Full Opinion

Affirmed and Memorandum Opinion filed September 15, 2026 In The Fifteenth Court of Appeals NO. 15-25-00001-CV CREATEAI HOLDINGS, INC. F/K/A TUSIMPLE HOLDINGS, INC., Appellant V. BOT AUTO TX INC., Appellee On Appeal from the Business Court Division 11A Harris County, Texas Trial Court Cause No. 24-BC11A-0007 MEMORANDUM OPINION This trade secrets cases involves two technology startups involved in autonomous trucking. Appellant CreateAI1 was formerly an autonomous vehicle company that has since pivoted to AI-enhanced entertainment. Appellee Bot Auto is still actively engaged in the development of autonomous trucking technology, but 1 Formerly known as TuSimple Holdings, Inc., the petitioner company changed its name to CreateAI on December 11, 2024. We use the current name for ease of reference. stands accused of misappropriating CreateAI’s trade secrets in furtherance of that project. CreateAI sought a pretrial temporary injunction to restrain Bot Auto from transferring or disseminating key components of the contested technology. The trial court denied the motion for temporary injunction, and CreateAI filed this interlocutory appeal. We affirm. BACKGROUND At its founding in 2015, CreateAI focused on autonomous trucking, specifically to develop 18-wheelers capable of fully driverless “hub-to-hub” transport.2 It successfully completed an autonomous hub-to-hub trip in just two years, and achieved a full “driver-out” hub-to-hub trip in late 2021. But trouble lay ahead. Less than a year after its driver-out milestone, during a period of significant internal turmoil and external challenges the company fired its co-founder and CEO, Dr. Xiaodi Hou. Dr. Hou remained on the board of directors for several months, but severed all ties with the company in March of 2023. Within the month, Dr. Hou started Bot Auto by launching an initial round of fundraising to pursue commercially viable autonomous trucking. By the summer of 2023, Bot Auto had accomplished a fully-autonomous hub-to-hub trip—though not a driver-out trip—in approximately fifteen months. Just as Bot Auto was revving up, CreateAI was seeking an off-ramp for its American autonomous trucking operations. In June of 2023, the company publicly disclosed that it was considering selling the U.S. portion of its business. 3 Unable to 2 “Hub-to-hub” transport broadly refers to transportation between two logistics hubs, and in this context implies that a human driver is present only as a failsafe. “Driver-out” hub-to-hub transport refers to an autonomous truck travelling from a distribution center, onto a highway, and to another distribution center, all without a human driver present. 3 TuSimple Holdings Inc., Current Report (Form 8-K) (June 27, 2023). 2 find a buyer, it instead closed down its U.S. operations and turned its attention to the Asia-Pacific region, specifically its Chinese and Australian subsidiaries. In December of 2024, it redirected its business to AI-powered digital entertainment crafted primarily for Asian markets.4 But shortly before this pivot, the company sued Bot Auto under the Texas Uniform Trade Secrets Act (“the Act”). 5 Its petition alleged misappropriation of proprietary autonomous vehicle technologies and sought damages, fees, costs, and temporary and permanent injunctive relief. The very next day, the parties agreed and the trial court granted a temporary restraining order designed to maintain the status quo until a temporary injunction hearing could be held. That agreement and order prevented Bot Auto from transferring or disseminating any contested technology, though not from “working in the ordinary course of business with its suppliers and contractors under customary non-disclosure agreements.” The trial court renewed its order on October 29. Following a discovery period and numerous discovery disputes, both parties presented evidence and witnesses at a temporary injunction hearing held on November 18–19, 2024. After the hearing, the trial court denied the temporary injunction and dissolved the TRO. CreateAI immediately filed a notice of accelerated interlocutory appeal. CreateAI asserts three claims of misappropriation. First, theft of its proprietary sensor suite—the collection of cameras, radar, and lidar 6 that act as the “eyes and ears” of a self-driving vehicle so it can “perceive” its surroundings. Second, theft of aspects of its proprietary decision-making technology—the “brain” 4 Evelyn Cheng, Chinese self-driving trucking company pivots to generative AI for video games, CNBC (Dec. 18, 2024, 7:00 PM), https://www.cnbc.com/2024/12/19/china-self-driving- truck-company-tusimple-pivots-to-genai-for-games.html; see also CREATEAI, https://iamcreate.ai /en-US/animation and https://iamcreate.ai/en-US/games (last visited Sept. 10, 2026). 5 See generally TEX. CIV. PRAC. & REM. CODE ch. 134A. 6 Lidar operates essentially like radar but uses light instead of radio waves. 