Industry Rating: Cautiously Bullish | Time Horizon: 12-24 months | Confidence: Medium Report Date: 2026-11-12 | Type: Full-coverage baseline report (initial)
Core Conclusion: The AI data center power supply chain is one of the steepest "volume-price resonance" tracks in tech hardware today. Demand is driven by both hyperscaler capex (2026E ~$730 billion, +67% YoY) and a generational leap in rack power density (120kW → 600kW/rack), while supply-side bottlenecks measured in "years" have emerged in large power transformers, premium diesel gensets, and BBU cells. The 800VDC architecture transition (starting with the 2027 Kyber rack) is triggering a generational repricing of value across the entire chain. We estimate the global data center power supply system market (full power chain scope) will grow from ~$38 billion in 2025 to ~$95-115 billion in 2030 (central value ~$104.5 billion, CAGR ~22%). But buoyancy ≠ returns: profits are highly stratified — the real money is made by players positioned within the NVIDIA ecosystem and in bottleneck segments, while domestic midstream procurement-driven segments already exhibit the classic symptom of "volume up, profit flat." Moreover, power supply orders are second-order derivatives of AI capex; the 2027 growth downshift and valuation overextension in A-share theme stocks are the two primary risks. Buy "ecosystem positioning and bottlenecks"; guard against "procurement-driven commoditization and theme valuations."
Research Scope: The full power supply chain of AI/cloud data centers — the power conversion and assurance system from grid interface to chip power intake. The chain is divided into nine segments: ① medium-voltage distribution and dry-type transformers (10kV/35kV); ② backup diesel gensets; ③ UPS (traditional double-conversion AC, including transformer-based/modular); ④ high-voltage DC (HVDC) (China's 240V/336V, overseas -48V, and the new-generation 800VDC); ⑤ busway/precision power distribution; ⑥ rack power shelves (AC/DC rectifier modules, server PSUs); ⑦ BBU battery backup units (cells + integration); ⑧ power semiconductors (SiC/GaN/IGBT); ⑨ passives such as capacitors/busbars. Liquid cooling is outside the scope of this report (cross-referenced only).
One-line industry chain map: Grid/green power (PUE and grid-connection constraints) → medium-voltage transformation & distribution + diesel gensets (facility level) → UPS or HVDC/SST (architecture layer) → busway/precision distribution → rack power shelf rectification → BBU/supercap backup → DC-DC → chip. Value is migrating from "facility-level centralized UPS" toward three segments: "architecture-layer HVDC/SST + rack-level power shelves/BBU + device-layer SiC."
Three generations of power architectures compared (the key to understanding this industry):
| Dimension | Traditional 2N AC UPS | Panama power / 240V·336V HVDC (China-led) | 800VDC + rack-level/building-level (NVIDIA-sanctioned, from 2027) |
|---|---|---|---|
| Conversion stages | 4-5 | 2-3 | 2 (endgame grid-to-chip) |
| System efficiency | ~88% | ~90-92% | +5pp end-to-end (NVIDIA figures) |
| Current/copper loss | 1MW at 54V busbar needs ~12,500A | Medium-voltage direct connection reduces low-voltage losses | At 800V, 1MW ~750A, copper loss ~-45% |
| Value center | Centralized UPS + lead-acid battery room | Medium-voltage direct-connected rectification + DC distribution | Power shelf + BBU + SiC + SST |
| Lifecycle | Mainstream in installed base, shrinking in new build | Already deployed at scale by Chinese internet players | Mandatory from 2027 Kyber; 40-60% penetration by 2030 |
| End Market | Scale and Growth | Logic |
|---|---|---|
| Big Four North American cloud providers | 2025 ~$437 billion → 2026E ~$730 billion (+67%; Amazon ~$200bn, Alphabet $195-205bn, Meta $135-145bn, Microsoft ~$190bn, FY26 basis) | Frontier model training + inference; self-built multi-GW "AI factories" pull through full-chain procurement |
| Neoclouds/model providers | Stargate plans ~7GW; xAI Colossus 2GW; CoreWeave 2026E power >1.85GW, capex $35-39bn | Power lead times (transformers 2+ years) have become the core siting constraint, favoring integrated rack/power-room solutions |
