During its 2026 autumn strategy conference held from September 2 to 3, Huatai Securities Co., Ltd. outlined its core view that the AI industry remains in an early development stage. If this AI industrial revolution were compared to a baseball game, the current phase would roughly correspond to the second inning of a nine-inning match, though the industry's focus is gradually shifting from early-stage computing power investment toward the question of how to monetize AI. The forum connected key segments including data centers, autonomous driving, and embodied intelligence, featuring Liu Yang, Chairman and CIO of West Result Investments, discussing AI-era investment logic; Wu Gansha, CEO of UISEE, exploring autonomous driving commercialization; and Yang Ruda, Co-founder of Poke Robotics, presenting advancements in embodied intelligence. Huang Mincong, Senior Partner at Venture Hub, concluded the session by sharing perspectives on investing in and operating Southeast Asian data centers, examining how AI computing power generates value from an infrastructure standpoint.

One key takeaway is that token pricing may come to dominate the future AI investment framework. Liu Yang shared how the investment framework is evolving in the AI era, drawing on her new book titled "Silicon-Based — The Jerusalem of the Computing Power Era." As computing power and data become core corporate assets, the applicability of traditional PE and DCF valuation methods has declined, with the investment framework shifting from profits and cash flow toward computing power and tokens. Liu Yang believes AI is moving toward application deployment and commercial monetization, and that investment opportunities will transition from model capabilities and computing power scale to critical monetization abilities such as industry position and traffic entry points. On the application side, she recommends focusing on companies with existing demand and earnings support; on the infrastructure side, scarce resources such as energy and data centers warrant attention; and on the hardware side, high-certainty directions like domestic substitution, semiconductor equipment, and advanced packaging are preferable. Meanwhile, AIDC investment still requires monitoring risks related to electricity supply, hardware iteration, and financing leverage, avoiding the trap of evaluating project value solely by planned scale.

Data center ROI is another focal point for the industry. Both Liu Yang and Huang Mincong examined the investment returns of AIDC from different angles. Liu Yang approached it from an investment framework perspective, advocating that AI investment should extend from traditional financial metrics to token value creation. Huang Mincong, drawing on specific projects, outlined the investment and monetization path of AI infrastructure across four dimensions: assets, computing power, operations, and applications. The view is that data center projects are highly non-standardized, and ROI must comprehensively account for electricity costs, construction and operational efficiency, customer absorption capacity, financing structures, and exit arrangements. GPU financing also requires attention to lease terms, hardware depreciation, and residual value. As AIDC investment scales up, the importance of project screening and return assessment will increase further.

Physical AI represents a significant direction worth monitoring. While past market attention on autonomous driving focused mainly on open-road scenarios such as Robotaxi, UISEE has taken a different path by entering closed environments like airports and factories, achieving large-scale unmanned operations first. It then leverages a general autonomous driving foundation to replicate across more vehicle types and scenarios. Similarly, Poke Robotics challenges the traditional notion that robot applications are mainly confined to industrial settings, focusing instead on household scenarios and exploring the possibility of embodied intelligence moving from industry into homes. The company employs its WAM world action model to develop general manipulation capabilities in complex home environments. As technology matures, physical AI is moving toward multi-scenario deployment, with commercialization pathways expected to broaden.

Looking back at historical industrial revolutions, the current AI wave driven by large models and computing power follows a similar evolutionary pattern to the previous three. The first industrial revolution solved the problem of power sources; the second used electricity to achieve production scale and standardization; and the third established the critical positions of semiconductors, computers, and the internet. This AI wave integrates electricity, production standardization, massive data, computing power, and networks into a deep fusion, building a new type of industrial carrier known as the "AI factory," which continuously converts these inputs into intelligent production units called tokens, ultimately driving AI toward commercial monetization. Regarding concerns in the capital markets about investment returns, technological innovation from its infancy to truly releasing productivity inevitably requires a long period of infrastructure construction and business model evolution — this cannot happen overnight.

