Key Points

• A Texas-based technology company filed patent infringement actions in May 2026 in the Eastern and Western Districts of Texas, the District of Delaware, and the International Trade Commission, asserting patents directed at memory technologies against companies including Nvidia, SK Hynix, Kioxia, Corsair Gaming, and Western Digital, according to Law360. 
• With global semiconductor sales reaching a record $795.6 billion in 2025, according to the Semiconductor Industry Association, the concentration of value in a small number of advanced AI devices is the kind of condition that tends to drive patent assertion. 
• If routing UCIe signals over organic package substrates becomes a common way to connect compute dies with HBM stacks, patents covering that approach could function similarly to standard-essential patents, with FRAND licensing implications for anyone building UCIe-compliant AI accelerators. 
• As of October 2026, the typically short window before large operating companies build defensive moats through cross-licensing and inter partes review campaigns appears open to universities, research institutions, independent inventors, and monetization funds holding patents in HBM and co-packaged optics, and it is opening in chiplet interconnects. 
• Because companies that treat intellectual property strategy as a compliance function tend to manage crises reactively, those that conduct proactive freedom-to-operate analyses and prepare litigation-readiness plans are better positioned to protect their AI infrastructure investments.

---

Introduction: Why This Matters Now

The global semiconductor industry reached record sales of $795.6 billion in 2025, according to the Semiconductor Industry Association’s “2026 State of the U.S. Semiconductor Industry” report, with AI chips now driving roughly half of total revenue while representing less than 0.2% of total unit volume. That concentration of value in a small number of advanced devices is the kind of condition that tends to drive patent assertion. When billions of dollars of revenue flow through technology that only a handful of companies produce, and every producer must solve the same underlying engineering problems, patent rights over those solutions become valuable enough to fight over.

Three structural forces are now converging to make AI-specific semiconductor hardware the next major front in patent litigation: (1) the sheer scale of capital being invested in AI infrastructure; (2) a set of well-defined physical and engineering bottlenecks that every competitor must overcome; and (3) a patent landscape that is reaching maturity across all of the relevant technology areas simultaneously.

The Memory Wall: HBM Litigation Is Already Here

The first major front is already open. The fundamental constraint on AI accelerator performance is not compute; it is memory bandwidth. Graphics processing unit (GPU) compute density has scaled far faster than memory bandwidth, leaving even the most powerful AI accelerators starved of data. The industry’s solution, high-bandwidth memory (HBM), stacks dynamic random-access memory (DRAM) dies vertically using through-silicon vias (TSVs) directly atop or beside the GPU die, delivering bandwidth measured in terabytes per second. HBM is no longer optional; it is now standard architecture in every leading AI accelerator in production today.

The economics of this market make litigation likely. The HBM market was valued at $2.93 billion in 2024 and is projected to reach $16.72 billion by 2033, growing at a 21.35% compound annual growth rate driven almost entirely by generative AI demand, according to Patsnap. HBM demand grew 130% year-over-year in 2025, and global HBM spend is projected to grow 58% year over year in 2026 to reach $54.6 billion, approaching $100 billion by 2028, a figure that would surpass the size of the entire 2024 DRAM market, according to Global X ETFs.

The supply side is equally concentrated. SK Hynix holds approximately 50% to 62% of the HBM market (depending on report date), Micron approximately 18% to 21%, and Samsung approximately 17% to 33%. That concentrated supply structure means a single valid patent on TSV stacking, base die design, or HBM controller architecture could reach nearly every AI accelerator shipped worldwide.

The litigation has already begun. DRAM patent disputes rose 119% in U.S. district courts in 2025, with NAND disputes up 56% year over year. In May 2026, a Texas-based technology company filed patent infringement actions in the the Eastern and Western Districts of Texas, the District of Delaware, and the International Trade Commission (ITC) targeting Nvidia, SK Hynix, Kioxia, Corsair, and Western Digital, asserting patents directed at memory technologies, as reported by Law360.

