Deep Robotics

1.1 Identity, stage, and business model

Deep Robotics presents itself as an embodied-AI robotics company built around hazardous and industrial use cases rather than a consumer robotics story. Official materials describe a Hangzhou-based high-tech enterprise focused on quadruped robots, humanoid robots, and core robotic components, with a full-stack “Perception–Decision–Execution” framework aimed at energy, emergency response, industrial operations, and education. Independent reporting lines up with that framing: the company was founded in November 2017, remains private, and is now best understood as a late-private, pre-IPO industrial robotics company rather than an early research spinout. The business model visible in the public pack is hardware-and-solution sales into inspection, rescue, logistics, and research scenarios, with later humanoid work layered on top of an already commercial quadruped base. The core diligence caveat is disclosure depth. Basic identity and product scope are clear, but audited financials, current customer count, and a precise lifetime capital tally are still not publicly reconciled.[CO001, CO002, CO003, CO005, CO007, CO041]

1.2 Founders, public leadership bench, and governance visibility

The founding story is a genuine part of the underwriting case because Deep Robotics still looks founder-led in both technology direction and public narrative. Zhu Qiuguo and Li Chao came from the Zhejiang University robotics ecosystem, and public interviews portray Zhu as the strategic, commercial, and mission-setting voice while Li covers technical commercialization, industrial applicability, and sector positioning. That founder-market fit is reinforced by Zhu’s associate-professor role and the company’s repeated claims of full-stack self-development, patent depth, and industrial deployment discipline. What is less reassuring is the thin public governance layer outside the founders. Reviewed sources name several public-facing operators—such as product and overseas-sales leads—but do not provide a clean board roster, independent-director list, or a fully itemized post-reform executive lineup. That means leadership depth is visible mainly through public spokespeople and customer-facing operators rather than through a standard late-stage governance disclosure package.[CO003, CO004, CO031, CO032, CO033, CO034]

1.3 Capital base, private status, and stakeholder map

Deep Robotics’ financing story is substantial, but not yet fully disclosure-grade. Public reporting supports a clear sequence of recent financings: a nearly CNY500 million pre-Series C round in July 2025, a more than CNY500 million Series C in December 2025, and an early-2026 pre-IPO round described only as hundreds of millions of renminbi. The company also completed a November 2025 shareholding reform into a non-listed joint-stock company and then filed for IPO tutoring on December 23, 2025 after signing its advisory package on December 8. Taken together, those steps strongly support the view that the company is still private but actively preparing for a mainland listing. The public record also supports a media-reported valuation around CNY8 billion, but not a prospectus-grade ownership picture. Exact lifetime capital raised, debt or credit exposure, and any secondary-sale history remain unresolved, so investors can understand momentum and funding direction but not yet reconstruct a clean cap-table or mark history from public documents alone.[CO005, CO006, CO007, CO008, CO009, CO010]

1.4 Products, customer proof, and scale signals

The strongest part of the company-overview story is that Deep Robotics did not jump straight into humanoid marketing; it first built a usable quadruped business in power inspection and other high-risk environments. Public evidence shows an industrial product ladder: the X20 established the all-weather quadruped baseline, the X30 became the export-grade inspection flagship, the DR01 marked the first humanoid debut in 2024, the Lynx M20 widened the portfolio into wheeled-legged mobility in April 2025, and the DR02 pushed the company into industrial humanoids by October 2025. Customer and deployment proof is strongest in power infrastructure. The Singapore Power Group SPock deployment is independently corroborated and operationally specific, including tunnel inspections, localization work, and an estimated 480 annual labor hours saved. Management interviews also point to broadening application depth across emergency response, security patrol, research, and industrial sites, with more than 600 projects reported and a 10,000-unit shipment target for 2025. That is a meaningful commercial signal even though audited customer-count and backlog disclosure remain absent.[CO017, CO018, CO019, CO020, CO021, CO022]

1.5 Milestones and the main adverse tension

The milestone record makes Deep Robotics look materially more mature than a pure concept-stage humanoid startup. From the 2018 first-generation quadruped and the 2020 power-grid pilot, the company progressed into productized industrial quadrupeds, overseas deployment, new product categories, repeated fundraising, and formal IPO preparation. Public sources also show that 2024 revenue more than doubled year over year, headcount reached roughly 400, and management is willing to say the company is not operating at a loss. The adverse tension is not lack of activity but lack of full proof. Public commentary from investors and sector analysts still questions whether robotics leaders can scale ROI-positive deployments fast enough to justify current marks, and the DR02 launch itself left autonomy, battery-life, and cost questions unanswered. In other words, Deep Robotics appears commercially ahead of many humanoid peers, but the public pack still supports a cautious reading: real traction, real milestones, and real IPO preparation, paired with incomplete governance and financial disclosure plus unresolved commercialization skepticism.[CO013, CO014, CO015, CO016, CO017, CO025]

