DCT
1:25-cv-01021
Xockets Inc v. Amazon.com Inc
I. Executive Summary and Procedural Information
- Parties & Counsel:
- Plaintiff: Xockets, Inc. (Texas)
- Defendant: Amazon.com, Inc. and Amazon Web Services, Inc. (Delaware)
- Plaintiff’s Counsel: Susman Godfrey LLP
- Case Identification: 1:25-cv-01021, W.D. Tex., 06/30/2025
- Venue Allegations: Plaintiff alleges venue is proper in the Western District of Texas because Defendant maintains established places of business in the district, including its Annapurna Labs facility in Austin, where it allegedly researches, designs, develops, and tests the accused instrumentalities.
- Core Dispute: Plaintiff alleges that Defendant’s cloud computing server systems, which utilize Data Processing Units (DPUs) like the AWS Nitro System and NVIDIA NVLink Switch systems, infringe patents related to a novel cloud fabric and switching plane architecture.
- Technical Context: The technology involves specialized processors (DPUs) and network architectures designed to offload data-intensive tasks from central server processors, thereby accelerating performance for large-scale distributed computing, particularly in the fields of artificial intelligence and machine learning.
- Key Procedural History: The complaint details a significant history between the parties, alleging that Amazon first learned of the patented technology at a 2015 industry conference. It further alleges that in a May 2017 "Deep Dive" meeting, held under the premise of a potential acquisition, Xockets disclosed detailed technical and design information to Amazon, including to the head of its Austin-based Annapurna Labs. The complaint asserts Amazon subsequently launched its infringing AWS Nitro product within a year of this meeting. Finally, it notes that in March 2024, Xockets formally notified Amazon of its patents during a sales process, which Amazon declined to join.
Case Timeline
| Date | Event |
|---|---|
| 2012-05-22 | Earliest Priority Date ('297, '092 Patents) |
| 2013-01-17 | Earliest Priority Date ('161, '640 Patents) |
| 2015-09-01 | Amazon allegedly first became aware of Xockets’ technology at Strata Conference |
| 2016-06-28 | U.S. Patent 9,378,161 Issues |
| 2016-09-06 | U.S. Patent 9,436,640 Issues |
| 2017-05-01 | Xockets and Amazon hold "Deep Dive" meeting |
| 2017-12-31 | Amazon allegedly begins deploying Xockets' DPU technologies on Nitro v3 |
| 2019-02-19 | U.S. Patent 10,212,092 Issues |
| 2019-03-05 | U.S. Patent 10,223,297 Issues |
| 2024-03-01 | Xockets allegedly puts Amazon on notice of its patents during sales process |
| 2025-06-30 | Complaint Filing Date |
II. Technology and Patent(s)-in-Suit Analysis
U.S. Patent No. 10,223,297 - Offloading of Computation for Servers Using Switching Plane Formed by Modules Inserted Within Such Servers (Issued Mar. 5, 2019)
The Invention Explained
- Problem Addressed: The patent addresses the performance bottlenecks in conventional cloud data centers where data-intensive workloads overwhelm server processors (CPUs) and the standard communication network, leading to inefficiencies and reduced speed (Compl. ¶8, ¶72).
- The Patented Solution: The invention proposes a system of servers, where each server contains not only a main processor but also a separate "computation module" (e.g., a Data Processing Unit or DPU). Each computation module includes its own switch. The key innovation is that these "second switches" within the computation modules of different servers connect to form a new, independent "switching plane." This secondary network fabric operates independently of the main processors and the primary communication network, allowing data-intensive tasks to be offloaded from the CPUs and brokered directly between the computation modules, thereby accelerating distributed computing (Compl. ¶25, ¶73-74; ’297 Patent, Abstract, col. 7:15-32).
- Technical Importance: This architecture fundamentally changes data flow in a data center, enabling massive performance gains for applications like AI model training by creating a high-speed, specialized data path that bypasses traditional chokepoints (Compl. ¶5, ¶75).
Key Claims at a Glance
- Independent Claim Asserted: Claim 1 (Compl. ¶149).
- Essential Elements of Claim 1:
- A system comprising a plurality of first server modules interconnected via a communication network.
- Each server module includes a first switch, at least one main processor, and at least one computation module coupled to the main processor by a bus.
- Each computation module includes a second switch and a plurality of computation elements.
- The second switches of the server modules form a switching plane for ingress and egress of network packets independent of any main processors.
- Each computation module is insertable into a physical connector of the first server module.