The AI industrial revolution is still in its early stages, and its evolutionary logic closely aligns with the five-layer cake theory detailed by NVIDIA CEO Jensen Huang in a March 2026 post: energy, chips, infrastructure, models, and applications together form the complete AI factory production system. The top application layer is the final stage where the AI factory outputs tokens and enables commercial value realization. If this AI industrial revolution is compared to a baseball game, the current phase is roughly the second inning. The period from 2023 to 2025 could be viewed as the first inning, characterized by large model training with capital expenditure highly concentrated in GPUs and HBM. Entering 2026, as inference begins to achieve commercial monetization and low-latency agents accelerate into enterprise production processes, the second inning of the AI industrial revolution has officially begun. In this phase, hardware demand is no longer dominated solely by GPUs and HBM. The overall chip architecture is evolving toward heterogeneous computing, with tech giants driving in-house ASIC deployment, CPUs regaining an important position, and low-latency inference chips such as Groq LPU participating in reshaping the computing landscape. Meanwhile, the storage system is also expanding from an HBM-dominated approach to a multi-tiered architecture encompassing DRAM, NAND, HDD, and on-chip SRAM.

In the second layer of the five-layer cake, the Agentic AI era is accelerating the shift of computing architectures toward heterogeneous computing and multi-tiered storage. In the first inning, GPUs excelled in training tasks thanks to their outstanding parallel computing capabilities. But in the inference-dominated second inning, workloads are highly diversified with low-latency requirements, covering real-time inference, agent planning, tool invocation, and long-term memory management. A single GPU architecture can no longer cover all these tasks. NVIDIA is building a full-stack computing platform: the Vera CPU serves as the hub for general computing and agent orchestration, while the Rubin GPU continues to tackle high-throughput tasks. Interconnect technologies like NVLink 6 and ConnectX-9 enhance coordination efficiency, and the adoption of Groq's LPU accelerates the transition from training to real-time inference. Google is similarly building a full-stack AI computing matrix with TPUs, Axion ARM CPUs, low-latency chips, and OCS interconnects. On the storage side, multi-tiered architectures are emerging simultaneously: on-chip SRAM handles low-latency tasks, HBM provides high-bandwidth support, DRAM works closely with CPUs, CXL memory enables resource pooling expansion, and large-capacity SSDs and HDDs together support the data lake foundation. NVIDIA's latest platform not only features HBM4 on the Rubin GPU but also uses LPDDR5X-based SOCAMM modules in the Vera CPU, fully confirming the trend toward storage architecture heterogeneity.

In the fifth layer of the five-layer cake, continued growth in inference demand drives AI factories to convert computing power into token output and further into commercial monetization, fully reflecting the concept of "Compute is Revenue" proposed by NVIDIA CEO Jensen Huang at Computex. Looking back at the training phase of the first inning, tech giants expanded capital expenditure following the Scaling Law. But in the inference-focused second inning, as agents gradually penetrate enterprise production processes, AI is formally transitioning from a capital consumption phase to a value creation phase. Scenarios such as code generation, enterprise office tools, intelligent customer service, and software development are accelerating toward monetization, with payment mechanisms evolving toward subscription models based on token consumption, API call frequency, and agent service capabilities. Data centers are also shedding their traditional cost-center identity and accelerating their transition toward revenue centers. Under this business model, tokens are the fundamental monetization unit. The key metrics determining an AI factory's revenue-generating ability include token output per unit of power consumption, time to first token, average uptime without interruption, and equipment service life. These four dimensions respectively measure energy efficiency, system response experience, infrastructure stability, and long-term returns on computing assets — the key points for evaluating AI factory ROI.

The sustainability of capital expenditure and cash flow pressure remain focal points for the market. However, Meta's decision to lease out some of its computing power represents an asset optimization strategy of "making room for new capacity," rather than a signal of computing power oversupply or spending cuts. Some market views suggest that Meta may release older H-series chips to inference workloads or external leasing, while concentrating Blackwell and Rubin generation chips on frontier model training, potentially aiming to maximize asset utilization and boost operating cash flow. Over the past three years, AI capital expenditure has relied mainly on the自有资金 and operating cash flows of tech giants like Google, Amazon, Meta, Microsoft, and Oracle. But as infrastructure scale expands rapidly, traditional financing channels are approaching their carrying capacity limits. The financing paradigm for computing power has undergone profound changes, with off-balance-sheet financing models led by strategic financing gradually emerging. Google's $25 billion financing platform with Blackstone, Broadcom's credit backing for NeoCloud and Anthropic, NVIDIA's planned up to $250 billion project financing guarantee for OpenAI, and the computing power backstop models adopted by NVIDIA and AMD to support emerging computing platforms all indicate that capital markets should break through the traditional analytical framework of merely tracking tech giants' capital expenditure and expand their vision toward emerging infrastructure service providers and credit markets.

