Power architecture patents are becoming the new moat for AI hardware.
As competition in artificial intelligence servers stretches from GPUs all the way to switches, optical modules, liquid cooling and power supplies, some technologies that rarely stood in the capital markets spotlight are starting to look very expensive. The latest example is "vertical power delivery."
On October 11, Foxconn Industrial Internet Co.,Ltd. (601138) announced that the United States International Trade Commission (USITC) had formally launched a Section 337 investigation on October 9 local time. American power technology company Vicor Corporation accused Foxconn Industrial Internet, along with FII USA, Ingrasys Technology and Foxconn Ingrasys, of infringing one of its patents covering a "vertical power delivery system." Foxconn Industrial Internet said that after an internal review, the products in question are still at the internal verification and evaluation stage and pose no material impact on the company's current production, operations or financial performance.
A day earlier, Luxshare Precision Industry Co.,Ltd. (002475) also disclosed the same investigation. The company and its controlled subsidiary Dongguan Luxshare Technology were named as respondents, in connection with U.S. Patent No. 10,903,734, with the products mainly used for integrated power delivery schemes for data center power modules and processors. Luxshare Precision stated that the relevant products are still at the customer verification stage and have not yet entered mass production, and are not expected to have a major adverse impact on its current operations or performance for the period.
Judged purely on the basis of the financial statements, this does not appear to be an imminent crisis. But viewed from the direction in which the AI server industry is evolving, it looks more like a bill that has arrived early.
Over the past two years, the biggest task for Chinese supply chain companies has been figuring out how to enter the AI server order system. Now, as more and more companies push from chassis, connectors and switches deeper into complete systems, liquid cooling, power supplies and rack-level solutions, a new question has surfaced: after winning AI server orders, how much more must be spent on R&D, patents and customer certification?
The investigation launched by Vicor offers one answer.
Section 337 Hits AI Servers Right in the "Power Belly"
This is not a lawsuit aimed only at Foxconn Industrial Internet and Luxshare Precision. The list published by the USITC shows that the respondents in investigation 337-TA-1526 also include Delta Electronics, Infineon, Monolithic Power Systems and its Chinese subsidiaries, Flex, Celestica, Quanta Computer, Wiwynn, Hon Hai Precision and other companies. The list spans power semiconductors, power supplies, server ODMs and complete-system manufacturers, covering almost the entire industrial chain from power conversion to AI computing system integration.
Vicor filed its complaint on September 9 and supplemented the materials the next day. The USITC decided to open the case on October 9. Vicor is asking the commission to issue a limited exclusion order and a cease and desist order.
It must be emphasized that the opening of an investigation does not mean the USITC has already found Foxconn Industrial Internet or Luxshare Precision guilty of infringement. The USITC itself made clear that initiating a Section 337 investigation does not mean the commission has made any decision on the substantive issues of the case. The case will subsequently go through administrative law judge proceedings, an initial determination and commission review. The commission will set a target completion date within 45 days of instituting the investigation.
But what truly makes global hardware companies wary of Section 337 investigations is not mainly the size of damages. Under USITC rules, if a violation of Section 337 is ultimately found, the commission can issue an exclusion order directing U.S. Customs to block the infringing products from entering the United States. It can also issue cease and desist orders restricting the respondents from selling the relevant imported products in the United States. The Section 337 process itself does not provide monetary damages in the ordinary sense.
For an ordinary component in the consumer electronics era, that would already be troublesome enough. For companies scrambling for orders on the next generation of AI server platforms, it is even more sensitive. Once a power architecture enters the design of a GPU, ASIC or server platform, the problem is no longer simply about swapping out a supplier. PCB layout, power conversion, thermal management, structure, electromagnetic compatibility and processor power delivery interfaces may all need to be re-verified. In other words, in the age of AI hardware, patents have already become bound up with customer certification.
