AI Data Center Power Chain Strategy Report: From Power Plants to Chips, New Bottlenecks in the 800V and Rubin Era

AI data center power chain, Rubin / 800VDC architecture, and near-chip power delivery strategy report

Analysis Date: 2026-06-03

Chapter 1: Opening: What AI data centers really lack is not a wire, but an inspectable power delivery path

Standing in front of a row of AI servers, the easiest things to see are the GPUs, network cards, liquid-cooling tubes, and rows of black racks. What actually determines whether these machines can work is a less visible path: electricity leaves the power plant, passes through step-up transformation, transmission, substation equipment, switchgear, uninterruptible power supplies (UPS), busbars, power cabinets, racks, server motherboards, and finally becomes the low-voltage, high-current power near the GPU cores, at less than 1V.

Every time voltage drops along this path, current rises; every time current rises, heat and losses increase exponentially. As AI rack power moves from tens of kW to hundreds of kW, and even to the 1MW class, the traditional data center power supply model cannot be solved by simply thickening the cables. Thicker copper bars, connector heating, larger switchgear, squeezed power distribution unit (PDU) space, and VRMs that must handle more violent transient current all become changes in project delivery, product certification, bill of materials (BOM) growth, and supplier profit pools.

Rubin is part of NVIDIA's next-generation AI computing platform, and it can be understood as a new generation of GPU server systems with higher power, higher density, and greater dependence on rack-level coordination. The previous generation of AI servers was already no longer like ordinary servers, where individual machines are simply placed side by side in racks; instead, GPU, high-bandwidth memory (HBM), high-speed networking, liquid cooling, and power were integrated into a single system. Rubin pushes this one step further: if a single rack must carry more GPUs, higher bandwidth, and higher thermal density, then the power system can no longer keep using the old low-voltage, high-current approach.

800VDC emerged in this context. DC means direct current, and 800VDC means 800-volt DC power delivery. It is not meant to let the GPU use 800V directly; the GPU cores still ultimately use low voltage of less than 1V. The purpose of 800V is to deliver large power to the vicinity of the rack at a higher voltage first, reducing current, copper loss, cable heating, and space usage, and then step the voltage down in stages near the servers and GPUs. In this way, the power cabinet, backup power, current protection, connectors, liquid cooling, and VRMs all have to be redesigned.

For readers unfamiliar with power systems, 800VDC can be thought of as a highway. For long-distance transport, using large trucks to carry more goods per trip is more efficient; once you reach the city and the entrance to the residential area, you switch to smaller vehicles to deliver to each building. Power delivery in AI data centers follows a similar logic: the far end and the area near the rack need high voltage and low current, while the chips themselves need low voltage and high current. The higher Rubin's power density becomes, the more complex this power delivery path from "highway to capillaries" will be.

To make the technical terms that follow feel less unfamiliar, it helps to remember a few analogies first. A power rack is like the "central power station" next to the rack, responsible for centralized management of power that used to be scattered across servers; a power shelf is like a "drawer" for power modules, which can be pulled out and replaced when broken, or used in parallel with multiple drawers for power delivery; a battery backup unit (BBU) is like an "emergency power bank" next to the rack, not necessarily able to run for very long, but able to give the system a buffer of seconds to minutes when power fluctuates; a DC busway is like a "power highway" in the rack, responsible for safely distributing high-voltage DC power; DC circuit breakers and electronic fuses are like the "emergency brakes" on that power highway, and must cut power quickly when a fault occurs; a VRM is like a "precision faucet" next to the GPU, converting higher voltage into the low voltage the chip can consume; a multiphase controller is the "traffic command center" for the VRM, coordinating dozens of current paths to share the load; and an MLCC is like a "small instantaneous reservoir" next to the chip, quickly supplying power when the GPU suddenly needs it to prevent voltage droop.

For investors, the most important question is not "which company says it benefits from 800V," but four other questions:

• Is the bottleneck grid interconnection, power-on, backup power, protection, conversion, cooling, certification, or GPU core power delivery?
• Is its revenue project-based, system-based, component-based, material-based, or service-based?
• Will demand first show up in orders, backlog, pricing, gross margin, or will it first consume inventory and capital expenditure?
• Is this company being carried by the architecture trend, or does it occupy a locked-in position in standards, certification, design-in, or on-site delivery?

When judging this industrial chain, the key is not to label every company as "AI power," but to place each one back into its real position: some companies sell long-distance power, some deliver transformers and switchgear, some work on near-rack power and liquid cooling, and some only make a very small but hard-to-replace power chip next to the GPU.

---

Chapter 2: Why Long-Distance Power Transmission Cannot Do Without High Voltage

The AI power chain can be understood from three relationships.

First, power equals voltage multiplied by current. Put simply, the same amount of “electrical work” can be delivered with high voltage and low current, or with low voltage and high current.

Second, heat in lines and devices mainly comes from losses generated when current passes through resistance. This loss is roughly equal to current squared times resistance. When current doubles, losses do not double; they become four times as large. When current falls to one-tenth, losses fall to one-hundredth.

Third, the voltage actually used by the chip at the end of the chain is very low. A GPU core does not run at 800V, and it does not run at 48V either; it operates at a low voltage of less than 1V. But because GPU power consumption is huge, the final stage of power delivery is inevitably an extreme scenario of low voltage, high current, and rapid variation.

This creates the fundamental contradiction of the power chain: long-distance transmission wants high voltage and low current, while near the chip it must be low voltage and high current. The entire power-delivery architecture of AI data centers is, in essence, solving this contradiction.

Here is a rough example. If 1MW of power is transmitted at 800V, the current is about 1,250A; if transmitted at 48V, the current is about 20,833A. The current differs by more than 16 times, and line heat losses scale with the square of current, so the engineering difficulty is on an entirely different level. Going one step closer to the chip core, if a GPU-related workload requires 1000W, and the core voltage is only 0.8V, the current is 1,250A. This explains why high-voltage transmission, 800V racks, and near-chip VRMs all become important at the same time.

Traditional power grids have long followed this logic. Power plants typically generate electricity at relatively low voltage, then use step-up substations to raise the voltage, transmit power over long distances through high-voltage or extra-high-voltage lines, and then step the voltage down again in stages near cities, campuses, and data centers. Grid materials from the U.S. Department of Energy and the EIA both emphasize that the purpose of high-voltage transmission is to reduce current and losses during long-distance transmission.

AI data centers simply push the same physical logic closer to the rack and the chip: the farther away you go, the more economical higher voltage becomes; the closer you get to the chip, the lower the voltage but the higher the current; every layer in between, including conversion, protection, connection, and cooling, becomes a new engineering problem.

---

Chapter 3: Electricity Must Pass Nine Hurdles to Get from the Power Plant to the GPU

An AI data center does not run a line directly from the power plant to the GPU. The power path can be divided into at least nine layers.

The first layer is generation and long-term power contracts. Nuclear, gas, hydropower, renewable energy, and energy storage determine whether a data center has a stable power supply. Large cloud providers are not just buying cheap electricity; they are also buying dispatchable, long-term lock-in, grid-connectable power resources that can satisfy decarbonization commitments. Companies such as Talen Energy (TLN), Constellation Energy (CEG), and Vistra (VST) enter AI power research not because they manufacture racks, but because they may control scarce dispatchable power and locations well suited for data center interconnection.

The second layer is transmission and transformation. Power plants require transmission lines, substations, main transformers, and protection systems to reach the campus. Large transformers, switchyards, interconnection windows, and transmission expansion cycles are often slower than server procurement. Extended transformer lead times, rising prices, and locked-in factory capacity are very early signs of supply tightness in the AI power chain.

The third layer is medium-voltage campus interconnection. After a data center campus receives power, it still needs medium-voltage switchgear, protective relays, circuit breakers, metering, distribution automation, and on-site construction. The companies in this layer are typically heavy electrical and power equipment platforms such as Eaton (ETN), ABB, Siemens, Schneider, GE Vernova, Hitachi Energy, Powell Industries (POWL), and Hubbell (HUBB).

The fourth layer is electrical buildings, prefabricated electrical houses (E-houses), and on-site distribution. Many AI campuses use prefabricated electrical rooms, modular switchgear, rectification, and distribution systems. POWL's E-house, ETN's electrical systems, and ABB / Siemens / Schneider's on-site distribution and automation all belong to this layer.

The fifth layer is UPS, battery energy storage systems (BESS), battery backup units (BBU), and power quality. AI loads are not smooth; GPU clusters create power fluctuations. Batteries and backup systems no longer just "keep things running for a few minutes during an outage"; they also provide transient buffering, power quality, black start, frequency support, and load smoothing. The positions of Fluence Energy (FLNC), Vertiv Holdings (VRT), ETN, Schneider, Delta Electronics (2308.TW), LITE-ON Technology (2301.TW), Tesla Megapack, and Bloom Energy should be understood in this layer.