Future tracking focus should shift toward project financing accessibility, long-term service agreement quality, credit enhancement mechanisms, and the enabling effects of innovative tools such as REITs and private credit. As off-balance-sheet financing continues to catalyze hardware procurement and data center expansion, a positive cycle of "computing power — financing — applications — revenue — profit" is expected to be established. In terms of industry chain investment opportunities, wafer foundry and advanced packaging directly benefit from AI industry development. Yield rates and capacity bottlenecks in advanced process nodes and advanced packaging directly constrain AI chip supply, with demand spilling over from TSMC. Intel's 18A node and EMIB-T packaging technology are gradually maturing, and Samsung's foundry business has recently achieved important breakthroughs, making progress in projects such as Anthropic's custom chips, Tesla AI5, NVIDIA LPU, and Google TPU. Although Intel and Samsung are improving their technological capabilities and customer expansion in advanced process and advanced packaging, whether they can achieve significant market share gains still depends on multiple variables including yield improvement, capacity ramp-up, policy implementation, geopolitical dynamics, and ecosystem development.

As Agentic AI workloads shift toward inference, CPUs regain an important position in the second inning, becoming GPU's strongest complement. x86 leverages its single-core performance advantage in handling complex control instructions, while ARM addresses inference workloads with low power consumption and multi-core parallelism. Both excel in the new Agentic AI battlefield, and the CPU renaissance directly drives traditional DRAM demand. On the storage front, HBM remains a medium-to-long-term priority, DRAM works in deep coordination with CPUs to carry working memory, large-capacity NAND handles the task of invoking model weights, HDD serves as the massive storage foundation, and on-chip SRAM handles low-latency inference. All tiers of storage have become essential. Currently, overseas DRAM and NAND markets are dominated by an oligopoly of SK Hynix, Micron, and Samsung, but with CXMT listing on the capital market and releasing capacity, and YMTC's technology iteration gaining momentum, the global storage market share landscape may be reshaped.

In summary, as the AI industrial revolution enters the inference-dominated second inning, tokens have become an important monetization unit. From the perspective of the AI five-layer cake framework, the value creation path of AI factories is becoming increasingly clear. To hit a "home run" in this inning, capital market investment logic should move beyond simply betting on GPUs and their supply chains, embracing heterogeneous computing architectures and multi-tiered storage collaboration, while broadening investment research horizons to include the emerging computing ecosystem driven by off-balance-sheet strategic financing.

AI data centers are undergoing a transformation from traditional computer rooms to gigawatt-level token power plants. NVIDIA has redefined AI data centers as AI factories, using electricity and data as raw materials and tokens as products. The unit of capacity measurement has evolved from rack count to gigawatt level, and the delivery method has upgraded from individual chips to full-stack AI factory platform output. Liu Yang, Chairman and CIO of West Result Investments, noted that since 2026, this trend has accelerated: xAI's Colossus 2 has entered GW-level training cluster territory, continuing the rapid construction capability demonstrated by Colossus 1 (100,000 H100 GPUs completed in just 122 days). OpenAI and NVIDIA plan to deploy at least 10GW of computing clusters together, and NVIDIA and LG reached a cooperation agreement in June 2026 targeting physical AI and robotics, extending AI factories from large model training further into physical industries. On the power supply side, 800VDC may become the standard architecture for next-generation high-density AI factories, improving power supply efficiency for GW-level clusters by reducing power conversion stages, copper usage, and electrical equipment footprint.

