AI Data Center Physical Stack — Supplier Landscape
When 1 MW of AI capacity costs $20–30M in capex, who actually takes home each dollar in that supply chain? This section walks through 33 representative companies in the L1 physical-compute layer (data centers / power / backup / electrical / cooling / networking) by capex share, market share, and recent M&A.
This module supports the tokenmap framework — specifically the L1 layer of the AI Token Economics value chain.
1. The Money Math — 1 MW AI DC Capex Breakdown
| Sub-item | Share | Tier-1 suppliers |
|---|---|---|
| Electrical (UPS / PDU / switchgear / distribution) | 40–50% | vertiv · schneider-electric · eaton · abb · siemens · legrand |
| Cooling (liquid-cooling primary) | 15–20% | vertiv · coolit-systems · boyd-corporation · motivair · nvent |
| Networking (switches + optical modules + NICs) | 12–18% | arista · broadcom · nvidia · coherent · innolight |
| Building shell | 15–20% | equinix · digital-realty · qts · iron-mountain |
| Backup power (diesel gensets / gas turbines) | Included in electrical | caterpillar · cummins · ge-vernova · generac · rolls-royce-power · kohler |
| Design / management / permitting | 5–10% | General contractors |
Key baseline data (2026 estimates, source: JLL / CBRE / Wood Mackenzie composite):
- US hyperscale DC construction: ~$10–15M/MW (normal), AI-optimized high-density ~$15–25M/MW
- PUE: 1.1–1.3 (liquid-cooling-dominant)
- Power price: $0.05–0.12/kWh (Virginia / Texas mainstream parks)
- Depreciation cycle: infrastructure 20–30 years vs GPU 5–6 years — within the same building's lifetime, GPUs are swapped 4–5 times
- AI rack power density: 30–100+ kW/rack (traditional DC ~10 kW)
2. Sub-Layer A.a — Colocation (data centers)
Tier-1 players all publicly listed, AI demand drove accelerating 2025-2026 revenue:
| Player | 2026 signal |
|---|---|
| equinix | Q1 2026 $2.44B revenue (+9.8%); 8/10 top AI labs + 4/5 neoclouds expanding in its data centers |
| digital-realty | Q1 $1.6B (+16%); landmark 200 MW AI inference single-tenant lease (Charlotte) |
| qts | Blackstone-owned private ($10B take-private 2021); 400 MW → 3 GW commissioned; $25B development pipeline |
| iron-mountain | DC revenue >$1B (+25%); transitioning from records-management business to DC, 507 MW → 1.4 GW developable |
Economics: DC assets depreciate over 25-30 years (vs GPU 5 years), the slowest-depreciating portion of L1; in the 1 MW total capex it's 15-20%. AI customer connections mean colocation providers can secure long-dated leases + higher power-density unit pricing.
3. Sub-Layer A.b — Power & Backup (power + backup power)
Gas turbines are the real rate-limiting step — the AI DC "Bring Your Own Power" trend has turned generators from commodity to scarce goods.
| Player | Type | Key signal |
|---|---|---|
| crusoe-energy | Off-grid self-build + flare-gas | Series E $1.375B @ >$10B; 4.5 GW contracted; main developer for OpenAI Stargate |
| caterpillar | Diesel genset oligopoly | Backlog $63B (+79% YoY); power segment +44%; reciprocating engine capacity expanded to 3× of 2024 |
| cummins | Diesel + natural gas | Q1 2026 Power $2.0B (+19%); record 29.5% EBITDA margin; DC genset lead time 12-36 months extending into 2028 |
| ge-vernova | Gas turbines + grid | Backlog $163B; 100GW gas-turbine backlog stretching to 2030 (5-year lead time) |
| generac | Expanding from residential to DC | DC backlog grew from $400M → $700M+ in one quarter; $600M LOI from an unnamed hyperscaler |
| rolls-royce-power | Large diesel gensets (mtu brand) | FY2025 €5.72B (+19%); DC genset sales +50% YoY for two consecutive years; 2026 launches 45-second quick-start gas genset |
| kohler | KD Series 4 MW + integrated stack | Private, Platinum Equity $3B+ acquisition (2024-05) → renamed Rehlko; the only one integrating genset + UPS + microgrid |
Wood Mackenzie data:
- Gas-turbine prices up 195% vs 2019
- Large-frame turbines lead time stretched to 5 years
- Diesel genset lead time 12-36 months
- US DC power equipment 2026 market size ~$65B
→ Structural conclusion: power is harder to get than chips. AI investment is shifting from a GPU bottleneck to a Power bottleneck, which is the source of revaluation for 100-year industrial giants like GE Vernova / Cat / Cummins.