The sixth layer is the power rack, external power sidecar cabinets, and 800VDC distribution. The core change in the Rubin / Kyber era is to move part of the power supply out of the compute rack and into a separate power rack or sidecar cabinet. The Open Compute Project (OCP) Diablo design and NVIDIA's 800VDC architecture both point in this direction. The goal is to let the compute rack carry as many GPUs, networking components, and cooling structures as possible instead of being packed with low-voltage, high-current power supplies.

The seventh layer is DC-to-DC conversion inside the server or tray. 800V does not go directly into the GPU core; it must first be stepped down to 48V, 12V, 6V, or another intermediate voltage. This layer involves high-ratio LLC resonant conversion, isolated intermediate bus converters (IBC), intermediate bus conversion, hot-swap, isolation, drivers, silicon carbide (SiC) / gallium nitride (GaN) / metal-oxide-semiconductor field-effect transistors (MOSFETs), controllers, and high-end passive components.

The eighth layer is the VRM near the GPU / ASIC. VRM stands for voltage regulator module, and it is responsible for converting the intermediate voltage into the low-voltage, high-current power required by the GPU core. What matters here is not high voltage, but many phases, high current, fast transients, and low noise tolerance. Multi-phase controllers, integrated drivers and MOSFET power stages (DrMOS), smart power stages (SPS), transient voltage regulation with coupled inductors (TLVR), multilayer ceramic capacitors (MLCC), silicon capacitors, current sensing, and telemetry all become critical.

The ninth layer is decoupling and the last few millimeters of power delivery inside and around the chip package. The closer you get to the chip, the shorter the path should be, and the more important capacitors, inductors, package substrates, vertical power delivery, and low-loss materials become. The investment mapping here is not necessarily large equipment companies; it may instead be passive components, magnetic components, materials, testing, and high-end power IC companies.

The commercial characteristics of these nine layers are completely different. Generation and long-term power purchase agreements (PPAs) are long-duration assets; transformers and switchgear are long-cycle equipment; E-houses are project deliveries; power racks are system products; BBUs are about power quality and safety certification; VRMs are design-ins; MLCCs and magnetic components are high-end passive supply; and testing and certification determine whether customers will accept delivery and whether a company can recognize revenue. Calling all of them "AI power beneficiaries" obscures the most important differences.

---

Chapter 4: Rubin and 800V Shift the Bottlenecks Toward the Rack and the Chip

The biggest change in the Rubin era is not that the GPU got a little faster; it is that the rack itself has become a high-power, high-speed interconnect, high-heat-density system. Public BOM models show that the unit value of the VR200 NVL72 rack is materially higher than that of GB300, with the increase coming mainly from GPUs, HBM, networking, and PCBs. Power delivery does not look like the biggest incremental item inside the rack BOM, but that can easily mislead investors, because most of the value in the 800V power system is not fully reflected in the server-internal power line item.

The new budget will spread outside the rack and around the rack. Power racks, power shelves, BBU / centralized backup power units (CBU), DC busways, high-voltage connectors, DC breakers, solid-state protection, 800V-to-12V/6V DC/DC, CDU liquid-cooling coolant distribution units, station-level transformers, and switchgear may all increase as Rubin power density rises.

The changes can be divided into three categories.

The first category consists of segments that already existed but have been amplified by AI. Transformers, switchgear, UPS systems, busways, PDUs, generation assets, backup power, liquid cooling, and connectors all belong here. They are not technologies that appeared only because of 800V, but AI data centers have pushed up requirements for power density, delivery speed, and reliability, turning them from ordinary infrastructure into bottlenecks for project start-up and energization.

The second category consists of segments that previously existed only in a narrow range of high-end systems and are now being pushed into standardized deployment in data centers. High-voltage DC distribution, rack-level BBU, solid-state breakers, high-voltage hot-swap, telemetry, insulation monitoring, customer witness testing, and load-bank testing all belong here. These may have previously existed in industrial, electric-vehicle, telecom, or specialized power systems, but in AI data centers they need to be scaled, modularized, and made serviceable.

The third category consists of areas that are newly added in a meaningful way, or where value per unit rises sharply, because of 800V and the higher power density of Rubin racks. External power side cabinets / decoupled power racks, 800VDC standard interfaces, 64:1 or high-ratio DC/DC, OpenVReg-related controllers, higher-phase-count GPU core-voltage (Vcore) VRMs, protected DrMOS / SmartClamp, TLVR, high-end MLCCs / silicon capacitors, and co-design of liquid cooling and power all belong here.

These three categories cannot be valued using the same logic. The first category is about backlog, lead times, pricing, and cash flow; the second is about certification, standards, on-site deployment, and customer acceptance; the third is about design wins, reference architectures, BOM lock-in, volume production, and gross margin.

The second and third categories are especially easy to confuse. The essence of the second category is to take high-end technologies that originally existed in industrial, EV, telecom power, or specialized power systems, move them into AI data centers, and turn them into reproducible, testable, maintainable standard solutions. System-level protection, rack-level backup power, high-voltage hot-swap, solid-state breakers, current sensing, insulation monitoring, load-bank testing, and customer acceptance all belong to this kind of “data-center-scale deployment of mature high-end technology.” This layer is closer to engineering delivery, and companies such as Eaton, ABB (ABBN.SW), Schneider Electric (SU.PA), Siemens (SIE.DE), GE Vernova (GEV), VRT, FLNC, Analog Devices (ADI), Texas Instruments (TXN), ON Semiconductor / onsemi (ON), Littelfuse (LFUS), Allegro MicroSystems (ALGM), Keysight Technologies (KEYS), and Chroma ATE (2360.TW) will appear at different points in it. Their commonality is not that a single component has enormous value, but that customers are willing to pay for safety, certification, less downtime, and maintainability.

The essence of the third category is that after Rubin / 800V, components that were once unremarkable or appeared only on a few high-end platforms suddenly move to a higher level of value, more phases, higher voltage, or higher reliability. Power racks and external power side cabinets will increase the system value for Flex Ltd. (FLEX), VRT, Delta, LITEON, and Schneider; high-ratio conversion from 800V down to 48V, 12V, and 6V will bring Navitas Semiconductor (NVTS), Power Integrations (POWI), ON, Infineon (IFX.DE), STMicroelectronics (STM), ADI, and TXN into a power path closer to the rack; and OpenVReg, high-phase-count VRMs, DrMOS, SPS, TLVR, and high-end MLCCs will place Alpha and Omega Semiconductor (AOSL), Monolithic Power Systems (MPWR), TXN, ADI, Renesas (6723.T), Infineon, ON, Bel Fuse (BELFB), Vishay Intertechnology (VSH), Murata (6981.T), TDK (6762.T), Taiyo Yuden (6976.T), Samsung Electro-Mechanics (009150.KS), and Yageo (2327.TW) in a more chip-adjacent value pool. Connectors, busbars, thermal management, and power co-design will continue to involve Amphenol (APH), TE Connectivity (TEL), nVent Electric (NVT), HUBB, BizLink (3665.TW), and Modine Manufacturing (MOD).

AOSL should be classified most fundamentally in the third category. It is not the leading player in rack-level BBU, DC breakers, or on-site acceptance platforms; its main thesis gap lies near GPU core power supply: OpenVReg, Vcore VRM, multiphase controllers, DrMOS / SPS, and SmartClamp. If it enters the AI server power path through high-voltage hot-swap, intermediate bus converters, or medium-voltage MOSFETs, that would only touch part of the second category; it should not be described as a core company in the second category. This distinction matters: second-category companies are usually judged first by orders, certification, and on-site delivery, whereas third-category companies such as AOSL are judged first by design wins, motherboard BOMs, sustained shipments, gross margin, and the repair of operating cash flow.

Investors also need to include the timeline. 2026 is more like a window for architecture confirmation and early projects, with the key question being whether NVIDIA, OCP, ODMs, power system vendors, and data center customers move 800VDC from reference architecture into prototypes, certification, and small-batch projects. Starting in 2027, if Rubin / VR200-related racks enter larger-scale deployment, Power Racks, BBU, DC protection, CDU, connectors, and high-end VRMs will begin to contribute to revenue more visibly. Beyond 2028, what really matters is whether 800V moves from a handful of premium rack solutions to a common configuration for high-power AI racks. If, by then, only a small number of customers and platforms adopt it, the revenue elasticity of the related companies will be below the most optimistic assumptions.

