Tesseract Systems LLC v. NVIDIA Corp

I. Executive Summary and Procedural Information

• Parties & Counsel:
  • Plaintiff: Tesseract Systems LLC (New Mexico)
  • Defendant: Nvidia Corporation (Delaware)
  • Plaintiff’s Counsel: Rabicoff Law LLC
• Case Identification: 7:25-cv-454, W.D. Tex., 10/07/2025
• Venue Allegations: Plaintiff alleges venue is proper in the Western District of Texas because Defendant maintains an established place of business in the district and has allegedly committed acts of patent infringement there.
• Core Dispute: Plaintiff alleges that Defendant’s unspecified products infringe a patent related to a specific neural network architecture designed to facilitate easier training of deep networks.
• Technical Context: The patent addresses methods for optimizing deep neural networks, a foundational technology for modern artificial intelligence and machine learning applications, particularly in fields like image recognition and data processing where Defendant is a market leader.
• Key Procedural History: The complaint does not mention any prior litigation, Inter Partes Review (IPR) proceedings, or licensing history related to the patent-in-suit.

Case Timeline

Date	Event
2016-05-02	’320 Patent Priority Date
2021-04-20	’320 Patent Issue Date
2025-10-07	Complaint Filing Date

II. Technology and Patent(s)-in-Suit Analysis

U.S. Patent No. 10,984,320 - Highly trainable neural network configuration

The Invention Explained

• Problem Addressed: The patent’s background section describes the difficulty of training "deep" neural networks (those with many successive computation layers). While deeper networks are theoretically more powerful, their optimization is "considerably more difficult" than that of shallower networks, a problem that can hinder performance. (’320 Patent, col. 1:36-41).
• The Patented Solution: The invention proposes a novel neural network architecture, termed a "highway network," that uses a "learned gating mechanism for regulating information flow." (’320 Patent, col. 2:35-36). Each neuron in this architecture includes gates that can dynamically control how much of the input signal is transformed versus how much is passed through unchanged (carried). This allows information to flow across many layers "without attenuation," which in turn "may enable the optimization of networks with virtually arbitrary depth." (’320 Patent, col. 2:33-40; Fig. 2).
• Technical Importance: This gating architecture was designed to overcome the "vanishing gradient" problem that often stalls the training of very deep traditional networks, thereby making it practical to build and train much deeper and more powerful AI models using standard optimization methods. (’320 Patent, col. 12:30-34).

Key Claims at a Glance

• The complaint asserts "one or more claims" of the ’320 Patent, and references "Exemplary ’320 Patent Claims" in a non-provided exhibit (Compl. ¶11). The following analysis focuses on independent claim 1.
• Independent Claim 1:
  • A computer-based method of facilitating training in a computer-based neural network, the method comprising:
  • receiving an electrical input signal at a neuron in the computer-based neural network that comprises a plurality of neuron layers;
  • applying a first non-linear transform to the input signal at the neuron to produce a plain signal;
  • applying a second non-linear transform to the input signal at a first gate in the neuron to produce a transform signal;
  • applying a third non-linear transform to the input signal at a second gate in the neuron to produce a carry signal; and
  • calculating a weighted sum of a non-transformed first component of the input signal and the plain signal at the neuron...
• The complaint does not explicitly reserve the right to assert dependent claims.

III. The Accused Instrumentality

Product Identification

The complaint does not identify any specific accused products by name. It refers generally to "Defendant products identified in the charts incorporated into this Count below (among the 'Exemplary Defendant Products')" (Compl. ¶11). However, the referenced charts (Exhibit 2) were not filed with the complaint.

Functionality and Market Context

The complaint does not provide sufficient detail for analysis of the accused instrumentality's specific functionality or market positioning. It makes only a conclusory allegation that the "Exemplary Defendant Products practice the technology claimed by the ’320 Patent" (Compl. ¶16).

IV. Analysis of Infringement Allegations

The complaint incorporates by reference an "Exhibit 2" containing claim charts that allegedly compare the asserted claims to the accused products (Compl. ¶16-17). As this exhibit was not provided, a detailed claim chart summary cannot be constructed. The complaint alleges that Defendant’s products "satisfy all elements of the Exemplary ’320 Patent Claims" (Compl. ¶16). Without the specific product identifications and element-by-element mappings from the missing exhibit, the precise theory of infringement remains undefined.

