Deep|LLM: Computer Use - White-Collar Work’s "Spinning Jenny" Moment; How Many Jobs Will AI Replace?

A research note from FUNDA estimates that AI computer-use agents could automate work equivalent to roughly 180 million full-time jobs, about 18% of global white-collar employment, representing a long-term opportunity of around $1 trillion in additional annual revenue for AI providers. The analysis credits Astra's improved performance on benchmarks such as Agents' Last Exam and OSWorld 2.0, which test sustained multi-application desktop workflows, with bringing GUI-based automation closer to practical office adoption, though it expects task redistribution before headcount cuts.

Astra’s advances in complex desktop applications and extended workflows across multiple programs warrant a fresh assessment of how much white-collar work AI could automate. We believe computer use may be approaching a point where automation becomes practical for some office tasks, with repetitive, rules-based administrative and operational work likely to see the earliest adoption. This would allow AI providers to compete for a share of corporate labor spending as well as software budgets. We estimate a long-term opportunity of ~$1 trillion in additional annual revenue . This estimate reflects the market AI providers could realistically capture as revenue instead of treating the entire white-collar wage bill as potential revenue. The work computer use could automate is equivalent to ~180 million full-time jobs, or ~18% of global white-collar employment . This would mark the upper bound on job losses only if all saved hours translated into headcount reductions; we expect task redistribution to come first. More reliable execution, faster and cheaper models such as Jev-style architectures, and improvements in enterprise workflows and execution harnesses should broaden adoption and make deployment at scale more feasible. I. Astra improves at complex tasks across desktop applications 1. Strong performance on public benchmarks AI can interact with websites and applications in three main ways: 1 calling APIs to retrieve data or perform actions; 2 using structured information, such as the web document object model DOM or accessibility trees, to identify and manipulate interface elements; or 3 interpreting screenshots of a graphical user interface GUI and operating a mouse and keyboard as a person would. Protocols such as MCP can provide access to tools that support these methods. We focus on the third approach, in which the model observes the interface, takes an action, and verifies the result. GUI interaction works across websites, desktop applications, and workflows spanning several programs. Where usable APIs or structured interface data are unavailable, it can reduce the need for custom integrations and allow AI to complete more of a task within the user’s existing software environment. In an earlier report link https://fundaai.substack.com/p/deepllm-gpt-6-astra-opens-new-markets , we described Astra’s computer use capabilities as a potential “Claude 3.7 moment” for white-collar work more broadly , echoing the point at which Claude 3.7 made AI coding agents practical. Once models cross a capability threshold, users become more willing to delegate complete tasks, while improvements in supporting tools expand the range of viable applications. Astra’s benchmark results and public demonstrations show substantial progress in complex desktop software and multistep workflows. Relative to the previous generations, it offers a better balance of accuracy, execution time, and API cost, bringing computer use closer to everyday adoption. Among the officially released benchmarks, Agents’ Last Exam and OSWorld 2.0 are the most noteworthy. Agents’ Last Exam covers 13 industry clusters and more than 1,000 professional tasks drawn from real projects. Models must combine desktop applications, browsers, and the command line to produce verifiable deliverables. OSWorld 2.0 comprises 108 extended workflows across web and desktop applications, covering both everyday and professional tasks and requiring an average of more than 250 actions per task. Both benchmarks test sustained execution across applications. They more closely resemble actual white-collar work than short browser tasks and provide a more useful test of practical capability. OSWorld 2.0 offline results Agents’ Last Exam Source: OpenAI; compiled by FUNDA. Anthropic was the first major developer to publicly release computer use capabilities in 2024, enabling models to interpret screenshots and operate computers through simulated mouse and keyboard input. Other leading developers followed. Anthropic’s Fable 5 and Opus 5 models remain among the leaders in benchmark task completion, but take longer and cost more per task than comparable OpenAI models. In our view, this reflects differences in development priorities and resource allocation rather than a fundamental capability gap that Anthropic would struggle to close. “Astra wins on raw multi-app speed and visual clicking tasks, but Anthropic remains focused and highly integrated for complex software engineering. When it comes to computer use, I think OpenAI has a