800V also carries route risk. It is more of a natural direction for high-power racks than a single standard already adopted uniformly by all customers. NVIDIA’s platform influence is strong, but AMD GPUs, Broadcom ASICs, Google TPUs, Amazon Trainium, and hyperscaler in-house systems may adopt different designs in voltage level, power shelf, BBU, protection method, and VRM architecture. A truly strong supplier should not be tied only to one advertised specification; it must be adaptable across multiple platforms, multiple customers, and multiple power architectures. For investors, this means the evidence for 800V-related companies cannot stop at “fits a certain roadmap”; it also has to show customer breadth, platform adaptability, and volume shipments.

---

Chapter 5: 800V Solves Transmission, While Low-Voltage High Current Remains Next to the Chip

Many people understand 800V as “using 800V directly at the rack.” That is only half right.

The value of 800V is that it delivers large amounts of power at a relatively close distance to the rack using high voltage, reducing current stress over long runs and on the busbars inside the rack. But the GPU core cannot consume 800V. In the end, the voltage must be stepped down to 12V, 6V, or even less than 1V. In other words, 800V does not eliminate low-voltage high-current delivery; instead, it compresses that contradiction into a place closer to the chip and requiring much more precise control.

That is why VRM, DrMOS, SPS, TLVR, MLCC, silicon capacitors, and current sensing become more important.

The power-delivery path in the 800V era can be roughly divided into four stages.

The first stage is long-distance and campus-level power transmission, where the goal is high voltage, low current, and low loss. The key components here are generation, transmission, transformers, and medium-voltage switchgear.

The second stage is power distribution near the data center and rack, where the goal is to safely deliver high power to the rack or the external power-side cabinet. The key components here are switchgear, UPS/BBU, Power Rack, DC busway, connectors, circuit breakers, and protection.

The third stage is intermediate-voltage conversion inside the rack or on the server tray, where the goal is to convert 800V into a lower intermediate voltage. The key components here are SiC/GaN/MOSFET, LLC, IBC, isolated drivers, hot-swap, telemetry, and high-voltage capacitors.

The fourth stage is GPU/ASIC core power delivery, where the goal is to turn the intermediate voltage into a stable, low-noise, fast-responding Vcore. The key components here are multiphase controllers, DrMOS/SPS, inductors/TLVR, MLCC, current sensing, and firmware telemetry.

The further upstream you go, the larger the single-ticket value, the longer the cycle, and the stronger the project nature. The further downstream you go, the value per chip may not be large, but design wins and specification lock-in are stronger, gross margins may be higher, and switching costs are also higher.

---

Chapter 6: Near-Rack Equipment Is the First to Make Its Way into Orders

Near-rack technology refers to the stretch between site-level power and the GPU board: Power Rack, Power Shelf, BBU, DC busway, connectors, protection, CDU, and field service. It sits closer to customer procurement, so it usually shows up in orders, backlog, and project revenue earlier than upstream semiconductors.

The reason Power Rack / external power-side cabinets emerged is to move power supplies out of the compute rack. Traditional servers put many server power supplies (PSUs) in each server or rack; in high-power AI racks, space is increasingly precious, and low-voltage high-current delivery consumes too much copper busbar and too many power modules. Centralizing the power supply in a side cabinet or power cabinet can increase the effective compute density of the compute rack while also centralizing backup power, protection, and maintenance.

A Power Shelf is the "drawer" for power modules. A high-power rack needs multiple power modules in parallel to provide redundancy, hot-swappability, and maintainability. The key issue here is not the specification of a single power module, but system efficiency, power density, redundancy strategy, thermal design, service safety, and customer qualification.

BBU / Central Backup Unit (CBU) provides near-rack backup power and energy buffering. Traditional UPS systems are more often deployed at the data-center level; high-power AI racks move part of that buffering capability closer to the rack. Its purpose is not long-duration energy storage, but power quality, transient fluctuation handling, and safe shutdown on a seconds-to-minutes time scale. The commercial value of a BBU depends on how many racks the customer will equip, qualification, battery management system (BMS), thermal runaway testing, fire-code compliance, field replaceability, and customer acceptance.

This layer is also the one most likely to be extrapolated too far by the market. If an energy storage system company such as FLNC appears in an AI data center reference architecture, it indicates that near-rack backup power and power quality are being reassessed by customers, but a reference architecture is not an order. What truly changes the conclusion is the project contract, delivery window, customer economics, fire protection and grid-interconnection responsibility, warranty terms, and cash conversion. If you only see an architecture diagram and do not see customer signatures and delivery responsibility, it is still only an early opportunity, not a confirmed financial increment.

DC busway, connectors, and cable assemblies are the backbone of high-voltage DC transmission. 800V reduces current, but increases the demands on insulation, safety distance, blind mating, contact resistance, maintenance isolation, and arc risk. The value here is not only in copper and plastic, but in safety design, reliability, low loss, and field maintainability.

DC circuit breakers, hot-swap, smart fuses, and solid-state protection are safety issues that must be solved before deployment. DC does not have the natural zero-crossing point that AC has, so arcs are harder to extinguish. An 800VDC protection system needs faster detection, faster interruption, higher insulation, and more accurate current and voltage monitoring. Protection-related products may not generate the largest revenue, but they can determine whether a project can pass acceptance and be replicated.

CDU and liquid-cooling systems are the other side of the power equation. Nearly all electricity entering the GPU must eventually be turned into heat and removed. The higher the rack power, the less liquid cooling is a performance optimization and the more it becomes a deployment requirement. Cold plates, distribution manifolds, quick connectors, fluid connectors, pumps, valves, leak sensors, coolant, CDUs, and on-site commissioning will all scale up accordingly.

The main company mapping for near-rack technology includes VRT, Delta, LITEON, FLEX, Schneider, ETN, ABB, Siemens, NVT, APH, TEL, MOD, CoolIT / Boyd, Advanced Energy Industries (AEIS), BizLink, and others. Their business profiles differ: VRT and Schneider are more like system platforms, Delta and LITEON are more like power manufacturing and modular systems, FLEX is a structural re-rating of CPI / SpinCo, ETN / ABB / Siemens are heavy electrical and protection platforms, NVT leans more toward racks, connectivity, protection, and system integration, APH / TEL lean more toward connectors and the interface layer, and MOD / CoolIT lean more toward liquid-cooling capacity lock-up.

The best validation signals in the near-rack segment are not slogans, but customer prepayments, capacity reservations, reference architectures, UL/IEC certification, field commissioning, order-to-shipment ratio, contract liabilities, segment margins, and normalized free cash flow.

---

Chapter 7: The voltage regulation module (VRM) is the last mile of GPU power delivery

If Power Rack solves the problem of “getting electricity near the rack,” VRM solves the problem of “delivering electricity safely, accurately, and quickly into the GPU core.”

You can think of the VRM as the GPU’s voltage-regulation heart. The GPU core voltage is very low, but power consumption is very high, current is enormous, and load changes are extremely fast. Under AI workloads, a GPU may switch from low load to high load in a very short time. The power delivery system must respond in microseconds or even less, without letting voltage collapse or overshoot and burn out the chip.

A typical GPU VRM consists of five types of components.

First is the multiphase controller. It is the brain of the VRM, deciding when each phase switches, how current is shared, how load changes are handled, and how overcurrent, overtemperature, negative current, and short-circuit conditions are protected against. The current in AI GPUs is too large for single-phase power, so dozens of phases must be paralleled. The controller’s phase count, algorithms, telemetry, protocol compatibility, and customer specifications determine whether it can enter high-end platforms.

Second is DrMOS / SPS. This is the execution layer, integrating the high-side MOSFET, low-side MOSFET, driver, current sensing, temperature sensing, and protection logic into a single package. AI GPUs require higher efficiency, smaller size, better thermal performance, and more reliable protection, so the importance of DrMOS / SPS rises accordingly.

Third are power devices such as MOSFET / GaN / SiC. Low-voltage, high-current power delivery close to the chip still mainly relies on silicon MOSFETs and high-performance power stages; higher-voltage conversion stages will use more SiC / GaN. These power devices should not all be treated as one category.

Fourth are inductors, TLVR, and magnetic components. Inductors handle energy storage and filtering, while new magnetic structures such as TLVR help improve transient response and current sharing. The rapid load swings of AI GPUs turn magnetic components from ordinary passive parts into part of system performance.

Fifth are MLCCs, silicon capacitors, and decoupling networks. They sit close to the chip and provide or absorb charge during transient load changes to prevent voltage fluctuation. Although high-end MLCCs account for much less value in AI servers than GPUs and HBM, they can see specification upgrades, longer lead times, and pricing opportunities in high-power, high-frequency, high-reliability scenarios.

This is also where AOSL’s position needs to be reinterpreted. AOSL is not a full rack-power company like VRT, nor is it a traditional small-power discrete MOSFET company that only sells discrete MOSFETs. It is closer to the final stage of GPU / system-on-chip (SoC) core power delivery: multiphase controllers, DrMOS / SPS, MOSFETs, and protection circuits.