No probative visual evidence provided in complaint.

Identified Points of Contention

• Specificity Question: The primary point of contention will be identifying which specific Nvidia products, software libraries (e.g., CUDA, cuDNN), or hardware architectures (e.g., specific GPU series) are accused and what specific technical operations within them allegedly correspond to the claim limitations.
• Technical Question: A key question will be whether any accused functionality performs the distinct steps of applying three separate non-linear transforms to generate a "plain signal," a "transform signal," and a "carry signal," and then combines them as required by the claim. The defense may question whether standard neural network operations in its products can be disaggregated into the specific gating structure claimed by the patent.
• Scope Question: A potential dispute may arise over whether standard components of modern deep learning frameworks, which may have evolved independently, perform the specific "gating mechanism" described and claimed in the patent, or if they operate on different principles that fall outside the claim scope.

V. Key Claim Terms for Construction

The Term: "non-transformed first component of the input signal"

Context and Importance

This term is critical because it requires that the final "weighted sum" include a part of the original input signal that has not been subjected to the "first non-linear transform." The infringement analysis will depend on whether an accused process truly "carries" a component of the original, unmodified input signal forward, as opposed to using only transformed or modified versions of it.

Intrinsic Evidence for Interpretation

• Evidence for a Broader Interpretation: The specification suggests this is a flexible concept, describing it as a "copy of a corresponding component of the input signal" that is passed to the output, which could be read to cover any data that retains the essential information of the original input. (’320 Patent, col. 6:62-64).
• Evidence for a Narrower Interpretation: The patent’s mathematical formulation, y = H(x)·T(x) + x·(1-T(x)), explicitly shows the original input x being multiplied by the carry gate (1-T(x)), suggesting the "non-transformed component" must be the mathematically unaltered input vector itself, not merely a representation of it. (’320 Patent, col. 11:37-39).

The Term: "plain signal"

Context and Importance

This term defines the output of the first non-linear transform (H). Differentiating this "plain signal" from the final output and the other intermediate signals ("transform signal," "carry signal") is essential to the claimed method. Infringement will require identifying a discrete, corresponding signal within an accused process.

Intrinsic Evidence for Interpretation

• Evidence for a Broader Interpretation: The term is not explicitly defined, which may support an argument for its plain and ordinary meaning as the main processed signal before it is gated or mixed with the carried-through input.
• Evidence for a Narrower Interpretation: The detailed description consistently presents the "plain signal" as the output of the transform H(x, WH), which is the equivalent of a standard feedforward network layer. (’320 Patent, col. 11:15-22). This could limit the term to the output of a traditional non-linear activation function, as opposed to other types of processed signals.

VI. Other Allegations

• Indirect Infringement: The complaint alleges inducement of infringement based on Defendant's distribution of "product literature and website materials" that allegedly instruct end users on how to use the accused products in an infringing manner. Knowledge is alleged to have begun upon service of the complaint and its accompanying (but unfiled) claim charts (Compl. ¶14-15).
• Willful Infringement: Willfulness is alleged based on Defendant’s continued infringement after gaining "actual knowledge" of the patent and its alleged infringement from the service of the complaint (Compl. ¶13-14). The allegation appears to be based entirely on post-suit conduct.

VII. Analyst’s Conclusion: Key Questions for the Case

1. A central procedural question will be one of evidentiary specificity: which of Defendant's vast portfolio of hardware and software products are actually accused of infringement, and what specific technical evidence does Plaintiff possess to support the allegation that these products implement the claimed three-transform gating architecture?
2. The case will likely involve a core issue of claim scope: can the elements of Claim 1—specifically the distinct "plain," "transform," and "carry" signals generated by three non-linear transforms—be mapped onto the complex, integrated operations of a modern GPU or deep learning library, or is the architecture described in the patent a specific implementation not practiced by the accused products?
3. A key legal and technical question will be one of functional distinction: does the accused technology perform the claimed function of creating a "weighted sum of a non-transformed...component...and the plain signal," or does it achieve a similar outcome (e.g., mitigating vanishing gradients) through fundamentally different, non-infringing technical means?