- The complaint asserts dependent claim 7 and reserves the right to assert others (Compl. ¶148, ¶195).
U.S. Patent No. 10,212,092 - Architectures and Methods for Processing Data in Parallel Using Offload Processing Modules Insertable into Servers (Issued Feb. 19, 2019)
The Invention Explained
- Problem Addressed: The patent addresses the challenge of efficiently executing distributed computing tasks, such as MapReduce workloads, where data must be processed across many different nodes in a cluster, a process that is often constrained by network latency and processor overhead (’092 Patent, col. 3:46-57).
- The Patented Solution: Building on the concept of a secondary fabric formed by offload modules, this invention details a specific distributed computing architecture where tasks are divided between the main server CPUs and the offload modules. "First computing operations" (e.g., the "Map" step) are performed on the main CPUs, generating "first processed data." This data is then sent to the offload modules, which form a "switch fabric" and execute "second computing operations" (e.g., the "Shuffle" and "Reduce" steps) in parallel directly within the network fabric. The virtual switches on the offload modules form a switch fabric for exchanging this data, enabling "in-network computing" (Compl. ¶26; ’092 Patent, Abstract, col. 3:35-57, FIG. 3-1).
- Technical Importance: This architecture accelerates large-scale parallel processing by moving parts of the computation itself into the network, reducing the data that must be sent back and forth to central processors and minimizing bottlenecks (Compl. ¶73).
Key Claims at a Glance
- Independent Claim Asserted: Claim 1 (Compl. ¶215).
- Essential Elements of Claim 1:
- A distributed computing architecture for executing at least first and second computing operations in parallel.
- It comprises a plurality of servers, each with at least one CPU and at least one offload processing module.
- The offload processing module has computation elements configured to operate as a virtual switch and execute the second computing operations.
- The virtual switches form a switch fabric for exchanging data between the offload modules.
- Critically, the first computing operations generate first processed data and are not executed by the offload processing modules, while the second computing operations are executed on a plurality of the offload processing modules in parallel.
- The complaint reserves the right to assert other claims.
U.S. Patent No. 9,378,161 - Full Bandwidth Packet Handling With Server Systems Including Offload Processors (Issued June 28, 2016)
- Technology Synopsis: The patent is directed to a rack-level server system where offload processor modules (DPUs) are connected to each other directly via their respective I/O ports. This direct connection forms a novel switching architecture or fabric that enables active traffic management and stream processing to accelerate distributed computing (Compl. ¶27, ¶267).
- Asserted Claims: At least Claim 1 (Compl. ¶271).
- Accused Features: Amazon's AWS Compute Server Systems, particularly AWS UltraCluster, equipped with NVIDIA NVLink Switch DPUs are accused of infringing (Compl. ¶271, ¶273).
U.S. Patent No. 9,436,640 - Full Bandwidth Packet Handling With Server Systems Including Offload Processors (Issued September 6, 2016)
- Technology Synopsis: The patent describes a rack server system using offload processor modules (DPUs) for in-network reduction/combining operations, particularly for map/reduce data processing. The architecture uses a "midplane switch" formed by directly connecting offload processors, allowing one module to execute "map" steps and provide data directly to a second module for executing "reduce" steps (Compl. ¶28, ¶292).
- Asserted Claims: At least Claim 9 (Compl. ¶296).
- Accused Features: Amazon's AWS Compute Server Systems, including AWS UltraCluster with NVIDIA NVLink Switch DPUs, are accused of infringing (Compl. ¶295).
III. The Accused Instrumentality
Product Identification
The accused instrumentalities are primarily Amazon’s cloud server systems, categorized as (1) AWS Compute Server Systems and (2) AWS Storage Server Systems (Compl. ¶147, ¶213). These systems are alleged to be infringing when they incorporate either the AWS Nitro System (which includes the Nitro DPU, Elastic Fabric Adapter (EFA), and Elastic Network Adapter (ENA)) or NVIDIA's NVLink Switch DPU technology (Compl. ¶147). Specific examples cited include AWS Trainium Servers and the AWS UltraCluster (Compl. ¶147).