Liu Yang stated during the event that an AI factory is essentially a type of computing real estate that can be repeatedly upgraded. She pointed out that 1GW of load approaches the electricity consumption scale of a medium-sized city, and modularization and factory prefabrication have compressed the traditional 18-24 month delivery cycle to approximately 6-9 months. While the design life of plant buildings and power grids is about 20 years, GPU iteration cycles are only 2-3 years, meaning land and electricity can be reused long-term, but each chip generation update will trigger new capital expenditure. The development of AI data centers is driven by three factors: demand, economics, and strategy. On the demand side, inference is taking over from training as AI shifts from one-time model training to high-frequency daily usage. Agents and multimodality further extend token consumption from human-machine interaction to autonomous machine execution. Liu Yang believes that declining unit token costs will expand application scenarios and, through the Jevons paradox, further drive up total token demand. On the supply side, GW-level clusters can amortize electricity, network, and cooling costs, with scale expansion helping reduce unit token costs. On the strategic front, countries are accelerating the construction of sovereign AI computing power, while tech giants lock in electricity and computing resources through long-term contracts, continuously enhancing the strategic infrastructure attributes of data centers.

Liu Yang also noted that with the rapid growth of AI data center electricity demand, PJM capacity market clearing prices have surged significantly in recent years. The auction for the 2028/29 delivery year, announced in July 2026, has already hit the FERC-approved price cap of $325/MW-day, reflecting that electricity supply is becoming an important constraint on AI infrastructure expansion. Liu Yang believes AIDC construction is an important starting point for demand expansion across computing, storage, optical interconnect, electricity, and packaging sectors. However, capital expenditure growth does not mean all business models benefit equally. High-leverage computing leasing and capacity wholesale models still need to watch GPU iteration, financing costs, and customer bargaining risks. Liu Yang emphasized that along the demand transmission chain of AIDC construction, computing power, storage, optical interconnect, electricity, and advanced packaging will benefit, but investment logic is diverging. In the computing segment, NVIDIA's competitive advantage is built on the CUDA ecosystem and developer usage habits; ecosystem control and continuous iteration capabilities are as important as hardware performance. In storage, AI continues to drive bandwidth and capacity demand, with HBM becoming an important companion to high-end computing; attention should focus on new capacity and market share gains. In optical interconnect, large-scale GPU clusters continue to raise high-speed interconnect demand; order flow, customer validation, delivery cycles, and product mix typically lead the income statement in reflecting industry prosperity, though FCC policy uncertainty in the US still requires attention. In electricity, the ability to secure power is increasingly influencing data center site selection and construction progress, so project evaluation should focus on actually available electricity rather than planned capacity alone. In advanced packaging, as transistor scaling costs rise, it becomes an important path to sustain computing growth; semiconductor equipment, consumables, and advanced packaging simultaneously carry both AI expansion and domestic substitution logic. Overall, AIDC investment requires moving research focus forward, paying more attention to leading indicators such as orders, validation, capacity, electricity prices, and delivery cycles.

The AI data center era also places new demands on traditional valuation systems. Liu Yang believes that as computing power and data become core corporate assets, the applicability of traditional PE and DCF valuation methods has declined. Some value creation processes in the AI industry are difficult to fully capture with existing financial metrics. First, GPUs and data centers can be capitalized, but R&D, talent, data, and ecosystem building are largely expensed in the current period, making long-term capabilities difficult to reflect in income statements. Second, large capital expenditure combined with rapidly changing competitive dynamics increases the difficulty of long-term cash flow forecasting. Third, moats change with algorithm and technology shifts, making competitive advantages and risk premiums more dynamic. Fourth, long-term competitiveness factors such as data flywheels, developer ecosystems, core talent, and workflow lock-in are difficult to measure through traditional balance sheets. While growth metrics like ARR and NRR can supplement the revenue dimension, their explanatory power for long-term competitive advantage remains relatively limited.

Based on this, Liu Yang proposed in her book "Silicon-Based — The Jerusalem of the Computing Power Era" that a new pricing anchor should be built around tokens. Every economic form has core observation metrics: the energy economy looks at levelized cost of electricity, the advertising economy looks at eCPM, and the AI economy can observe production and consumption through tokens. From this perspective, the industry chain can be divided into three layers: computing power manufacturers such as NVIDIA and TSMC determine the physical ceiling of token supply; token manufacturers such as OpenAI and Anthropic convert computing power into intelligence; and token integrators such as Cursor and Genspark transform model capabilities into products users are willing to pay for. Evaluating AI companies can focus on three key dimensions: token scale reflects size and bargaining power; token efficiency measures the effective value created per unit of computing power or token investment; and ecosystem control reflects developer, customer, and user workflow stickiness, as well as a company's ability to continuously participate in value distribution. These three form a mutually reinforcing flywheel: scale expansion drives efficiency optimization, efficiency improvement enhances ecosystem stickiness, and ecosystem growth in turn feeds back into scale.

