Patent Cases, NETWORK-1 TECHNOLOGIES, INC., Plaintiff-Appellant v. GOOGLE LLC, YOUTUBE, LLC Defendants-Appellees, U.S. Court of Appeals, Federal Circuit, (Apr. 23, 2026)

Lourie, Circuit Judge.

Network-1 Technologies, Inc. (“Network-1”) filed suit in the United States District Court for the Southern District of New York, asserting that two separate versions of Google LLC and YouTube, LLC's (collectively, “Google”) Content ID system infringed several claims of its U.S. Patents 8,010,988 (“the ′988 patent”), 8,205,237 (“the ′237 patent”), 8,904,464 (“the ′464 patent”) (collectively, “the asserted patents”). Network-1 Techs., Inc. v. Google LLC, No. 14-cv-PGG-02396, 2024 WL 1814296 (S.D.N.Y. Apr. 24, 2024) (“Decision”). The district court issued a combined claim construction and summary judgment decision. Id. The court first determined that the asserted claims of the ′988 and ′464 patents are invalid as indefinite. Id. at *1. It then determined that neither version of Google's Content ID system infringed the asserted claims of the ′237 patent as a matter of law and thus Google was entitled to summary judgment of noninfringement as to that patent. Id.

Because a genuine issue of material fact exists as to whether one version of Content ID infringes the ′237 patent, we reverse and remand as to that issue. We affirm the district court's conclusion of invalidity of the ′988 and ′464 patents for indefiniteness and grant of summary judgment of noninfringement of the ′237 patent as to the other version of Content ID.

Background

I

Network-1 owns the asserted patents, which are directed to methods for “linking traditional media to new interactive media, such as that provided over the internet,” and “identifying a work … without the need to modify the work.” See ′988 patent col. 1 ll. 23–28. [1] As an example, the patent describes “[c]ommerce opportunities” where television viewers could “place[] … direct orders for products” they see on screen via an “interactive capability,” i.e., an internet link to a website associated with the specific product. Id. at col. 1 ll. 36–60. The asserted patents facilitate this process by describing a technique for identifying work (such as content or an advertisement) without inserting an identification code, such as a bar code. See id. at col. 1 ll. 25–28, col. 3 ll. 8–23, col. 4 ll. 7–19.

In relevant part, claims 15 and 17 of the ′988 patent and claim 1 of the ′464 patent recite a method comprising identifying an electronic work or correlating an electronic work with an identifier via a “non-exhaustive search.” See ′988 patent col. 25 l. 65–col. 26 l. 6; ′464 patent col. 24 ll. 44–49.

In relevant part, independent claim 33 of the ′237 patent recites “[a] computer-implemented method comprising … [a] determin[ation] by the computer system, an identification of [a] media work using the media work extracted features to perform a sublinear approximate-nearest neighbor search of reference extracted features of reference identified media works.” ′237 patent col. 28 ll. 5, 10–14 (emphasis added).

Google operates the website YouTube, which allows users to upload content to the internet to be viewed by the public. Decision, 2024 WL 1814296 at *3. Google employs a “Content ID” system that allows content owners (e.g., copyright owners) to control how their content is used on YouTube. Id. at *4. The Content ID system generates matches by comparing an uploaded video, also known as a “query work,” to a database of reference works. Id. There are two versions of Google's Content ID system at issue in this appeal: an older version known as “LSH,” and a newer version known as “Siberia.” Id.

The LSH version of Content ID works by searching over an index of local sensitive hashing (“LSH”) bands, into which “subfingerprints” corresponding to short snippets of reference works are organized. Id. Subfingerprints for query works and reference works are generated in the same manner so that when a search is conducted for a particular query work, the LSH version of Content ID returns only the reference works associated with a matching LSH band. Id. These reference works are then further processed to eliminate candidates unlikely to be a match with the query work. Id.

