AI Datacenter Power Investment Map: From 765kV to 0.65V
Where Along the Chain Is the Real Opportunity?
Everyone says power matters for AI.
They’re right.
Global datacenter power demand is projected to more than double from roughly 59GW in 2023 to over 122GW by 2030.
But knowing that and knowing where to invest are entirely different problems.
Sell-side reports lump Eaton, Schneider, Vertiv, and MPS together as “power infrastructure plays.” This is no different from grouping NVIDIA and Applied Materials under “semiconductor stocks.”
A 765kV substation transformer that Eaton builds and a sub-1V GPU VRM that MPS designs just happen to share the word “power.” Peel back the surface and the technology, competitive dynamics, margin structure, and investment thesis are completely different businesses.
Power is not a single theme.
It is a chain with at least ten conversion stages from the power plant to the GPU die, and each stage has its own moat characteristics, competitive intensity, and growth curve.
This article breaks down the entire chain, maps 24 public companies onto each segment, and tracks where the largest value shifts are happening from an investor’s perspective.
Disclaimer
This article does not constitute a recommendation to buy or sell any specific security. All content is based on publicly available information and is intended solely to help readers better understand the AI datacenter power chain. Investment decisions should be made independently based on your own research, risk tolerance, and investment horizon. This article is designed to help you build your own investment principles and strategy.
The Structure of the Power Chain: Three Voltage Tiers, Ten Layers
The most intuitive way to divide this chain for investment purposes is by voltage. Electricity leaving a power plant passes through three major voltage tiers before it reaches the GPU.
High-voltage tier (765kV → 13.8kV)
This is where power is sourced and delivered to the datacenter. Power plants, transmission lines, and substations live here. Once installed, this physical infrastructure runs for 20 to 30 years, and lead times of 2 to 7 years make it the slowest bottleneck in the chain.
Medium-voltage tier (13.8kV → 48V)
This is where sourced power is distributed inside the datacenter building to the servers, while simultaneously removing heat. Switchgear, UPS (uninterruptible power supply), PDU (power distribution unit), server PSU (power supply unit), and cooling systems all belong here. This tier is also where the largest architectural shifts (800V DC, 48V direct-to-load) are underway.
Low-voltage tier (48V → 0.65V)
The “last 1cm,” where voltage undergoes its final conversion right next to the GPU. GPU core voltage varies between 0.65V and 0.85V depending on operating conditions; this article uses 0.65V as the reference point to illustrate the full conversion range of the power chain. VRMs (voltage regulator modules), next-gen power semiconductors (GaN/SiC), passive components like MLCCs (multilayer ceramic capacitors), and power connectors all belong here. Growth rates are the highest but moat durability is the lowest.
How each tier’s moat, growth rate, and margin structure differ is laid out layer by layer below.
High-Voltage Tier: The Layers That Secure Power
① Generation / Grid
The starting point of power. For AI datacenters to secure GW-scale electricity, the fight begins with long-term power purchase agreements (PPAs) with generators. In key regions like Virginia, applying for grid interconnection and actually receiving power takes 4 to 7 years. Three distinct businesses are mixed inside this layer.
First, grid generation and PPA contracts. Constellation Energy closed its acquisition of Calpine in January 2026, becoming the largest private power producer in the US (55GW). Vistra has signed a 20-year PPA with AWS (up to 1,200MW) and a 2,600+MW PPA with Meta from its Comanche Peak nuclear plant. 2026 EBITDA guidance: $6.8 to 7.6B. Talen Energy signed a 1,920MW PPA with Amazon (through 2042) from its Susquehanna nuclear facility.
Second, grid infrastructure construction. Quanta Services is the key contractor that builds transmission lines and substations. Q1 2026 revenue $7.87B (up 26% YoY), backlog $48.5B.
Third, on-site generation. Bloom Energy uses SOFC (solid oxide fuel cells) to generate power directly at the datacenter site. This is a head-on bypass of grid interconnection queues. Oracle Project Jupiter up to 2.45GW, Brookfield $5B partnership. Q1 2026 revenue $751M (up 130% YoY), annual guidance $3.4 to 3.8B. Total backlog approximately $20B per company IR.
② Substation / High-Voltage Conversion (765kV → 13.8kV)
The first step-down, converting high-voltage transmission power (345 to 765kV in the US) to levels the datacenter can use. Transformers and high-voltage switchgear are the core equipment. Lead times of 2 to 3 years make this the physically slowest bottleneck.
Eaton holds the dominant installed base in North American datacenters.
