For AI patents asserted against generative AI products, the patent's age can be decisive in determining whether the U.S. Patent and Trademark Office (USPTO or the Office) will review the patent's validity at all. On July 31, 2025, the USPTO declined to review a neural-network computing patent asserted against NVIDIA because the patent had been in force for more than eleven years.

In February 2025, NVIDIA petitioned for inter partes review (IPR) of U.S. Patent No. 8,648,867, "Graphic Processor Based Accelerator System and Method," and reissue patent RE49,461, both asserted against NVIDIA in Neural AI, LLC v. NVIDIA Corp., No. 7:24-cv-00221 (W.D. Tex.). In deciding whether to institute the review, Acting Director Coke Morgan Stewart denied both petitions in IPR2025-00606 and IPR2025-00608, and the grounds differed for each. The '867 patent "has been in force for over 11 years, creating strong settled expectations for Patent Owner." Settled expectations is the Office's term for the reliance interest a patent owner builds as a patent goes unchallenged over time. In contrast, the reissue patent had been in force for less than two and a half years. RE49,461 reissued on March 14, 2023 with claims different from the original 2015 patent, and the decision acknowledged that "ordinarily such circumstances would counsel against discretionary denial," that is, a refusal to hear a challenge for policy reasons without reaching the merits.

Denial of the reissue petition was grounded, instead, on notice and efficiency. Neurala, Inc., the original patent owner, had a commercial relationship with NVIDIA and in 2017 sent NVIDIA a presentation discussing Neurala's patent portfolio, so the Acting Director found NVIDIA "had actual notice of the challenged patents and Petitioner's failure to seek early review of the patents favors denial." Trial in the Western District of Texas was set for September 8, 2026, ahead of the dates by which the IPRs would have produced final written decisions projected as October 14 and 15, 2026, so the schedule put the district court on track to decide the validity of both patents before any final written decision would issue. The Office denied rehearing on October 31, 2025. The '867 patent claims priority to a 2006 application, more than a decade before the transformer architecture behind today's generative AI models was first published.

The NVIDIA denial illustrates a larger change in how the Patent Trial and Appeal Board (PTAB or the Board) handles IPR proceedings. Patent age, the petitioner's prior knowledge of the patent, and the timing of parallel litigation now decide whether many accused infringers get the opportunity of an IPR proceeding at all. Four petitions from accused infringers ask the Supreme Court whether any court may review a denial of institution, or the rules behind such denials, based on grounds the petitioner contends are contrary to the patent statutes. A fifth petition, from a patent owner, asks whether courts may review the Board's refusal to decide a statutory question in an IPR that went forward. A final rule that would convert the USPTO's discretionary grounds for denying IPR institution into mandatory regulatory bars has been under White House regulatory review since July 22, 2026. For AI companies defending patent suits, the stakes are magnified. Many patents now asserted against generative AI products were filed years before those products existed, and older patents are the ones the settled-expectations framework most protects from review by the Board.

Congress Gave the Director Discretion and Made the Institution Decision Nonappealable

IPR, created by the America Invents Act (AIA) of 2011, allows a petitioner to challenge an issued patent before the Board only on grounds of anticipation or obviousness, and only based on prior art "consisting of patents or printed publications." 35 U.S.C. § 311(b). Section 311(c) bars a petition until at least nine months after the patent issues, and the statute sets no maximum age. Section 314(a) bars institution of the IPR proceeding unless there is a reasonable likelihood the petitioner would prevail on at least one challenged claim. But nothing in Section 314(a) requires institution once the threshold is met. Section 314(d) provides that the Director's determination "whether to institute an inter partes review under this section shall be final and nonappealable."

The Supreme Court has construed Section 314(d) three times. In Cuozzo Speed Technologies, LLC v. Lee, 579 U.S. 261 (2016), the Court held that Section 314(d) bars review of questions "closely tied to the application and interpretation of statutes related to the Patent Office's decision to initiate inter partes review." The Court expressly declined to decide the effect of Section 314(d) on appeals "that implicate constitutional questions, that depend on other less closely related statutes, or that present other questions of interpretation that reach, in terms of scope and impact, well beyond 'this section,'" and noted that agency "shenanigans," such as acting "outside its statutory limits," may be reviewable on appeal from a final decision and under the Administrative Procedure Act (APA), the federal statute that authorizes courts to set aside unlawful agency action.