The ITC presents distinct strategic advantages for semiconductor patent disputes that explain why sophisticated patent holders increasingly file parallel actions there. Unlike district court litigation, ITC investigations proceed on a statutory 15-to-18-month timeline, delivering a final determination before most district court cases reach claim construction. The remedy — an exclusion order barring infringing products from entering the U.S. — can be devastating for companies dependent on global supply chains, particularly when the targeted component is manufactured exclusively overseas, as HBM and advanced packaging substrates largely are. For patent holders, even the credible threat of an exclusion order can accelerate settlement negotiations that would otherwise stall in district court. For respondents, ITC proceedings demand immediate resource allocation; the accelerated discovery schedule and limited opportunities for stays leave little room for delay. Companies with significant import exposure should assess their ITC risk profile now, particularly for products incorporating HBM, advanced packaging, or CPO components sourced from Asian fabs. For more on ITC practice, see Troutman Pepper Locke’s The Top Five Challenges of Pursuing Litigation at the ITC.

Chiplet Interconnects: The Standard-Essential Patent Risk

The second major front involves the architecture that allows modern AI accelerators to be built at all. No single silicon die can integrate all the logic, memory, and I/O functions needed for a frontier AI processor; reticle size limits and fabrication yield economics make monolithic design increasingly unviable. The industry has converged on a “chiplet” model: distinct compute, memory, and I/O dies fabricated separately and joined at the package level, through silicon interposer wafers, which are then mounted onto a larger organic substrate. The critical question, and the patent battleground, is how those dies communicate.

Three interconnect families are racing to become the industry standard: UCIe (Universal Chiplet Interconnect Express), CXL (Compute Express Link), and NVLink. Patent filings in this space are heavy and increasingly contentious. For example, multiple assignees, including Intel, Shanghai QuXin, and others, filed patents in 2025 and 2026 proposing to route UCIe signals directly over organic package substrates to connect compute dies with HBM stacks, bypassing expensive silicon interposers entirely, according to Patsnap’s UCIe patent landscape. If organic-substrate HBM routing becomes a common implementation path for UCIe-compliant chiplet designs, patents covering it could function similarly to standard-essential patents, with licensing/FRAND implications for anyone building UCIe-compliant AI accelerators.

Companies participating in standard-setting bodies carry patent disclosure obligations, and failing to manage them carefully can create real legal exposure. Companies that have not yet audited their portfolio against pending standard specifications should do so promptly. For companies that are implementers rather than contributors, the FRAND licensing framework provides a degree of protection, but only if the implementer understands which patents in a licensor’s portfolio are genuinely essential and which are merely declared. Engaging patent counsel to conduct SEP essentiality analysis before entering licensing negotiations is a necessary step that is routinely deferred until after the first demand letter arrives. On the other hand, companies looking to monetize or assert their patent assets should map their interconnect portfolios against industry standards and specifications such as UCIe.

Co-Packaged Optics and Silicon Photonics: The Next Memory Wall

The third, and arguably most consequential, front over the next five years involves the replacement of copper electrical interconnects with optical ones at the chip-package level. Traditional copper traces between a switch application-specific integrated circuit (ASIC) and its optical transceivers are reaching fundamental physical limits. At 800G and 1.6T data rates, signal loss in the electrical traces becomes severe enough to limit system performance and efficiency. Co-packaged optics (CPO) addresses this by integrating the optical engine directly onto the same package as the switch ASIC, eliminating the electrical bottleneck entirely. CPO is a key enabler for AI and data center scale-up as noted at Nvidia’s GTC 2025 conference.

The capital flowing into this space is substantial. The global silicon photonics and CPO market was valued at $2.16 billion in 2025 and is projected to reach $22.4 billion by 2034, at a compound annual growth rate of 29.5%, according to this report. IDTechEx projects the CPO market alone will exceed $20 billion by 2036, growing at a 37% CAGR from 2026 to 2036. In March 2026, Nvidia announced a $2 billion investment in Lumentum, including equity and multiyear laser-component purchase commitments, to scale silicon photonics for AI interconnects.

Two specific technical problems within CPO warrant particular attention from both operating companies and IP holders.

The Known-Good Die Problem

In a CPO system, the photonic integrated circuit (PIC) is permanently bonded to the switch ASIC at the package level. Unlike a pluggable transceiver that can be field-replaced, a defective PIC in a CPO assembly means the entire switch, potentially worth $100,000 or more, must be scrapped. Scaling CPO from current production volumes to the tens of millions of units implied by 2026 demand projections requires optical design-for-test (DfT) architectures, wafer-level probing methodologies, and automated calibration protocols that do not yet exist at commercial scale, according to TrendForce and Semiconductor Engineering.