The Siberia version of Content ID works by generating a sequence of embeddings corresponding to short snippets or frames of content. Id. The reference embeddings are further processed and stored in multiple reference indices for searching, organized by content type. Id. at *5–6. Each index is divided into smaller indices known as “shards” that can each fit on one computer. Id. at *6.

II

In April 2014, Network-1 sued Google for infringement of several claims of the asserted patents. [2] Decision, 2024 WL 1814296 at *7.

In June 2015, Google petitioned the United States Patent and Trademark Office Patent Trial and Appeal Board (“the Board”) for inter partes review (“IPR”), asserting that the ′237 and ′988 patents were unpatentable as anticipated and obvious. See id. The district court stayed the case pending resolution of the IPRs. Id. The Board instituted the IPRs, and the proceedings before the Board turned on the term “non-exhaustive search.” Id. The Board found that Google had failed to carry its burden of demonstrating the claims were not patentable. Google LLC v. Network-1 Techs., Inc., 726 F. App'x 779, 780 (Fed. Cir. 2018) (“IPR Appeal Decision”). Google appealed, and we vacated and remanded the Board's decision because it erred in its construction of “non-exhaustive search.” Id. at 787. Using the broadest reasonable construction standard, we construed “non-exhaustive search” as “a search that locates a match without conducting a brute-force comparison of all possible matches, and all data within all possible matches.” Id. at 786.

In January 2019, the Board terminated the proceedings on remand upon joint stipulation of the parties. Google, Inc. v. Network-1 Techs., Inc., No. IPR2015-00343, 2019 WL 104044 (P.T.A.B. Jan. 4, 2019). Following the Board's decision, the district court lifted the stay in this case and the parties narrowed the claims to claim 17 of the ′988 patent; claims 1, 8, 10, 16, 18, 25, 27, and 33 of the ′464 patent; and claims 33–35 of the ′237 patent. See Decision, 2024 WL 1814296 at *9, 11.

In April 2024, the district court published a combined claim construction and summary judgment order disposing of all asserted claims. Id. at *38. It concluded that the term “non-exhaustive search” recited in the asserted claims of the ′988 and ′464 patents was indefinite, rendering those claims invalid. Id. at *20. The district court also granted summary judgment of noninfringement of the asserted claims of the ′237 patent, concluding that there was no genuine issue of material fact that Google's LSH and Siberia Content ID systems met the “sublinear search” limitation. Id. at *34, 38.

Network-1 timely appealed, and we have jurisdiction under 28 U.S.C. §1295(a)(1).

Discussion

Network-1 appeals the district court's invalidity and summary judgment determinations. See Open. Br. 26–27. We address each in turn.

I

Patent claims must “particularly point[] out and distinctly claim[] the subject matter” regarded as the invention. 35 U.S.C. §112 ¶2. “A claim fails to satisfy this statutory requirement and is thus valid for indefiniteness if its language, read in light of the specification and prosecution history, ‘fail[s] to inform, with reasonable certainty, those skilled in the art about the scope of the invention.’” Interval Licensing LLC v. AOL, Inc., 766 F.3d 1364, 1369–70 (Fed. Cir. 2014) (quoting Nautilus, Inc. v. Biosig Instrs., Inc., 572 U.S. 898, 901 (2014)). Definiteness is measured from the viewpoint of a skilled artisan “at the time the patent was filed.” Nautilus, 572 U.S. at 908.

We review a determination of indefiniteness de novo. Interval Licensing, 766 F.3d at 1370. Because “[g]eneral principles of claim construction apply to indefiniteness,” “we review a district court's determinations of subsidiary facts based upon extrinsic evidence for clear error, and those based upon intrinsic evidence (the patent claims, specification, and prosecution history) de novo.” HZNP Medicines LLC v. Actavis Lab'ys UT, Inc., 940 F.3d 680, 688 (Fed. Cir. 2019) (citing Biosig Instrs., Inc. v. Nautilus, Inc., 783 F.3d 1374, 1377–78 (Fed. Cir. 2015)).

