AI Infrastructure Opportunities (2): Where Revenue Comes From in the 800V Power Chain

NVIDIA Rubin and the 800V AI Data Center Power Chain Series Report

Analysis Date: 2026-05-27

Chapter 1: One-Page Conclusion

The first report in this series already answered several foundational questions: why Rubin is not a single-GPU upgrade but a rack-level system upgrade; why 48V / 54V architectures begin to hit constraints in the face of several-hundred-kW to 1MW racks; and why 800V is better understood as an AI data center power-architecture shift rather than an isolated component. Therefore, this second report does not repeat the physics of 800V, nor does it fully restate the list of 29 companies.

This report answers one more specific question: Which links are Rubin and 800V actually pushing capital toward, and can these revenue pools support the commercialization progress and valuations of related companies?

There are four core conclusions.

First, value inside Rubin compute racks has indeed risen significantly, but the largest increments are mainly in GPUs, HBM, networking chips, NVLink, PCBs, and board-level materials. Widely cited public figures indicate that VR200 NVL72 value per rack is about USD 7.8 million, nearly double GB300's roughly USD 3.99 million per rack. This shows Rubin is a system-level bill-of-materials step-up, but it does not fully represent the 800V power-system revenue pool.

Second, the 800V revenue pool that is most easily underestimated is outside the compute rack. In-rack power increases by only about USD 18,000 from GB300 to VR200, but deploying 800V high-power racks also requires power cabinets, high-power PSU modules, backup power, DC conversion, protection, busbar interconnects, cooling distribution units, and station-level heavy electrical equipment. In other words, looking only at the "in-rack power" line item will materially underestimate opportunities for system vendors, heavy-electrical vendors, and liquid-cooling companies.

Third, this report uses an 880kW high-power rack as a unified estimation basis, allocating external power, liquid cooling, and station-level electrical systems back to each rack by power. Under this basis, the rack-adjacent 800V power package is about USD 655,000 per rack; after adding liquid-cooling distribution, the cumulative total is about USD 755,000 per rack; after adding station-level heavy electrical allocation, the cumulative total reaches about USD 1.195 million per rack. The focus here is the size of the external supporting-equipment revenue pool created by adding one high-power AI rack, not a data hall layout diagram.

Fourth, company assessment must move from "evidence" to "valuation." Delta, LITEON, Flex, Vertiv, and Advanced Energy are closer to power cabinets, high-power PSU modules, backup power, and cooling distribution units; Eaton, ABB, Schneider, GE Vernova, Hitachi Energy, and Siemens are closer to station-level heavy electrical and data-center power infrastructure; semiconductor companies such as TI, ST, Infineon, onsemi, Navitas, AOSL, POWI, MPS, and ADI are technically important but face a longer revenue path, requiring entry into system-vendor BOMs, certification, and mass-production shipments. For investment decisions, the most important issue is not whether a company is "adjacent," but which layer of revenue it can capture, at what gross margin, when it can recognize it, and how much the current share price has already priced in.

---

This report advances beyond Part 1 in four steps:

Step 1: Break down value inside Rubin compute racks. This answers how money is distributed inside the compute rack.

Step 2: Break down the 800V external revenue pool for compute racks. This answers how much power cabinets, liquid cooling, backup power, protection, and heavy-electrical equipment can capture.

Step 3: Break down the semiconductor revenue path. This answers why SiC / GaN, isolation, drivers, control, and sensing are important but may not be reflected in revenue first.

Step 4: Map the above breakdowns to company commercialization and valuation pressure. This answers which companies have the strongest evidence, which face the greatest valuation pressure, and which may still have expectation gaps.

Chapter 2: Core Evidence 1: Where Money Flows Inside Rubin Compute Racks

Start with the bill of materials inside the compute rack. The compute rack here refers to the rack that houses GPUs, CPUs, memory, high-speed interconnects, motherboards, board-level power, in-rack thermal components, and assembly/test, not the adjacent power cabinet that supplies it.

Widely cited public Rubin figures show VR200 NVL72 value per rack at about USD 7.8 million, nearly double GB300 NVL72 at about USD 3.99 million per rack. The core significance is not merely that GPUs are more expensive, but that the Rubin platform reallocates value toward a full-rack system with higher power, higher bandwidth, and more complex interconnects.

This table tells us three things.

First, Rubin's primary incremental dollars still come from compute and memory: GPU plus memory contributes the majority of absolute uplift. Second, interconnect and board-level materials are growing rapidly, showing value is no longer concentrated only in GPUs, while the complexity of NVLink, networking chips, PCBs, ABF, and assembly/test is all increasing. Third, in-rack power rises by only about USD 18,000, which is too small to judge the full external 800V power-system opportunity.

If the bill of materials is consolidated into several main lines, value migration becomes clearer:

Therefore, the investment implication of Evidence 1 is: Looking only at the Rubin in-rack BOM will overestimate non-GPU components that remain inside the compute rack, and underestimate external rack-adjacent power, liquid cooling, protection, backup power, and station-level electrical equipment.

---

Chapter 3: Core Evidence 2: Most Incremental 800V Value Sits Outside the Compute Rack

Evidence 1 focuses on the in-rack Rubin BOM: GPU, HBM, networking chips, PCB, in-rack power, and in-rack cooling. Evidence 2 answers a different question: to actually run high-power AI racks, how much additional external power, cooling, and heavy-electrical systems are required?

First, note a basis difference: the roughly USD 7.8 million BOM for VR200 NVL72 measures value inside the compute rack; the later 880kW high-power rack basis estimates the revenue pool of external supporting systems. They are presented together to compare revenue direction and order of magnitude; for project-level modeling, external allocations should be adjusted by actual rack power.

This is also why you cannot look only at the "in-rack power" line. Versus GB300, VR200 adds about USD 18,000 of in-rack power, but that only represents the internal power portion of the compute rack. Actual deployment of 800V high-power racks also requires power cabinets, sidecar power cabinets, high-power PSU modules, backup modules, DC conversion, DC protection, busbar interconnects, cooling distribution units, transformers, switchgear, grid-interface equipment, and on-site services.

Therefore, the main incremental 800V revenue pool is not in the small "in-rack power" line item, but in external rack-adjacent power, liquid cooling, and station-level electrical systems. Whoever can deliver deployable power-cabinet, backup-power, protection, interconnect, liquid-cooling, and heavy-electrical solutions is closer to this cycle's revenue realization.

The conclusion of this table is not that every row can be added mechanically. Many items overlap within system packages: for example, DC conversion and SiC / GaN devices are included in PSU modules or system-vendor BOMs; protection, busway, and interconnect may also be sold as integrated solutions by system vendors or heavy-electrical vendors. Therefore, this table is better for judging value-migration direction than for summing all rows into one per-rack total.

External power and cooling revenue pool per 880kW high-power AI compute rack.

This table is straightforward: Cumulative Basis means the coverage scope expands layer by layer; Increment vs Previous Layer means only the newly added revenue portion. When one new 880kW high-power AI rack is added, liquid-cooling allocation adds about USD 100,000 per rack; station-level heavy-electrical allocation adds roughly another USD 440,000 per rack.

The investment implication of Evidence 2 is: Per-rack revenue capture for system vendors and heavy-electrical vendors may be far higher than for any single semiconductor supplier. Semiconductor devices are critical, but usually capture only part of the value inside system packages; if system vendors can deliver full power-cabinet, backup-power, and liquid-cooling solutions, per-rack revenue-capture efficiency is higher.

---