In SAS Institute Inc. v. Iancu, 584 U.S. 357 (2018), the Court held 5 to 4 that once review is instituted, Section 318(a) requires the Board's final written decision to address every claim the petitioner challenged, ending the Office's practice of instituting on only some of the challenged claims. Section 314(d) did not bar that ruling, as the challenge asked whether the Office had exceeded its statutory limits, not whether review should have begun: "nothing in §314(d) or Cuozzo withdraws our power to ensure that an inter partes review proceeds in accordance with the law's demands." In Thryv, Inc. v. Click-To-Call Technologies, LP, 590 U.S. 45 (2020), the Court held that Section 314(d) bars appeal of the Board's application of the one-year time bar in Section 315(b), which requires a petitioner to file within one year after being served with an infringement complaint.

The petitioners now before the Court rely on the Cuozzo reservation and on SAS to argue that a denial resting on a ground the statute does not authorize, such as a patent-age limit or a second filing deadline, remains reviewable. The government relies on Thryv and on United States v. Arthrex, Inc., 594 U.S. 1 (2021), to argue that Section 314(d) covers the entire decision whether to institute, a decision committed to the Director's discretion, and on the principle that "an agency's decision not to take enforcement action should be presumed immune from judicial review," Heckler v. Chaney, 470 U.S. 821, 832 (1985).

Settled Expectations Became a Denial Ground Within Four Months

On February 28, 2025, the USPTO rescinded the June 2022 memorandum that had limited the application of discretionary denials based on parallel district court litigation over the same patent. On March 26, 2025, Acting Director Stewart issued Interim Processes for PTAB Workload Management, which separated discretionary considerations from the merits and invited briefing on "[s]ettled expectations of the parties, such as the length of time the claims have been in force."

The Office has not adopted a fixed age at which a patent becomes immune from IPR. In Dabico Airport Solutions Inc. v. AXA Power ApS, IPR2025-00408, Paper 21 (June 18, 2025), the Acting Director denied institution on a patent "in force almost eight years" and explained: "Although there is no bright-line rule on when expectations become settled, in general, the longer the patent has been in force, the more settled expectations should be." Dabico added that "actual notice of a patent or of possible infringement is not necessary to create settled expectations." In October 2025, newly appointed Director John Squires took direct control of institution decisions.

The framework has since expanded to consider factors beyond patent age. In Magnolia Medical Technologies, Inc. v. Kurin, Inc., IPR2026-00097, Paper 17 (May 14, 2026) (precedential), Director Squires denied institution after a jury had already found the patent not invalid. The decision reasoned that "[t]he purpose of AIA reviews was, and is, to provide a quick and cost-effective alternative to district court patent litigation for resolving disputes over patent validity," and that the petitioner was "using the IPR process to obtain a second bite at the apple after losing in district court."

Most IPR petitions now fail at the threshold, and challengers have begun turning away from IPR toward ex parte reexamination, where a USPTO patent examiner, rather than the Board, reconsiders a patent over prior art without further participation by the challenger. Under the USPTO's PTAB trial statistics for fiscal year 2026 through August 31, the Director denied 435 of 762 petitions at the discretionary stage, or 57 percent, and referred 327 to the Board for a decision on the merits. Including petitions sent directly to the Board because the patent owner raised no discretionary arguments, 466 petitions reached a merits decision. Counting those discretionary denials together with 110 denials on the merits, 356 petitions were instituted and 545 denied, a 40 percent institution rate by petition, against 50 percent in fiscal year 2025. The Electrical/Computer technology group, which includes most AI patents, had the lowest institution rate of any group at 35 percent (201 of 570). Industry data show that IPR petitions "fell to an all-time quarterly low of 57 in Q2 2026," while ex parte reexamination requests reached an all-time quarterly high of 336. The September 10 article, Eligibility Has One Forum for Accused Infringers, discussed the limits of reexamination as a substitute.