2.1 Market Boundary, Included Spend, and Substitutes

Deep Robotics should not be valued against the whole robotics economy or even the whole quadruped category. The company's own public evidence is much narrower: industrial inspection, underground navigation, emergency response, and research deployments anchored in hard-to-access environments. That boundary implies included spend around mobile inspection hardware, payloads, autonomy software, and deployment services for utilities and industrial operators. It also implies explicit exclusions. Defense-only UGV programs, consumer quadrupeds, fixed sensors, and drone-only budgets belong in adjacent or substitute pools, not in a Deep-specific TAM story. This distinction matters because the status quo is not empty; manual patrols, wheeled robots, fixed sensing, and drones already cover many inspection jobs. Deep Robotics only wins when stairs, confined spaces, irregular terrain, or contact-rich sensing make legged mobility materially better than those substitutes.[CM001, CM002, CM003, CM004, CM005, CM039]

2.2 Sizing Lenses and Contradictory Estimates

The public corpus does not support one clean headline TAM for Deep Robotics. The best broad market data comes from quadruped-market reports, but those estimates diverge materially. Precedence projects USD 3.71 billion in 2026 and USD 17.53 billion by 2035, while Emergen starts from USD 1.2 billion in 2024 and reaches only USD 4.8 billion by 2034. The contradiction does not mean the market is fake; it means category definitions differ, especially around whether defense, logistics, research, and inspection are bundled together. Deep Robotics therefore has to be sized through multiple lenses: the broad quadruped TAM, China's huge robot-installation base and policy support, and the narrower evidence-backed slice where utilities and industrial operators are already deploying inspection robots. Investors should treat the gap between broad TAM and reachable SAM as a core diligence issue, not as a rounding error. That gap is valuation-relevant because different publishers can make the same company look niche or huge.[CM006, CM007, CM008, CM009, CM010, CM011]

2.3 Buyers, Users, Payers, and Regional Beachheads

The best-evidenced near-term buyers are utility and energy infrastructure operators, not a generic "enterprise" bucket. Deep Robotics' own case studies point to substations, wind farms, UHV assets, and Singapore power tunnels, which implies users such as maintenance crews, field inspectors, and operations teams. The payer or sponsor, however, is more likely to sit with the grid operator, plant owner, or industrial operations function controlling maintenance, safety, and automation budgets. Research and education buyers operate under a different logic: they care more about affordability and experimentation than certified uptime. Competitive evidence sharpens the picture further. ANYbotics is aimed at hazardous industrial operators and has certification and channel strength in oil and gas, Ghost mainly serves defense agencies, and Unitree shows that Chinese vendors are converging on utility inspection while also covering low-cost research demand. Deep's public beachhead is therefore strongest in energy infrastructure, with China and nearby Asia better evidenced than the West. That split also affects sales motion: utility accounts need site integration and procurement patience, while research buyers can close faster but at much smaller contract values.[CM005, CM017, CM018, CM019, CM020, CM021]

2.4 Adoption Drivers, Constraints, and Valuation Relevance

The adoption case is strongest where robotics solves expensive, hazardous, or repetitive inspection work. Deep Robotics claims meaningful efficiency and accuracy gains in substations and wind farms, Emergen cites major inspection cost reductions from autonomous systems, and OSHA statistics keep worker-safety pressure visible for industrial payers. China adds a policy tailwind through the 15th Five-Year Plan and still-low robot density. But the constraints are equally important for valuation. ANYbotics' ATEX and IECEx certifications show that hazardous-site compliance is a real commercial moat. Unitree openly acknowledges autonomy limits, GasGoo highlights mass-production and reliability friction, and Boston Dynamics' pricing plus Orbit software illustrate how enterprise buyers must absorb both hardware and orchestration spend. Those constraints mean Deep Robotics can plausibly grow in utilities and selected industrial sites, but the market will not clear at full TAM unless certification, software, and reliability all improve together.[CM015, CM016, CM021, CM023, CM024, CM025]

2.5 Exhibits

3.1 Landscape and competitive lanes

Deep Robotics faces a layered field rather than one single reference rival. In its home lane, the company positions X30 and Lynx as industrial inspection robots for pipelines, substations, tunnels, and other hazardous infrastructure environments, then backs that positioning with case studies in Chinese substations, Ningxia wind assets, Singapore cable tunnels, and UHV stations. Unitree is the closest domestic hardware peer because it spans both civilian and industrial quadrupeds, but its public materials still disclose manual-operation and secondary-development limits on parts of the stack. ANYbotics is the most dangerous global industrial-inspection rival because it overlaps on power, energy, and utilities while also holding ATEX/IECEx hazardous-area credentials that Deep has not publicly shown. Boston Dynamics is still the category's best-known premium brand, yet its market role is a higher-cost platform-plus-software offer, not a low-cost China utility specialist. Ghost Robotics matters less in utility inspection but more as a reminder that quadruped demand also flows through defense and government budgets where Chinese-origin hardware can face structural procurement barriers. Substitute methods remain real: indoor drones, fixed sensors, and status-quo manual rounds still solve parts of the inspection job, especially where full quadruped mobility is unnecessary.[CP001, CP004, CP005, CP006, CP007, CP008]

