Editorial correction — September 22, 2026. This page previously gave a broad market-size range and claims about job growth without a specific supporting dataset. Those claims have been removed. The figures below identify their observation year and source; a report published in 2026 does not turn earlier observations into 2026 results.
Three measures that answer different questions
A funding announcement, a customer survey and an electricity forecast describe different parts of the AI economy. Adding them together produces neither revenue nor market size. Before comparing two headlines, write down the quantity each one measures and who was counted.
| Measure | Sourced observation | Limit |
|---|---|---|
| Investment | Stanford's 2026 AI Index reports that global corporate AI investment more than doubled in 2025. | Investment is capital committed to companies; it is not customer spending or profit. |
| Adoption | The same report puts AI adoption at 88% of surveyed organizations in 2025. | A survey response is not a census of every business, a measure of daily usage or evidence of return on investment. |
| Electricity | The IEA's 2026 update reports 485 TWh of data-centre electricity use in 2025 and projects about 950 TWh in 2030. | This covers data centres overall, not only AI. The 2030 number is a projection. |
Sources: Stanford AI Index 2026, Economy; IEA, Key Questions on Energy and AI (2026), executive summary. Sources checked September 22, 2026.
Read the scope before the growth rate
For a market-size claim, check whether the definition includes hardware, software, consulting, cloud hosting or several of these. A chip vendor's sales and a cloud provider's purchases can describe opposite sides of one transaction. A total assembled from both can double-count the same spending.
For an adoption percentage, record the survey population, geography and question. “Uses AI in at least one function” can include a small experiment. It does not establish how many employees use the system, whether it is reliable, or whether the project pays for itself.
A worked example: power and energy
Suppose a fictional facility draws a constant 100 megawatts throughout a non-leap year. Its annual electricity use would be 100 × 8,760 = 876,000 megawatt-hours, or 0.876 TWh. This is an arithmetic illustration, not a measured facility. If 100 MW is only its connection capacity, actual annual consumption cannot be calculated without utilization data.
This distinction matters when a proposed project is described as a gigawatt campus. A connection request is not proof that equipment is installed, that the facility runs continuously, or that all its consumption serves AI. Keep capacity, construction status and measured consumption in separate fields.
Keep forecasts separate from observations
A forecast is conditional. The IEA identifies constraints across grid connections, equipment and chip supply, and uncertainty about how quickly more demanding uses of AI expand. Its projected electricity demand should therefore retain the report date and scenario context when reused.
For a company presentation, a useful label is “2030 projection, IEA 2026 update,” rather than “current AI electricity use.” When a newer edition appears, keep the previous estimate in your notes so a changed forecast is not mistaken for a change in measured demand.
A reusable evidence worksheet
- Claim: copy the specific statement you intend to support.
- Metric and unit: revenue, investment, capacity, energy, adoption or another defined quantity.
- Time: observation period, publication date and forecast year, recorded separately.
- Coverage: geography, sample and activities included or excluded.
- Evidence: link the exact report section and record whether the number is measured, estimated, surveyed or projected.
- Decision: explain what the number changes in your plan. If it changes nothing, it may not belong in the headline.
For a small software team, successful task completion, support burden and cost per accepted result may be more actionable than a global market forecast. Those are measurements to collect in your own product, not numbers this article claims to have measured for you.
Continue reading
Compare AI models with a repeatable task worksheet · Five concrete data-integrity cases from our converter repair