The Federal Circuit Has Closed Every Route to Review Without Deciding Whether the Framework Is Lawful

In Apple Inc. v. Vidal, 63 F.4th 1 (Fed. Cir. 2023) (precedential), the U.S. Court of Appeals for the Federal Circuit held that Section 314(d) precludes APA challenges to the content of the Director's institution instructions to the Board about when to deny institution. Unlike an appeal from a single IPR decision, an APA suit asks a district court to set aside an agency policy itself. The court allowed a notice-and-comment claim, challenging the procedure used by the Office in adopting the instructions without public comment, to proceed. In In re Motorola Solutions, Inc., No. 2025-134 (Fed. Cir. Nov. 6, 2025) (precedential), the court described a request for APA relief on mandamus, a court order directing an official to act, available only where the right to relief is clear and indisputable, as "nothing but an attempted end run around § 314(d)'s bar on review."

In re Cambridge Industries USA Inc., No. 2026-101 (Fed. Cir. Dec. 9, 2025) (nonprecedential), was the first mandamus petition directed at a settled-expectations denial. The court denied relief with an express reservation: "We reiterate that we do not decide whether the PTO's actions are correct or whether the use of this [settled-expectations] factor is permitted under the statutes."

In Apple Inc. v. Squires, No. 2024-1864 (Fed. Cir. Feb. 13, 2026) (precedential), the court resolved the notice-and-comment claim left open in 2023. The Fintiv instructions, the Director's guidance on denying institution when a parallel trial is likely to conclude first, named for the decision in Apple Inc. v. Fintiv, Inc., fall "on the statement-of-general-policy side of the line" under 5 U.S.C. § 553(b), the APA's notice-and-comment provision, so notice and comment was not required, and "[t]he non-institution authority of the Director is statutorily insulated from judicial review, at least when, as relevant here, no constitutional claim is presented." Nonprecedential orders denying mandamus to Google (Jan. 27, 2026), Kahoot! (Feb. 25, 2026), and Tesla (Feb. 27, 2026) followed.

The Federal Circuit has therefore rejected every route to review presented to the court, while expressly stating that the court has not decided whether settled expectations is a lawful basis for denial.

Five Petitions Seek a Route to Judicial Review That the Government Says Congress Closed

Three petitions seek a writ of mandamus and ask whether a court may review an individual denial of institution. A fourth petition arrives from an APA suit that challenges the Director's institution rules themselves rather than any single decision. A fifth petition, filed by a patent owner, challenges the Board's handling of an IPR that was instituted.

In Google LLC v. VirtaMove Corp., No. 25-1230, the Office denied inter partes review of a patent that had "been in force for more than 14 years, creating strong settled expectations," even though no trial date had been set in a parallel district court case. Google asks "[w]hether the PTO lacks statutory authority to deny institution based on 'settled expectations' where the patent statutes allow for administrative review at any time during the life of a patent," and "[w]hether courts have power to review a PTO decision denying inter partes review on grounds that are contrary to statute." The petition was distributed for the Justices' June 25, 2026 conference, and the Court then requested a response. The Solicitor General filed a September 11, 2026 brief in opposition, recasting the question as "[w]hether the court of appeals abused its discretion in declining to grant a writ of mandamus," and arguing that Google has "various proper ways to obtain a ruling" on validity, "either in federal district court or in other administrative proceedings."

In Tesla, Inc. v. Granite Vehicle Ventures LLC, No. 26-136, the Office denied four petitions, challenging autonomous-driving patents, on the ground that a final written decision was unlikely before trial in the Eastern District of Texas. The Office denied institution even though Tesla had filed a broadened Sotera stipulation, a commitment not to pursue in court any ground raised or that reasonably could have been raised in the IPR. The case was then transferred to the Northern District of California, where trial is set for June 2028, well after an instituted IPR would have concluded. While Tesla's mandamus petition was pending, the USPTO granted Tesla's requests for ex parte reexamination of two of the challenged patents.

The government's two briefs make the same concession in nearly identical words. The September 25, 2026 brief in Tesla states: "To be sure, petitioner has no alternative means to obtain judicial review of the Director's decisions denying its petitions for inter partes review. But the absence of alternative review mechanisms with respect to that specific agency decision reflects a considered congressional judgment." The VirtaMove brief also contains the same two sentences, aimed at a single decision in that case. The government thus takes the position that district court litigation and reexamination supply sufficient routes for an accused infringer to obtain a validity ruling, and the lack of a route to judicial review of the IPR denial is by congressional design.