The claims of the ′988 patent, which are representative of the asserted claims of both that patent and the ′464 patent for purposes of analyzing the indefiniteness inquiry, see Open. Br. 25, read as follows:

15. A method for associating an electronic work with an action, the electronic work comprising at least one of audio and video, the method comprising:

a) electronically extracting features from the electronic work;

b) electronically determining an identification of the electronic work based on the extracted features, wherein the identification is based on a non-exhaustive search identifying a neighbor;

c) electronically determining an action based on the identification of the electronic work; and

d) electronically performing the action.

17. The method of claim 15, wherein the non-exhaustive search is sublinear.

′988 patent col. 25 l. 65–col. 26 l. 9; id. col. 26 ll. 14–15 (emphases added).

Network-1 submits the following construction for “non-exhaustive search” in contending that the term is not indefinite: “a search designed to locate a [near] neighbor without comparing to all possible matches (i.e., all records in the reference data set), even if the search does not locate a [near] neighbor.” Open. Br. 52. Google submits that the district court was correct and the term is indefinite. Resp. Br. 21. We conclude that the district court correctly determined that the “non-exhaustive” search limitation is indefinite.

We begin with the claim language. Claims 15 and 17, taken together, refer to a “non-exhaustive search identifying a neighbor” which can be “sublinear.” ′988 patent col. 26 ll. 5–6, 14–15. Other than demonstrating that “non-exhaustive searches” do not inherently identify a neighbor (i.e., a close match) or are sublinear, the scope of that language sheds little light on how a skilled artisan would understand this term.

The written description also does not inform a skilled artisan as to the scope of “non-exhaustive search” with reasonable certainty. Our previous analysis of the term from the IPR Appeal Decision, although not binding, is informative. See 726 Fed. App'x at 786. There, Network-1 argued that the written description of a materially similar patent reasonably conveyed the scope of the term because it differentiated between “exhaustive searches” and “non-exhaustive searches” because it “identifie[d] ‘a linear search of all N entries’ as an ‘exhaustive search’” and “[o]ther forms of matching,” such as “those based on clustering, kd-trees, vantage point trees and excluded middle vantage point forest,” as “non-exhaustive searches.” Id. at 784–85. We rejected that argument because the written description did not “draw a clear line between ‘exhaustive’ and ‘non-exhaustive’ searching in terms of how much data within a record a search must consider in order to qualify as one or the other.” Id. at 785 (emphasis added).

That same reasoning applies here. Network-1 argues that column 9, lines 24–32 of the written description informs a skilled artisan as to the scope of “non-exhaustive searches” with reasonable certainty. Open. Br. 52–53. But that language fails to do so. The written description does not contain the terms “exhaustive” or “non-exhaustive,” but instead contrasts “linear search[es]” which “can be computationally very expensive” with “[o]ther forms of matching,” including “kd-trees, vantage point trees and excluded middle vantage point forests.” ′988 patent col. 9 ll. 24–30. Although the written description refers to a “linear search,” nothing in the written description suggests that a skilled artisan would understand a “linear search” to be interchangeable with a “non-exhaustive search.” Nor does it inform a skilled artisan with reasonable certainty as to how much data within a record the “non-exhaustive search” must consider.

Indeed, the phrase “non-exhaustive search” appears nowhere in the original patent application that led to the ′988 and ′464 patents, but rather was added to the claims nine years after the filing of the provisional application and three years after the patent filing date which contained the “linear search” language, in part to overcome prior art. J.A. 3245–53; see J.A. 79. To narrow “non-exhaustive search” based on the “linear search” language used in the written description would thus be to “view[] matters post hoc”––a practice the Supreme Court has admonished. See Nautilus, 572 U.S. at 911–12 (“[T]he definiteness inquiry trains on the understanding of a skilled artisan at the time of the patent application.”). It does the patent system or the public no favor for this court to accept the distinguishing of prior art with vague language that only results in the uncertainty illustrated by this appeal and this case's long history.