4. Sub-Layer A.c — Power Distribution & Cooling (electrical distribution + liquid cooling)
4.1 Electrical distribution (Tier 1)
| Player | Q1 2026 signal |
|---|---|
| vertiv | FY2025 $10.23B (+27.7%); FY2026 guide $13.5-14B; backlog $15B (+109%); book-to-bill 2.9× |
| schneider-electric | FY2025 €40B+ (first time); DC business double-digit growth; 2025-10 acquired 75% of motivair |
| eaton | Q1 2026 $7.5B (+17%); DC orders +240% YoY; backlog $14.5B (+44%); 2026-03 $9.5B acquisition of boyd-corporation |
| abb | Q1 2026 record $11.3B orders (+24%); Electrification +44%; triple-digit DC growth |
| siemens | Smart Infrastructure Q1 FY2026 orders €7.2B (record, +22%); US SI orders +54% YoY |
| legrand | Q1 2026 +18.3% organic; DC offerings ~+30%; FY2025 DC ~€2.4B = 26% of group, potential 40% |
4.2 Liquid cooling · DLC and cold plate — 6/8 acquisition wave (2024-2026)
| Company | Acquirer | Price | Date |
|---|---|---|---|
| jetcool | Flex | undisclosed | 2024-11 |
| motivair | schneider-electric | $850M (75%) | 2025-02 |
| coolit-systems | Ecolab | $4.75B | 2026-03 |
| chilldyne | Daikin | undisclosed | 2026-03 |
| boyd-corporation | eaton | $9.5B (22.5× EBITDA) | 2026-03 |
| liquidstack | Trane | undisclosed | 2026 |
Independent remaining: nvent (public, FY2026E +26-28% rev), grc-cooling, submer, asperitas (three immersion-cooling, mostly strategic investments)
Structural conclusions:
- The era of independent pure-play AI cooling is over. Industrial conglomerates swept the market in 12 months.
- Schneider signed earliest in 2024-10, getting a mid-single-digit revenue multiple (vs Eaton paying 22.5× EBITDA in 2026-03) — early movers got cheap tickets.
- Eaton + Schneider + Vertiv + nVent now control the full DLC stack; immersion cooling (GRC / Submer / Asperitas) is the next direction to be bought.
4.3 Liquid cooling · immersion (independent specialists)
| Player | Path | Estimated funding |
|---|---|---|
| grc-cooling | Single-phase immersion (pioneer) | $43M+ cumulative (Samsung Ventures / SK strategic lead) |
| submer | Single-phase immersion + InferX AI cloud vertical integration | $131M+ cumulative (M&G lead 2024) |
| asperitas | Active-passive combined (unique) | <$50M estimated (STECON + Invest-NL 2025) |
5. Sub-Layer A.d — Networking
GPU-to-GPU communication is the AI DC's nervous system. A single GPU needs ~6 optical modules; for a million-GPU cluster, the optical modules alone consume ~180 MW.
