AI-Infrastructure Market Map

Provenance note: This page is the standalone 2026-05-12 market-map view, with per-archetype economics from the 2026-05-24 DC-archetype BOM model. Both vintages have since been superseded by the unified 2026-05-29 token-and-capex model; the figures here predate that consolidation.

This page is the static snapshot of the AI-Infrastructure Market Map model — a driver-based projection of where AI-data-center capex actually flows across the supply chain, anchored to Goldman Sachs' Tracking Trillions (April 2026) capex pool and decomposed into seven categories: Compute Silicon, Memory, Networking + IC, Fiber + Optics, Power Infrastructure, Cooling, and Land + Shell + EPC. The visual encoding is an L-shape Marimekko — a vertical Compute column on the left whose width encodes Compute's share of the cluster, plus six horizontal bands stacked on the right whose heights encode each non-compute category's share. Within each band, sub-cells carry the named players. Width and area always equal share; nothing is decorative.

The bottom-up sum of category spend equals the Goldman pool to the dollar by design — $765B in 2026 and $1,392B in 2029 (Base case) — but the mix shifts: Compute Silicon widens from 33% → 44% of pool while Power Infrastructure compresses 20% → 12%. The DC archetype variants on this page show why one number hides five different bills of materials.

Reference chart — Agentic AI archetype, per-cluster BOM

Of the five DC archetype views, Agentic AI is the most informative single chart: it sits between Training/Core and Inference on compute intensity (50.5% of cluster vs. 53.5% / 44.4%), and the networking band widens to 10.4% of cluster (+3.5pp vs. unified baseline) — the empirical hook behind the optical-networking thesis. Each band's vertical height is proportional to its share of the $5.19B reference cluster ($35M/MW × 148MW). Tiny bands (Cooling 2.5%, Fiber 3.8%) get a min-height floor so they stay legible; this is encoded in the data, not a layout cheat.

$100 of Agentic AI CapEx · DC Archetype BOM

L-shape Marimekko · area ∝ share of cluster · $5.19B cluster total · $35M/MW · 2026-05-24 BOM model

The widened Compute column (50.5% vs. unified baseline 42.5%) and the widened Networking band (+3.5pp) are the two material deltas — both push toward AI-native silicon and optical-interconnect exposure rather than shell+EPC contractors.

Methodology — bottom-up construction

Every category in the 14-tab Excel model is built from primary inputs on the 01_Assumptions tab — vendor revenue lines (NVDA DC, AMD, TSMC AI, AVGO AI semi, Vertiv DCPI, Eaton, GEV, etc.), driver inputs (US 2026 AI-DC under-construction 22.4 GW per Eaton Q1'26, 2026→2029 CAGR 28% Base per Eaton anchor), and explicit judgment cells (highlighted yellow). The seven categories sum to exactly the Goldman pool by design after a 2026-05-13 net-pool rework that strips supplier-revenue double-counts: HBM ($145B 2029) moved entirely to Memory; NVLink scale-up (~$25B) moved to Networking; AVGO Tomahawk switching ASIC (~$15B) moved to Networking; Vertiv/Schneider thermal-mgmt (~$21B) stripped from Cooling since it's already in Power Infrastructure DCPI. Three scenarios (Bear / Base / Bull) flex via cell 00_Cover!B5.

Table 1 — Category roll-up, Base case

Roll-up by category, 2026 and 2029, expressed in dollars and as share of the Goldman pool.

Category	2026 ($B)	2026 % of pool	2029 ($B)	2029 % of pool	Mix shift (pp)
Compute Silicon	251	32.8%	613	44.0%	+11.2
Memory	128	16.8%	258	18.5%	+1.7
Power Infrastructure	150	19.6%	167	12.0%	−7.6
Land + Shell + EPC	144	18.8%	216	15.5%	−3.3
Networking + IC	53	6.9%	96	6.9%	0.0
Fiber + Optics	31	4.0%	28	2.0%	−2.0
Cooling	8	1.1%	14	1.0%	−0.1
Total — 7 categories (Net AI-DC capex)	765	100.0%	1,392	100.0%	—

Roll-up totals reflect Goldman Tracking Trillions (Apr 2026) pool anchors. Category shares reflect the 2026-05-13 net-pool rework. Mix shift is in percentage points of pool share, 2026 → 2029.

Table 2 — Per-archetype cluster BOM (the five DC variants)

One headline number per archetype, then category shares of that archetype's reference cluster. This is the "what does $100 of this kind of AI-DC capex buy" view. Cluster totals and $/MW intensities come from the 2026-05-24 DC-archetype BOM model; the unified baseline column shows the cross-archetype average.

Category	Training / Core	Inference	Agentic AI	Edge	Legacy Ent.	Unified base
Cluster size ($B)	6.52	6.26	5.19	2.91	3.21	—
Intensity ($M/MW)	43	42	35	19	21	—
Compute Silicon	53.5%	44.4%	50.5%	25.5%	19.7%	42.5%
Memory	11.5%	18.7%	17.0%	20.3%	21.4%	16.9%
Power Infra	5.9%	5.9%	5.4%	15.0%	9.8%	4.3%
Land + Shell + EPC	15.4%	19.5%	10.4%	24.4%	37.2%	25.7%
Networking + IC	8.7%	7.2%	10.4%	9.1%	7.2%	6.9%
Fiber + Optics	3.3%	2.8%	3.8%	2.9%	1.5%	1.2%
Cooling	1.7%	1.4%	2.5%	2.9%	3.3%	2.5%

Read the columns vertically — each is a separate L-shape Marimekko. Training and Inference share the high-compute-share profile; Legacy Enterprise inverts it (shell-heavy, compute-light); Edge spreads cluster spend almost evenly across Power, Memory, and Shell. Networking band peaks under Agentic and Training, where east–west bandwidth is the bottleneck.

Per-archetype mix at a glance

Five 100% stacked bars — same data as Table 2, visual form. Each bar is one archetype; segment widths are category shares.

Training / Core

$6.52B cluster · $43M/MW · highest compute-share

Compute 53.5% · Memory 11.5% · Land 15.4% · Networking 8.7%

Inference

$6.26B cluster · $42M/MW · memory-heavy

Compute 44.4% · Memory 18.7% · Land 19.5% · Networking 7.2%

Agentic AI

$5.19B cluster · $35M/MW · networking widens

Compute 50.5% · Memory 17.0% · Networking 10.4% · Fiber 3.8%

Edge

$2.91B cluster · $19M/MW · spread profile

Compute 25.5% · Memory 20.3% · Power 15.0% · Land 24.4%

Legacy Enterprise

$3.21B cluster · $21M/MW · shell-heavy

Compute 19.7% · Memory 21.4% · Land 37.2% · Power 9.8%

Unified baseline (avg)

cross-archetype average · reference for delta calcs

Compute 42.5% · Memory 16.9% · Land 25.7% · Networking 6.9%

How to read the L-shape Marimekko

Width = Compute share. The left column is Compute Silicon. Its width as a fraction of total canvas = Compute's share of the cluster. Wide left column ⇒ compute-heavy archetype.
Height = non-compute share. Each horizontal band on the right is one non-compute category. Its vertical height as a fraction of canvas height = that category's share of the (non-compute) right side.
Min-height floors apply. Cooling at ~2.5% and Fiber at ~3.8% would render as illegible slivers without a floor; they are widened deliberately so the named players stay readable. This is in the data spec — see min_height_px and css_class fields.
Players are not decoration. Each cell lists 2–3 named public companies (with tickers where applicable) — these are the suppliers we believe capture the dollars in that category. Replace them as the BOM evolves.