More notably, this is not even the first time Vicor has done this this year. In February, the USITC had already launched investigation 337-TA-1484 based on another complaint filed by Vicor, targeting data center power converters, circuit board assemblies and computing systems containing those components. At that time, Luxshare Precision, Delta, MPS, Quanta and others were already on the respondent list. That case is still in the administrative law judge stage, and USITC materials show that the initial determination is scheduled for May 20, 2027, with a target completion date of September 20, 2027.
From "Power Converters" in February to "Vertical Power Delivery Systems" directly in October, Vicor's intellectual property offensive is advancing toward the power delivery link that sits closer to the processor. That is precisely the position in AI servers that is becoming more and more valuable.
Why a "Single Wire" Becomes a Technological Barrier
To understand this patent war, one must first understand a change taking place in AI servers: GPUs are getting faster and faster, but powering them is becoming more and more difficult.
Official Nvidia materials show that a GB200 NVL72 rack contains 72 Blackwell GPUs and 36 Grace CPUs, with a fully loaded rack power consumption of about 120 kilowatts. Its power shelf supplies about 50V to 51V of direct current to rack equipment through busbars. That is no longer power consumption in the traditional server sense.
More importantly, the power level continues to rise. In the next-generation data center power roadmap published in 2025, Nvidia proposed promoting an 800V high-voltage direct current architecture starting in 2027 to support AI racks of 1 megawatt or even higher. Nvidia said the current 54V in-rack power system begins to encounter physical constraints once rack power exceeds 200 kilowatts.
The problem can be explained by a simple physical relationship. Power equals voltage multiplied by current. If the final processor operating voltage is close to 1V, then just 1 kilowatt of power implies a theoretical current of about 1,000 amperes. The larger the current, the harder it is to control losses in copper busbars, PCBs and packaging. The I²R loss in conductors in turn increases with the square of the current.
Power engineers therefore face a contradiction: to reduce transmission losses, they want to deliver electricity to a point as close as possible to the processor at a relatively high voltage. But the processor itself needs extremely low voltage and extremely high current.
This is the backdrop against which vertical power delivery technology emerged. Under the MGX rack specification system published by the Open Compute Project, the 48V DC busbar hardware supplied by mainstream manufacturers can already reach a rated capacity of 1,400A. Vicor's own technology route is to deliver 48V as close as possible to the processor and then convert it at the point of load.
According to Vicor's official technical materials, its factorized power architecture splits DC-DC conversion into two stages, regulation and transformation, using current multipliers to convert the 48V bus voltage into an approximately 1V AI processor power rail. It emphasizes placing power modules as close as possible to, or even beneath, the processor to shorten the high-current transmission distance.
So what "vertical power delivery" is really fighting over is not an ordinary power board. It is fighting over the last few centimeters of electricity. Whoever can efficiently convert tens of volts into low voltage and high current at the position closest to the XPU may be able to help customers reduce copper loss, lower heat, free up PCB area and increase the power density of the entire AI system.
This explains why the US 10,903,734 patent asserted by Vicor is titled "Delivering power to semiconductor loads," and why the overseas intellectual property dispute response guidance system of the Ministry of Commerce explicitly classifies it as vertical power delivery technology used in high-performance computing scenarios such as artificial intelligence computing.
Competition in power supplies in the AI era has evolved from "who can build a more efficient power module" into questions such as: Where should the power supply be placed? At what voltage should it be delivered? Where should the step-down occur? How should capacitors be arranged? How should the module connect to the chip? How can transient response, thermal management and high current density all be balanced?
Once the answers to these questions are solidified into circuits, packaging and system structures, they easily fall within the scope of patent protection. As a result, an industry that has looked mature for decades has once again become an intellectual property-intensive industry.
Winning AI Orders Is Only the Beginning of the Cost
For Foxconn Industrial Internet and Luxshare Precision, what is truly worth watching in this Section 337 investigation may not be the immediate financial impact. It is the timing.