| Chinese internet giants | Alibaba RMB 380bn over three years (RMB 126bn already invested in FY2026); Tencent 2026Q1 single quarter RMB 31.9bn (+16%); ByteDance introducing 800V HVDC in tenders for the first time in 2026 | Daily token call volume grew a thousandfold in two years (National Data Bureau figures: 100bn in early 2024 → 140 trillion/day in March 2026) |
| Big Three telecom operators | 2026 total capex declining but tilting toward compute: China Mobile computing network RMB 37.8bn (+62.4%) | Mainly 240V/336V HVDC and Panama-style procurement; the base of domestic HVDC demand |
| Traditional enterprises/Colo | Global UPS market 2024 ~$29.55bn → 2032E ~$39.1bn (CAGR only 3.5%, aggregator figure pending verification); China UPS 2026E ~RMB 12.7bn (CAGR 9.2%, CCID) | Replacement of the installed base of traditional AC+UPS architecture; slowly ceding ground to HVDC |
2025 ~$38bn → 2030E central value ~$104.5bn, with each of the five drivers traceable (Chart 1):
Cross-validation and scope statement: This report's headline scope is "full power chain" (including medium-voltage distribution/transformers/diesel gensets/UPS/HVDC/busway/rack power/BBU). Dell'Oro's DCPI scope (including cooling and racks) at ~$120bn in 2030 (CAGR 22%), when split into its power portion, aligns in magnitude with our scope; whereas MarketsandMarkets' pure "data center power" scope at only ~$50.5bn in 2030 (CAGR 7.5%, aggregator figure, pending verification) — scope differences exceed 100%, so investors must verify scope boundaries before citing any "market size."
Structural (80%+ of increments): AI-driven GW-scale expansion + generational density leap + 800VDC architecture repricing. Cyclical: GPU generational transition years (GB300→Rubin→Rubin Ultra) create single-quarter order pulses and digestion periods; traditional UPS replacement fluctuates with the macro cycle. Key risk characterization: power supply orders are second-order derivatives of AI capex; once capex is cut, transmission is faster than for IT hardware itself.
| Segment | Status | Key Evidence (as of 2026-11) |
|---|---|---|
| High-power power shelves/rectifier modules | Tightest (2026-2027) → easing 2028 | NVIDIA's 800VDC component-layer reference list has only 5 names (Delta/Flex/Lead Wealth/LiteOn/Megmeet); Vertiv organic orders +100% vs own capacity expansion +45%; Delta publicly says capacity is tight (AI ~50% of its revenue) |
| BBU cells | Shortage (2026) → tight balance 2027 | 2026H1 global shipments ~276m cells (~50m in prior-year period, 5.5x); Panasonic ~40% + Samsung SDI 37% (Top2=77%, >80% including Murata, industry citations pending verification); domestic full-tab production lines in transition |
| Medium-voltage/dry-type transformers | Divergent: US large power transformers tightest (lead time ~128 weeks, Wood Mackenzie); data center dry-type transformers eased back to ~30-50 weeks | Large transformer prices +77% vs 2019; Jinpan Technology order backlog RMB 10.697bn (of which data centers RMB 3.869bn) ≈1.5x annual revenue |
| Premium diesel gensets | Tight | CAT backlog ~$78bn (nearly doubled), orders booked into 2027-2028 |
| SiC devices/substrates | Balanced → loosening | Wolfspeed resumed operations after completing Chapter 11 restructuring in Sept 2025 (debt cut ~$4.6bn); 8-inch substrate price 29,417 → 6,172 RMB/wafer (-79%), device side benefits while substrate side suffers |
| Traditional UPS/general power supplies | Oversupply (shrinking segment) | Global demand CAGR 3.5-9.2% << AI chain 30%+; FSP (Evertech) 9M26 gross margin 17.0%, net profit -98.4% |
Vertiv four plants in the Americas (South Carolina capacity ~7x, Mexico power conversion +45%), Eaton's acquisition of COL Group (European medium-voltage), Jinpan Technology convertible bond-funded power module expansion + second core production line (production Aug 2026), Egor Mexico plant (production Jun 2026), Megmeet Thailand AI power plant, Evertech (FSP) Vietnam plant (small batches late 2026/scale-up 2027). Electronics/complete-unit expansion cycles are 12-18 months, heavy equipment 2-3 years, SiC 2-4 years. Conclusion: 2026-2027 full-chain tight balance (manifested as delivery backlogs); from 2028 electronics segments turn to balance first — the first to be oversupplied will be commoditized traditional UPS and general power supplies, followed by domestic dry-type transformer capacity; heavy equipment (large transformers/diesel gensets) has the longest-lasting buoyancy.