Furthermore, the publications incorporated by reference into the written description do not meaningfully narrow the scope of “non-exhaustive search.” See ′988 patent col. 7 ll. 37–43, col. 9 ll. 32–38. The works cited–– Duda & Hart, [3] Fukunaga, [4] and the Yianilos papers [5] –– either make passing references to an “exhaustive search” without any further information, see J.A. 2872 (Duda & Hart), J.A. 2961 (Fukunaga), or conflict with Network-1's proposed “non-exhaustive searches,” see J.A. 1403–04 (Yianilos describing a vantage point forest search possibly being “exhaustive”).

Even if the intrinsic record were not dispositive on its own, the extrinsic evidence further supports a conclusion of indefiniteness. The district court considered the declaration of Dr. Michael Mitzenmacher, Network-1's expert witness, as well as several academic papers which provided definitions of “exhaustive” versus “non-exhaustive searches.” Decision, 2024 WL 1814296 at *15–18; see J.A. 2764–95 (Mitzenmacher declaration). It ultimately concluded that the extrinsic evidence did not support Network-1's construction but instead “highlight[ed] the vague nature of ‘exhaustive search’ and ‘non-exhaustive search.’” Decision, 2024 WL 1814296 at *18.

We find no clear error in the district court's analysis. Mitzenmacher's declaration simply repeats the same failed points above regarding the written description. See J.A. 2781–86. The declaration also relies on academic papers Denny [6] and Orwant. [7] Id. at 2783, 2786–87. But these do not inform a skilled artisan as to the meaning of “non-exhaustive search.” Orwant states that “the definition of exhaustive search is vague.” J.A. 3020. And Denny defines a “non-exhaustive search strategy” as one that “traverse[s] the search space more or less at random and thus certain states may never be examined.” J.A. 3011. But as the district court observed, this is inconsistent with Network-1's proposed construction because the proposed construction does not have such a requirement to “traverse the search space more or less at random.” Decision, 2024 WL 1814296 at *17.

Network-1's argument to the contrary is unpersuasive. It argues that our analysis of “non-exhaustive search” in the IPR Appeal Decision should not inform our decision here because it was under the broadest reasonable interpretation (“BRI”) standard. Open. Br. 57–58. We disagree. It is true that we utilized the BRI, not Phillips, standard in the IPR Appeal Decision. See 726 Fed. App'x at 786. But the analysis regarding the written description's lack of identification of “non-exhaustive search” is the same––i.e., regardless of standard, the written description does not draw a clear line between “exhaustive” and “non-exhaustive.” See id. at 785.

In sum, neither the intrinsic nor extrinsic evidence offers sufficient guidance to a skilled artisan as to the scope of the term “non-exhaustive search” with reasonable certainty. The asserted claims of the ′988 and ′464 patent are thus invalid for indefiniteness.

II

We next consider the district court's grant of summary judgment of noninfringement of the ′237 patent. [8] “We review the district court's grant of summary judgment under the law of the regional circuit in which the court sits, here, the Second Circuit.” See Medgraph, Inc. v. Medtronic, Inc., 843 F.3d 942, 947 (Fed. Cir. 2016) (citation omitted). “The Second Circuit reviews a grant of summary judgment without deference, construing the evidence in the light most favorable to the nonmovant and drawing all reasonable inferences in that party's favor.” Id. (citing Kuebel v. Black & Decker Inc., 643 F.3d 352, 358 (2d Cir. 2011)). “Summary judgment may only be granted when no ‘reasonable jury could return a verdict for the nonmoving party.’” Id. (quoting Anderson v. Liberty Lobby, Inc., 477 U.S. 242, 248 (1986)).

