The two complaints
On November 13, 2026, Inference Stack Inc. filed Inference Stack Inc. v. Advanced Micro Devices, Inc., No. 7:26-cv-00453, and Inference Stack Inc. v. NVIDIA Corporation, No. 7:26-cv-00454, both in the Midland-Odessa Division of the Western District of Texas. Each is a complaint for patent infringement, ECF No. 1, on the same two patents, and each demands a jury trial under Rule 38 of the Federal Rules of Civil Procedure. Both complaints say Inference Stack Inc. is a corporation organized under the laws of the Republic of Korea with its principal place of business in Seoul, Republic of Korea. The AMD complaint alleges, on information and belief, that Advanced Micro Devices, Inc. is a Delaware corporation with its principal place of business in Santa Clara, California. The NVIDIA complaint alleges, on information and belief, that NVIDIA Corporation is a Delaware corporation with its principal place of business in Santa Clara, California.
The exclusive license
Both complaints say the patents arose from research at a government-funded research institute in Daejeon, Republic of Korea, which they identify as the assignee of record of both patents. They allege that the institute and Inference Stack executed a Patent Rights Exclusive License Agreement on January 22, 2026, granting Inference Stack an exclusive license to both patents, and that the agreement was recorded with the United States Patent and Trademark Office on March 23, 2026, at Reel/Frame 075226/0256. They allege that a Supplementary Agreement dated as of November 9, 2026 gives Inference Stack "the sole right" to file suit and to recover damages for past infringement, including infringement before January 22, 2026, and that the license agreement does not reserve to the institute any right to practice the patents, to grant further licenses under them, or to sue for their infringement. On that basis each complaint alleges that Inference Stack is the exclusive licensee of the patents-in-suit and holds all substantial rights in them, including the right to enforce them, to exclude others, and to recover damages for all past infringement.
The patents and claim 1
Both complaints identify U.S. Patent 11,003,985, titled "Convolutional Neural Network System and Operation Method Thereof," issued May 11, 2021, and U.S. Patent 11,068,394, titled "Neural Network System Including Data Moving Controller," issued July 20, 2021. The NVIDIA complaint says the patents generally relate to hardware acceleration of neural-network computation. The AMD complaint describes the '985 patent as directed to reducing the computation a convolutional neural network performs, and the NVIDIA complaint describes it as directed to reducing the convolution computations of a convolutional neural network on the basis of sparse weights. Both complaints describe the '394 patent as directed to moving multidimensional data between an accelerator's external and internal memories.
Claim 1 of the '985 patent covers a convolutional neural network system comprising a data selector that outputs an input value corresponding to a position of a sparse weight from among input values of input data, on the basis of a sparse index indicating the position of a nonzero value in a sparse weight kernel, and a multiply-accumulate computator that performs a convolution computation on that output input value using the sparse weight kernel. The sparse weight kernel comprises at least one weight value of 0. The data selector comprises a switch circuit and a plurality of multiplexes (MUXes). The switch circuit provides the respective input values to the MUXes on the basis of the sparse weight kernel, and each MUX selects and outputs, on the basis of the sparse index, the input value corresponding to the position of the sparse weight from among the input values the switch circuit provides.
Claim 1 of the '394 patent covers a neural network system for processing data transferred from an external memory, comprising an internal memory that stores input data transferred from the external memory, an operator that performs a multidimensional matrix operation using that input data and transfers a result to the internal memory as output data, and a data moving controller that controls an exchange of the input data or the output data between the external memory and the internal memory. For the multidimensional matrix operation, the data moving controller reorders a dimension order with respect to an access address of the external memory to generate an access address of the internal memory, and it is configured to generate physical addresses for reading input data of a first multidimensional array from the external memory, storing that data in the internal memory in the form of a second multidimensional array, reading output data of the second multidimensional array from the internal memory, and storing that output data in the external memory in the form of the first multidimensional array.
The accused AMD products
As to the '985 patent, the AMD complaint accuses AMD Instinct GPUs whose matrix hardware performs 4:2 structured-sparsity acceleration under a sparse matrix multiply-accumulate instruction that carries a two-bit sparse index, including the V_SMFMAC instruction family of AMD's CDNA 3 and CDNA 4 architectures and the V_SWMMAC instruction family of its CDNA 5 architecture. It names the Instinct MI300X, MI300A, and MI325X (CDNA 3), the MI350X and MI355X (CDNA 4), and the MI455X (CDNA 5) and any other MI400 Series or MI450 Series GPU that supports the V_SWMMAC instruction family, together with the boards, modules, platforms, rack-scale systems, and servers incorporating them, including the AMD Helios rack-scale solution, and the AMD software through which the accused sparse execution path is prepared and invoked, including the ROCm software platform, hipSPARSELt, and MIOpen. As to the '394 patent, it accuses the Instinct MI455X and any other MI400 Series or MI450 Series GPU built on the CDNA 5 architecture with a Tensor Data Mover, the systems incorporating them, including Helios, and the AMD software that programs and invokes the accused data transfers and matrix operations, including ROCm, the HIP programming interface and its compiler, the Triton compiler and Gluon language support that AMD provides for CDNA 5, and AMD's aiter and MIOpen libraries. The complaint excludes any product, or any use of a product, to the extent it is licensed under, or covered by a covenant not to sue on, the relevant patent.