Applying this three-dimensional framework at key nodes can further determine whether computing investment can translate into commercial value. Liu Yang believes that TSMC's advanced process capacity directly affects global high-end AI chip supply, and its value comes not only from wafer foundry profitability but also from the scarcity of its critical industry node position. In contrast, xAI, despite having significant computing power reserves and Grok entering the first tier of mainstream models, had 2025 revenue of only $3.2 billion and an operating loss of $6.4 billion. Its valuation sustainability depends on whether X platform users, data, and distribution entry points can transform into a stable ecosystem and commercialization capability. Computing investment alone cannot replace commercialization outcomes; if user stickiness, revenue growth, and sustainable models fail to form over the long term, asset value needs to be reassessed. Sun Microsystems once possessed leading SPARC performance and the Java ecosystem but was later disrupted by x86 and Linux; Intel similarly missed opportunities in the mobile era due to insufficient power efficiency and mobile ecosystem development. Therefore, rather than statically judging advantages in scale, efficiency, and ecosystem, what matters more is continuously tracking trend changes across these three dimensions.

Based on the above analysis, the AIDC theme still warrants continued tracking, but the investment focus is shifting from broad prosperity toward high-certainty segments and genuine commercialization. AI is gradually extending from "stacking computing power and competing on models" toward application deployment. AI applications remain a long-term direction, but greater emphasis must be placed on real revenue, customer quality, and closed business loops. Liu Yang believes that the judgment that AI will comprehensively replace traditional software remains unverified. Enterprise software and database vendors' existing customer relationships, product capabilities, and workflow stickiness still hold certain value. On the risk side, focus should be on tracking electricity access, GPU iteration impairment, computing utilization rates, and policy changes, with particular attention to verifying whether projects have actually secured power, started construction, and secured customer commitments, rather than simply equating planned capacity with future revenue. Overall, AI-era investment needs to move from simply comparing technical performance toward judging whether companies control scarce industry nodes, can continuously improve efficiency, and can build stable ecosystems and commercialization capabilities.

UISEE has undergone three strategic iterations over its decade of development, extending from autonomous driving into Physical AI. CEO Wu Gansha stated at the conference that over the past ten years, core AI industry investment has mainly concentrated on the energy, chip, infrastructure, and model layers. As end applications begin large-scale deployment, the industry is gradually entering a next phase that places greater emphasis on closing business loops, with autonomous driving being one of the earliest directions in Physical AI to achieve real-world scenario validation. Wu Gansha believes autonomous driving serves both as an incubator for Physical AI — with data flywheel methodologies and talent accumulated over years migrating toward embodied intelligence — and as an important testing ground for Physical AI technologies and business models. The company's strategy has undergone three adjustments accordingly: in 2019, it focused on "truly unmanned" operations, treating the removal of safety drivers as an important prerequisite for large-scale commercial operations; in 2021, it proposed "all-scenario" coverage, hoping to reuse a unified technology base across different vehicle types and application scenarios to amortize the high R&D costs of autonomous driving; and in 2024, it further proposed the Physical AI direction, aiming to replicate the intelligent capabilities developed through high-cost training at lower marginal cost across more physical scenarios.