Functionality and Market Context
The complaint alleges these systems use DPUs to offload and accelerate infrastructure functions like networking, storage, and security from the main server processors (Compl. ¶153, ¶177). The AWS Nitro DPU is alleged to operate as a virtual switch that creates a secondary switching plane to facilitate high-speed inter-node communications, particularly for ML/AI workloads (Compl. ¶158, ¶160). The complaint includes a diagram from a Xockets presentation illustrating how its "StreamSwitch" transforms cluster architecture, a concept it alleges Amazon adopted. (Compl. p. 26). Similarly, the NVIDIA NVLink Switch DPUs are alleged to form a "non-blocking compute fabric" that performs "in-network computing operations" such as reductions for ML/AI training (Compl. ¶170, ¶172). The complaint alleges these technologies are central to Amazon's AI strategy and provide massive financial benefits and cost savings (Compl. ¶20-21, ¶102, ¶109-111). The complaint includes a photograph of a chip from Annapurna Labs, the Amazon subsidiary alleged to be central to the design of the accused technology (Compl. p. 14).
IV. Analysis of Infringement Allegations
’297 Patent Infringement Allegations
| Claim Element (from Independent Claim 1) | Alleged Infringing Functionality | Complaint Citation | Patent Citation |
|---|---|---|---|
| a plurality of first server modules interconnected to one another via a communication network | Amazon EC2 instances (server modules) are interconnected by AWS Cloud, using networking like Elastic Fabric Adapter (EFA) or Elastic Network Adapter (ENA). | ¶151-152, ¶166 | col. 7:15-17 |
| each first server module including ... at least one main processor, and at least one computation module coupled to the main processor by a bus | Each EC2 compute instance includes a host processor (main processor). It also includes an AWS Nitro DPU or NVLink Switch DPU (computation module) coupled to the host via a PCIe bus. | ¶156-157, ¶168-169 | col. 7:19-22 |
| each computation module including a second switch, and a plurality of computation elements | The AWS Nitro DPU operates as a virtual switch and includes ASICs and ARM cores (computation elements). The NVLink Switch DPU includes a programmable switch and SHARP hardware accelerator engines (computation elements). | ¶158-159, ¶170-171 | col. 7:23-25 |
| the second switches of the first server modules form a switching plane for the ingress and egress of network packets independent of any main processors of the first server modules | The virtual switches in the Nitro DPUs form a switching plane for ML/AI packet flows that offloads communication from the host processor. The switches in the NVLink DPUs form a non-blocking compute fabric independent of the main superchips. | ¶160-161, ¶172-173 | col. 7:26-30 |
| each computation module is insertable into a physical connector of the first server module | The AWS Nitro DPU is inserted into physical PCIe interfaces. The NVLink Switch DPU is insertable into a physical connector of the NVIDIA server module. | ¶162, ¶174 | col. 7:31-33 |
Identified Points of Contention
- Scope Questions: A central question may be whether a "computation module" described in the patent, which is shown in embodiments as a DIMM-based module, can be construed to read on the accused PCIe-based Nitro cards and other DPU form factors. The definition of "insertable into a physical connector" will be key.
- Technical Questions: The analysis will likely focus on the degree to which the alleged switching plane is "independent of any main processors." A court may need to determine if the data plane's independence is sufficient, even if the main processors are involved in control plane functions like establishing and managing the virtual switches.
’092 Patent Infringement Allegations
| Claim Element (from Independent Claim 1) | Alleged Infringing Functionality | Complaint Citation | Patent Citation |
|---|---|---|---|
| A distributed computing architecture for executing at least first and second computing operations executed in parallel | Amazon EC2 instances, including Trainium instances in UltraClusters, form a distributed computing architecture for ML/AI workloads. | ¶217-218, ¶230 | col. 3:19-22 |
| a plurality of servers, including first servers that each include at least one central processing unit (CPU), and at least one offload processing module | EC2 compute instances (servers) include host processors (CPUs) and an AWS Nitro DPU or NVLink Switch DPU (offload processing module). | ¶219-220, ¶233, ¶235 | col. 1:49-53 |
| the virtual switches form a switch fabric for exchanging data between the offload processing modules | The Nitro DPUs, interconnected by EFA networking, form a switching fabric. The NVLink Switch DPUs form a non-blocking compute fabric. | ¶224-225, ¶239-240 | col. 3:35-39 |
| the first computing operations generate the first processed data and are not executed by the offload processing modules | Customer applications on EC2 compute instances (main CPUs) generate requests (e.g., for storage) that initialize microservices. In ML/AI training, GPUs (on the server) perform arithmetic operations to generate partial local gradients. These are not executed by the DPUs. | ¶226, ¶241 | col. 3:52-57 |
| the second computing operations are executed on a plurality of the offload processing modules in parallel | The AWS Nitro DPU executes microservices (e.g., mapping NVMe functions) in parallel. The NVLink Switch DPUs use SHARP engines to perform in-network reduction operations in parallel on the gradient data. | ¶227, ¶242 | col. 3:58-62 |
Identified Points of Contention
- Scope Questions: The term "computing operations" will be critical. A court will need to decide whether a host CPU making a request that initiates an I/O function on a DPU qualifies as the "first computing operations generating first processed data," or if it requires a more substantive computational step like the map/reduce examples in the specification.