AOSL’s website shows that the company offers hybrid-digital and fully digital multiphase controllers covering notebooks, desktops, graphics cards, servers, communications equipment, FPGA, and ASIC, and is compatible with Intel IMVP, AMD SVI3, AVS, and other requirements. This shows that it is working on the “brain” of the VRM. AOSL’s AOZ73016QI is a 16-phase, dual-output controller aimed at NVIDIA OpenVReg16 / OVR16 specifications, supports DrMOS and SPS, and can support designs of up to 48 phases. This evidence is more valuable than vaguely saying “AI data centers,” because it has entered the specification language of NVIDIA GPU / AI server power rails.

AOSL’s SmartClamp protected DrMOS, launched in April 2026, places the company in the execution layer. The product targets AI servers, data centers, and high-end graphics cards, and is used to protect multiphase VRMs by handling peak current, overcurrent, and negative current issues under AI workloads. This shows that AOSL covers both controllers and power stages in close-to-chip power delivery.

Therefore, AOSL can no longer be viewed simply as an ordinary MOSFET company. It looks more like an early design-in candidate in close-to-chip power delivery, positioned around OpenVReg, multiphase VRM controllers, DrMOS, SPS, and SmartClamp. The significance of this position is that GPU core power delivery is not merely about sending power over; it is about distributing, protecting, and monitoring dozens or even more phases of current within an extremely short time. There are not many suppliers that can participate in this layer, and they are mainly concentrated among a small number of companies such as MPWR, AOSL, Renesas, Infineon, TXN, and ADI; the number of suppliers that can truly enter AI GPU mainboards and remain in customer BOMs for the long term is even smaller. AOSL’s incremental upside comes from here: if it moves from specification language and product launches to continuous shipments on mainstream AI server mainboards, its revenue and gross margin slope could be materially higher than that of a traditional discrete-component business.

But the investment judgment must remain restrained. Having a product position does not mean AOSL has already secured a large AI server BOM; entering the OpenVReg specification language does not mean it exclusively owns the NVIDIA platform; and supplying SmartClamp in mass production does not mean the related revenue has already reshaped the company’s gross margin. The correct way to describe AOSL is: it is an early design-in candidate in the AI GPU / AI server VRM supply chain, and the most important validation is whether AOZ73016QI, AOZ73004CQI, and SmartClamp DrMOS enter mainstream GPU / AI server reference designs or ODM production BOMs, while later quarters bring improvements in advanced-computing revenue mix, gross margin, and cash flow.

Compared with AOSL, MPWR looks more like a mature AI GPU power-management platform; Infineon, TXN, Renesas, and ADI are large-platform players covering controllers, drivers, isolators, sensing, power stages, and reference designs; Vicor (VICR) is betting on a more aggressive path of modularization and vertical power delivery. AOSL’s upside comes from its market cap and old label, but it also bears the risks of insufficient evidence, low gross margin, unrepaired cash flow, and opaque customer share.

AOSL’s competitive landscape needs to be described more precisely. Its opportunity is not simply to “replace” MPWR, but to win more design-ins in high-phase-count VRMs, OpenVReg, DrMOS / SPS, and protected power stages. MPWR’s strengths are a mature platform, high customer trust, and a complete product portfolio; TXN, ADI, Infineon, and Renesas have the advantages of broad coverage across analog, power, isolation, sensing, and system customers; AOSL’s advantages lie in a more focused close-to-chip product line, earnings leverage from a smaller market cap, and the emergence of NVIDIA OpenVReg specification language. What it needs to prove most is not that it can “do it technically,” but whether it can win stable share in mainstream GPU / AI server mainboards.

This also explains why AOSL cannot be judged only by product launches. The real positive signals are a rising share of advanced-computing revenue, continuous shipments of VRM / DrMOS-related products, gross margin improvement with the mix, no abnormal inventory buildup, and operating cash flow keeping pace with revenue. The real negative signals are products remaining only in reference designs, customers failing to commit to production, or revenue growth being swallowed by low margins and working-capital consumption.

---

Chapter 8: The Sources of Profit Are Completely Different for Different Bottlenecks

In the AI power chain, there are very few monopolies in the legal sense. More common are five forms of "quasi-monopoly" or lock-in effect.

The first is resource-location lock-in. Nuclear power, gas, grid interconnection points, available interconnection capacity, substation locations, and PPA contracts cannot be replicated quickly. If companies like TLN, CEG, and VST form long-term ties with data centers, what gets locked in is not a single component, but electricity availability.

The second is manufacturing-capacity and delivery lock-in. The capacity and on-site engineering ability for large transformers, medium-voltage switchgear, E-houses, power pods, circuit breakers, and busways are limited. The advantages of POWL, ETN, GE Vernova, ABB, Siemens, Schneider, and Hitachi Energy come more from delivery windows, certifications, engineering teams, and customer relationships.

The third is standards and architecture lock-in. NVIDIA 800VDC, OCP Diablo, OpenVReg, ORv3 HPR, UL/IEC certifications, and customer reference architectures all pull suppliers into a specific design language. Once a customer platform is designed around a particular specification set, the cost of replacing the supplier rises.

The fourth is design-in lock-in. After devices such as VRM controllers, DrMOS, power stages, hot-swap components, current-sensing devices, MLCCs, TLVR, and connectors enter the motherboard or power shelf, replacement requires revalidation, retesting, and recertification. This process is time-consuming and risky, so even small components can form a high-quality profit pool.

The fifth is on-site service and acceptance lock-in. 800V, BBU, liquid cooling, and high-power racks do not finish generating revenue when they ship. They still require factory acceptance testing, customer witness testing, load-bank testing, on-site commissioning, maintenance training, and fault-liability allocation. The on-site service capabilities of companies such as VRT, Schneider, ETN, Siemens, ABB, NVT, and MOD affect revenue recognition and customer repeat business.

These lock-in effects show up differently in financial results. Resource-location lock-in is reflected in long-term contracts and capacity revenue; manufacturing and delivery lock-in is reflected in backlog and pricing; standards-and-architecture lock-in is reflected in reference designs and customer certifications; design-in lock-in is reflected in BOM share and gross margin; and on-site service lock-in is reflected in deferred revenue, service attach, and cash quality.

Therefore, even if a company is a bottleneck, it still needs further classification:

• C1 Resource / power-entry type: generation, PPA, interconnection capacity.
• C2 Long-cycle heavy-electric type: transformers, switchgear, E-house, busway.
• C3 Near-rack system type: Power Rack, Power Shelf, BBU, CDU, on-site service.
• C4 Safety, protection, and certification type: DC breaker, electronic fuse, hot-swap, BBU safety certification, testing.
• C5 Near-chip design-in type: VRM controller, DrMOS, SPS, TLVR, MLCC, sensing.
• C6 Materials / passive-component supply type: MLCC, magnetic components, silicon capacitors, SiC substrates, high-voltage capacitors, and insulating materials.
• C7 Software / control / monitoring type: power management, digital twin, telemetry, energy management platforms.

The case of NVT is a classic example. It is indeed constrained at points such as data-center electrical protection, rack enclosures, connectivity, cooling adjacency, and on-site delivery, but it is more of a service and systems-integration bottleneck than a product-shortage bottleneck. Its growth does not necessarily depend on rapid price increases and scalable replication, but on project mix, system delivery, customer relationships, on-site service, and M&A integration. Therefore, its gross margin, contract assets, receivables, and FCF conversion are more important than simple order growth.

AOSL is a different case. It is not a service-type bottleneck, but a near-chip design-in bottleneck. Its focus is not backlog, but whether its products are locked into motherboards, GPUs, ODMs, and reference designs. Its validation metrics are the share of advanced-computing revenue, gross margin, design wins, inventory, and operating cash flow.

MOD is another example, of a liquid-cooling capacity lock-in type. Customer prepayments and long-term capacity agreements prove that delivery value is high, but if the valuation has already prepaid for years of growth, then one must look at 2027-2029 revenue, gross margin, and normalized FCF.

This kind of classification is more useful than a single score, because it tells investors why a company is constraining the system, and how that bottleneck will flow into the financials.

---

Chapter 9: The inside and outside of the rack are two different books

There are two easy mistakes to make when breaking down the BOM for Rubin / VR200.

The first mistake is to look only at the BOM inside the rack. The internal rack BOM shows GPU, HBM, networking, and PCB as the biggest incremental items, which is certainly important. But much of the value in 800V sits outside the rack and around the rack, including the Power Rack, BBU, CDU, site-level heavy electrical, protection, and interconnect. If you focus only on the server internal power-supply line, you will underestimate the opportunity for power systems companies.

