USITC Launches VPD Patent Investigation Affecting AI Server Supply Chain

Table of Contents

• You might want to know
• Main Topic
• Key Insights Table
• Afterwards...

You might want to know

• What does the USITC investigation mean for companies supplying AI server power systems?
• Could the case disrupt entire server shipments, or might manufacturers avoid the impact by changing specific power modules?

Main Topic

The U.S. International Trade Commission (USITC) voted last Friday (9th) to institute a Section 337 investigation into “certain vertical power delivery systems (VPD), components thereof, and computing systems containing the same.” The respondents include Luxshare Precision, Foxconn Industrial Internet, Hon Hai, Delta Electronics, Quanta, Flex, Celestica, Infineon, and Monolithic Power Systems (MPS). Taken together, the list reaches across much of the global AI server contract-manufacturing and power-supply chain.

The complainant, Vicor, is a Nasdaq-listed power-module company. It filed its complaint on September 9, and the investigation was instituted within the same month, a relatively fast pace. That speed may draw attention from investors and suppliers, but it should not be confused with a final finding. Opening a Section 337 investigation does not establish that any respondent infringed a valid patent. The evidence, patent claims, and questions of validity will still be considered in proceedings before an administrative law judge.

The technology at the heart of the dispute addresses a power-delivery challenge created by increasingly demanding processors. A high-end GPU can require more than a kilowatt of power. In conventional designs, current travels laterally across the motherboard before reaching the chip. The longer route and associated impedance can increase power losses and make efficient delivery more difficult. VPD places power-conversion components beneath the GPU package, allowing power to travel vertically over a much shorter distance. The route is reduced from inches to millimeters.

This “last inch” of power delivery is becoming more consequential as AI accelerators draw more power and system designers seek to control losses, heat, and performance constraints. A shorter path can be valuable in a densely integrated server, but it also makes the underlying circuit architecture and packaging choices strategically important. The investigation therefore concerns not only individual power modules, but also components and computing systems that may incorporate the challenged technology.

Experts say Vicor has spent a decade building its patent portfolio. In October last year, the company expanded its licensing program, and in September this year it began offering VPD licenses to major AI original equipment manufacturers (OEMs). At the same time, it has pursued companies that it says are using the technology without authorization. That combination of licensing and litigation has led some observers to characterize the approach as an effort to establish a patent tollbooth around an important part of the AI power chain. The claim remains a matter for the legal process, rather than a conclusion about the merits of the patents or the conduct of the respondents.

Section 337 investigations are conducted under the Tariff Act of 1930. The most significant potential remedy is an exclusion order, which can bar infringing products from entering the United States. Such an order is not the same as a monetary penalty: it does not impose fines, close factories overseas, or automatically restrict sales in other markets. The potential trade remedy is tied to the importation of products into the United States, rather than a blanket prohibition on a company’s global operations.

A Section 337 case usually takes one year to one and a half years to resolve. During that period, affected companies may continue shipping products unless an interim measure applies. They also have several possible responses: redesigning a product to avoid the asserted claims, obtaining a license, reaching a settlement, or challenging the validity of the patents. Which path is practical will depend on the specific product, the scope of the patent claims, the evidence, and the commercial relationship between the parties.

The current case follows an earlier dispute, 337-TA-1370. In February 2025, a final determination found that certain Delta modules infringed, and a limited exclusion order was issued. Foxconn’s U.S. subsidiary was exempt after relying on a license secured through procurement terms. That history offers context, but it does not predetermine the outcome of the new investigation. The present respondent list includes companies from the United States, Germany, and Canada, indicating a supply-chain commercial patent dispute rather than a country-specific tariff action.

Market timing may also affect the initial response. Because of the time difference, share prices did not immediately reflect the news. On Monday (12th), related contract manufacturers, server-power suppliers, and power-management chip companies could open lower or face pressure. Yet the scale of any lasting effect depends on the final scope of a remedy. If an eventual exclusion order covers only specific modules rather than complete servers, manufacturers may be able to switch power designs or suppliers. In that scenario, the impact could differ substantially from one company or product line to another.

For that reason, a broad sell-off across all AI server-related shares is not the only possible outcome. Some firms may have direct exposure to the challenged technology, while others may be able to substitute components or rely on licensed designs. The likely risk is therefore differentiated across the supply chain, not necessarily a uniform shock to every AI server manufacturer. Investors and industry participants will need to follow the claims at issue, the products identified, and any subsequent decisions on remedies.

The dispute also highlights a broader shift in the constraints facing AI infrastructure. Attention has already extended beyond processors to optical modules and liquid cooling; power delivery is now another potential bottleneck. High-voltage direct current at 800V is being considered for entry into server racks, while VPD places conversion closer to, or beneath, the processor package. As power architecture becomes more advanced, the value associated with power systems can increase. At the same time, critical patents and technical know-how may be held by companies outside the domestic supply base.

The AI power ecosystem can be viewed in three layers. The materials layer includes soft magnetic materials, power components, high-frequency capacitors, and copper-clad laminates. The device layer includes DC-DC converters, power racks, and vertical power modules. The system layer includes AI servers and 800VDC data-center infrastructure. These layers are connected, but they do not carry the same exposure to a patent dispute. A materials supplier, a module maker, and a server assembler may each face different commercial and legal risks.

Experts emphasize that competitive advantage is increasingly determined by core topologies and patent rights, not assembly capacity alone. The investigation may therefore encourage companies to review licensing, design alternatives, and ownership of key power technologies. For investors, a useful distinction is whether a company develops its own power-delivery solutions, occupies a strategic position in essential materials, or simply applies an AI-power label to its products. Assessing technological capability and actual supply-chain exposure is more informative than reacting to the headline alone. This approach may also help investors avoid panic-driven trades while evaluating opportunities associated with domestic substitution.

Key Insights Table

Afterwards...

The next stage will be the examination of patent scope, validity, and the specific products identified in the complaint. Until decisions clarify those issues, the investigation is a source of uncertainty rather than a final judgment against the named companies. The most useful signals will be developments in the administrative proceedings, licensing discussions, and any evidence that manufacturers are changing designs or supply arrangements.

Over the longer term, the case underscores how power architecture is becoming a central part of AI infrastructure strategy. Companies that can develop efficient designs, secure access to key intellectual property, or provide hard-to-replace materials may be better positioned as server power requirements evolve. For market participants, distinguishing genuine technical capability from broad thematic branding will remain essential as the AI supply chain adapts.