| Segment | 2026Q4 | 2027 | 2028 | Evidence |
|---|---|---|---|---|
| High-power power shelves/modules | Tight | Tight→Balanced | Balanced | Orders +100% vs capacity +45% (Vertiv) |
| BBU cells | Shortage | Tight balance | Balanced | Shipments 5.5x, Top2=77% |
| Large transformers/diesel gensets | Tightest | Tightest | Slow easing | Lead time 128 weeks, backlog nearly doubled |
| UPS/HVDC complete units | Tight on AI side/oversupplied on traditional side | Same | Converging | Eaton b2b=1.2 |
| SiC | Balanced | Loosening | Loose | Substrate -79% |
Vertiv backlog ~$19-20bn (~13.6 months of shipments, secondhand citation pending verification), customer prepayments $4.1bn (3.6x prior-year period); Eaton Electrical backlog ~$20bn (+46%), data center orders +101% vs sales +54%. Order growth broadly exceeds shipment growth; the industry is in the early-to-mid stage of order acceleration and backlog release. But counter-evidence must be recorded: Gartner (Oct 2026) directly called current capex "unsustainable," with potential impairments exceeding $3.5 trillion; UBS expects hyperscaler capex growth to slow to ~25% in 2027; Alphabet posted negative FCF for two consecutive quarters as of 2026Q3, with shares -18% 30 days post-earnings; OpenAI/Oracle terminated Abilene 600MW expansion negotiations in March 2026 — backlog is the second derivative of capex, carries cancellation clauses, and the prepayment model binds delivery risk to the customer financing chain. Historically, order-driven equipment chains (optical communications 2000-2002, crypto miners 2022-23) took no more than 18 months from order peak to revenue decline; the orders peaking in 2026 are precisely the source of 2027H2-2028 digestion risk for this industry.
Using "global data center power supply system per-MW blended value" as the baseline metric (traditional baseline $1.9m/MW × (1 - penetration) + 800VDC full configuration $3.2m/MW × penetration, layered with facility-level inflation and SST increments, hedged against 10-20% annual module price declines):
| Quarter | Range ($10k/MW) | Basis |
|---|---|---|
| 2026Q4 | 195-205 | Mostly traditional orders + first 800V deliveries (Vertiv Q4 organic guidance +49-51%) |
| 2027Q1 | 198-212 | Rising Rubin rack power share, BBU still tight |
| 2027Q2 | 205-222 | Kyber small batches, medium-voltage price pass-through |
| 2027Q3 | 210-230 | Kyber volume ramp (if on schedule) |
| 2027Q4 | 205-235 | Two scenarios for 800V penetration 10-25%; delayed scenario takes lower bound |
Supporting path: AI power module ASP ~$0.28-0.40/W (2026Q4) → $0.22-0.32/W (2027Q4), 10-20% annual decline (100kW generational dilution + Taiwan/mainland competition).
Market size path (full power chain scope, consistent with the baseline scenario):
Growth: 2026E +24%, 2027E +30%, 2028E +21%, downshifting year by year thereafter — 2027 is the growth peak; the slope converges from 2028 onward.
Traditional UPS/general power supplies (already falling, ASP -3-5% per year with price wars continuing) → SiC substrates (-79% already happened, transmitting to devices) → AI power modules (10-20% annual decline after concentrated capacity deployment in 2027) → dry-type transformers (price increases fade to inflation levels in 2028, absolute prices not falling). Large power transformers and premium diesel gensets face no substantive price-cut risk before 2028.
Disappearing/shrinking: Traditional line-frequency UPS, lead-acid battery rooms, low-voltage distribution cabinets, STS static transfer switches — the redundant components of 2N architectures are structurally deleted under 800VDC; per-GW power equipment value rises from ~RMB 300-500 million under traditional HVDC to ~RMB 400-600 million under 800V (Guosen, cited pending verification), with the increment not in legacy segments.