Q1 2026 results announced May 2026: revenue $7.45B (up 17% YoY), datacenter orders up 240% YoY, datacenter revenue up 50%, backlog $22.8B.
That said, earnings growth (+2%) lagged significantly behind revenue growth (+17%). Capacity expansion costs are pressuring margins in this phase.
Expanding into a full power-plus-cooling stack through the Fibrebond ($1.45B) and Boyd Thermal ($9.5B) acquisitions.
GE Vernova is a full-stack player spanning layers ①②③, from gas turbines (generation) to transformers/switchgear (grid) to 800V DC architecture. Q1 2026 datacenter Electrification orders $2.4B (already exceeding full-year 2025). Company-wide revenue guidance $44.5 to 45.5B, backlog $163B.
ABB announced a collaboration with NVIDIA on 800V DC power architecture and 1MW server rack support, positioning itself across layers ②③④.
Powell Industries specializes in datacenter medium-voltage switchgear (5 to 38kV). Q2 FY2026 backlog $1.8B, $400M+ datacenter mega-order booked.
Medium-Voltage Tier: The Layers That Distribute Power and Remove Heat
③ Medium-Voltage Distribution (13.8kV → 480V AC)
The distribution segment inside the building. Medium-voltage switchgear and distribution panels are the key equipment. Because specifications are locked in at the datacenter design stage, switching vendors after selection is extremely difficult. Eaton and ABB compete in this segment as well, but the most differentiated position belongs to Schneider Electric.
Schneider’s strength lies in EcoStruxure, a DCIM (datacenter infrastructure management) software platform, combined with digital twin technology that virtually simulates datacenter power and cooling without physical prototypes. It is a software moat that captures design decisions inside the Schneider platform.
④ AC-DC Conversion / UPS (480V AC → 48V to 800V DC)
The critical conversion point from AC to DC. UPS (devices that maintain server power during outages), PDU (devices that distribute power to multiple servers), and rectifiers (AC-to-DC converters) are the core equipment. Conversion efficiency at this layer directly determines the datacenter’s overall energy efficiency (PUE).
Schneider leads through design lock-in. Vertiv has differentiated on fast delivery and grown rapidly. Revenue guidance $13.5 to 14.0B, Q1 2026 revenue $2.65B (up 30% YoY), order backlog roughly $9.5B. What makes Vertiv unique is that it is the only large company covering both power (UPS/PDU) and cooling simultaneously. Delta Electronics is expanding on cost competitiveness.
Legrand positions across layers ④⑩ with Raritan/Server Technology-based intelligent rack PDUs and busway systems.
The 800V DC power architecture ecosystem that NVIDIA first introduced at GTC 2025 and expanded in 2026 includes ABB, Eaton, Schneider, Vertiv, Infineon, STM, and Navitas. If 800V DC gains traction, this layer becomes the intersection where equipment companies and power semiconductor companies meet.
⑤ Rack / Server PSU (48V ↔ 12V)
The PSU layer that feeds power to servers. 48V direct-to-load and 800V DC have the potential to reshape this layer. This is the first contact point where architectural transitions reach actual datacenter designs.
Delta Electronics holds a dominant position on cost competitiveness. Advanced Energy is targeting 12V/48V output power solutions.
⑥ Cooling
Every voltage conversion generates heat. With GPU TDP (thermal design power, the maximum power consumption and heat output during peak GPU operation) reaching 1,400W according to supply chain reports, air cooling has hit its limits. The latest GPUs (B300/GB300) require liquid cooling.
Vertiv covers both power and cooling.
nVent specializes in rack-level thermal management.
Eaton entered cooling through the $9.5B Boyd Thermal acquisition.
Cooling is not technically a layer on the voltage conversion chain, but CDUs (coolant distribution units) sit alongside UPS equipment and cold plates share space with PSUs inside server racks, so it moves in lockstep with the medium-voltage tier’s infrastructure.
Low-Voltage Tier: The Layers That Perform Final Conversion
⑦ VRM / Power Stage (12V/48V → 0.65V)
The “last 1cm,” converting voltage right next to the GPU. A VRM (voltage regulator module) steps down 12V or 48V to GPU core voltage (0.65 to 0.85V).
GPU TDP rose from 700W on H100 to 1,000W on B200, and supply chain reports estimate B300 will reach 1,400W. Industry estimates suggest converting this to sub-1V requires handling over 2,000A of current, and assuming 60 to 70A per phase, more than 30 VRM phases must be arranged around the GPU. That said, advanced power stages (90A+ per phase) that reduce the phase count are also under parallel development.