“Summary judgment of noninfringement is appropriate where the patent owner's proof is deficient in meeting an essential part of the legal standard for infringement, since such failure will render all other facts immaterial.” Telemac Cellular Corp. v. Topp Telecom, Inc., 247 F.3d 1316, 1323 (Fed. Cir. 2001). “[I]n order for a court to find infringement, the plaintiff must show the presence of every … [limitation] or its substantial equivalent in the accused device.” Wolverine World Wide, Inc. v. Nike, Inc., 38 F.3d 1192, 1199 (Fed. Cir. 1994).

Claim 33 of the ′237 patent, which is representative for purposes of analyzing summary judgment, reads as follows:

33. A computer-implemented method comprising:

a) obtaining, by a computer system including at least one computer, media work extracted features that were extracted from a media work, the media work uploaded from a client device;

b) determining, by the computer system, an identification of the media work using the media work extracted features to perform a sublinear approximate nearest neighbor search of reference extracted features of reference identified media works; and

c) determining, by the computer system, an action based on the determined identification of the media work.

′237 patent col. 28 ll. 5–16 (emphasis added).

The parties' dispute centers on whether the LSH or Siberia versions of Content ID meet the “sublinear search” limitation as recited in claim 33 of the ′237 patent. Open. Br. 30; Resp. Br. 41. The parties agree that LSH or Siberia meet the “sublinear” limitation if its “execution time scales with a less than linear relationship to the size of the data set to be searched, assuming computing power is held constant.” Open. Br. 44; Resp. Br. 41; see also Decision, 2024 WL 18142962024, at *28. The district court concluded that Network-1 did not create a material issue of fact as to whether either version met this limitation and thus granted summary judgment of noninfringement. Id. at *34, *38. We address each version in turn.

A

We begin with the LSH version of Content ID. Network-1 relied on three pieces of evidence in opposing summary judgment before the district court: (1) the academic work of Google research scientist Dr. Shumeet Baluja (“the Baluja papers”) and his testimony; (2) a Google 2010 draft document; and (3) Mitzenmacher's report. See Decision, 2024 WL 18114296 at *28. The district court concluded that none of the three created a genuine issue of material fact as to whether the LSH version was sublinear. Id. at *34. After de novo review of each, we agree with the district court.

The Baluja papers and his testimony do not create a genuine issue of material fact. See J.A. 9690–9703 (paper titled “Waveprint”); J.A. 9714–51 (paper titled “Learning to Hash”); J.A. 7225–26, 7321–24, 7361–63, 7371–74, 7398–7400, 7403, 7406 (Baluja testimony). Network-1 argues that the papers contain statements which a reasonable trier of fact could interpret to describe LSH systems as sublinear. See Open. Br. 32–33. But neither paper describes the LSH version of Content ID. See generally J.A. 9690–703; J.A. 9714–51. Rather, they describe a system called Waveprint, which uses computer-vision techniques for identifying audio, see J.A. 9690, and describe hash functions only generally without addressing the LSH version, see J.A. 9716–18. And at his deposition, Baluja confirmed that he was not involved in implementation, but rather only research. J.A. 7399. Network-1's argument is thus unavailing, as there is insufficient evidence connecting Baluja's papers or testimony to Google's LSH version to create a genuine issue of material fact.

The Google 2010 draft document suffers from a similar flaw: there is insufficient evidence connecting it to the implementation of the LSH version. Network-1 argues it creates a genuine issue of material fact because it describes Google's matching infrastructure as having LSH tables that will scale sublinearly. Open. Br. 34–35 (citing J.A. 9387). We disagree. There is no evidence in the record that Google implemented such tables from the 2010 draft document. Indeed, the document contemplates two types of LSH tables: one that scales sublinearly and one that does not. J.A. 9387. But Network-1 points to no evidence that one was implemented over the other. See Open. Br. 34–36. Rather, Network-1 is essentially arguing that Google was capable of infringing the ′237 patent. But even with all reasonable inferences in Network-1's favor, there is no evidence of what Google actually implemented to present a genuine issue of material fact. See Miken Composites, L.L.C. v. Wilson Sporting Goods Co., 515 F.3d 1331, 1340–41 (Fed. Cir. 2008) (affirming summary judgment of noninfringement when there was no record evidence of literal infringement, even though the possibility of infringement existed).