The company is currently enhancing its large-scale replication capabilities across three dimensions: civil aviation-grade safety, scenario generalization capabilities, and all-scenario vehicle platform layout. Airports represent UISEE's benchmark scenario for large-scale commercial operations and serve as an important source for its safety systems and data capabilities. Autonomous driving in the physical world directly involves personnel, vehicles, and aircraft, imposing significantly higher requirements on system reliability, exception handling, and continuous operation compared to digital applications. The company has long used civil aviation safety and service systems as important references for its product standards, addressing the final portion of long-tail problems through long-term real-world operations. As of August 2026, the company's market share in airport unmanned driving exceeds 90%, with operations covering 21 domestic and international airports, cumulative fully unmanned safe operation mileage exceeding 6 million kilometers, and more than 1,000 related patents. Domestic second- and third-tier airport projects have achieved break-even on unit economics. Since receiving approval for unmanned operations in 2019, the company has invested over $150 million, iterated approximately 200 versions, and processed tens of thousands of issues, building proprietary airport-specific autonomous driving data accumulation. Long-term operational capability extends beyond vehicle autonomous driving itself, gradually expanding toward full-process unmanned airport operations. The company's products have been validated in complex conditions including high temperature and humidity, heavy rain, salt spray, blizzards, sandstorms, weak networks, and nighttime operations, with systems established for real-time dispatch, multi-vehicle coordination, vehicle diagnostics, and predictive maintenance. Building on this foundation, the company is further advancing unmanned operations for automatic trailer hooking/unhooking, automatic charging, and automatic loading/unloading, reducing the manual intervention steps required after vehicle unmanned operation, further lowering customers' full-lifecycle operating costs, and expanding apron applications from luggage and cargo towing to shuttle services, perimeter patrol, cleaning, bird control, and meal delivery scenarios. The company has also launched the "Zaofu" open-source AI programming delivery platform, converting existing industry experience into programming and business orchestration tools that support airport control system integration and customer custom development. As civil aviation unmanned driving industry standards gradually improve, UISEE expects its airport autonomous driving solutions to move from a few benchmark projects into broader replication.

The U-Drive all-scenario autonomous driving operating system is the core technology base for UISEE to reduce cross-vehicle and cross-scenario replication costs. The company connects different vehicle hardware and application scenarios through its unified U-Drive platform, covering passenger vehicles, long-haul and urban logistics, buses, airports/factories/parks, ports, mines, agriculture and animal husbandry, sanitation, and inspection. After ten years of iteration, U-Drive has evolved from version 1.0 for low-speed passenger transport, 2.0 for truly unmanned all-weather logistics, 3.0 for logistics and passenger transport integration, and 4.0 for complex open-road scenarios, to the current 5.0 all-scenario self-learning version, featuring generalizable algorithm libraries, millions of scenarios in its library, standard hardware templates, and adaptive toolchains. The upcoming 6.0 is planned to upgrade to a general autonomous driving foundation model, improving complex long-tail scenario handling through world models, while 7.0 will further explore model integration involving end-to-end approaches, world models, and reinforcement learning. UISEE aims to gradually transform traditional project-based delivery into productized capabilities that can be replicated at scale through platformization. The company targets approximately 3 months for new product productization, about 1 week for new scenario adaptation, and further simplification of new customer deployment complexity. From existing projects, early delivery of 25 vehicles at Hong Kong Airport took about one year, while delivery of 40 vehicles at Urumqi Airport has been reduced to less than one month. Factory scenarios have also shortened from 5 vehicles in half a year initially to delivery of tens to hundreds of vehicles in approximately 3 months. New scenarios currently require only small teams of product managers, architects, developers, and testers for adaptation. Through new joint development processes, the company has reduced new vehicle development costs to approximately one-tenth of the original model and development time to about one-quarter.

On the business model front, the company plans to increase the proportion of AI driver subscription revenue from the current vehicle and solution sales. Currently, the company offers various options including vehicle sales with maintenance, vehicles with AI driver subscriptions, financial leasing, autonomous driving models, and retrofit kits, all aimed at reducing customers' upfront capital expenditure while ensuring that the full-lifecycle cost of unmanned solutions is lower than manual driving. Long-term, UISEE hopes to further reduce its own participation in vehicle manufacturing and frontline operations, forming a platform model of "AI driver + ecosystem partners." The company plans for commercial vehicle OEMs to provide certified vehicle bodies with pre-installed AI drivers, system integrators, solution providers, and operators to handle specific industry deployments, while UISEE provides subscription billing, health monitoring, and data security compliance services through its AI driver management cloud, concentrating its capabilities more on autonomous driving software and continuous iteration. On globalization, the company has established four systems covering quality, access certification, data security, and local services. Three vehicle models have passed CE certification, and cooperation is progressing with European and Middle Eastern airlines, airports, and ground service companies. The company expects to expand business coverage from 6 countries and regions to 50 within five years, with more channel partners developing new scenarios. Short-term focus remains on validated high-value scenarios, with medium-term expansion into commercial vehicles and broader Physical AI applications. Over the next 1-2 years, the company will focus primarily on airport and factory towing logistics, campus unmanned shuttles, passenger vehicle assisted driving, and specialized scenarios in ports, chemicals, power, and mining; in 3-5 years, it plans to expand into bus, sanitation, heavy truck, and urban distribution markets; and in 5-10 years, Robotaxi, long-haul logistics, and generalized embodied intelligence are forward-looking directions. The company's long-term vision is to deliver 1 million reliable AI drivers globally, becoming the world's largest provider of AI drivers and new workforce. It also noted that China possesses combined advantages in open-source AI capabilities, manufacturing and supply chains, and industry application depth, suggesting that Chinese new workforce robot products could become one of the important directions for the globalization of Chinese technology products.