3 of a self-driving vehicle—specifically the terms used to label data fed into this autonomous decision process. Third, misappropriation of its proprietary safety technology, in particular a dual “nerve system” for auxiliary braking and the technical parameters of its steering system. STANDARD OF REVIEW “A temporary injunction is an extraordinary remedy and does not issue as a matter of right.”7 To obtain a temporary injunction, the party seeking it “must show: (1) a cause of action against the party to be enjoined; (2) a probable right to recover on that claim after a trial on the merits; and (3) a probable, imminent, and irreparable injury absent the temporary injunction.” 8 We review a trial court’s decision to deny a temporary injunction for abuse of discretion.9 Though trial courts have no discretion to “incorrectly analyze or apply the law,” on evidentiary matters we will affirm whenever “some evidence reasonably supports the court’s ruling.”10 The text of the Act defines misappropriation to include both improper acquisition and improper disclosure or use of trade secrets.11 It encompasses a broad array of information that has been kept secret and has independent value as such.12 7 Abbott v. Anti-Defamation League Austin, Sw., & Texoma Regions, 610 S.W.3d 911, 916 (Tex. 2020) (quoting Walling v. Metcalfe, 863 S.W.2d 56, 57 (Tex. 1993) (per curiam)). 8 Harley Channelview Properties, LLC v. Harley Marine Gulf, LLC, 690 S.W.3d 32, 37 (Tex. 2024) (citing Butnaru v. Ford Motor Co., 84 S.W.3d 198, 204 (Tex. 2002)). 9 State v. Hollins, 620 S.W.3d 400, 405 (Tex. 2020). 10 Abbott, 610 S.W.3d at 916. 11 TEX. CIV. PRAC. & REM. CODE § 134A.002(3). 12 Id. § 134A.002(6) (“‘Trade secret’ means all forms and types of information … if: (A) the owner of the trade secret has taken reasonable measures under the circumstances to keep the information secret; and (B) the information derives independent economic value, actual or potential, from not being generally known to, and not being readily ascertainable through proper means by, another person who can obtain economic value from the disclosure or use of the information.”). 4 CreateAI alleges Bot Auto actually used its trade secrets, not the more-difficult-to- prove allegation that Bot Auto merely possesses them improperly. Thus, for each alleged trade secret we must determine whether the record supports the trial court’s determination that CreateAI likely cannot show misappropriation of them. That analysis is inherently more evidentiary than legal. DISCUSSION On each of the three contested areas of autonomous vehicle technology, we hold that some evidence reasonably supports the trial court’s decision not to impose a temporary injunction. First, CreateAI’s sensor array is public and therefore not a trade secret, and Bot Auto’s sensor array is too distinct from CreateAI’s to warrant a misappropriation claim. Second, the decisional processes of each company’s offerings are so markedly different as to eschew comparison. Third, CreateAI’s claims of misappropriated safety technology fail both by overbreadth and by the details in evidence—the mere existence of a redundant braking system is not a protectable trade secret, and the technical parameters of CreateAI’s steering system do not match those of Bot Auto’s. Further, we find that CreateAI is unlikely to suffer imminent injury that cannot be compensated by money damages because it no longer maintains U.S. autonomous trucking operations. I. Sensor Array A trade secret can only exist if its owner “has taken reasonable measures under the circumstances to keep the information secret” and if the information is not “generally known” or “readily ascertainable” by others who may profit from it.13 CreateAI’s sensor array—like other self-driving vehicles presently in existence—is generally not secret since each sensor is housed on the outside of the truck in plain 13 TEX. CIV. PRAC. & REM. CODE § 134A.002(6)) (emphasis added). 5 sight. This may be unavoidable because a sensor may be likely to “sense” much less if entirely shielded from external view. To be sure, a layperson observing an autonomous 18-wheeler on a public highway would likely be unable to identify the various sensors in its array. But CreateAI publicized the exact locations and varieties of its sensors in an investor presentation in 2022, complete with color-coded highlighting and visualizations of how the sensors operate in tandem. In short, an external sensor array on a self-driving vehicle may be no more secret than side-view mirrors of a traditional vehicle. Additionally, record evidence shows that several respects of Bot Auto’s sensor array are markedly different from CreateAI’s. Bot Auto uses high resolution lidar sensors that were unavailable during CreateAI’s developmental stage. Bot Auto also positions its lidar sensors high off the ground—on the roof of its trucks—while CreateAI positions its lidar sensors at the height of the engine hood, about four and a half feet off the ground. These differences in positioning facilitate differences in function: Bot Auto’s higher placement allows lidar to be used for long-range object detection, while CreateAI’s lower placement optimizes lidar’s mid-range peripheral perception and leaves long-range perception to cameras. More broadly, Bot Auto uses different makes and models of sensors than does CreateAI, mounts them in different locations on different brands of 18-wheelers, and processes differently- formatted sensor data through a different type of neural network.14 In sum, even if sensor positioning were a protectable trade secret, the record reflects no misappropriation by Bot Auto. II. Decision-making Technology CreateAI claims that Bot Auto misappropriated its “semantic regime for the 14 See infra at 8–10 (discussing the differences between CreateAI’s convolutional neural networks and Bot Auto’s transformer neural network). 