The second mistake is to double count system content and semiconductor content. The prices of Power Shelf, BBU, DC/DC, and protection modules already include SiC, GaN, drivers, controllers, isolators, sensors, MLCCs, and magnetic components. Revenue at the system vendor and content value at the semiconductor vendor cannot simply be added together. From an investment standpoint, you should first look at which type of system package the customer is buying, and then examine how the profit within that package is allocated among the system vendor, power semiconductors, passive components, connectors, testing, and services.

The external power and cooling budget for an 880kW high-power rack can be understood in three layers.

The first layer is the rack-adjacent 800V power package, including Power Rack, Power Shelf, BBU, DC/DC, DC protection, interconnect, and telemetry. This layer is the closest to 2026-2027 orders and small-volume revenue, with major companies including Delta, LITEON, FLEX, VRT, AEIS, BizLink, ETN, ABB, and Schneider.

The second layer adds CDU and liquid-cooling allocation on top of the power package. High-power racks must upgrade liquid cooling at the same time; liquid cooling is not purely 800V electrical revenue, but it scales together with the 800V rack. Major companies include VRT, Delta, Schneider, MOD, CoolIT / Boyd, APH, TEL, and others.

The third layer adds site-level heavy electrical allocation, including transformers, switchgear, substations, grid interconnection, and field engineering. The dollar amount of each project is larger here, but revenue recognition is slower, and what shows up first is backlog, pricing, and lead times. Major companies include ETN, ABB, Schneider, GE Vernova, Hitachi Energy, Siemens, Mitsubishi Electric, POWL, and HUBB.

Semiconductors and passive components need to be broken out one more level. SiC/GaN/MOSFETs, drivers, isolators, controllers, sensing, MLCCs, inductors, TLVR, and silicon capacitors usually enter the customer's BOM through the system package. Their dollar value is not necessarily the largest, but their gross margins, design-in lock-in, and switching costs may be better.

This breakdown explains why the stock-price reaction of system vendors and semiconductor vendors often differs. System vendors get the order first, ship first, and recognize revenue first, but gross margin and cash flow depend on project execution; semiconductor vendors see revenue later, but once the design win is secured, product life cycles and gross-margin quality may be better.

---

Chapter 10: Companies Must First Be Placed Back in the Power Path

Companies in the AI power chain cannot be mixed together in a single table. A more reasonable approach is to first position each one within the hierarchy of the power path.

Power generation and long-duration power entry points: TLN, CEG, VST, certain regulated utilities, and merchant power generators. Their key variables are PPAs, capacity markets, power prices, grid interconnection, fuel, regulation, balance sheets, and long-term cash flow.

Heavy electrical equipment and grid interconnection equipment: GE Vernova, ETN, ABB, Siemens, Schneider, Hitachi Energy, Mitsubishi Electric, POWL, HUBB, Hammond Power Solutions. Their key variables are transformers, switchgear, busway, E-houses, substations, backlog, pricing, lead times, and working capital.

Station-level heavy electrical equipment is the earliest layer that can be financially validated. For an AI campus to break ground, it often first secures transformers, switchgear, electrical rooms, busway, protection, and grid interconnection work. Here, volume and price both rising is not about any single news item, but about whether three things are happening at the same time: backlog continues to grow faster than revenue, lead times do not shorten materially, and pricing and project gross margin are not being eaten by raw materials, labor, and delay costs. If companies such as Hammond Power Solutions, POWL, ETN, GE Vernova, and Hitachi Energy continue to disclose strong orders and stable gross margins, that indicates AI power demand has already entered engineering production scheduling rather than remaining only a capital expenditure story.

Near-rack systems and services: VRT, Delta, LITEON, FLEX, NVT, Schneider, ETN, AEIS, FLNC, MOD. Their key variables are Power Rack, Power Shelf, BBU, UPS, CDU, field services, reference architectures, customer prepayments, contract liabilities, and normalized FCF.

Connectors, busway, and liquid-cooling interfaces: APH, TEL, Molex, BizLink, nVent, HUBB, ETN. Their key variables are high-voltage connections, safety clearance, blind-mate insertion, liquid-cooling quick connectors, rack density, customer qualification, and margins.

Near-chip power delivery and semiconductors: MPWR, AOSL, TXN, ADI, Renesas, Infineon, ON, STMicroelectronics, NVTS, POWI, ROHM, VICR. Their key variables are OpenVReg, multiphase controllers, DrMOS/SPS, SiC/GaN, drivers, isolators, sensing, hot-swap, DC/DC, customer design wins, product models, and gross margin.

Protection, sensing, and passive components: LFUS, ALGM, VSH, BELFB, Murata, TDK, Taiyo Yuden, Samsung Electro-Mechanics, Yageo / KEMET. Their key variables are fuses, electronic fuses, current sensing, magnetic sensing, MLCCs, inductors, high-voltage capacitors, lead times, price increases, and customer qualification.

Testing and acceptance: Keysight Technologies (KEYS), Chroma, load banks / field commissioning service providers, systems integrator service teams. Their key variables are high-power DC system validation, load testing, customer witness testing, on-site acceptance, customer acceptance, and revenue recognition. Network and optical communication testing outside the power chain should be compared separately in optical and interconnect industry research.

Liquid-cooling systems and thermal management: VRT, MOD, Ecolab / CoolIT, Boyd, Delta, Schneider, nVent, APH, TEL. Their key variables are cold plates, CDUs, distribution manifolds, quick connectors, leak detection, reserved capacity, site acceptance, and FCF.

The most common source of expectation gaps here is not necessarily the most familiar giant. Large platform companies are more certain, but their incremental growth is diluted by scale; smaller component and materials companies have more elasticity, but the evidence is weaker; project-based companies have strong order optionality, but cash flow may lag; service companies grow more steadily, but they may not have aggressive pricing power.

---

Chapter 11: New Procurement Items Brought by 800V

800V does not conjure an entirely new power chain out of thin air. Power generation, transmission, transformers, switchgear, UPS, PDU, liquid cooling, connectors, and VRM have all existed before. But 800V pushes some of these links from “standard supporting equipment” into “must be redesigned.”

The closest thing to a zero-to-one creation is the decoupled Power Rack / external power side cabinet. In traditional rack power systems, power is more distributed within servers and racks; 800V makes the external power cabinet a key configuration for high-power AI racks. It brings a new rack structure, new safety standards, new maintenance paths, and new system vendors.

The second new link is the rack-level 800VDC protection chain. High-voltage DC interruption, hot swapping, maintenance isolation, ground fault detection, insulation monitoring, and solid-state circuit breakers all have corresponding technologies in low-voltage or AC systems, but 800V AI racks require new combinations and a higher level of validation.

The third is near-rack BBU / CBU. What AI racks need is not just backup during a power outage, but transient buffering for high-power loads, power quality, and safe shutdown. This pulls batteries, BMS, fire protection, thermal management, and power electronics closer to the rack.

The fourth is high step-ratio DC/DC. Converting 800V to 12V / 6V / 48V will increase the value of SiC/GaN, isolated drivers, LLC, controllers, magnetics, and high-voltage capacitors. Revenue here first enters the system vendor BOM, then flows through to semiconductors and passive components.

The fifth is the specification language of OpenVReg and high-phase-count VRM. VRMs have existed before, but the current, transients, and power density of AI GPUs are driving upgrades in the specifications for controllers, DrMOS, SPS, TLVR, and MLCCs. AOSL’s OVR16 controller and SmartClamp DrMOS are a product-level reflection of this shift.

The sixth is the coordination of liquid cooling and power delivery. In the past, power and cooling could be designed relatively separately; in high-power AI racks, the power cabinet, BBU, CDU, cold plates, connectors, sensors, and field services must all be considered together. The reference architecture clues from NVT with Siemens, FLNC, and NVIDIA show that electrical systems, liquid cooling, and energy storage are becoming integrated delivery packages.

The common thread across these new opportunities is this: they are not simply a matter of “more demand,” but of customers having to redesign, recertify, and re-accept the system after an architectural change. Only companies that enter this redesign path are likely to win a genuine re-rating.

---

Chapter 12: The Largest Order Elasticity Usually Comes from Newly Added Procurement Items

If we look only at the 800V technology migration, we cannot fold in all the other AI hardware themes. The incremental demand discussed here comes from one thing only: rack power is moving from low-voltage, high-current architectures toward higher-voltage DC power architectures placed closer to the rack. What customers therefore need to buy more of, or redesign, are power cabinets, backup power, protection, interconnects, step-down conversion, near-chip power delivery, and field testing.

The companies most likely to see large percentage order elasticity are those with a relatively small existing business base that happen to sit directly in these newly added procurement items. Large platform companies will also benefit, but because their existing revenue base is already so large, the incremental orders may not represent the biggest percentage change in company-wide revenue.

