Introduction: Power architecture patents are becoming a new moat for AI hardware.
As the competition for AI servers extends from GPUs to switches, optical modules, liquid cooling, and power supplies, some technologies that rarely appeared in the spotlight of the capital market are becoming increasingly expensive.
A recent example is "vertical power delivery."
On October 11, Foxconn Industrial Internet announced that the United States International Trade Commission (USITC) officially initiated a Section 337 investigation on October 9, local time. US power technology company Vicor Corporation accused Foxconn Industrial Internet, FII USA, Hongbai Technology, and US Hongbai Technology of infringing its "vertical power delivery system" patent. Foxconn Industrial Internet stated that following an internal review, the products involved are still in the internal validation and evaluation stage, which will not have a substantive impact on the company's current production, operations, and performance.
A day earlier, Luxshare Precision also disclosed the same investigation. The company and its controlling subsidiary, Dongguan Luxshare Technology, were listed as respondents, involving US Patent No. 10,903,734. The products are mainly used for integrated power delivery solutions for data center power modules and processors. Luxshare Precision stated that the related products are currently still in the customer validation stage and have not yet entered mass production. It is expected that this will not have a significant adverse impact on the current production, operations, and current period performance.
Looking solely at the financial statements, this does not seem to be an imminent major issue.
However, from the perspective of the evolution direction of the AI server industry, it is more like a bill that arrived in advance.
Over the past two years, the biggest task for Chinese supply chain enterprises has been to find a way to enter the AI server order system. Now, as more and more companies begin to delve deeper from chassis, connectors, and switches to complete machines, liquid cooling, power supplies, and rack-level solutions, a new question has surfaced: After securing AI server orders, how much more R&D, patent, and customer certification costs will need to be incurred?
The investigation initiated by Vicor provides an answer.
The Section 337 Investigation Strikes at the "Power Heartland" of AI Servers
This is not a lawsuit targeting only Foxconn Industrial Internet and Luxshare Precision.
The list published by the USITC shows that the respondents in the 337-TA-1526 investigation also include Delta Electronics, Infineon, Monolithic Power Systems and its Chinese-funded subsidiaries, Flex, Celestica, Quanta Computer, Quanta Cloud Technology, and Hon Hai Precision, among other enterprises.
The list spans power semiconductors, power supplies, server ODMs, and complete machine manufacturers, almost covering the entire industrial chain from power conversion to AI computing system integration.
Vicor filed the complaint on September 9 and supplemented the materials the next day. The USITC decided to institute the investigation on October 9. Vicor requested the Commission to issue a limited exclusion order and a cease and desist order.
It needs to be emphasized that the institution of the investigation does not mean that the USITC has already determined that Foxconn Industrial Internet or Luxshare Precision has infringed any patents.
The USITC itself also explicitly stated that the initiation of a Section 337 investigation does not represent that the Commission has made any decision on the substantive issues of the case. The case will subsequently go through procedures such as hearing by an Administrative Law Judge, initial determination, and Commission review. The Commission will determine the target completion date within 45 days after the institution of the investigation.
However, what truly makes global hardware enterprises wary of Section 337 investigations is not primarily the amount of damages.
According to USITC rules, if a violation of Section 337 is ultimately determined, the Commission can issue an exclusion order, directing U.S. Customs to block the relevant infringing products from entering the United States; it can also issue a cease and desist order, restricting the respondent enterprises 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, this is already troublesome enough.
For enterprises competing for orders for the next-generation AI server platforms, it is even more sensitive.
Because once a power architecture is integrated into the design of a GPU, ASIC, or server platform, the issue is no longer just about replacing a supplier. PCB layout, power conversion, thermal management, structural design, electromagnetic compatibility, and processor power delivery interfaces may all need to be re-validated.
In other words, in the AI hardware era, patents have begun to be bound together with customer certification.
What is even more noteworthy is that this is not even the first time Vicor has done this this year.