6 annotation of road trip data” and its “annotation process design”—in plain terms, accusing Bot Auto of copying its labels and categories for organizing data to “teach” an autonomous vehicle how to drive safely. The record before us supports the trial court’s rejection of that allegation at this point in the litigation. CreateAI offers, and the record contains, only one concrete example of possible semantic misappropriation: Bot Auto classifies possible scenarios as either “green cases” (in which the vehicle could perform a liability-free maneuver based on generally applicable rules) or “red cases” (in which there are no liability-free options and the vehicle’s maneuvers must be dictated by situation-specific rules). Those color-coded terms were developed by CreateAI. But this sole example proves very little of CreateAI’s claim. An identical color-coding scheme for the broadest level of data categorization is a far cry from the wholesale theft of an entire semantics regime that CreateAI alleges. CreateAI’s own exhibits demonstrate that its semantic regime is far more comprehensive and detailed than a single “red versus green” distinction, covering contingencies ranging from “heavy traffic” and “potholes” to “small moving animals” and “emergency landing plane.” CreateAI’s own CEO refused to assert that the red/green dichotomy is itself a trade secret, merely alleging that color scheme is an “indication of the series of logical flow[s] of how to solve or think about” the problems inherent to self-driving vehicles. By contrast, the record contains evidence that CreateAI’s semantics regime and data annotation conventions are of no use to Bot Auto. Apart from the probabilistic nature of the conclusory “red/green” terms,15 there are fundamental 15 See CAE Integrated, L.L.C. v. Moov Techs., Inc., 44 F.4th 257, 263 (5th Cir. 2022) (“CAE contends that Moov could never have succeeded without CAE’s data, claiming that ‘the ‘use’ of this data can reasonably be inferred from Moov’s results.’ This inference is insufficient to support a finding that Moov used CAE’s trade secrets.”) (emphasis in original). 7 distinctions between the parties’ technological foundations. CreateAI’s decision- making technology is (or was when it was still operating) based on “convolutional neural networks.”16 The term “convolutional” references the specific mathematical operation that this type of neural network performs, but might also serve as a vernacular term for the convoluted process by which these networks are trained.17 Training a convolutional neural network requires vast quantities of visual data, and each image from that data must be manually annotated before it can be input to the network—meaning that human workers must collect images from thousands of miles of road data obtained by test vehicles, draw boxes around any potential hazards (such as persons or cars), and label each hazard according to its type. This process is expensive and time-consuming; it also must be repeated for every form of data that an autonomous vehicle is expected to recognize, since convolutional neural networks must be individually trained to parse a single type of data. CreateAI’s decisional technology required at least five distinct convolutional neural networks, one apiece for camera data, lidar data, “fusion” between camera and lidar data, object tracking, and prediction. Given this immense reliance on and investment in the labor- intensive data annotation process, it is no wonder that CreateAI is fiercely protective of its proprietary conventions and deeply suspicious of any company appearing to 16 The record seems to support that CreateAI exclusively—or almost entirely—relied on convolutional neural networks up until 2023, at which point it began to steer away from autonomous trucking entirely. Dr. Hou testified that only convolutional networks were in use on its trucks for the entirety of his time at the company, which ended in March of 2023. CreateAI’s current CEO testified that CreateAI began developing transformer technology in early 2022, and that he supervised a “migration” of that technology from CreateAI’s Chinese subsidiary to its U.S. operations sometime in “early 2023.” It began seeking a buyer for its U.S. operations in June of 2023. See TuSimple Holdings Inc., Current Report (Form 8-K) (June 27, 2023). 17 See Convolution, Wolfram MathWorld, https://mathworld.wolfram.com/Convolution.html (last visited Sept. 10, 2026) (“A convolution is an integral that expresses the amount of overlap of one function 𝑔𝑔 as it is shifted over another function 𝑓𝑓. It therefore ‘blends’ one function with another.”); see also Convoluted, MERRIAM-WEBSTER’S COLLEGIATE DICTIONARY (11th ed. 2003). 