Poke Robotics was founded in April 2026 and, from its inception, has focused on world models and household scenarios rather than VLA models and industrial scenarios. The company's core team comes from Tsinghua University and Shanghai Qi Zhi Institute. Leveraging its globally leading WAM (World Action Model) technology, the company has established three core R&D divisions: world action models, full-chain reinforcement learning, and robot hardware, while developing ecosystem partnerships with well-known domestic hotel and real estate industry players. On the financing front, existing investors including Yunqi Capital, Shunwei Capital, Matrix Partners China, Xiaomi Strategy Investment, and Honghui Fund have continued to add capital, placing the company's financing pace and investor lineup at the leading level among early-stage embodied intelligence startups. In terms of technology route selection, the company believes the WAM world action model is the technical path closest to robot decision-making, balancing prediction and control. Although WAM is more complex in training and deployment, it can directly model actions, observations, and future states, achieving a better balance between physical correctness and robot decision-making. Therefore, the company chose to self-develop its WAM world action model, aiming to build a manipulation foundation model with high physical consistency on top of powerful pre-training infrastructure and large-scale data. The company believes that moving from robot manipulation to Physical AGI requires more than just "training fast" — models need to truly understand the physical world. The company's self-developed Poke WAM produces prediction results closer to real-world performance, demonstrating its advantage in physical consistency.

Service scenarios impose higher requirements on data, models, and hardware, with large-scale embodied intelligence deployment still facing challenges. On the data front, traditional teleoperation collection costs are high, with million-hour-level dataset investments potentially reaching hundreds of millions of yuan. Fine manipulation remains primarily "humanoid operation" rather than true "hands-on operation," and large-scale data collection and management costs are significant. On the model front, rapid technology route iteration demands continuous R&D investment from companies, while existing models still have considerable room for improvement in dexterity and generalization. Breakthroughs are needed in fine manipulation, long-horizon tasks, and tool use, along with cross-scenario, cross-task, and cross-embodiment generalization. On the hardware front, robot hardware requires continuous cost reduction and reliability improvement, with costs needing to decline from hundreds of thousands to tens of thousands or even less, while stable operation time needs to increase from hours to days and months. To address these bottlenecks, the company has built a "trinity" embodied intelligence flywheel around its WAM world action model, full-chain reinforcement learning, and high-quality real data. Through the mutual reinforcement of data, models, and reinforcement learning, the company accelerates progress toward Physical AGI. Poke WAM enhances understanding and prediction of the physical world through the world action model; Poke UMI continuously acquires high-quality real data to support model training and iteration; and Poke RL uses reinforcement learning to form better policies and more stable outputs. Together, they form a closed loop of "real data — world model — reinforcement learning," with each training and deployment round continuously feeding the next iteration, driving continuous improvement in robot dexterity, generalization capabilities, and complex task execution.

Venture Hub is positioned in the Southeast Asian market as a technology investment and industry empowerment platform covering the full AI industry chain. Its core team's capabilities span the complete chain from underlying infrastructure to upper-level applications. On the infrastructure side, it covers AI data center site selection, power supply design, engineering construction, and ongoing operations. On the computing and platform layer, it achieves task orchestration, model selection, token economics, and load optimization through intelligent routing. On the application side, it helps partners convert computing resources into deployable, commercializable AI applications through investment and industry consulting. The company also screens data center construction and operation teams with unique technological advantages and global commercialization potential through its accelerator program, accelerating projects toward product-market fit and early commercialization through systematic incubation, industry expert mentoring, and real customer scenario validation, laying the foundation for subsequent financing rounds.