little edge, because OpenAI and Microsoft are like brothers and sisters … you always get the ecosystem benefits. … When deep technical coding, text logic, or reasoning comes into play, Anthropic is better. This reflects a difference in strategic priorities.” --A head of sales at a frontier AI lab Read full transcript at FUNDA expert platform https://funda.ai/zh/interviews/cmub38vns000o01hsaffmoalu “Anthropic focuses more on high-value enterprise applications in software engineering, finance, law, life sciences, and security. These customers primarily need to connect systems and tools through APIs and similar methods. Computing resources are limited, and investment needs to earn a return, so resources have to go where the economic and social value is greatest. In my view, computer use currently serves more consumer needs and relatively low-value, repetitive business tasks. That is a different emphasis from Anthropic’s.” --A cluster and infrastructure engineer at a frontier AI lab 2. Broader coverage of software and workflows Official demonstrations cover 3D modeling, video production, software testing, finance and office tasks, and engineering and scientific research. Community examples extend this range and show longer periods of sustained execution. Models can operate a growing number of applications and increasingly link actions across programs to complete a workflow. Creating a model in Blender and then building a scene in Unreal Engine, for example, requires a handoff between applications. Reconciling transactions in Xero and configuring CRM workflows bring these capabilities closer to routine business operations. However, public demonstrations are subject to substantial selection bias. Visually impressive 3D scenes attract attention, whereas finance, CRM, and internal enterprise workflows often involve sensitive information and produce results that are harder to showcase. The examples that gain the most attention on social media may therefore have little bearing on the size of the commercial opportunity. Frequent, repetitive tasks that require people to navigate business systems may offer greater economic value. Astra computer use: selected official and community examples Source: Compiled by FUNDA. 3. Hands-on testing on investment research tasks We tested Astra on a Mac with reasoning effort set to high, on a set of basic investment research tasks across websites and desktop applications. It completed every task accurately, with careful attention to detail and consistent adherence to formatting requirements. Astra investment research tests Mac, high reasoning effort Source: FUNDA. Sample outputs from Astra computer use tests Source: FUNDA. II. What drives computer use capabilities 1. The observe-act-verify loop The model breaks the goal into steps, establishes their order and dependencies, and tracks progress as it moves between pages and applications. It interprets the interface and selects an action for the harness to execute. It then uses fresh screenshots and environment feedback to verify the result and adjust its plan. When something goes wrong, it corrects the error or retries the action before continuing. The computer use execution loop Source: OpenAI; compiled by FUNDA. Reliable execution requires three distinct capabilities. - First, the model must understand what it sees. It needs to read text, map the user’s intent to the right location on the screen, and understand the relationships between buttons, windows, menus, and other controls. It must remain accurate when the layout or window size changes. - Second, it needs to understand the application and the business context. Finding the right button is only the starting point. The model must understand how a financial field is defined, what a design parameter means, and which conditions must be met before taking an action. It also needs to judge whether the result satisfies the task requirements. More specialized software and business rules demand deeper domain knowledge. - Third, it must plan and manage long sequences of interdependent actions. The model needs to track completed work, unresolved requirements, and earlier errors while handling pop-ups, delays, and exceptions. It must avoid losing sight of the goal or reporting success prematurely. Reliability over an extended workflow depends on continuous feedback and verification. 2. Data, training, and the harness We attribute Astra’s computer use capabilities to the combination of training data, staged training, and the execution harness. Data and training teach the model to interpret interfaces, plan actions, and recover from errors. The harness translates those actions into operations in real systems. Failures in live tasks then inform further training and engineering improvements. - Data: According to industry experts, GUI data is used throughout Astra’s training, with its share increasing in later stages. Screenshots and screen recordings first teach the model to recognize common controls and navigation patterns. Data from office applications, e-commerce, design, and other domains then builds its understanding of specific software and tasks. This data covers not only different applications but also