5.1 InfiniBand vs RoCE Ethernet
| Dimension | InfiniBand | RoCE Ethernet |
|---|---|---|
| Lead | nvidia (Mellanox) | broadcom + arista + UEC |
| Latency | <1μs | ~2-5μs |
| Ecosystem | Closed | Open (UEC 1.0 released 2025) |
| Share | ~60-70% | ~30-40%, faster growth |
5.2 Switches & switching chips
| Player | Type | 2026 signal |
|---|---|---|
| arista | Ethernet switch leader | Q1 2026 $2.7B (+35%); 2026 AI fabric target $3.5B; non-GAAP op margin 47.8% (approaching SaaS) |
| broadcom | Switching chip + custom ASIC | Q1 FY2026 AI revenue $8.4B (+106% YoY); AI backlog $73B; 4 top labs' custom ASIC (Google TPU / Meta MTIA / OpenAI 10GW / Anthropic 1→3GW) |
| nvidia | Quantum IB + Spectrum-X | InfiniBand bundled by default with H200/B200; Spectrum-X counter-strike against the Ethernet route |
5.3 Optical modules · 1.6T → CPO transition
| Player | Positioning | 2026 signal |
|---|---|---|
| coherent | Lasers + optical modules vertical integration | Q3 FY2026 $1.81B (+21%); NVIDIA $2B strategic stake (2026-03); DC&Comm = 75% rev |
| lumentum | Silicon photonics + tunable lasers | Q3 $808M (+90% YoY); NVIDIA $2B strategic stake (concurrent with Coherent); 50-60% global 200G/lane EML share |
| innolight / Zhongji Innolight | #1 global shipment volume | Q1 2026 ¥19.5B (+192%); FY2025 ¥38.24B; filed for HK IPO; CPO transition is a structural risk |
NVIDIA double-betting hedge: 2026-03 simultaneously invested $2B each in Coherent (vertical integration) and Lumentum (pure silicon photonics player) — NVIDIA also doesn't know who wins the 1.6T → CPO transition, so it buys both sides.
6. The Bottleneck Cascade
TSMC CoWoS (sold out through end of 2026)
↓
Broadcom / NVIDIA (AI accelerators + switching chips)
↓
Arista (switch integration)
↓
Coherent / Lumentum (InP fab → laser chips)
↓
Innolight (module assembly)
Every node has demand > supply for 12-24 months; Q1 2026 earnings statements across all companies are consistent on this.
L1 is the second pricing-power concentration point on the AI value chain after NVIDIA. This is the structural insight from this section that should be amplified.
7. AI Revenue Concentration (investor view)
By AI exposure share in total company revenue, high to low:
| High exposure (≥75%) | Mid exposure (30–60%) | Low exposure (<30%) |
|---|---|---|
| innolight (~95%) | broadcom (~50%) | cummins (~25% Power Systems) |
| coherent (~75%) | arista (~40-50%) | ge-vernova (~15-25%) |
| lumentum (~75%) | equinix, digital-realty, iron-mountain | schneider-electric, abb, siemens |
| vertiv (~80%+ Tier-1 hyperscaler) | eaton | caterpillar |
Economic implications:
- High exposure = very high earnings elasticity, but also very high CPO transition / price-deflation risk (Innolight is the sharpest example)
- Low exposure = AI is incremental upside; share price less levered; suited for the "find a second curve" steady allocation
8. Investment Takeaways
- Vertiv is the purest L1 "picks-and-shovels" play — full-stack power + cooling, hyperscalers + neoclouds all pass through its gate
- 5-year gas-turbine lead time means the BYO Power trend is structural — Crusoe / GE Vernova / Cat directly benefit
- The window for liquid-cooling specialists has closed — the three remaining immersion-cooling players (GRC / Submer / Asperitas) are the next wave of acquisition targets
- Arista's 47.8% op margin is close to SaaS — a hardware-company outlier; the EOS operating system is the real moat
- NVIDIA simultaneously bet $2B each on Coherent + Lumentum = a double hedge on the CPO transition timeline
- TSMC CoWoS is the true source of the entire chain — all the downstream demand-supply gaps are locked by it
Sources
- Company Q1 / FY 2026 earnings releases for: Vertiv, Schneider Electric, Eaton, ABB, Siemens, Legrand, Equinix, Digital Realty, Arista, Broadcom, Coherent, Lumentum, Caterpillar, Cummins, GE Vernova, Generac, Rolls-Royce
- Wood Mackenzie 2026 Power Equipment Outlook (gas turbine pricing / lead time)
- JLL / CBRE 2026 Data Center Cost Index
- Dell'Oro Group 2026 Data Center Switch Forecast
- LightCounting Q1 2026 Quarterly Update (optical module market)
- IDC / Synergy Research 2026 Colocation Capacity Report
- Local pre-research:
~/Desktop/tokenresearch/data/L1a_equipment_suppliers.md
Compiled 2026-05-11. Cross-link to tokenmap framework for the L1 layer chip references; this module covers the long tail and analysis the main framework omits.