The first category is Power Rack / external power side cabinets. In the past, server power was more fragmented inside racks and servers; after 800V, external power cabinets or side cabinets will become an important procurement item for high-power AI racks. This is not simply about selling a few more power chips. It means the customer needs to procure an entirely new rack-level power delivery system. The elasticity here is more likely to show up in systems vendors such as FLEX, VRT, Delta, LITEON, and Schneider. FLEX is distinctive because the market has often treated it as a plain-vanilla EMS provider, but if CPI / Power Rack / integrated systems continue to scale, the re-rating potential from incremental orders versus its legacy label will become much more obvious.

The second category is BBU / CBU near-rack backup power. High-power AI racks need local energy buffering and power-quality management, not just a traditional UPS that merely keeps things running for a short time during an outage. After 800V, BBU must be considered together with Power Rack, protection, liquid cooling, and field acceptance. FLNC, VRT, Delta, LITEON, Schneider, and ETN are all on this path. FLNC's elasticity may come from reference architectures or data-center energy storage solutions turning into actual project orders, but this must be treated with great caution: a reference architecture is not an order, and only when project contracts, delivery windows, customer economics, and cash collection appear does certainty improve.

The third category is DC circuit breakers, hot-swap, e-fuses, and protection devices. Protection under 800VDC is much more difficult than in low-voltage systems. Direct current does not have the natural zero-crossing of AC, so arcing and fault isolation are harder. Customers must redesign protection around safety, maintenance, disconnect, and telemetry. The dollar amount of any single item in this segment may not be the largest, but it determines whether the system can be deployed safely. ETN, ABB, Schneider, and Siemens are system-level protection and circuit-breaker platforms; ADI, TXN, ON, and LFUS are more focused on board-level protection, sensing, e-fuses, and high-reliability small components. This segment is more likely to see both volume and price rise together, because customers are not paying for "cheap"; they are paying for fewer outages, fewer incidents, and certification approval.

The fourth category is DC/DC conversion from 800V down to 48V, 12V, and 6V. 800V solves the problem of delivering high power close to the rack, but it cannot be used directly by the GPU. In between, high-ratio, high-efficiency, high-reliability step-down conversion is required. That increases content across SiC, GaN, MOSFETs, drivers, controllers, isolators, magnetics, and high-voltage capacitors. NVTS, POWI, ON, STMicroelectronics, Infineon, AOSL, MPWR, ADI, and TXN could all participate at different points. Smaller companies have greater elasticity, but only if they get designed into the system vendor's BOM, rather than remaining on a demo board or on an ecosystem list.

The fifth category is VRM, DrMOS, SPS, and TLVR. 800V has not eliminated the low-voltage, high-current rails next to the GPU; instead, it has pushed the pressure on the final power stage even higher. The higher Rubin's power density, the more GPU/ASIC Vcore delivery depends on multiphase controllers, DrMOS / SPS, TLVR, high-end MLCCs, and current sensing. AOSL, MPWR, TXN, ADI, Renesas, and Infineon are important companies in this layer. AOSL's elasticity comes from near-chip product lines such as the OVR16 controller and SmartClamp DrMOS, but it still needs to prove that these products have entered mainstream AI server BOMs and are driving higher-computing revenue, gross margin, and cash flow improvement.

The sixth category is DC busway, high-voltage connectors, and cable assemblies. 800V reduces current, but it raises requirements for insulation, creepage and clearance, contact reliability, blind mating, and maintenance. Connectors are no longer just ordinary cables; they are part of safe deployment for high-power racks. APH, TEL, nVent, HUBB, ETN, and BizLink all sit in this layer. The advantage here is that customer qualification and reliability requirements are high; the drawback is that company scale and business diversification differ widely, so the elasticity of incremental orders versus total revenue can vary a great deal.

The seventh category is high-end MLCCs, magnetics, current sensing, and small protection devices. These items are not large on a per-unit basis, but the higher demands for transients, noise, protection, and reliability in high-power AI racks will drive specification upgrades and improved high-end product mix. Murata, TDK, Taiyo Yuden, Samsung Electro-Mechanics, LFUS, ALGM, BELFB, VSH, and Yageo / KEMET can all be placed in this category. They may not be the earliest orders to take off, but if lead times lengthen, average selling prices (ASPs) rise, and gross margins improve, the quality of the volume-and-price expansion will be better.

High-end passive components in particular should not be treated as ordinary "small parts." The higher Rubin's rack power density becomes, the greater the transient current, noise-control, and decoupling pressure around the GPU / ASIC, and the specifications for MLCCs, magnetics, TLVR, current sensing, and protection devices will move up accordingly. The investment signal here usually does not begin with a big order right away; it first appears as longer lead times for high-end models, customers locking inventory earlier, rising ASPs, and product-mix improvement, before eventually reaching margin expansion. If passive-component leaders such as Murata, TDK, Taiyo Yuden, and Samsung Electro-Mechanics see server-related products growing materially faster than the company average, that shows the impact of 800V and high-power racks has already propagated from the system layer down to the component layer.

If we look only at the possible elasticity of "incremental orders relative to the original business base," the companies most sensitive to the 800V migration may not be the largest ones, but rather the smaller-base companies whose legacy labels are older and who are now entering new procurement items. FLEX may be reinterpreted because of CPI / Power Rack / SpinCo; FLNC may gain project elasticity because near-rack BBU / BESS enters AI data-center reference architectures; AOSL may achieve a steep ramp if its OpenVReg controllers and SmartClamp DrMOS enter GPU power rails; NVTS and POWI may gain early option value from 800V DC/DC and GaN / SiC routes; and although POWL is not a pure 800V component company, the large switchgear / E-house orders driven by high-power AI data centers still create very strong percentage elasticity relative to its size.

If we look only at "volume and price rising together," the ranking changes somewhat. DC protection, Power Rack / BBU, VRM / DrMOS, high-voltage connectors, and high-end passive components are more worth watching, because these areas have both unit growth and spec upgrades as well as certification barriers. By contrast, if pure systems integration scales too quickly under competition, revenue may grow very fast, but gross margin may not rise at the same pace.

Therefore, the 800V incremental opportunity can be divided into three groups:

• System order elasticity: FLEX, FLNC, POWL, VRT, Delta, LITEON.
• Component and design-in elasticity: AOSL, NVTS, POWI, ALGM, LFUS.
• Quality platform elasticity: ETN, APH, nVent, MPWR, ADI, TXN.

These three groups answer different questions. System order elasticity asks whether there will be projects, prepayments, and reference-architecture conversions to orders over the next 3-9 months; component elasticity asks whether the products enter customer BOMs over the next 4-8 quarters; quality platform elasticity asks whether multi-year orders, gross margins, and cash flow continue to improve. Only by separating them can we avoid conflating "order percentage elasticity" with "a good long-term company."

---

Chapter 13: Which Bottlenecks Are More Likely to Turn Into Profit

One important link in the chain does not mean the company can make money; a company being on an ecosystem list does not mean orders will follow. Quality can be judged with six questions.

First, where does the system get stuck if this item is missing? If the missing item is a transformer, the project may not be able to connect to the grid; if it is switchgear, the electrical room cannot be delivered; if it is BBU certification, high-power cabinets may not be deployable; if it is a VRM controller, the GPU board cannot be powered stably; if it is an MLCC, transients and yield may be affected. The closer the blockage is to customer acceptance and system irreplaceability, the higher the evidentiary value.

Second, has the customer already started paying for time? Prepayments, capacity reservations, long-term contracts, PPAs, take-or-pay agreements, and factory slot reservations matter more than ordinary partnership news. When customers are willing to lock in capacity early, it means the bottleneck has moved from a technical story into a commercial constraint.

Third, where will the revenue show up first? For system vendors, it shows up first in orders and revenue; for semiconductors, it shows up first in design wins and design-in; for passive components, it shows up first in lead times and ASP; for heavy electrical equipment, it shows up first in backlog; for liquid cooling, it shows up first in capacity lock-in and field acceptance. “Good news” looks different at different stages.

Fourth, can gross margins improve? If project-based revenue is merely a pass-through of costs, revenue growth may not equal profit growth. If a component company increases the share of high-end products, gross margins should recover. If a services company expands its installed base, service attachment should strengthen. Revenue alone is not enough; look at gross margin, segment margin, working capital, and FCF.

Fifth, is it easy to be diluted by a second supplier? Hyperscalers and the NVIDIA ecosystem usually do not like sole-source suppliers. The real moat is not that only one company can do it, but that even if the customer introduces a second supplier, the original supplier still maintains share and margins through design expertise, certification, field service, or installed base.