the pages, states, and interface hierarchies within each one - Training: Expert demonstrations teach the model how to break down goals and sequence actions by linking screenshots, actions, and coordinates to subsequent states. The model also repeatedly executes tasks in sandbox environments, using reinforcement learning to improve its strategies based on task outcomes. Successful action sequences and demonstrations of error recovery provide further training material, strengthening its ability to sustain long workflows and correct mistakes. Media reports indicate that OpenAI has purchased tens of thousands of Mac mini and Mac Studio devices in recent months to build desktop environments. These allow models to practice tasks, collect interaction data, and refine their strategies. - Harness: The harness is the execution layer connecting the model to the operating system. It translates instructions such as clicks and text entry into system calls and returns feedback on the outcome. Frontier AI developers are improving this layer in two areas. The first is control: permission boundaries, safety checks, audit logs, and rollback mechanisms keep execution within defined limits. The second is coordination: the harness combines GUI actions with search, code, and software or database tools connected through MCP, allowing some operations to run in parallel or asynchronously. For example, code tools can organize Excel data and generate charts for a subsequent design task, replacing many individual clicks. 3. Computer use is harder to distill than text capabilities Distillation of computer use typically has a teacher model perform various tasks in software environments, recording the interfaces it observes, the actions it takes—such as clicking and typing—and the feedback it receives. These records are then filtered to select high-quality interaction trajectories, which are used to train a student model to choose its next action based on the task goal, current interface, and interaction history, so that it gradually learns to complete multi-step tasks. The challenge with this approach is that the student model must learn to act reliably in an environment that changes after every action. Each of the teacher’s actions fits a specific page and state. If the student deviates early in the workflow, it may encounter a different interface, making the remaining actions in the teacher’s sequence inapplicable. Long workflows amplify small errors. A mistake may cause failure only many steps later, making it difficult to trace the cause from the final outcome. Action sequences often contain substantial repetition but relatively few signals that explain success or failure. Behavior learned for one interface may also break when the software version, window layout, or controls change. Replicating reliable performance across complete tasks therefore requires extensive interactive training, examples of error recovery, and validation in real environments. The need to build and maintain those environments, collect high-quality action sequences, and obtain feedback from long workflows makes the full capability costly to reproduce. III. Current limitations of computer use 1. Unfamiliar applications and proprietary workflows remain difficult Public demonstrations and industry feedback suggest that models can now operate a wide range of desktop applications. They can use general knowledge of interfaces to adapt to different layouts, resolutions, and pop-ups, and sometimes operate software on which they have received no specific training. Reliability nevertheless declines with heavily customized interfaces, unfamiliar specialist software, and proprietary business rules. Accurate execution requires an understanding of what parameters mean in context, how actions constrain one another, and how to recover when something goes wrong. Progress may resemble the development of autonomous driving: corner cases encountered in practice feed into further data collection, rule refinement, and training. For computer use, these cases include unfamiliar interfaces and failures in real workflows. Model developers can identify specialized tasks their models cannot yet handle and collect or generate data for targeted training. They can also use failure cases from customer workflows to guide further training, but must first obtain customers’ explicit consent to use the relevant data and apply appropriate de-identification and data protection measures. “Astra continues to add interface and specialist software data throughout pretraining and post-training, and its coverage is still expanding. My rough estimate is that computer use can now operate more than 70% of the 100 most widely used enterprise applications. But being able to operate an application does not mean reliably completing every task in it. Only about one-third of those applications achieve end-to-end task success rates above 90%.” -- A senior forward-deployed engineer at a frontier AI lab Read full transcript at FUNDA expert platform https://funda.ai/zh/interviews/cmudkvzn800hh010ecapvwtei Applications and websites featured in selected Astra demonstrations Source: OpenAI; YouTube; Reddit; X; LinkedIn; compiled by FUNDA. 