Migration destinations (by elasticity): Rack power shelves/rectifier modules (per-MW value from USD 0.30M to ~1.05M, +3.5x/MW) > BBU/supercapacitors (0→1 rack-standard) > SST/medium-voltage direct connection (small batches 2028, ~25% US penetration / ~15% China by 2030) > SiC/GaN devices (1200V/3300V volume step-up) > busbars/connectors (high-voltage premium at 800V) > large transformers/diesel gensets (volume and price both rising but not an architecture dividend).
Who captures it: NVIDIA is restructuring the supplier system via its reference-design list — system level Vertiv/Eaton/Schneider, component level Delta/Lite-On/Megmeet, chip level Infineon (GaN 800VDC partnership signed with NVIDIA 2026-10). Who cedes it: legacy integrators without ecosystem positioning, the lead-acid chain, BBU pure assemblers using outsourced cells, low-end domestic PSUs. One correction that must be made: the NVIDIA ecosystem list is a "continuously expanding admission system" rather than a closed club — MGX partners showcased at OCP 2026 have grown to 102 (including Sungrow, Midea, Innoscience), so both "on the list = exponential growth" and "off the list = deletion" are oversimplifications; the list is a ticket, but profit still depends on power density/efficiency technology and capacity delivery.
| Company | Position | Key Figures | One-Line Read |
|---|---|---|---|
| Vertiv | DCPI duopoly, 800VDC system level | 3Q26 sales USD 4.31B (+61%), GM 38.1%, FY26 guide +43%, prepayments USD 4.1B | The purest 800VDC systems play, validated by both orders and margins |
| Delta Electronics | Full stack: power shelves/PSU/BBU/SST | 3Q26 revenue TWD 197.1B (+31%), GM 34.1%, ~70% Blackwell PSU share (est.) | Biggest winner of value migration; per-unit value rises rather than falls |
| Eaton | Full medium/low-voltage distribution | Electrical backlog ~USD 20B (+46%), DC orders +101% | A beneficiary but distribution is partially simplified under 800V; less pure than Vertiv |
| Infineon | #1 global power semiconductors | GaN partnership with NVIDIA, joint SST development with Eaton | Default collector of profits at the 800VDC device layer |
| Cummins | High-end gensets | 28% share, #1 in China DC gensets; DC revenue growth 30-35% | Strongest pricing power at the bottleneck |
| Jinpan Technology | Dry-type transformers, overseas | Backlog RMB 10.697B, GM 27.2% | A real beneficiary of transformer volume+price rises; US tariffs are the key headwind |
| Zhongheng Electric | Domestic HVDC/Panama leader | 9M26 revenue +30.1% but net margin 3.9%, OCF -RMB 320M | The most "pure-play" domestic 800V concept with the weakest earnings quality |
| Kstar | UPS/HVDC + IDC | FY25 revenue RMB 8.16B (+5.2%), GM 24.6% | Steady domestic substitution but muted elasticity |
| Megmeet | NVIDIA component list (earliest domestic) | 9M26 revenue RMB 9.04B (+33%), GM 20.8%, adj. net margin ~1.2%, TTM PE ~111x | Positioning is genuinely scarce, but current financials and valuation severely over-discount delivery |
| Improve Dynamics | Second-tier server PSU | 9M26 revenue +14.8%, GM 17.0%, net profit -98.4% | Specimen of a boom illusion: revenue rising, profit eaten by price war |
| EVE Energy | BBU cells | 9M26 revenue RMB 73.39B (+63.1%), GM 14.5% | BBU is a 0→1 option; in-house cell makers earn real money, pure assemblers get thin margins |
Return quality: Overall "medium". Segments earning real money — NVIDIA-ecosystem-positioned OEMs (Vertiv GM 38% + orders doubling), Taiwan component dual champions (Delta PSU share + GM 34%), high-end devices and gensets (bottleneck pricing power). Boom illusions of revenue growth without profit — low-to-mid domestic server PSUs (Improve Dynamics net profit -98%), domestic HVDC centralized-procurement system segments (Zhongheng net margin 3.9%, negative cash flow). 2026-2027 is a volume-price window; from 2028 onward, capacity release + list expansion + cloud providers' direct sourcing will push profit further toward segments with "standard-setting power + customer qualification + in-house devices".