For the '985 patent, the complaint alleges, citing AMD's instruction set guides, that AMD's matrix hardware uses a two-bit index to select, from the dense matrix, the input value at the position of each non-zero value of the sparse matrix. It says AMD has not published the circuit design of its matrix hardware and alleges, on information and belief, that the hardware includes a switch circuit that delivers groups of four candidate input values to a plurality of multiplexers, each of which selects, on the basis of the two-bit index, the input value at the position of a non-zero weight. For the '394 patent, it points to the Tensor Data Mover that AMD's CDNA 5 architecture adds to move tiles of tensor data of up to five dimensions between memory and the Local Data Share, a scratch-pad memory in each workgroup processor, and alleges that convolution kernels AMD added to the development branch of its MIOpen library program the Tensor Data Mover to place a tile in the Local Data Share with its batch and height dimensions exchanged. It alleges, on information and belief, that each Tensor Data Mover is configured to store a tile it has loaded with reordered dimensions back to memory in the tile's original dimension order.
The accused NVIDIA products
As to the '985 patent, the NVIDIA complaint accuses NVIDIA graphics processing units that incorporate the NVIDIA Sparse Tensor Core and its fine-grained structured-sparsity acceleration, including its 2:4 structured-sparsity pattern and the 4:8 structured-sparsity pattern used with NVIDIA's NVFP4 format. It names the A100, H100, H200, GH200, and B200, together with the boards, modules, systems, and platforms incorporating them, and the NVIDIA software through which that sparse execution is prepared and invoked, including cuSPARSELt and NVIDIA TensorRT. As to the '394 patent, it accuses NVIDIA's Hopper and Blackwell data center GPUs and the products and systems incorporating them, including the H100 and H200 GPUs and the GH200 Grace Hopper Superchip, and the B200 GPU, the GB200 Grace Blackwell Superchip, and the GB200 NVL72 and GB300 NVL72 rack-scale systems, alone and as configured and operated with the NVIDIA software that programs and invokes their Tensor Memory Accelerator transfers and Tensor Core matrix operations, including the CUDA Toolkit, cuBLAS, CUTLASS, and TensorRT-LLM. The complaint excludes any product, or any use of a product, to the extent it is licensed under, or covered by a covenant not to sue on, the relevant patent.
For the '985 patent, the complaint alleges, citing NVIDIA's documentation, that the Sparse Tensor Core uses stored metadata, two 2-bit indices for each four-wide chunk, to select the input values that match the non-zero weights. It says NVIDIA has not published the circuit design of the Sparse Tensor Cores of its Hopper and Blackwell GPUs and alleges, on information and belief, that those Sparse Tensor Cores use the stored index as the A100's does, with circuitry delivering each group of input values to its own multiplexer and the index controlling which input values each multiplexer passes to the multiply-accumulate computation, and that the B200 applies the index of its 4:8 pattern in the same way. For the '394 patent, it points to the Tensor Memory Accelerator that NVIDIA's Hopper and later GPUs use to transfer multidimensional tensors between global memory and shared memory, and alleges that, under its swizzle modes, the Tensor Memory Accelerator stores a tile in shared memory in an array order different from its order in global memory and unswizzles the data when copying it back to global memory.
The counts
Each complaint pleads two counts, one for each patent. The AMD complaint alleges that AMD has directly infringed and continues to directly infringe each patent under 35 U.S.C. § 271(a), literally and equivalently, naming the Instinct MI300X, MI300A, MI325X, MI350X, MI355X, and MI455X GPUs for the '985 patent. For the '394 patent it alleges, on information and belief, that since launching the MI455X on July 23, 2026, AMD has offered it for sale in the United States, imported it, and sold and delivered it, including in Helios racks, to customers in the United States. It alleges induced and contributory infringement under 35 U.S.C. §§ 271(b) and (c), with knowledge of each patent since at least the filing and service of the complaint, pointing to AMD's documentation, software libraries, customer-engineering support, and system partners. It identifies as a material part of each claimed invention what it calls the Sparse Selection Stage for the '985 patent and the TDM Component for the '394 patent, and alleges, on information and belief, including a reasonable investigation, that each has no substantial use that does not infringe.