MPS has established itself as the most prominent GPU VRM supplier with clear design wins in the H100/B200 generation. 2025 revenue $2.8B (up 26.4% YoY).
Vicor takes a different approach with its proprietary FPA (Factorized Power Architecture) and VPD (Vertical Power Delivery, a technology that places power components vertically directly beneath the GPU). Q1 2026 backlog $300.6M (up 70% QoQ), revenue guidance ~$570M.
Infineon raised its AI DC revenue target to €1.5B for FY2026.
STMicroelectronics co-announced the 800V DC-to-6V architecture with NVIDIA.
In parts of the supply chain, there are signs that NVIDIA is diversifying power component sourcing for its next-generation GPU platform.
⑧ Next-Gen Power Semiconductors (GaN/SiC)
The mainstream Si MOSFETs used in today’s VRMs generate more heat and power loss as switching speed increases.
GaN handles the same power while being smaller, faster, and cooler.
SiC excels in high-voltage segments, while GaN is strongest in medium-to-low voltage, high-frequency segments like server VRMs. GaN adoption is a likely direction, but other paths including advanced Si-based designs are also being explored.
Navitas unveiled an 800V-to-6V board aligned with NVIDIA’s 800V DC direction at GTC 2026. 2025 revenue $46M, AI pipeline $165M. Infineon acquired GaN Systems to secure its GaN portfolio. STM straddles both SiC and 800V DC.
⑨ Passive Components (MLCC, Inductors)
Industry estimates indicate AI servers consume 10 to 15 times more passive components than standard servers. Supply chain reports suggest Murata maintains a share lead in high-spec AI server MLCCs.
Samsung Electro-Mechanics ranks second, with TDK also a major supplier. Lead times for high-spec MLCCs have stretched to 24 weeks per supply chain reports.
⑩ Power Connectors / Cables / Busbar
Everything that physically connects power belongs in this layer. As GPU power rises, connectors must handle higher currents. As server density increases, cable and busbar quantities and specifications scale alongside.
Because connectors go into servers, networking, and power infrastructure everywhere, this layer has leverage on the entire chain’s growth.
Amphenol is showing explosive growth in this space. 2025 revenue $23.1B (up 52% YoY), IT Datacom segment alone $8.3B (up 128%), Communications Solutions OPM 32.7%.
Four Architectural Transitions Underway on This Chain
Once you understand the ten layers, you need to recognize the four structural transitions simultaneously unfolding across them. These transitions are reshaping each layer’s competitive dynamics and demand structure.
48V direct-to-load. Most server internal power currently travels at 12V. 48V direct-to-load skips this 12V step, delivering power from 48V straight to the GPU side. Removing one conversion stage improves efficiency and reduces component count. If this transition accelerates, layer ⑤ (PSU) designs change fundamentally, and layer ⑦ (VRM) architectures are affected as well.
GaN (gallium nitride) material transition. Covered in detail at layer ⑧, the key point is that a material has emerged that pushes past the physical limits of traditional Si-based power semiconductors. As GPU TDP exceeds 1,400W per supply chain reports, Si-based VRMs are running out of physical board space for placement. GaN handles the same power in a smaller footprint, solving this problem.
800V DC distribution. Today’s datacenter interiors receive 480V AC and convert it to DC through multiple stages. 800V DC changes this by supplying 800V direct current into the building from the start. Fewer AC-to-DC conversion stages improve efficiency and reduce copper cable usage. First introduced by NVIDIA at GTC 2025, this architecture has the potential to fundamentally alter layer ④ (UPS/PDU) design. Equipment companies like ABB, Schneider, and Vertiv are positioning for this transition alongside power semiconductor companies like Infineon and STM.
On-site generation. Instead of pulling power from the grid, this approach generates it directly at the datacenter site. Bloom Energy’s SOFC is the leading example. As a solution that bypasses the 4 to 7 year grid interconnection bottleneck, it operates independently of the other three transitions while potentially reshaping the relationship between layer ① (generation) and layer ④ (UPS).
These four transitions are unfolding simultaneously across different layers, and the speed of each transition will determine each layer’s winners and losers.
That covers the topography of the power chain: how 24 companies are positioned across ten layers and how the investment characteristics of the three tiers differ.
From here on, I aim to answer the following questions in greater depth. “What framework should be used to evaluate these 24 companies?” “At this point in the cycle, which tier is most favorable?” And “if I were buying this chain, how would I structure the portfolio?”
For anyone just starting to invest in AI datacenter power, I am confident this will serve as an excellent foundation for building your own investment strategy.