Valuation position: SW Other Power Equipment II (801733.SI) PE 42.7x (40th percentile since 2014-07), PB 2.59x (50th percentile) — the index overall sits near decade medians, not in extreme bubble territory, but structural divergence is severe: overseas leaders are priced on order and earnings delivery, while A-share thematic names (Megmeet TTM PE ~111x / PS ~93x, Zhongheng, Improve Dynamics) are valued far above their earnings-delivery capacity — "median index + extreme single-stock premium" coexist. Sector upturn × median index ≠ sector upturn × single stocks at 100x PE; the risk-reward of the two groups is entirely different.
| Scenario | Prob. | Narrative | Implications |
|---|---|---|---|
| Bear | 30% | 2027 capex growth <+20% or Kyber slips to 2028; 800V penetration only ~10%; orders peak 2027H2, negative growth 2028 | Per-MW value stuck at USD 1.95-2.10M; market growth +5-15%; A-share 100x PE thematic valuations collapse |
| Base | 50% | capex growth slows to ~25% (UBS); 800V penetration 15-20%; bottlenecks keep premium, modules see 10-20% annual price cuts | Per-MW ~USD 2.15-2.35M; market +20-30%; ecosystem-positioned names deliver earnings |
| Bull | 20% | GW-scale projects ramp early + Kyber mandated on schedule + SST commercializes early; penetration 25-30% | Per-MW USD 2.40-2.60M; market +35%+ |
Calling out the value-trap logic: A positive sector view ≠ a buy. The lower-right quadrant's Zhongheng Electric and Megmeet — sector momentum up (x>50) but current return quality poor (y<50): Zhongheng grows revenue without profit (net margin 3.9%, negative cash flow), while Megmeet's positioning is scarce but adj. net margin ~1.2% and TTM PE ~111x already discount flawless 2027-2028 execution — both are classic "theme priced first, earnings unproven" names, and any capex or penetration miss opens a straight double-kill path. Improve Dynamics in the lower-left is the boom-illusion specimen of the generic PSU red ocean. The true "momentum × quality" double winners sit upper-right: Vertiv, Delta, Infineon, Cummins, Jinpan; EVE Energy's BBU is an option-style position (lower-right edge), with 2027-2028 in-house cell volume determining whether it migrates upper-right or stalls.
Preferred beneficiaries:
Cautious / avoid:
(For any of the above names, dedicated files with target prices and valuation bridges are available on request.)
| Scenario | Best Positioned | Worst Hit |
|---|---|---|
| Bear (30%) | Kstar (low valuation + IDC cash-flow defense), Jinpan (long order visibility) | Megmeet/Zhongheng (valuation collapse), Vertiv (order cancellations + multiple compression) |
| Base (50%) | Delta, Vertiv (ecosystem delivery), Jinpan | Improve Dynamics (price war persists), Schneider (slow AC brownfield transition) |
| Bull (20%) | Megmeet (biggest list-delivery elasticity), EVE (BBU standardization), Delta | Traditional UPS/lead-acid chain (accelerated exit) |
① AI capex slows faster than expected in 2027 (second-derivative drawdown elasticity exceeds cloud revenue itself); ② Kyber/800V transition delayed or standards fragment; ③ Taiwan/mainland capacity concentrated release in 2027-2028 triggers module price wars spreading upmarket; ④ US-China trade truce expires 2027-01-10 and tariffs snap back; ⑤ large-transformer/interconnection bottlenecks ease slower than expected, leaving demand "visible but unattainable"; ⑥ the gap between A-share thematic valuations and earnings delivery closes abruptly (downward) on any negative catalyst.
Key Sources (primary/front-line sources cited in the text): NVIDIA 800VDC technical blogs and ecosystem announcements; Dell'Oro DCPI forecasts and quarterly reports; Vertiv/Eaton/Caterpillar 3Q26 earnings releases; Wood Mackenzie transformer lead-time survey; McKinsey data center demand; Morgan Stanley power industry research (via citation, pending verification); SemiAnalysis "Inside the 800VDC Revolution"; KGI power industry reports; NDRC green low-carbon special action plan and green-power direct connection documents; FERC/PJM/ERCOT official filings; periodic reports and investor relations records of the listed A-share companies.