The Laser Supply Chain and Hybrid Bonding

Even though silicon is excellent for routing, modulating, and detecting light, it cannot natively produce the light efficiently due to its indirect bandgap. Therefore, every silicon photonic system requires an external or integrated laser source. Integration of efficient laser sources (typically III-V materials like indium phosphide) onto silicon photonic integrated circuits enables high density bandwidth scale with low-loss at high volume manufacturing required for AI data center applications. Laser materials cannot be grown directly on silicon due to crystal lattice mismatches. Hybrid bonding for laser materials onto silicon photonic substrates is a technique that physically joins (fused or pressed) the laser light emitting material to silicon chips containing optical pathways. Patents covering hybrid III-V/silicon bonding techniques are being pursued by multiple independent filers simultaneously, according to Patsnap’s heterogeneous integrated silicon photonics landscape. Whichever party secures a dominant position in hybrid bonding could hold significant leverage over the CPO supply chain.

Compute-in-Memory: The Next Wave After HBM

Even HBM is insufficient for the most demanding AI workloads. Moving data between any separate memory and compute die consumes an estimated 62-80% of total system energy for data-intensive AI workloads. The architectural solution, performing computation directly inside the memory array and eliminating data movement, encompasses SRAM-based compute-in-memory (CIM), DRAM-based processing-in-memory (PIM), ReRAM, FeFET, and in-memory computing. The 2020-2025 era saw a clear inflection point in patent filings across all of these approaches.

Advanced Cooling: A Secondary but Rapidly Monetizable Front

Thermal limits in AI hardware are as structurally significant as the memory wall. The global data center liquid cooling market will grow from $4.68 billion in 2025 to an estimated $25.80 billion by 2035, at an 18.61% CAGR. The Liquid Cooling for AI Act of 2025 (H.R. 5332, 119th Congress) explicitly acknowledges that “traditional air-cooled systems are reaching limits to effectively remove heat from AI chips.”

Data center cooling can account for up to 40% of total facility energy consumption, and companies are filing patents in AI-driven cooling control architectures. Companies deploying direct-to-chip or immersion cooling in AI data center builds should assess their cooling architecture designs against this landscape, particularly in the AI-controlled coolant distribution unit and two-phase immersion fluid formulation spaces.

Patents in direct-liquid-cooling manifold design, two-phase flow boiling in confined microchannels, and AI-driven predictive cooling control filed before 2020 may cover techniques now being adopted at scale in AI data centers. Portfolio holders in this space should evaluate whether their claims are broad enough to cover the specific geometries being used in server-grade direct-to-chip cooling manifolds, and whether reissue or continuation prosecution is warranted to bring claim scope current with commercial implementations.

Looking Ahead

The semiconductor industry appears to be entering a patent litigation cycle that resembles the smartphone patent wars of the 2010s: heavy capital concentration, shared architectures that every competitor must adopt, and a patent landscape that is only now reaching the assertion stage. Unlike the smartphone wars, however, this cycle touches infrastructure the global economy depends on, and the stakes, measured in hundreds of billions of dollars of annual capital expenditure, are larger.

For operating companies, prior technology patent cycles offer a clear lesson: companies that treat IP strategy as a compliance function tend to manage crises reactively. Those that build portfolios, conduct proactive freedom-to-operate (FTO) analysis, and prepare litigation-readiness plans are better positioned to protect the capital they are investing in this space.

For patent innovators, including universities, research institutions, independent inventors, and monetization funds, the lesson is similar: the window between when a technology reaches commercial scale and when large operating companies build defensive moats through cross-licensing and inter partes review campaigns tends to be short. That window appears open now in HBM and CPO, and is opening in chiplet interconnects. These windows will not stay open indefinitely.

How Troutman Pepper Locke Can Help

Troutman Pepper Locke’s Intellectual Property practice group has deep experience in semiconductor patent litigation, including ITC Section 337 investigations, district court patent trials, PTAB proceedings, and strategic IP counseling for companies across the AI hardware value chain. Our team advises chipmakers, memory manufacturers, hyperscalers, OEMs, and technology investors on freedom-to-operate analysis, patent portfolio development, SEP licensing strategy, and litigation preparedness. Whether you are an operating company seeking to derisk your AI infrastructure investments or a patent holder evaluating enforcement options, we can help you navigate the complexities of this rapidly evolving landscape. To discuss how these developments may affect your IP strategy, contact Gregory D. Len.