Finally, the Mitzenmacher report does not create a genuine issue of material fact. Indeed, it relies significantly on the Baluja papers, Baluja's testimony, and the Google 2010 draft document. See J.A. 6598–99. But as we explained above, those do not create a genuine issue of material fact because there is insufficient evidence connecting them to the implementation of the LSH version. Furthermore, Mitzenmacher's review of Google's code does not demonstrate that the LSH version is sublinear; it rather steps through the process of the content identification without any analysis as to how it is sublinear. See J.A. 6599–6602. Because Mitzenmacher's report contains no other support, it does not create a genuine issue of material fact. See Minkin v. Gibbons, P.C., 680 F.3d 1341, 1352 n.5 (Fed. Cir. 2012) (“It is well-established that unsupported expert opinions do not create a genuine issue of material fact.” (citations omitted)).

In sum, no piece of evidence considered by the district court, and which Network-1 relies on before us, creates a genuine issue of material fact. Google is thus entitled to judgment as a matter of law of noninfringement of the ′237 patent as to the LSH version.

B

We next consider the Siberia version. The district court granted summary judgment of noninfringement because, in its view, the evidence indicated that the Siberia version did not scale sublinearly. See Decision, 2024 WL 1814296 at *36.

Upon de novo review, we conclude that there is a genuine issue of material fact regarding the Siberia version so as to preclude summary judgment of noninfringement of the ′237 patent.

First, there is a fundamental dispute as to how the Siberia version functions. Network-1 argues that the Siberia version is sublinear because the algorithm has different “tunable knobs,” which Google adjusted to lower resource costs. See Open. Br. 45–47 (citing J.A. 9812; J.A. 10395–96). Google counters that the Siberia version is undisputedly linear because it “search[es] a fixed fraction of the data set.” Resp. Br. 54 (citing J.A. 3953). And it further argues that “[e]ven if Google reduced the number of searched [components], Siberia would still search a predetermined proportion of all partitions and thus scale linearly.” Id. at 60 (citing J.A. 4447–48).

We conclude that this dispute is best left to a factfinder. With all reasonable inferences in Network-1's favor, a factfinder could find that the Siberia version as an overall system was designed to have an execution time which scaled in less than a proportional relationship to the size of the reference set because the “tunable knobs” allowed the system to adapt to a growing dataset size. See J.A. 9812; J.A. 10335; J.A. 10378–79; J.A. 10395–96. For purposes of summary judgment, it is immaterial that the Siberia version, once that “knob” is turned, may search that new “fixed fraction” linearly.

Furthermore, an internal Google document regarding the Siberia version states that Google “will need … a sublinear search.” J.A. 10265 (emphasis added). That is also enough to create a factual issue; with reasonable inferences in Network-1's favor, a factfinder could consider this document to describe the implementation of the Siberia version as sublinear. Google argues this document “is referring to a potential strategy, not Siberia as implemented.” Resp. Br. 56. That argument is unpersuasive. A factfinder could reasonably read the phrase “will need” as imposing a requirement of a sublinear search on any future Siberia implementation. J.A. 10265. It thus creates a genuine issue of material fact regarding whether the Siberia version is sublinear.

Because there is a genuine issue of material fact, summary judgment of noninfringement of the ′237 patent as to the Siberia version is inappropriate.

Conclusion

We have considered the remainder of the parties’ arguments but find them unpersuasive. We reverse the district court's grant of summary judgment of noninfringement of the ′237 patent as to the Siberia version of Content ID and remand for further proceedings. We affirm the district court's conclusion of indefiniteness as to the ′988 patent and ′464 patent. We affirm the district court's grant of summary judgment of noninfringement of the ′237 patent as to the LSH version of Content ID.

AFFIRMED IN PART, REVERSED IN PART, AND REMANDED

Costs

No costs.