Sixth, has the valuation already been pre-paid? MOD’s evidence of locked-in liquid-cooling capacity is extremely solid, but if the valuation has already priced in realization in 2027-2029, then gross margins and FCF must keep up; AOSL has high product optionality, but it still needs customer BOM, revenue, and margin confirmation; near-rack system vendors are growing orders quickly, but they also need to prove project margins and cash conversion. High necessity and a good stock are not the same thing.

---

Chapter 14: The Signals Most Worth Watching Over the Next Few Quarters

First, watch whether customers move from reference architectures to orders and prepayments. NVIDIA, OCP, Siemens, FLNC, VRT, ETN, and Schneider reference architectures are only the starting point; what truly changes the financial picture is customer purchasing, capacity reservations, contract liabilities, and delivery windows.

The path from reference architecture to financial recognition usually goes through four steps. The first step is that the product gets written into the architecture diagram or specification language, showing that the technical direction has been accepted by customers; the second step is entry into prototypes, testing, certification, and field pilots, showing that the system can actually run; the third step is the appearance of customer orders, prepayments, capacity reservations, or a clear delivery window, showing that customers have started paying for time and certainty; the fourth step is that revenue, gross margin, and operating cash flow improve in sync, showing that the company has truly captured profit. Most early-stage high-beta companies get stuck between the second and third steps, and investors should not treat the second step as if it were the fourth.

Second, watch whether orders for power equipment continue to concentrate in transformer, switchgear, E-house, busway, and substation-level distribution. If the backlog in these areas continues to grow faster than revenue and pricing and gross margin remain stable, that indicates AI power capex has not remained at the theme level.

Third, watch whether near-rack Power Rack, BBU, and CDU products show clear projects and certifications. Power Rack and BBU are key to turning 800V from a roadmap into equipment purchases; CDU and liquid-cooling site acceptance determine whether high-power racks can be replicated.

Fourth, watch whether product news on VRM and power semiconductors turns into customer BOMs. For AOSL, MPWR, TXN, ADI, Infineon, Renesas, ON, NVTS, and POWI, the key is not another product launch, but whether the product enters the mass-production path for OpenVReg, GPU motherboards, ODM platforms, Power Shelves, or distribution boards.

Fifth, watch whether MLCCs, magnetic components, and protection devices show lead-time pressure and price increases. High-end passive components and protection devices are not large on a per-item basis, but if Rubin / VR200 quickly raises per-system value, the orders and gross margin from companies such as Murata, TDK, Taiyo Yuden, Samsung Electro-Mechanics, LFUS, ALGM, VSH, and BELFB will provide early signals.

Sixth, watch cash flow. Many companies in the AI power chain will see revenue rise first, inventories and contract assets rise first, and cash flow lag behind. Project-based companies should pay particular attention to operating cash flow / EBITDA (OCF/EBITDA), free cash flow / EBITDA (FCF/EBITDA), inventory days, days sales outstanding (DSO), contract assets, and customer prepayments. Growth without a cash bridge should not be placed near the top of the core main list.

---

Chapter 15: Investor Quarterly Monitoring Table

The 800V power chain cannot be judged by looking at a single metric. Each type of company has different early signals, and the timing of revenue and cash generation also differs. During quarterly reviews, you can inspect it layer by layer using the table below.

The core use of this table is simple: for systems companies, first look at orders and delivery; for component companies, first look at design wins and gross margin; for passive-component companies, first look at lead times and pricing; for liquid cooling and service companies, first look at site acceptance and cash. If you use the same yardstick for different types of companies, you will wrongly eliminate good early-stage companies and also rank too highly companies that are hot but lack sufficient evidence.

---

Chapter 16: Several Types of Companies That Are Most Easily Misread

The first category is service and system-integration bottlenecks. A portion of NVT, FLEX, VRT, Schneider, and ETN's businesses fall into this group. They do indeed constrain deployment, but they do not necessarily raise prices as quickly as shortage-constrained components. Their quality comes from system delivery, customer relationships, field service, and platform capabilities; their risks come from project gross margin, working capital, and competitive capacity expansion.

The second category is materials, packaging, or high-end passive-component companies with high physical necessity but slow monetization. Rogers Corporation (ROG), Amkor Technology (AMKR), and some high-end MLCC / magnetic component / advanced packaging companies all belong here. They may sit very low in the stack, but revenue recognition is slower, and customer qualification and path-selection risks are higher. Non-power-chain materials and communication devices should be compared separately in their respective reports.

The third category is high-upside but unproven power semiconductors. AOSL, NVTS, POWI, and some SiC/GaN companies belong here. Their technical positioning may be excellent, but without adoption by system vendors, BOM lock-in, and mass-production revenue, the roadmap cannot be treated as a financial outcome.

The fourth category is companies with strong orders but valuation already pulled forward. MOD and some highly watched liquid-cooling and near-rack power names fall into this group. The issue is not weak logic; rather, the stock price may already be pricing in consecutive delivery, leaving less room for error.

The fifth category is large platforms with high certainty but diluted optionality. ETN, GE Vernova (GEV), ABB, Schneider, TXN, ADI, and Infineon may all belong here. They are suitable for assessing long-term quality and order visibility, but they should not be compared with small-cap early-stage options using the same slope ranking.

---

Chapter 17: Orders Move First, Revenue and Cash Appear in Stages Thereafter

The impact of Rubin and 800V will flow into the financial statements in stages, not all at once. What the market sees first is often not the underlying technology itself, but the system-level components that customers must procure first, schedule first, and sign contracts for first.

The first wave usually appears in heavy electrical and substation-level equipment. Transformers, switchgear, E-houses, busbars, UPS systems, substations, and grid interconnection projects all have very long procurement and delivery cycles. Once a customer plans an AI data center campus, it has to secure factory capacity and construction windows for these equipment types well in advance. So this layer is first about orders, backlog, lead times, quotes, and customer capex, not about waiting for revenue to be fully recognized. If companies such as ETN, ABB, Siemens, Schneider, GE Vernova, POWL, HUBB, and Hitachi Energy continue to disclose rising data center electrical orders or backlog, that means AI power demand has moved from concept into engineering scheduling.

The second wave appears in near-rack systems. Power Rack, Power Shelf, BBU, CDU, liquid cooling, connectors, and field services usually do not scale until customers have clearly adopted a particular rack-level solution. The signals here are closer to revenue than heavy electrical equipment, but they still require qualification and on-site acceptance. Companies such as VRT, FLEX, Delta, LITEON, Schneider, ETN, NVT, MOD, APH, and TEL should be watched for whether reference architectures are converting into customer orders, whether prepayments or capacity reservations are appearing, whether segment margins are improving, and whether working capital is being stretched.

The third wave is where semiconductor and discrete-component revenue really ramps. VRM controllers, DrMOS, power stages, hot-swap devices, e-fuses, current sensing, isolators, MLCCs, inductors, TLVR, and similar components must first enter the designs of system vendors or ODMs, then pass qualification, BOM lock-in, mass production, and customer shipments. For companies such as AOSL, MPWR, TXN, ADI, Renesas, Infineon, ON, STMicroelectronics (STM), NVTS, POWI, LFUS, ALGM, Murata, TDK, and Taiyo Yuden, what really matters is whether design wins turn into recurring revenue and whether a higher-end product mix lifts gross margin.

The fourth wave is service, testing, and operations revenue. A high-power 800V data center is not finished once it is installed. It requires factory acceptance, site acceptance, customer witness testing, load-bank testing, on-site commissioning, troubleshooting, liquid-cooling maintenance, and electrical safety training. Service revenue may lag initial equipment revenue, but it better reflects customer stickiness. VRT, Schneider, ETN, Siemens, ABB, NVT, KEYS, Chroma, and the service teams of system integrators will all be affected by this line of business.

The financial transmission of Rubin / 800V broadly unfolds in the following order:

1. First come the AI campus and power plans, and generation and grid-interconnection assets are repriced.
2. Then come backlog in substation-level equipment and heavy electrical gear, with lead times and pricing moving first.
3. Next come orders for Power Rack, BBU, CDU, and near-rack systems.
4. Then semiconductors, discrete devices, and passive components move into mass production after design wins.
5. Finally, long-term cash flow accumulates through services, testing, maintenance, and the installed base.

This sequence explains why some companies already have financial validation today while others only have technical optionality. POWL's switchgear / E-house products may show up in revenue earlier than AOSL's VRMs, but if AOSL gets into the GPU power rail, its product life cycle and margins could be more resilient. NVT does indeed constrain data center deployment complexity, but it is more of a systems and services bottleneck and should not be treated as a pure supply-constrained component company. MOD's liquid-cooling customers are locking capacity very aggressively, but it still has to prove that 2027-2029 revenue, gross margin, and FCF can keep up with valuation. MPWR's VRMs are high quality, but the valuation already reflects very high expectations. Different companies need to be read with different financial clocks.