2. Desktop control may limit near-term adoption on Windows Computer use currently supports background execution on Mac. On Windows, however, it takes control of the active desktop’s mouse and keyboard, making it difficult for users to continue working on the same desktop while the AI runs a task. This reduces both convenience and the productivity gain. A separate virtual machine, remote desktop, or dedicated device can work around this, but adds deployment, maintenance, account management, and software licensing costs. Until reliable background execution becomes widely available, this constraint could slow adoption among Windows users. 3. Speed and cost still weigh on adoption, but faster models such as Jev may help - Many community users report that computer use is slow and costly on extended workflows, and our tests produced similar findings. With Astra’s reasoning effort set to high, investment research work that would take an experienced professional one hour cost approximately $30 to $60 in API charges. At these settings, the cost advantage over human labor is limited. BLS data put average annual pay for US financial analysts and advisors at approximately $130,000 in 2025, equivalent to roughly $50 an hour on the assumption of a 50-hour workweek. - These results depend partly on the test setup. We used high reasoning effort without optimizing prompts or the execution harness for individual tasks; enterprises could adjust both to reduce costs. Performance also varied by task. Email drafting, simple information compilation, and scheduling were substantially quicker and cheaper. On these administrative tasks, execution times were close to or faster than those of an experienced professional. Tasks with clear boundaries and relatively standardized processes therefore already appear suitable for adoption. Astra execution time and cost Mac, high reasoning effort Notes: 1. Each task was run twice; the figures show average execution time, token usage, and cost. 2. Human time estimates reflect how long an experienced investment research professional would take to complete the same tasks. Labor costs are based on BLS data showing average annual pay of approximately $130,000 for financial analysts and advisors in 2025, assuming a 50-hour workweek. 3. All prompts specified GUI interaction. The Excel task is excluded because it was better suited to automation without a GUI. Source: FUNDA analysis Read full table on FUNDA’s institutional platform https://funda.ai/zh/reports/computer-use-the-spinning-jenny-moment-for-white-collar-work--ccscj400hrpy . - Fast decision models such as Jev could reduce execution time and cost, making computer use economical for a wider range of tasks. In a two-tier architecture, a more capable model interprets the goal and business rules, breaks down the task, and handles exceptions, while a faster model handles high-frequency operations such as recognizing interface elements, checking the current state, and selecting the next action. The open-source browser-use/jev-ultrafast project applies a version of this approach to web tasks. Jev makes rapid decisions on frequent, well-defined actions such as clicking, scrolling, and selecting elements, while a language model generates content only when text entry is required. Tibo, the head of Codex, has publicly expressed support for Jev. Frontier AI developers are reportedly exploring this approach, with some already conducting internal tests. Payroll reconciliation: testing a finance workflow at a frontier AI lab Source: Compiled by FUNDA Read full report on FUNDA’s institutional platform https://funda.ai/zh/reports/computer-use-the-spinning-jenny-moment-for-white-collar-work--ccscj400hrpy . - We believe adoption would be easier if frontier AI developers offered a ready-to-use product that integrated task planning, model routing, state transfer, and error recovery. This would reduce the installation, configuration, and debugging work required of users. Our own experiment illustrates the challenge. We installed the relevant open-source project and let Codex dynamically divide web tasks between models, but execution actually slowed. The difficulty lies in both dividing responsibilities and coordinating handoffs. Gmail’s autosave, for example, changes the page state and repeatedly forces the lower-level executor to return control to the higher-level model, offsetting the time saved by faster decisions. A production harness needs to handle these state changes and recover from errors more effectively. - The commercial implications of lower token consumption depend on the pricing model. Under usage-based pricing, fewer tokens per task reduce revenue if task volumes and token prices remain unchanged. Under subscriptions, lower inference costs directly improve margins. Providers must weigh lower revenue per task against the potential for greater usage. As models become more reliable at completing entire tasks, pricing may also shift toward fees per task or per outcome. “Frontier AI developers are already testing similar two-tier architectures and want to release them once they are ready. But there are commercial considerations. Lower costs benefit subscription