THE COMPANIES SHAPING
THE INFRASTRUCTURE BEHIND AI.

Candidate universe — 223 records researched across the five layers, each with sourced evidence, status-tagged figures and an assigned confidence level.

Deduplication — 215 companies. Eight records appeared in two layers. Each was assigned a single primary layer, the higher subscore per dimension retained, the other layer recorded as secondary, and source lists merged.

Layer allocation. GRID 21 · DATA CENTRE 22 · NETWORK 19 · COMPUTE 19 · CHIP 19. Selection within layer was by raw score, followed by a five-company watchlist per layer.

Five independent research passes produced five different score distributions. A single global cut would have reflected which layer got the more generous scorer, not which companies matter.

Raw scores are not comparable across layers. A derived chainScore is published for cross-layer display only: the company’s position within its own layer’s distribution, expressed on a common scale (50 + 12z, bounded 0–100). Use rawScore inside a layer and chainScore across layers.

HIGH: primary source, or two or more strong corroborating sources. MEDIUM: one strong secondary source, or credible indirect evidence. No LOW-confidence company was published into the final 100.

Every capacity and capital figure carries exactly one primary status: OPERATIONAL · UNDER CONSTRUCTION · PLANNED · COMMITTED CAPITAL · CONTRACTED · ANNOUNCED · PROPOSED · TARGETED. Compound statuses are preserved where the evidence spans phases.

One override was applied. Ajinomoto was promoted over a one-point difference because the rubric under-weights materials monopolies. The override is recorded rather than hidden.

Inclusion is never sold.

Market relevance is asymmetric by design. Private companies are systematically under-scored on deployed footprint because they disclose less. Hyperscaler custom silicon is captured through its suppliers to avoid double-counting.