In February this year, the USITC already initiated the 337-TA-1484 investigation based on another complaint filed by Vicor, targeting data center power converters, circuit board assemblies, and computing systems containing these components. At that time, Luxshare Precision, Delta, MPS, Quanta, and others were already on the respondent list. The case is currently still in the Administrative Law Judge hearing stage. USITC data shows that the initial determination date is planned for May 20, 2027, and the target completion date is September 20, 2027.
From "Power Converters" in February to directly targeting "Vertical Power Delivery Systems" in October, Vicor's intellectual property offensive is advancing towards the power delivery stage closer to the processor.
This is precisely the increasingly valuable position in AI servers.
Why "A Single Wire" Becomes a Technical Barrier
To understand this patent war, it is necessary to first understand a change occurring in AI servers: GPUs are getting faster and faster, but powering GPUs is becoming increasingly difficult.
Official NVIDIA data shows that a GB200 NVL72 rack contains 72 Blackwell GPUs and 36 Grace CPUs, with a full-load power consumption of approximately 120 kilowatts for the entire rack. Its power shelf provides approximately 50V to 51V DC power to the rack equipment through busbars.
This is no longer power consumption in the sense of traditional servers.
More importantly, the power is continuing to rise.
In the next-generation data center power delivery roadmap published in 2025, NVIDIA proposed promoting an 800V high-voltage DC architecture starting from 2027 to support AI racks with 1 megawatt or even higher power. NVIDIA stated that the current 54V in-rack power delivery system begins to encounter physical constraints when the 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 theoretically means a current of approximately 1000 amperes.
The greater the current, the harder it is to control the losses in busbars, PCBs, and packaging; the I²R losses on conductors increase with the square of the current.
Thus, power engineers face a contradiction: to reduce transmission losses, they want to deliver power as close to the processor as possible using the highest possible voltage; but what the processor itself requires is extremely low voltage and extremely high current.
This is precisely the background against which vertical power delivery technology emerged.
Under the MGX rack specification system published by the Open Compute Project, the rated capacity of the 48V DC busbar hardware provided by mainstream manufacturers can already reach 1400A.
Vicor's own technical route is to deliver 48V as close to the processor as possible, and then convert it at the point of load.
According to Vicor's official technical data, its Factorized Power Architecture splits DC-DC conversion into two stages: regulation and transformation. Through a current multiplier, it converts the 48V bus voltage to an approximately 1V AI processor power rail, and emphasizes placing the power modules as close as possible, or even directly underneath the processor, to shorten the high-current transmission distance.
Therefore, what "vertical power delivery" is truly competing for is not an ordinary power board.
It is competing for the last few centimeters of power delivery.
Whoever can efficiently convert tens of volts into low voltage and high current at the position closest to the XPU can potentially help customers reduce copper losses, lower heat generation, free up PCB area, and improve the power density of the entire AI system.
This explains why the US Patent No. 10,903,734 asserted by Vicor this time is named "Delivering power to semiconductor loads," and also explains why the National Overseas Intellectual Property Dispute Response Guidance System of the Ministry of Commerce explicitly classifies it as a vertical power delivery technology applied in high-performance computing scenarios such as AI computing.
The power competition in the AI era has evolved from "who can build more efficient power modules" to further questions: Where should the power supply be placed? How high a voltage should be delivered? Where should the step-down be completed? How should the capacitors be arranged? How should the modules be connected to the chips? How to balance transient response, thermal management, and high current density?
The answers to these questions, once solidified into circuits, packaging, and system structures, can easily fall into the scope of patent protection.
As a result, a power industry that appeared to have matured for decades has once again become an intellectual property-intensive industry.
Securing AI Orders Is Just the Beginning of Costs
For Foxconn Industrial Internet and Luxshare Precision, what is truly worthy of attention in this Section 337 investigation may not be the immediate financial impact.
But rather the timing of its occurrence.
In the first half of 2026, Foxconn Industrial Internet achieved an operating revenue of CNY 557.861 billion, a year-on-year increase of 54.63%; net profit attributable to the parent company was CNY 23.74 billion, a year-on-year increase of 95.99%. Among them, cloud computing business revenue increased by 75.7% year-on-year, cloud service provider AI server revenue increased by 2.3 times year-on-year, and GPU AI rack shipments increased by 3.2 times year-on-year.
