responsiveness.
  - Mitigate price surges.
  - Reduce emission impacts from AI-driven electricity demand.
- Without sufficient expansion of renewables and T&D investments, AI-driven electricity demand can produce substantial price pressures (up to 9 percent in the U.S. in adverse cases) and notable increases in GHG emissions.
- Balancing AI expansion and climate objectives requires supply-side measures alongside investments in transmission and distribution infrastructure to avoid constraining economic activity and exacerbating emissions accumulation.

*Source: IMF Working Paper content unit "3.3    Scenarios" from the provided PDF chapter.*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2025/english/wpiea2025081-print-pdf.pdf_