Four Forces.
One view of power.
The AI Power Index is exmxc’s judgment of who controls AI’s direction and deployment. We set the scores, rankings, and forecasts. Research informs those judgments.
Locked September 18, 2026. Rankings refresh monthly; the methodology remains the same.
What we score
Control over the infrastructure and consequential bottlenecks needed to train and run AI.
Control over where people and institutions encounter AI, express intent, and act.
Authority and capacity to govern what AI may do: permissions, containment, monitoring, interruption, accountability, and recovery.
Control over electricity, continuity, location, and enabling power infrastructure.
The doctrine comes from the Four Forces of AI Power.
Dominance and breadth
Power Score = 50% strongest force + 50% average of all four.
A company can hold major power through one critical bottleneck. Control across several forces adds to that power. The strongest force deliberately appears in both terms: 62.5% effective weight for that force and 12.5% for each other force. No force receives a fixed preference.
Example: force scores of 80, 60, 40, and 20 produce a Power Score of 65.
Fixed scoring anchors
Each force is scored in five-point increments. We judge scale, indispensability, enforceable control, and difficulty of substitution.
- 0: No material independent control.
- 25: Limited control; substantial dependence.
- 50: Meaningful control with credible substitutes.
- 75: Major control position difficult to replace.
- 100: Defining global bottleneck or platform position.
These are anchored editorial judgments, not percentiles or market shares. Multiple companies may hold defining positions in different indispensable parts of a force. A maximum score does not mean invulnerability.
Ranks and ties
Rank by the unrounded score, highest first. Break exact ties by comparing the four force scores sorted strongest to weakest, then by company name in alphabetical order. Display scores to one decimal, rounded half up. Small differences represent close editorial calls.
Company boundaries
Score operating company groups and consolidate controlled subsidiaries. Alphabet includes Google; Microsoft includes GitHub; SpaceX includes xAI and X; Tesla is separate. Samsung Electronics excludes other Samsung group businesses. Partnerships and investments do not transfer control of a partner’s full stack. Pending acquisitions are not current ownership.
Chip efficiency alone is not Energy control. Model distribution alone is not authority over downstream permissions. Alignment measures the capacity to govern AI actions, not moral virtue or a claim that models cannot misbehave. Energy includes scarce electrical and thermal infrastructure, with suppliers distinguished from generation owners.
Three years, monthly editions
2026 is the inaugural current assessment. 2027 and 2030 are forecasts made at each edition’s research cut-off. Announced capacity informs forecasts; it is not treated as already deployed. Every edition stores the target date and assessment type. The named years do not roll forward automatically; completed-year assessments are preserved.
The same 50-company cohort provides continuity. A company’s score can rise while its rank falls. Each release records its date, methodology version, source notes, and changes. Prior editions and forecasts remain available. Corrections create a linked successor rather than silently replacing history.
Our research and judgment
The cohort spans public and private companies worldwide, selected for consequential positions across the four forces. It is an editorial selection, not an exhaustive statistical census. Current primary disclosures, technical documentation, and research inform the company notes. Reported capabilities, our interpretation, and forward assumptions remain distinguishable. Forecasts compare this cohort rather than hypothetical future entrants.
Sources establish factual premises. exmxc owns the judgments. The index is not a valuation, probability, or return forecast.
Series history
September 2026 begins the ranked longitudinal series. Earlier v1 exposure scores and the v2 profile pilot are separate historical methods, with no numerical conversion into this index.