The NVIDIA complaint alleges that NVIDIA has directly infringed and continues to directly infringe each patent under 35 U.S.C. § 271(a), literally and equivalently, naming the A100, H100, H200, GH200, and B200 GPUs for the '985 patent and the H100, H200, GH200, B200, GB200 Grace Blackwell Superchip, GB200 NVL72, and GB300 NVL72 for the '394 patent. It alleges induced and contributory infringement under 35 U.S.C. §§ 271(b) and (c), with knowledge of each patent since at least the date of service of the complaint, pointing to NVIDIA's documentation, software tools, advertising, technical support, and partner network. It identifies as a material part of each claimed invention what it calls the Sparse Selection Circuitry for the '985 patent and the Swizzle Datapath for the '394 patent, and alleges, on information and belief, formed after a reasonable investigation, that each has no substantial non-infringing use.
In both complaints, each count says Exhibit 2 (for the '985 patent) or Exhibit 4 (for the '394 patent) provides a description of the accused instrumentalities and a chart showing an example of how they infringe at least one claim, which Inference Stack provides without the benefit of discovery. Each count alleges that the defendant's infringement after it receives notice of the complaint is willful, entitling Inference Stack to enhanced damages under 35 U.S.C. § 284, seeks damages no less than a reasonable royalty and an injunction, and alleges that Inference Stack's damages claims are not limited under 35 U.S.C. § 287, at least because there are no unmarked patented articles subject to a duty to mark.
Jurisdiction and venue
Both complaints plead subject matter jurisdiction under 28 U.S.C. §§ 1331 and 1338(a) and venue under 28 U.S.C. § 1400(b). The AMD complaint pleads personal jurisdiction because AMD is registered to do business in Texas, maintains a registered agent there, has admitted that it has places of business in Austin, Texas, is staffing and, on information and belief, employs personnel in Austin who work on the accused Instinct GPUs, and has committed acts of infringement in the district. For venue it alleges acts of infringement and a regular and established place of business in the district, citing AMD's answer in another patent action in the district, in which it says AMD admitted that it has places of business in Austin, and AMD's statement in its annual report that it has "significant operations in Austin, Texas." It alleges, on information and belief, that AMD uses the accused products in the district and offers them for sale from there. It also alleges that platforms incorporating the MI355X are made in the district, citing a manufacturing company's announcement that it assembles them in Austin, and alleges, on information and belief, that AMD supplies the GPUs for those platforms.
The NVIDIA complaint pleads personal jurisdiction based on alleged acts of infringement in the district and minimum contacts with the forum. For venue it alleges that NVIDIA maintains a regular and established place of business in Austin, Texas, citing NVIDIA's admissions in other Western District of Texas cases that it maintains a physical facility and a regional office there, that it has hundreds of employees in the district, and that it "makes, uses, sells, offers to sell, and/or imports systems and components throughout the United States, including in this District." It also alleges that the accused products are deployed and in use in the district, including at an Austin research computing center.
The revenue figures the complaints plead
Each complaint pleads the defendant's own reported revenue in a section on the commercial significance of the accused instrumentalities. The AMD complaint says AMD reported Data Center segment revenue of $6.7 billion for the second quarter of 2026, up 107% year over year, on demand it attributed to its EPYC processors and Instinct GPUs, and that the Data Center business represented 58% of AMD's total revenue for the quarter. It says AMD's quarterly report attributes the growth in Data Center revenue for the first half of 2026 "primarily" to "strong demand for our AMD EPYC processors and AMD Instinct MI350 Series GPUs." The NVIDIA complaint says NVIDIA reported record revenue of $215.9 billion for fiscal year 2026, up 65%, and record Data Center revenue of $193.7 billion, up 68%. It says NVIDIA's annual report on Form 10-K shows Data Center revenue growing from $47.5 billion in fiscal 2024 to $115.2 billion in fiscal 2025 and $193.7 billion in fiscal 2026, of which $162.4 billion was Compute revenue. It says NVIDIA reported revenue of $96.2 billion for the second quarter of fiscal 2027, up 106% from a year ago, and Data Center revenue of $89.0 billion, up 117% from a year ago, and that NVIDIA's quarterly report attributes that Data Center growth to "the ramp of our Blackwell Ultra infrastructure."
The relief sought
The two prayers are the same apart from the defendant's name. Each asks for a judgment that the defendant has infringed, directly and indirectly, both patents, and that each patent is valid and enforceable; past and future damages in an amount no less than a reasonable royalty, together with costs, expenses, and pre- and post-judgment interest under 35 U.S.C. § 284; an accounting and supplemental damages; compulsory ongoing licensing fees for continued infringement after judgment; a judgment that infringement after notice of the complaint is willful and an award of enhanced damages under § 284; a permanent injunction under 35 U.S.C. § 283; a finding that the case is exceptional and an award of reasonable attorneys' fees and costs under 35 U.S.C. § 285; and any other relief to which Inference Stack may be entitled.
What to watch next
Watch for each defendant's answer or Rule 12 motion, a scheduling order, and any order relating or consolidating the two cases. Either case can end at any time, including by settlement or voluntary dismissal.