In Kahoot! AS v. Interstellar Inc., No. 26-198, Kahoot! filed its IPR petition on March 7, 2025, when U.S. Patent No. 10,339,825 had been issued for less than six years. The Office denied institution because the patent "has been in force for over six years," measured at the date of the decision, and despite a stay of the district court case during the pending IPR. As Professor Dennis Crouch observes, "the agency reached six only by measuring age at the date of its own decision."

The fourth challenger petition takes a different route. In Intel Corp. v. Squires, No. 26-73, Intel and Google seek review of Apple v. Squires, the APA suit challenging the institution instructions themselves. The petition asks "whether 35 U.S.C. § 314(d) . . . applies even when no institution decision is challenged to preclude review of PTO rules that set standards governing institution decisions." That is, Intel asks whether the bar on appealing institution decisions also bars a lawsuit challenging the Director's institution rules themselves, when no particular IPR denial is at issue. Seven amicus briefs support the Intel petition, including a joint brief from Samsung, Microsoft, TSMC, SAP, SAS, EchoStar, and Acushnet, and separate briefs from Tesla, Unified Patents, the PTAB Bar Association, Askeladden, the US*MADE coalition, and Teodor Holmberg.

The fifth petition comes from the opposite side of a patent dispute. In Federal Express Corp. v. Qualcomm Inc., No. 26-170, FedEx asks whether Section 314(d) bars review of "the Patent Office's refusal to determine whether a petition identifies all real parties in interest when § 312(a)(2) authorizes the Office to consider petitions 'only if' they identify 'all' such parties." Real parties in interest, the entities that fund or control a challenge, are required to be identified in an IPR petition. In a precedential decision on April 29, 2026, the Federal Circuit held a challenge to the sufficiency of a petition's identification of the real parties in interest was barred by Section 314(d), while vacating the Board's obviousness ruling as to certain claims and remanding. As Prof. Crouch notes, "Director Squires has reversed the underlying policy, but still argued that it was unreviewable on appeal." The petition was initially distributed for the September 28, 2026 conference. The Court then requested a response, and the Solicitor General sought and obtained an extension to October 28.

The petitions are not interchangeable. In Prof. Crouch's words, the four challenger petitions "occupy different cells of a grid." Two challenge the time-to-trial rule (Tesla and Intel) and two challenge settled expectations (VirtaMove and Kahoot!). Three arrive by mandamus from a denial of institution, and one, Intel, arrives from an APA suit that challenges no denial at all. The government's lead argument in the mandamus cases is the mandamus burden itself, which would allow the Court to deny VirtaMove, Tesla, and Kahoot! on the mandamus standard alone, without construing Section 314(d). The Intel petition carries no mandamus burden, arrives from a precedential opinion, and presents the Section 314(d) question directly.

As of today, none of the four challenger petitions has yet been distributed for an upcoming conference. Google filed its reply in VirtaMove today, which typically precedes redistribution. The federal responses are due October 14 in Intel and October 16 in Kahoot!. A grant of review this fall would likely produce a decision by the end of June 2027.

A Final Rule Would Move the Institution Bars From Policy Into Regulation

The USPTO's proposed rule, Revision to Rules of Practice Before the Patent Trial and Appeal Board, 90 Fed. Reg. 48335 (Oct. 17, 2025), would codify several institution bars in the Code of Federal Regulations at 37 C.F.R. § 42.108. Proposed § 42.108(d) would bar institution unless each petitioner stipulates that, if trial is instituted, the petitioner and its real parties in interest and privies (closely related parties) "will not raise grounds of invalidity or unpatentability with respect to the challenged patent under 35 U.S.C. 102 or 103 in any other proceeding." Other proposed provisions would bar institution where a court or the Board has already upheld a challenged claim, or where a parallel trial or decision is more likely than not to come first, and would prohibit the Board from waiving those requirements. According to the rulemaking notice, as summarized by practitioner commentary, the purpose is to "focus inter partes review proceedings on patent claims that have not previously been challenged in litigation or where prior litigation was resolved at an early stage."

The comment period closed December 2, 2025. The Office of Information and Regulatory Affairs (OIRA), the White House office that reviews significant agency rules before publication, received the final rule on July 22, 2026, and the review record for regulation identifier number (RIN) 0651-AD89 shows the review still pending. No final rule has yet been published.