products. For API and enterprise customers charged by token usage, revenue could fall as well, so providers still have to weigh the trade-off.” --A senior forward-deployed engineer at a frontier AI lab Read full transcript at FUNDA expert platform https://funda.ai/zh/interviews/cmudkvzn800hh010ecapvwtei 4. Compliance and vendor responses could constrain adoption Agents that read screenshots, access enterprise accounts, and execute tasks across systems may encounter personal information, trade secrets, and other sensitive data. Authorization to operate a computer does not automatically authorize the model provider to transmit, store, or retain that data, or use it for training. Technical access also does not resolve questions about platform terms, software licenses, and data use. Enterprises and model providers need clear data-handling and operating permissions, access controls, human confirmation for high-stakes actions, and audit trails. Software vendors may struggle to block agents that run locally through an existing browser and an authorized user account. Commercial agreements may be needed to establish access conditions, permitted actions, and the allocation of liability. Where no agreement can be reached, disputes could escalate to litigation, raising compliance costs and restricting adoption in some use cases. IV. Sizing the computer use market 1. Where computer use adds value We see greater commercial potential in automating enterprise white-collar work than in everyday consumer applications. Businesses can measure returns through hours saved, shorter processing cycles, and higher throughput. Once execution is sufficiently reliable and deployment and review costs are manageable, these benefits can support recurring spending. Computer use enables AI to operate software and carry out complete tasks, extending its role beyond assisting people with individual steps. AI providers can therefore tap into labor budgets, not just software budgets. We assess this opportunity along two dimensions: the share of work AI could automate and, within that share, the proportion requiring computer use GUI interaction. Administrative operations and software development both have substantial automation potential, but through different means. Administrative workflows often span legacy systems with limited APIs, making computer use more important. Software development can largely be handled through code, terminals, and APIs. Research, analysis, and content creation have moderate automation potential, with some steps still requiring GUI interaction. Strategy, management, and complex communication depend more heavily on contextual judgment, interpersonal trust, and goal setting. These activities are the hardest to automate fully, so AI is more likely to assist people in the near term. Automation potential and reliance on computer use by task category Source: FUNDA analysis. “Instead of sizing the market based on traditional enterprise software or cloud subscription budgets, model providers are basing a multi-trillion-dollar TAM projection on the entire global knowledge-work labor market. So there’s going to be a fundamental shift in the TAM economics. Historically, enterprise software acted as a tool to make human work more productive. The computer-use agent represents a paradigm shift, where the software directly executes a job role end-to-end.” --A head of sales at a frontier AI lab see full transcript at FUNDA expert platform https://funda.ai/zh/interviews/cmub38vns000o01hsaffmoalu “Computer use can expand the market, but most of the additional opportunity lies in the long tail of software without good APIs. If a workflow already has a well-developed API, I would still use that first: it is faster, cheaper, and more reliable. Computer use connects legacy systems, desktop applications, and tools that can only be operated through their interfaces, enabling back-office and white-collar workflows that were previously difficult to automate. Repetitive administrative work is a good starting point because processes are structured, volumes are high, and returns are relatively clear. But the opportunity extends beyond low-value tasks to professional work in analysis, reporting, and finance.” --A director of AI systems at a frontier AI lab 2. Estimating the addressable market - Methodology: Our estimate of the total addressable market TAM starts with employment, average pay, and the share of working time spent on different tasks for each occupational group. We then apply assumptions about the share of work AI could automate and the proportion requiring computer use to estimate potential labor savings. Finally, we estimate how much of those savings AI providers could capture as revenue. Framework for estimating the computer use addressable market Source: FUNDA analysis - Key assumptions and supporting - White-collar employment and salary: Given the available data, we divide white-collar employment into four groups: managers; professionals in finance, law, education, architecture, life sciences, physical sciences, social sciences, and related fields; clerical support workers; and service and sales workers. We estimate employment and pay for each