In the first half of 2026, Foxconn Industrial Internet posted operating revenue of 557.861 billion yuan, up 54.63% year on year, and net profit attributable to shareholders of 23.740 billion yuan, up 95.99% year on year. Cloud computing business revenue rose 75.7% year on year, revenue from AI servers for cloud service providers rose 2.3 times, and GPU AI rack shipments rose 3.2 times. In other words, AI servers are no longer a new business on the edge of Foxconn Industrial Internet's financial statements. They have become one of its most important growth engines.
Luxshare Precision is also moving rapidly in the same direction. Half-year report data show that Luxshare Precision generated revenue of 174.504 billion yuan in the first half of 2026, up 40.16% year on year. Revenue from communications and data center business reached 16.609 billion yuan, up 49.66% year on year, and 800G optical modules have already achieved large-scale mass production.
This kind of expansion is not cheap. According to Luxshare Precision's half-year report, the company's R&D investment in the first half of this year reached 6.573 billion yuan, up 43.09% year on year. As of the end of 2025, the company had invested a cumulative 28.173 billion yuan in R&D over the previous three years and owned 9,367 patents. The key R&D directions disclosed by the company already include high-speed electrical connections, optical connections, thermal management and power management.
Foxconn Industrial Internet's R&D expenses in the first half of 2026 also reached about 4.998 billion yuan. In the same period, the company had 37,700 R&D personnel. Its half-year report shows that the company is moving further from single-product manufacturing toward full-stack solutions and joint development with core customers.
These figures reveal a reality of the AI hardware supply chain that is easily overlooked. In the traditional contract manufacturing era, the most important moat for a manufacturer might have been scale, yield and supply chain efficiency. In the AI server era, those alone are no longer enough. Companies first need to make large upfront R&D investments. They then need to pass customer verification. After entering a platform, they must also handle freedom-to-operate analysis, supplier intellectual property warranties, licensing negotiations and potential litigation. If a patent dispute arises, they may also need to revise circuit and structural designs and redo part of the certification.
In an investor relations event in April this year, Luxshare Precision even said that more than half of its R&D investment is for preliminary technology reserves for businesses that have not yet landed. That may be the most important sentence for understanding this Section 337 investigation.
The capital markets usually see only revenue after orders land. Engineers see years of R&D investment before the orders. Intellectual property lawyers see patent risks that only emerge after products actually head to the U.S. market. Customers see a supply chain that must ultimately satisfy performance, cost, reliability, delivery capability and intellectual property security all at the same time.
Foxconn Industrial Internet also stressed in its announcement that the company has a dedicated intellectual property department and requires suppliers to provide non-infringement guarantees as well as relevant technology licenses and authorizations when screening them. This means intellectual property is no longer just a legal department issue. It is entering procurement and supply chain management systems.
The Patent War Has Only Just Begun
There is currently no evidence that this Section 337 investigation will change the operating trends of Foxconn Industrial Internet or Luxshare Precision in 2026. Both companies made clear that the products in question have not yet entered formal mass production.
But precisely because of that, this dispute has a certain forward-looking significance. While a product is still in internal evaluation and customer verification, the patent war has already arrived. This is very likely to become the norm in the next stage of AI hardware competition.
As GPU power consumption rises and racks move from the 100-kilowatt level toward hundreds of kilowatts or even megawatts, the value of servers is spreading to links that were once inconspicuous: power supplies, busbars, connectors, liquid cooling, PCBs and packaging are all beginning to determine how much computing power an entire computing system can accommodate.
In the past, Chinese supply chain companies were good at answering a manufacturing question: how do you make something faster, cheaper and more reliable? After entering the AI era, they also need to answer another question: how do you prove that this technology is your own and does not tread into someone else's patent territory?
That may be harder than winning the first AI server order. It is still too early to draw conclusions about how the Section 337 investigation between Vicor, Foxconn Industrial Internet and Luxshare Precision will ultimately be decided. But one point is becoming increasingly clear: in AI servers, electricity is no longer just a cost. It is becoming performance. And once a technology can determine performance, it usually soon becomes a patent as well.
Comments