The rule raises the stakes of the Intel petition. Apple v. Squires rejected the procedural challenge to the current instructions because the instructions were deemed to be a policy statement. A rule adopted through notice and comment would eliminate the procedural challenge altogether, leaving only a challenge to the content of the rule, and Apple v. Vidal holds content challenges to institution standards precluded by Section 314(d). The Intel petition therefore decides whether the codified bars can be tested in any court. If review becomes available, Loper Bright Enterprises v. Raimondo, 603 U.S. 369 (2024), which ended the practice of deferring to an agency's reading of an ambiguous statute, would require a court to apply Sections 314(a) and 316(a) to IPR institution challenges without deferring to the Office's interpretation. No court has yet applied Loper Bright to the IPR institution rules.

A codified rule would also outlast the officials who built the current framework. Acting Director Stewart, who issued the settled-expectations memorandum, announced her departure from the Office in late July 2026. IP legal commentators reported on September 16, citing unnamed sources, that Director Squires has told Board judges he will return institution decisions to the Board. The USPTO has not announced the change. Director Squires is scheduled to testify today at a Senate Judiciary Subcommittee on Intellectual Property oversight hearing.

Settled Expectations Sorts AI Patents by Age, Often Within a Single Case

The Office has not said how long a patent must be in force before settled expectations favors denial. Secondary sources describe a six-year benchmark, tied to the six-year limit on recovering past damages in 35 U.S.C. § 286, but no Director decision adopts a bright line. One practitioner review of 249 decisions issued between June and August 2025 found that patents in force six years or more were mostly denied, patents four years old or younger almost all proceeded to the merits, and five-year-old patents split. But age is not dispositive in either direction.

In Home Depot U.S.A., Inc. v. H2 Intellect LLC, IPR2025-00480, Paper 11 (Sept. 4, 2025), the Acting Director referred a petition to the Board despite a patent "in force for over twelve years," because the patent had not been "commercialized, asserted, marked, licensed, or otherwise applied" in the petitioner's technology space. In Apple Inc. v. Allani, IPR2025-00856, Paper 11 (Sept. 5, 2025), the Acting Director referred a petition on a patent "in force for thirteen years" because the patent owner waited eleven years after contacting Apple to assert the patent. Neither decision instituted review. Rather, each declined to deny the petition on discretionary grounds and sent the petition to the Board to decide whether the petition met the merits threshold for institution.

The 2026 AI docket shows how the doctrine sorts patents asserted against generative AI products, often in the same complaint:

  • Many Worlds 2T Innovations LLC v. OpenAI OpCo, LLC, No. 2:26-cv-00774 (E.D. Tex.), filed August 31, 2026 against ChatGPT features, asserts five patents claiming priority to a 2011 application: two issued in 2014, one in 2020, and two in 2025.
  • University of Tennessee Research Foundation v. Anthropic PBC, No. 1:26-cv-00887 (D. Del.), filed July 20, 2026 against Claude Code, asserts two patents issued in 2018, now about eight years in force.
  • Semantic Engines LLC v. Microsoft Corp., No. 2:26-cv-00339 (E.D. Tex.), filed April 23, 2026 against Copilot, asserts patents issued in 2012, 2015, and 2020. The complaint alleges that in September 2007 the inventor "reached out to Microsoft to discuss his invention," the kind of pre-suit contact the NVIDIA decision treated as actual notice.

Generative AI products are recent by patent standards, as most reached the market after late 2022. Many patents asserted against those products claim priority from the pre-transformer era, when machine-learning claims were often written in broad functional terms. Such older patents are the patents most likely to have claims that could be read broadly onto new generative AI products, the patents most exposed to an early motion challenging patent eligibility, which requires that a claim recite more than an abstract idea, and the patents settled expectations most protects from Board review. For the oldest AI patents being asserted, both the prior-art defense and the eligibility defense now tend to end up in district court. The same complaint frequently pairs those older patents with recent continuations, later-filed patent applications in the same family, which fall on the other side of the age line and remain candidates for IPR, or for the broader post-grant review, available only within nine months of grant and the only Office proceeding that can consider eligibility.