8 rapidly gain ground by drafting in CreateAI’s slipstream. But Bot Auto’s decisional technology employs a “transformer neural network” that has no need to engage in large-scale manual data annotation and no use for data conventions developed by CreateAI. Transformer networks rely on different mathematics than do convolutional networks, using tensors or matrices to process several large sequences of data at once instead of “using convolutions to process smaller subsets of input data one piece at a time.”18 In consequence, transformer neural networks are better able to discern more attenuated correlations (such as between pixels that are not directly adjacent) and can be trained “on unprecedentedly massive datasets through self-supervised learning.” 19 This improved correlative reasoning and facility with larger datasets, in turn, allows transformer networks to be “pretrained”—the entire point of which is to minimize 18 Cole Stryker & Dave Bergmann, What is a transformer model?, IBM THINK, https://www. ibm.com/think/topics/transformer-model (last visited Sept. 10, 2026); see also id. (“Transformer models such as relational databases generate query, key, and value vectors for each part of a data sequence, and use them to compute attention weights through a series of matrix multiplications) (emphasis in original); see also Ziad Salloum, Behind the Magic: How Tensors Drive Transformers, TOWARDS DATA SCIENCE (Apr. 25, 2025) https://towardsdatascience.com/behind- the-magic-how-tensors-drive-transformers/ (“Transformers have changed the way artificial intelligence works, especially in understanding language and learning from data. At the core of these models are tensors (a generalized type of mathematical matrices that help process information).”); see also Testimony of Dr. Xiaodi Hou (discussing the mathematics undergirding transformer networks). 19 Stryker & Bergmann, supra note 18. Self-supervised learning means that instead of “relying on labeled datasets for supervisory signals,” the neural network will “generate implicit labels from unstructured data”—eliminating much of the slow, expensive manual annotation required for CreateAI’s model. See Dave Bergmann, What is self-supervised learning?, IBM THINK, https://www.ibm.com/think/topics/self -supervised-learning (last visited Sept. 10, 2026). The record is not explicit that Bot Auto is using self-training. But it is explicit that Bot Auto’s training program markedly differs from CreateAI’s and requires significantly less manual data annotation. See Testimony of Dr. Xiaodi Hou (noting Bot Auto’s dramatically reduced reliance on human annotation). 9 the labor required to train a network.20 While Bot Auto’s network still requires some human annotation, it requires nowhere near as much as CreateAI’s. And because transformer networks can simultaneously process different varieties of sensor data, Bot Auto needs to develop only a single network, not five separate ones. Moreover, CreateAI’s sensors are configured differently than Bot Auto’s, and the resulting difference in data formats means it is not at all clear that Bot Auto could derive any benefit from CreateAI’s data annotation techniques. At a minimum, the trial court had discretion to rely on Dr. Hou’s testimony that “anything pertaining to convolutional neural network is intrinsically incompatible to the technology that we use today for transformer neural networks.” III. Safety Technology CreateAI alleges that Bot Auto misappropriated two elements of its safety technology: (1) a dual-CAN bus system to create a redundant braking method, and (2) key technical parameters of its steering system. The trial court did not abuse its discretion by deciding otherwise. Redundancy itself is not a trade secret; it is merely a general principle derived from the fact that no one wants robot trucks with faulty brakes. CreateAI alleges that it and Bot Auto both achieve redundancy with a “dual-CAN bus system” that transfers signals between parts of the automobile, like a nervous system for the vehicle. In the especially important context of braking, CreateAI and Bot Auto deploy two of these systems, so that if the primary CAN bus malfunctions, the auxiliary CAN bus will still transmit the signal “telling” the brakes to engage. Such a system is one of several ways to achieve redundancy, but that CreateAI and Bot Auto both employed the same method does not inherently mean that Bot 20 See Cole Stryker, What is a pretrained model?, IBM THINK, https://www.ibm.com/think/ topics/transformer-model (last visited Sept. 10, 2026); see also Testimony of Dr. Xiaodi Hou (discussing the use and usefulness of the pretraining methodology). 