In other words, AI servers are no longer a marginal new business in Foxconn Industrial Internet's financial statements, but have become one of the most important growth engines.
Luxshare Precision is also moving rapidly in the same direction.
Semi-annual report data shows that in the first half of 2026, Luxshare Precision achieved a revenue of CNY 174.504 billion, a year-on-year increase of 40.16%; communications and data center business revenue was CNY 16.609 billion, a year-on-year increase of 49.66%, and 800G optical modules have achieved scaled mass production.
This kind of expansion is not cheap.
According to statistics from Luxshare Precision's semi-annual report, the company's R&D investment in the first half of this year reached CNY 6.573 billion, a year-on-year increase of 43.09%; as of the end of 2025, the company's cumulative R&D investment over the past three years was CNY 28.173 billion, possessing 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 approximately CNY 4.998 billion. During the same period, the company's R&D personnel reached 37,700. Its semi-annual report shows that the company is further shifting from single product manufacturing to full-stack solutions and joint R&D with core customers.
These figures reveal an easily overlooked reality in the AI hardware supply chain.
In the traditional contract manufacturing era, the most important moat for manufacturing enterprises might have been scale, yield rate, and supply chain efficiency.
In the AI server era, relying solely on these is no longer enough.
Enterprises first need to conduct extensive upfront R&D; subsequently, they need to pass customer validation; after entering the platform, they must also handle patent freedom-to-operate analysis, supplier intellectual property guarantees, licensing negotiations, and potential litigation; if patent disputes are encountered, they may also need to revise circuit and structural designs and re-complete partial certifications.
Luxshare Precision even stated in an investor relations event in April this year that more than half of the company's R&D investment is for upfront technical reserves for businesses that have not yet been implemented.
This is perhaps the most important sentence for understanding this Section 337 investigation.
The capital market usually only sees the revenue after orders are landed.
Engineers, on the other hand, see the R&D investment of several years prior to the orders.
Intellectual property lawyers see the patent risks that only begin to emerge after the products truly enter the US market.
And customers see a supply chain that must ultimately satisfy performance, cost, reliability, delivery capability, and intellectual property security simultaneously.
Foxconn Industrial Internet also specifically emphasized in this announcement that the company has a dedicated intellectual property department and requires non-infringement guarantees as well as relevant technical licenses and authorizations when screening suppliers.
This means that intellectual property is no longer just an issue for the legal department.
It is entering the procurement and supply chain management system.
The Patent War Has Just Begun
Currently, there is no evidence to suggest that this Section 337 investigation will change the business trends of Foxconn Industrial Internet or Luxshare Precision in 2026.
Both companies have explicitly stated that the products involved have not yet entered the formal mass production stage.
But precisely because of this, this dispute has a certain forward-looking significance.
When a product is still in the internal evaluation and customer validation stage, the patent war has already arrived in advance.
This is very likely to be the norm in the next stage of AI hardware competition.
As GPU power consumption rises and racks advance from the 100-kilowatt level to hundreds of kilowatts or even megawatt levels, the value of servers is spreading to previously inconspicuous links: power supplies, busbars, connectors, liquid cooling, PCBs, and packaging are all beginning to determine how much computing power the entire computing system can hold.
In the past, what Chinese supply chain enterprises were good at answering was a manufacturing question: How to make things faster, cheaper, and more reliable?
After entering the AI era, they also need to answer another question: How to prove that this set of technologies is both their own and does not step into others' patent territories?
This may be even harder than securing the first AI server order.
It is still too early to draw a conclusion on how the Section 337 investigation between Vicor, Foxconn Industrial Internet, and Luxshare Precision will ultimately be adjudicated.
But one thing is becoming increasingly clear: on AI servers, power is no longer just a cost.
It is becoming performance.
And once a technology can determine performance, it usually quickly becomes a patent.