- Technical Questions: Infringement may turn on whether the accused functionality matches the claimed division of labor. For the Nitro system, the question is whether handling I/O microservices constitutes the claimed "second computing operations." The infringement theory against the NVIDIA system, which allegedly performs in-network mathematical reductions, may present a more direct mapping to the patent's description of in-network computing.
V. Key Claim Terms for Construction
"computation module" ('297 Patent, Claim 1)
- Context and Importance: The definition of this term is foundational, as the AWS Nitro DPU and NVIDIA NVLink Switch DPU must fall within its scope for infringement to be found. Practitioners may focus on this term because the patent's embodiments often depict modules in a DIMM form factor (’297 Patent, FIG. 4-1), whereas the accused products are primarily PCIe-based cards or other integrated components.
- Intrinsic Evidence for Interpretation:
- Evidence for a Broader Interpretation: The specification describes the modules functionally as containing "offload processors" which can include a variety of commodity processors like ARM cores or FPGAs, suggesting the term is not limited to a specific physical form factor but rather to its role in offloading computation (’092 Patent, col. 6:47-55).
- Evidence for a Narrower Interpretation: The abstract of the ’297 Patent and the language of claim 1 emphasize that the module is "insertable into a physical connector of the first server module," which, when read in light of figures showing DIMM-like modules, could be argued to limit the scope to components that plug into memory sockets, not general-purpose I/O buses like PCIe.
"switching plane ... independent of any main processors" ('297 Patent, Claim 1)
- Context and Importance: This term defines the core architectural feature of the invention. The dispute will likely center on the required degree of "independence."
- Intrinsic Evidence for Interpretation:
- Evidence for a Broader Interpretation: The patent's abstract states the plane is for "ingress and egress of network packets independent of any main processors," suggesting independence relates to the data path for network traffic, not necessarily the control path. The complaint alleges the accused systems offload data plane functions, which aligns with this interpretation (Compl. ¶161).
- Evidence for a Narrower Interpretation: A defendant could argue that since the main processor is typically responsible for configuring, managing, and directing tasks to the DPUs, the resulting switching plane is not truly "independent" in an operational sense, but is instead a subordinate, managed component.
VI. Other Allegations
Indirect Infringement
The complaint alleges that Amazon makes and sells hardware and software components (e.g., its compute and storage server systems with Nitro or NVLink DPUs) that are "especially made or especially adapted to practice the invention" and are not staple articles of commerce suitable for substantial non-infringing use (Compl. ¶205, ¶262).
Willful Infringement
The complaint alleges willful infringement based on both pre- and post-suit knowledge. Pre-suit knowledge is alleged to stem from Amazon's attendance at a 2015 conference where Xockets presented its technology; a 2017 "Deep Dive" meeting where Amazon allegedly received extensive technical details from Xockets under the premise of an acquisition; and Amazon's own patents citing Xockets' patents (Compl. ¶117, ¶82-86, ¶122). The complaint provides a floor plan of the 2015 conference showing Amazon's booth located near Xockets' booth (Compl. p. 36). Post-suit knowledge is alleged based on direct notice provided to Amazon in March 2024 as part of a potential IP sales process (Compl. ¶124).
VII. Analyst’s Conclusion: Key Questions for the Case
- A core issue will be one of definitional scope: Can the term "computation module," which is illustrated in patent embodiments with a DIMM form factor, be construed to cover the modern, PCIe-based DPU architecture of the accused AWS Nitro System? The outcome of this claim construction question may be dispositive for a significant portion of the infringement allegations.
- A key evidentiary question will be one of functional division: For the '092 patent, does the accused system's workflow—where a host application makes an I/O request that is fulfilled by a DPU—meet the claim requirement of "first computing operations" on the host generating data that is then processed by "second computing operations" on the DPU? The case may require a detailed technical analysis of whether this constitutes the specific two-step computational process claimed, or merely a standard offloading of I/O functions.
- A central theme of the case will be willfulness and intent: Given the detailed allegations of Amazon's early awareness and deep technical review of Xockets' technology years before launching the accused products, a key question for the fact-finder will be whether Amazon's alleged infringement was willful, potentially exposing it to enhanced damages.