10 Auto misappropriated trade secrets. CreateAI itself publicized its use of a CAN bus system in an investor presentation in 2022. Merely deploying a second one of these systems to act as a failsafe is not evidence of misappropriation. More generally, the record contains testimony that Bot Auto uses a “different brand of … braking components” with “completely different” interfacing protocols and failure modes than those used by CreateAI. As to its steering system, CreateAI asserts that Bot Auto’s technical parameters for the constituent elements installed on its vehicles are identical to CreateAI’s. But the steering requirements of the two companies are not identical. CreateAI’s maximum steering torque is 5,000 Newton-meters. Bot Auto has two options for its torque output: one greater than 20 Newton-meters, the other greater than 5,000 Newton-meters. CreateAI’s requirements include a maximum steering position error of plus or minus three degrees; Bot Auto’s acceptable range is within one degree. CreateAI specifies a response latency (the delay between issuing and implementing a steering command) of under 20 milliseconds; Bot Auto specifies a latency under 50 milliseconds. Even if identical requirements were sufficient to prove the misappropriation of trade secrets, they simply describe the standard to which a product is built, not the means by which a product reaches that standard. The trial court had discretion to reject this allegation of misappropriation. IV. Irreparable Injury Finally, it does not appear from the record that CreateAI will suffer any probable, imminent, and irreparable injury in the absence of a temporary injunction. First, it is not clear that CreateAI remains engaged in the autonomous trucking business at all, given its recent transformation into an AI-powered entertainment 11 company. CreateAI now makes animated films and video games, 21 and its U.S. entity responsible for autonomous trucking has only retained, in the words of its CEO, “roughly ten” employees. In consequence, it is not clear that CreateAI could suffer any injury at all from Bot Auto’s autonomous trucking endeavors, even if those endeavors were fueled by misappropriation, because the evidence on record strongly suggests that CreateAI’s business is AI entertainment, not autonomous trucking. Second, an injury is only irreparable “if the injured party cannot be adequately compensated in damages or if the damages cannot be measured by any certain pecuniary standard.”22 To the extent that CreateAI remains engaged in autonomous trucking at all, its engagement appears to be limited to efforts to license its technology to companies still active in that sector. The injury from lost licensing agreements is lost royalties, an injury that can be calculated and compensated with damages. 23 In consequence, even if Bot Auto misappropriated any of the contested technology, and even if any such misappropriation harmed CreateAI, the proper remedy is compensatory damages and not injunctive relief. 21 See CREATEAI, https://iamcreate.ai/ (last visited Sept. 10, 2026) (“We produce high-end video games that transform well-known intellectual property into immersive and interactive experiences.… We produce animated films and series that reintroduce impactful intellectual property to a global audience.”). 22 Harley Channelview Properties, LLC v. Harley Marine Gulf, LLC, 690 S.W.3d 32, 37 (Tex. 2024). 23 See Sw. Energy Prod. Co. v. Berry-Helfand, 491 S.W.3d 699, 711 (Tex. 2016) (cleaned up) (“Absent proof of a specific injury, the plaintiff can seek damages measured by a reasonable royalty. Because the precise value of a trade secret may be difficult to determine, the proper measure is to calculate what the parties would have agreed to as a fair price for licensing the defendant to put the trade secret to the use the defendant intended at the time the misappropriation took place. The royalty is calculated based on a fictional negotiation of what a willing licensor and licensee would have settled on as the value of the trade secret at the beginning of the infringement. A reasonable royalty is, in essence, a proxy for the value of what the defendant appropriated, but it is not simply a percentage of the defendant’s actual profits.”). 12 CONCLUSION The potential risks and rewards involved in successfully implementing self- driven 18-wheelers are colossal. The same is true of the risks for drivers and the difficulties of litigation resulting from trucking accidents. For example, a recent Texas Supreme Court opinion set aside a $16.8 million verdict due to numerous trial errors and remanded the case for a second trial about an accident that occurred 10 years earlier.24 If new technologies can reduce the incidence of such disasters, it should be pursued with all reasonable diligence for the good of the drivers, families, companies, and courts that currently must deal with the consequences. But it cannot do so at the cost of complaints by those who may claim that advances in implementation were achieved by using someone else’s proprietary technology. Balancing and resolving such disputes is a job for the courts. Recognizing that this case is before us solely on a preliminary record, we hold that the trial court did not abuse its discretion in finding that CreateAI did not show a probable right of recovery or a probable, imminent, and irreparable injury warranting a temporary injunction. Accordingly, we affirm. /s/ Scott A. Brister Scott A. Brister Chief Justice Before Chief Justice Brister and Justices Field and Farris. 24 Gregory v. Chohan, 670 S.W.3d 546, 546, 551–52 (Tex. 2023). 13