Chapter 18: Different Companies Earn Different Kinds of Money

The most common mistake for investors just getting into this supply chain is to see the four words "AI power" and put every company into one table. A clearer way is to first ask what kind of money this company makes.

Power generation and PPA companies earn long-term electricity availability. The key point for TLN, CEG, and VST is not how much power they sold to data centers in a given quarter, but whether they have scarce generation assets, whether they have long-term contracts, whether prices cover fuel and capital costs, whether regulators allow it, and whether customers are willing to pay a premium for low-carbon power and reliability. The risks for this type of company are equally clear: power prices, fuel, regulation, project approvals, customer contract economics, and the balance sheet.

Heavy electrical equipment companies earn project enablement. The common thread across ETN, ABB, Siemens, Schneider, GE Vernova, POWL, HUBB, Hitachi Energy, and Hammond Power Solutions is that they are not selling "smarter electricity," but the equipment and engineering capability that actually gets a project energized. To assess these companies, look at order continuity, backlog quality, whether pricing can be passed through, whether lead times are stretching, whether project gross margins are stable, and whether customers are willing to reserve factory capacity early.

Near-rack systems companies earn from the complexity of high-power deployments. Companies such as VRT, FLEX, Delta, LITEON, NVT, Schneider, ETN, AEIS, and MOD rely on combining power supplies, backup power, cooling, racks, interconnects, protection, and field services into systems that can be procured, installed, and accepted. These companies are the most likely to show revenue first, but they are also the most likely to face pressure from project gross margins, working capital, and competitive capacity expansion. They are not naturally high-margin software companies; segment margins and FCF have to be checked.

Semiconductor and power IC companies earn from design wins. If companies such as MPWR, AOSL, TXN, ADI, Infineon, Renesas, ON, STM, NVTS, POWI, and VICR get designed into GPU motherboards, power shelves, distribution boards, or OpenVReg specifications, customer switching costs rise significantly. The advantage for these companies is that gross margins and product lifecycles may be better. The downside is that revenue usually arrives later, and public evidence often stops at reference designs or demo boards, so you must avoid treating a technical path as an order too early.

The financial cadence of these two types of companies is completely different. Systems companies usually win the order first and recognize revenue first, but to deliver projects they must build inventory early, add engineering staff, and take on field debugging and warranty obligations, so cash flow may lag revenue. Component companies usually see revenue arrive more slowly, but once they are designed into a motherboard or power system BOM, a single product may ship with the platform for years, and gross margins are more likely to be supported by premium mix. If an investor only looks at revenue growth, it is easy to overestimate systems companies; if they only look at current profit, it is easy to underestimate component companies. The more rational approach is to evaluate systems companies on project gross margin and cash recovery, and component companies on design wins, gross margin recovery, and platform lifecycle.

Protection, sensing, and passive component companies earn reliability responsibility. The single-item revenue from LFUS, ALGM, VSH, BELFB, Murata, TDK, Taiyo Yuden, Samsung Electro-Mechanics, and Yageo / KEMET may not be large, but safety, transients, and noise control in high-power AI racks make these small components more important. To assess these companies, look at high-end product mix, lead times, price increases, customer qualification, and margins, rather than focusing only on revenue growth.

Testing and acceptance companies earn a necessary step before scale. KEYS, Chroma, load bank service providers, and systems vendor service teams sell validation capability. 800V, BBU, liquid cooling, high-power power cabinets, and GPU rack-level systems all require more load testing, electrical safety testing, and on-site acceptance. Their opportunity does not necessarily come from equipment shortages, but from increased test volume as complexity rises.

Liquid cooling companies earn thermal removal capability and field delivery. For companies such as VRT, MOD, CoolIT / Boyd, Ecolab, Delta, Schneider, NVT, APH, and TEL, what really needs to be proven is whether customers are locking capacity, whether systems pass field acceptance, whether gross margins are higher than traditional HVAC, and whether service revenue is recurring. Liquid cooling is a real bottleneck, but that does not mean every HVAC company has high-quality AI purity.

Once companies are clearly separated, many debates disappear naturally. Some companies are long-term infrastructure, some rely on order elasticity, some are in early design-win stages, and some depend on services and field delivery. NVT, AOSL, POWL, MOD, MPWR, and VST all seem related to AI power, but the money they make, their validation speed, and their risks are completely different.

Chapter 19: Conclusion: 800V shifts the bottleneck from power availability to safe delivery and near-chip power supply

The most important judgment in the AI data center power chain can be compressed into one sentence: long distance requires high voltage, near chip requires low voltage; 800V solves the high-power transmission in between, but it also creates new bottlenecks in conversion, protection, certification, and near-chip power delivery.

The biggest opportunity is not only in power generation, and not only in GPUs. Power generation and PPAs determine whether there is power; transformers and switchgear determine whether a project can connect to the grid; Power Racks, BBUs, and CDUs determine whether high-power racks can be deployed; DC protection and field testing determine whether acceptance can be completed; VRMs, DrMOS, TLVRs, and MLCCs determine whether the GPU core can operate stably.

What is most worth studying in the Rubin / 800V era is not a single industry, but a continuous profit chain from the power plant to the chip. The closer a company is to customer procurement and on-site delivery, the earlier it appears in orders and revenue, with typical companies including ETN, POWL, VRT, FLEX, NVT, MOD, APH, TEL, and FLNC; the closer it is to chips and specification design, the more likely it is to form high-margin design-in lock-in, with typical companies including AOSL, MPWR, TXN, ADI, Renesas (6723.T), Infineon (IFX.DE), ON, NVTS, and POWI; the closer it is to materials and passive components, the more likely it is to be underestimated by traditional financial models, but the more it needs customer qualification and a revenue bridge, with typical companies including LFUS, ALGM, BELFB, VSH, Murata (6981.T), TDK (6762.T), Taiyo Yuden (6976.T), Samsung Electro-Mechanics (009150.KS), Yageo (2327.TW), and ROG.

This is why a power chain report cannot just say "AI needs more power." Real investment research must answer: where does the power come from, how is it connected, how is it protected, how is it backed up, how is it cooled, how is it stepped down, how is it delivered to the GPU, who captures revenue at each layer, who can turn revenue into gross margin and cash, and who is only technically relevant but has not yet proven it commercially.

In the next few quarters, what is most worth watching first is not a single headline, but whether several signals appear at the same time: AI data center campus PPAs and grid interconnections continue to increase, heavy electrical equipment backlogs and pricing remain strong, Power Rack / BBU / CDU reference architectures turn into orders, VRM / DrMOS / protection components enter the mainstream BOM, high-end MLCCs / magnetic components see sustained price increases and longer lead times, and finally whether these orders can flow through to gross margin and free cash flow.

If these signals all materialize, 800V is not a short-term theme, but a long-term rerating narrative for AI data centers shifting from training centers to inference factories and from GPU procurement to system-level power infrastructure. If there is only promotion and stock price, but no orders, certification, margin, or cash, investors should treat it as a watchlist name, not as a confirmed bottleneck asset.

---

Chapter 20: Core Terminology Quick Reference

800VDC: An 800-volt DC power-supply architecture that uses a higher voltage to reduce current at the same power level, cutting copper losses and easing space constraints.
Power Rack / Sidecar: A "centralized power station" placed near the server rack that separates power conversion and backup power from the computing rack.
Power Shelf: The "drawer" for power modules, where multiple modules share the load in parallel, making redundancy and replacement easier.
Rack-level Backup Power Unit / Centralized Backup Power Unit (BBU / CBU): The "emergency power bank" beside the rack, used to address power quality, transient fluctuations, and safe shutdowns.
DC breaker / eFuse / hot-swap: The "emergency brake" on the power superhighway, preventing short circuits, arcing, surges, and fault propagation.
Voltage Regulator Module (VRM): The "precision faucet" beside the GPU that converts intermediate voltage into the low voltage required by the GPU, CPU, or ASIC core.
Multiphase controller: The VRM's "traffic command center," which determines phase switching, current sharing, protection, and telemetry.
Power stage / smart power stage with integrated driver and MOSFET (DrMOS / SPS): A high-density power stage that integrates MOSFETs, drivers, sensing, and protection.
Transient voltage regulator with coupled inductors (TLVR): The VRM's "shock absorber," used to improve the transient response of high-current VRMs.
Multi-layer ceramic capacitor (MLCC): The "instant mini reservoir" beside the chip, used for decoupling, filtering, and transient support in high-end AI servers.
Liquid cooling coolant distribution unit (CDU): Delivers coolant to the rack and cold plates.
Customer acceptance testing / on-site commissioning: Determines whether the system is accepted and whether revenue can be recognized.

---