The Proposed Stipulation Is More Cost-Effective for AI Defendants, If Institution Is Available

The stipulation in proposed new regulation § 42.108(d), discussed above, would require every IPR petitioner, as a condition of institution, to give up anticipation and obviousness defenses against the challenged patent in any other proceeding. The stipulation would reach well beyond current law. Statutory estoppel under 35 U.S.C. § 315(e)(2) bars a petitioner, after a final written decision, from raising in court any ground it raised or reasonably could have raised in the IPR, which means patent and printed-publication prior art. The proposed stipulation, by contrast, would reach every anticipation and obviousness ground, including prior use, prior sale, and pre-existing product and system art that cannot be raised in an IPR.

Two features of AI patent litigation lower the cost of that stipulation for an AI defendant. First, the stipulation leaves eligibility under Section 101, and written description, enablement, and definiteness requirements under Section 112, untouched. Indeed, eligibility is the ground that most often decides AI patent cases. Second, much of the prior art in machine learning consists of conference papers and preprints, which qualify as printed publications when publicly accessible, so much of the art the stipulation would surrender is art an IPR can already hear.

The counterweight is evidence outside the printed-publication category. A defendant whose strongest invalidity case relies on open-source code, deployed systems, prior use, or sales would give up considerably more under the stipulation than a defendant whose case relies entirely on published papers. For an older AI patent, settled expectations and the proposed parallel-proceeding bars decide whether IPR is available before the proposed stipulation could even be a factor.

The One-Year Deadlines for 2026 AI Defendants Will Arrive Before Any Supreme Court Decision

Section 315(b) bars a petitioner from filing an IPR petition more than one year after the petitioner is served with a patent infringement complaint. AI patent suits filed in spring and summer 2026 will produce petition deadlines in spring and summer 2027, before or near the end of June 2027, the earliest realistic date for a Supreme Court decision on these petitions. Every 2026 AI defendant must therefore decide whether to petition under the current framework, or under the final rule if the rule issues first.

Delay carries separate costs. The Office treats a petitioner's delay after notice as a reason to deny, as the NVIDIA decision shows, and the time-to-trial analysis grows less favorable as a district court case advances. The Tesla case also shows the limits of the fallback. Reexamination can proceed without institution of an IPR, but reexamination cannot resolve eligibility and does not guarantee a stay of the district court case.

The Calendar, Not the Court, Determines the Next Move

For companies facing assertion of an AI patent:

  • Treat pre-suit contact as the start of the clock. Licensing correspondence and technical presentations can later establish the actual notice that weighs against institution, as the 2017 Neurala presentation did in the NVIDIA denials.
  • Build the district court validity case from the first day, including eligibility, and treat IPR institution as an addition to that case rather than the plan.
  • Evaluate and file any IPR petition early in the one-year window, with any stipulation strategy decided before filing. Waiting for the Supreme Court strengthens the delay factor against the petitioner.
  • Preserve the reviewability argument where the budget allows by requesting rehearing of a denial, so the issue is preserved if the Court opens a path to review.

For companies monitoring competitors' AI patents:

  • Challenge strategically important patents while the patents are young. Post-grant review during the first nine months after issuance reaches every ground, including eligibility, and IPR is most readily available in the first several years of a patent's term.

For AI patent owners:

  • Document commercialization, licensing, marking, and investment associated with commercially important patents. Those facts inform the settled-expectations analysis, and their absence in the petitioner's field supported referral in Home Depot.
  • Evaluate Board resilience and district court eligibility exposure separately. The age that protects an older AI patent at the Board often coincides with the functional claim language most exposed to a motion for a determination of lack of patent eligibility in court.

Looking Ahead

The Federal Circuit has held the Director's institution framework insulated from judicial review, while expressly stating that the court has not decided whether settled expectations is a permissible basis for denial. The Supreme Court now has four petitions from accused infringers challenging denials of institution or the rules behind them, and one from a patent owner arising from an instituted proceeding, arriving by different routes and on different grounds. The government's position is that no court may review a denial of institution, and that Congress intended that result. A final rule pending at OIRA would place the institution bars in the Code of Federal Regulations. Unlike the current guidance, which the Office can rescind by memorandum, as the Office did with the June 2022 memorandum in February 2025, a regulation would bind future Directors.

For companies involved in AI patent disputes, the patent's age, the litigation calendar, prior notice, and the available prior art now determine whether an accused infringer receives Board review, while eligibility and the other district court defenses operate on a separate track. The result is that a validity strategy must be built across forums from the first day of a dispute. The Board may be available, and the district court should be ready.