group, drawing primarily on International Labour Organization ILO statistics. Pay estimates use weighted annual earnings across more than 70 countries. Because the sample is concentrated in Europe and North America and overrepresents developed economies, we assume the global average is 70% of the sample average. This produces estimated annual pay of approximately $26,000 for the average white-collar worker worldwide, broadly consistent with industry estimates. - Time spent by occupation and task: To estimate how workers allocate their time, we primarily use the paper Estimating time spent on work tasks , aggregated to our occupational groups. The authors combine task-frequency data from the Occupational Information Network O NET , AI estimates of task duration, and validation through human surveys to estimate time allocations across 876 occupations and more than 17,000 tasks. We cross-check and refine these estimates using the distribution of skills by occupation in McKinsey’s Agents, robots, and us: Skill partnerships in the age of AI , alongside surveys from Microsoft, Salesforce, Fyxer, ActivTrak, and relevant industry organizations. Source: Estimating time spent on work tasks ; McKinsey; FUNDA analysis Read full table on FUNDA’s institutional platform https://funda.ai/zh/reports/computer-use-the-spinning-jenny-moment-for-white-collar-work--ccscj400hrpy . - AI automation potential: For each task category, we estimate the share AI could automate based on how standardized the work is, its reliance on complex judgment and interpersonal interaction, and the reliability and economics of AI execution. Technical feasibility does not guarantee adoption. Deployment, process redesign, and the pace of uptake may leave technically automatable work in human hands during the forecast period, and our assumptions allow for these constraints. Internet adoption offers a useful comparison: the International Telecommunication Union ITU estimates that approximately 74% of the global population used the internet in 2025. Even after decades of adoption, cost, infrastructure, skills, and demand continue to limit its reach. - Computer use GUI reliance: Within the work AI could automate, we estimate the proportion that would primarily require computer use, such as clicking and typing through a GUI. These scenario assumptions reflect the software environment, API availability, native AI features and plug-ins, and how typical workflows are carried out. - Value captured by AI vendors: Only part of the resulting savings would become revenue for AI model and application providers. Enterprise customers would retain some of the gains or pass them on through lower prices, while integration, process redesign, and ongoing operations would absorb another portion. We use pricing examples from enterprise data platforms and cloud cost optimization services to inform our estimate of the provider share. Some cloud cost optimization providers charge a share of savings directly.Source: FUNDA analysis Read full table on FUNDA’s institutional platform https://funda.ai/zh/reports/computer-use-the-spinning-jenny-moment-for-white-collar-work--ccscj400hrpy - Long-term revenue potential and labor impact - Under these assumptions, computer use in white-collar work could generate approximately $1 trillion in additional annual revenue for AI providers over the long term . Enterprise customers would retain a portion of the cost savings and productivity gains. We regard the assumptions as reasonable rather than aggressive and believe this revenue opportunity is achievable over time. Note: Weighted by working time and average pay for each occupational group. Source: FUNDA analysis Read full report on FUNDA’s institutional platform https://funda.ai/zh/reports/computer-use-the-spinning-jenny-moment-for-white-collar-work--ccscj400hrpy - Measured in working hours, the white-collar labor that AI computer use could automate is equivalent to approximately 180 million full-time jobs, or about 18% of global white-collar employment. If all saved hours translated into headcount reductions, this would be the upper bound on job losses. For context, OpenAI and Anthropic in its most severe scenario put the share of jobs at risk of AI displacement at up to ~20%. This does not imply that jobs will be cut in the same proportion. We expect the initial impact to be a redistribution of tasks: AI performs routine operations, while people review outputs, handle exceptions, and coordinate work. Staffing needs may then decline as companies redesign workflows and consolidate the remaining tasks. The eventual effect on employment will depend on new demand, the creation of new roles, and the pace of adoption. Note: Weighted by working time for each occupational group. Source: FUNDA analysis Read full report on FUNDA’s institutional platform https://funda.ai/zh/reports/computer-use-the-spinning-jenny-moment-for-white-collar-work--ccscj400hrpy OpenAI estimates of the share of US jobs at risk of AI displacement Source: OpenAI. Anthropic scenarios for the labor market Source: Anthropic. As AI becomes more capable of operating complex software and completing workflows across applications, it is moving beyond