Venture Hub has built a full value-chain investment system around AI infrastructure covering "assets — computing — operations — applications," primarily through four components. First, data center investment: the underlying logic is similar to real estate investment but further layered with electricity resource value monetization — securing low-cost electricity and selling to downstream customers for the price spread, while enhancing revenue certainty through long-term consumption contracts and tiered price adjustment clauses. Projects can be divided into new construction and renovation. New projects must balance energy consumption and cost with scalability and flexibility, while also monitoring engineering delays, hidden land premium, and exit risks. Renovation projects primarily target idle industrial plants and former cryptocurrency mining sites, carrying out asset upgrades in mature markets such as Hong Kong and Japan, with relatively lighter capital requirements; the key lies in controlling ongoing operating costs and improving asset utilization efficiency. Second, GPU computing financing: unlike cloud vendors that mainly self-own servers and GPUs, new computing clouds are increasingly adopting asset-light models, supporting computing expansion through external capital, with typical financing structures of "equity + debt" using GPU computing assets as core collateral. Third, maintenance and value-added services: providing professional hosting, power management, liquid cooling, and heat dissipation optimization services around high-density AI clusters, improving infrastructure efficiency and asset value through ongoing operations. Fourth, token monetization: extending further into enterprise and application layers, improving computing utilization through model routing, token economic optimization, and exploring B2C transitions, dynamic pricing, and other business models to ultimately convert underlying computing power into sustainable application revenue.

The company highlighted its understanding of investment models, transaction structures, and target returns for data center and GPU computing financing projects. It should be noted that data center and computing financing projects have strong non-standardized characteristics, with significant differences across projects in development stage, asset conditions, customer structure, and financing arrangements. The return calculations and transaction structures primarily reflect Venture Hub's own project experience and investment framework. For data center investment, Venture Hub typically evaluates project value based on EBITDA multiples and sets corresponding target returns according to project development stage. According to the company, early-stage development projects target MOIC of approximately 3-7x, with early-stage investment target IRR of approximately 30%-40%. As projects mature and development risk declines, expected returns in later stages correspondingly converge. The company believes that from early development through construction to stable operations, different investor groups at different project stages correspond to different risk-return requirements. For GPU computing financing, the typical project structure described by Venture Hub uses an "equity + debt" capital structure with debt/equity ratios of approximately 30/70 or 40/60. Depending on the debt provider and transaction terms, the company observes comprehensive debt return targets of approximately 8%-15%. In the case studies presented, projects can secure GPU procurement and lease contracts of approximately 4 years, hosting GPUs in data centers operated by major cloud vendors, Neoclouds, or AI labs that consume the computing power, with rental income covering data center lease and financing costs. Some projects have payback periods of approximately 2 years, with target equity IRR exceeding 20% and MOIC of about 2x. Regarding the widely discussed GPU depreciation risk, the company believes that if consumption contract terms can be well-matched with hardware economic life, and GPU secondary market residual values exceed traditional straight-line depreciation assumptions, equity risk can be mitigated to some extent.

Risk warnings: industry and technology route uncertainty risk — AI terminal chips, embodied intelligence, and AI infrastructure are all emerging tracks with rapid technology iteration and unsettled competitive landscapes. Industry technology route choices, market demand growth rates, and competitive landscape evolution all carry significant uncertainty, potentially causing discrepancies between companies' technical judgments, product plans, and actual development paths. Early-stage operation and valuation fluctuation risk — AI chip and embodied intelligence-related companies and projects are all in early business development stages, without stable profit models and cash flows, primarily relying on continued external financing. Combined with the significant impact of market sentiment and capital environment on valuations in these tracks, there are risks of financing progress falling short of expectations and significant valuation fluctuations. Policy and macro environment risk — related businesses involve multiple regulatory domains including semiconductors, cross-border investment, electricity, and land resources. Changes in the macroeconomic environment, industrial policies, trade, and export control policies may all adversely affect companies' operating environments and industry development pace. Content in this research report involving unlisted companies or uncovered stocks represents only compilation of objective public information and does not constitute recommendations or coverage by this research team.