assistance toward direct task execution, with the potential to displace some white-collar work. The spinning jenny and the machinery that followed offer a historical parallel. They reduced the labor required per unit of output, while lower costs supported expansion in weaving, dyeing, and finishing and helped shift production from households to factories. New divisions of labor emerged as old jobs disappeared. Yet workers who lost their livelihoods did not automatically move into the new roles, and the gains from higher productivity were unevenly distributed. Technological displacement, market expansion, and the costs of adjustment can coexist; industry growth does not ensure that every worker benefits. What new demand and jobs this wave will create, and where displaced white-collar workers will fit, will become clear only as adoption and social adjustment unfold. Disclaimer and Terms of Use Analyst Independence Statement The views expressed in this report accurately reflect the personal and independent views of the analyst s regarding the securities and issuers discussed herein. No part of the analyst’s compensation was, is, or will be directly or indirectly related to the specific recommendations or views expressed in this report. The Company adheres to strict compliance procedures to ensure the integrity, security, and legality of the data sources and information utilized. About FUNDA Consensus AI Technology Pte. Ltd. the “Company” or “FUNDA” is an independent research firm incorporated in Singapore. The Company is not affiliated with any securities dealer, investment bank, or asset management firm, nor is it registered as an investment advisor or broker-dealer in any jurisdiction. The Company does not have any investment banking or underwriting relationship with the companies covered in this report, and does not accept sponsorship or funding from publicly listed companies. No Investment Advice and Report Neutrality The contents of this report and the information contained herein are for informational purposes only and do not constitute investment advice or a basis for business decisions. The views expressed herein represent solely the personal opinions of the respective authors at a specific point in time and do not reflect the position of FUNDA. As an independently operated platform, FUNDA maintains a neutral third-party status and makes no express or implied warranties regarding the accuracy, completeness, or applicability of such views. No Offer or Solicitation: This report does not constitute an offer, solicitation, recommendation, or advice to buy, sell, or hold any securities or financial instruments, nor does it provide any legal, tax, or accounting advice. Third-Party Data: All third-party data cited in this report is derived from public or authorized sources. The Company makes no representations or warranties, express or implied, as to the accuracy or completeness thereof. Non-Participation in Decision-Making: The Company does not participate, directly or indirectly, in the investment decision-making process of its clients or readers. Readers should make independent investment decisions based on their own investment objectives and financial circumstances, and consult professional financial or legal advisors prior to executing any transactions. Timeliness Statement Given the rapid changes in market environments and technological trends, this report reflects market conditions as of the date prior to its publication. Relevant information may become outdated over time. FUNDA has no obligation to update the contents of this report in real-time. Intellectual Property, Confidentiality, and Restrictions on Commercial Use All reports, data, models, and content provided by the Company constitute its exclusive proprietary intellectual property and are protected by copyright and intellectual property laws. Ownership of Rights: Unless otherwise expressly stipulated in a written agreement, all intellectual property rights including copyrights arising from or relating to this report are exclusively owned by the Company. Authorized Use: Institutional clients may use this report internally within the scope permitted by their internal compliance policies; however, disclosure or sharing with any third party is strictly prohibited. Prohibition on Commercial Use: Without the prior express written consent of the Company, no individual or entity may reproduce, republish, distribute, resell, repackage, or otherwise use this report or any part thereof for any commercial or profit-making purpose. Legal Recourse: The Company reserves the right to pursue all available legal remedies and claims against any unauthorized commercial use, redistribution, or infringement of its intellectual property rights. Forward-Looking Statements and Disclaimer This report may contain forward-looking statements, including forecasts, estimates, and analyses of industry trends. Such statements are subject to risks and uncertainties that may cause actual results to differ materially from those anticipated. The Company shall not be liable for any direct, indirect, or consequential losses arising from the use of or reliance on the contents of this report.