3.2 Capability, pricing, and buyer-fit comparison

On capability, Deep Robotics is strongest where inspection buyers care about terrain handling, utility workflows, and integrated sensing more than brand or defense pedigree. The X30 spec sheet emphasizes IP67 protection, multi-sensor fusion, and 2.5-4 hour missions, while the industrial-inspection page adds thermal, acoustic, and robotic-arm options that map cleanly to utility and plant maintenance workflows. Lynx M20 matters strategically because it shows Deep is not dogmatic about pure quadrupeds: when speed and corridor efficiency matter, the company is moving toward wheel-legged designs. Unitree's B2 is the more visibly aggressive general-purpose hardware challenger, with higher headline speed, load, and wheeled-leg optionality, but its public disclaimer on manual operation undercuts its case for fully lights-out inspection. Boston Dynamics remains the benchmark for managed inspection software via Orbit and preconfigured packages, although its historical Spot pricing reinforces why many buyers will test cheaper or quote-based Asian alternatives first. ANYbotics wins the best-documented hazardous-environment position through certified Zone 1 readiness, while Ghost Robotics is tuned to defense missions rather than civilian energy assets. Public pricing disclosure is still thin across this market, so the actionable comparison is less about exact list prices and more about which vendors have the best fit for specific buyer jobs, regulatory environments, and autonomy expectations.[CP002, CP003, CP004, CP010, CP011, CP012]

3.3 Switching costs, substitutes, and channel access

Deep Robotics does not appear to enjoy hard contractual lock-in from public evidence; instead, switching costs are operational and situational. Once a plant or utility has mapped routes, tuned inspection payloads, integrated thermal/acoustic readings into workflows, and trained staff on one robot family, changing vendors creates friction—but not the kind of exclusive data moat seen in some enterprise software markets. That makes multi-homing credible. Buyers can pilot one quadruped in substations, deploy a different one in hazardous petrochemical zones, and still use drones or fixed sensors for open-area or continuous-monitoring tasks. The strongest substitute is not another dog alone but the mixed stack: drones such as Flyability's Elios 3 for indoor/confined-space imaging, fixed sensors for always-on alerts, conventional industrial-automation vendors such as ABB for broader inspection-and-testing automation, and manual crews for low-frequency or low-complexity routes. Deep has an advantage where the job requires stairs, tunnels, clutter, contact-sensing, or frequent redeployment without heavy site modifications. The company is weaker where the buyer needs certified explosion-proof access, publicly demonstrated U.S. defense eligibility, or highly mature fleet software. Internationally, Singapore and GITEX show export intent, but channel power still looks earlier-stage than the regional enterprise reach claimed by ANYbotics or the government base claimed by Ghost Robotics.[CP003, CP007, CP008, CP020, CP021, CP024]

3.4 Moat durability and adverse evidence

The adverse evidence does not say Deep Robotics is weak; it says the category is getting harder to defend. Gasgoo's January 2026 coverage of MirrorMe's Apollo launch is the most important disconfirming signal because it argues that stable manufacturing and standalone innovation no longer form a durable moat by themselves—what matters is the feedback loop between iteration and industrial verification. That is relevant to Deep because its best differentiator today is exactly that deployment-feedback loop in utilities, not a publicly visible certification, software, or procurement monopoly. Unitree can push from below on flexible hardware and lower-end accessibility, ANYbotics can pressure from above on hazardous-area trust and global energy references, and Boston Dynamics can continue to define the premium software-led inspection benchmark. Meanwhile, Ghost Robotics shows that one large demand pool—U.S. national-security procurement—may remain relatively closed to Chinese-origin suppliers regardless of technical merit. Deep's wheel-legged Lynx response is therefore strategically positive: it suggests the company is adapting to site reality instead of overcommitting to one locomotion form. Still, the unresolved items matter for diligence: current public pricing, installed-base transparency, ATEX/IECEx status, and how much of the company's overseas activity is repeatable channel growth versus showcase-led market entry.[CP015, CP021, CP025, CP028, CP031, CP034]

4.1 Revenue model and pricing signals now look industrial and productized, but realized pricing is still opaque

Public evidence now shows a real revenue engine rather than a concept-stage robotics story. The prospectus summaries reviewed through Sina, CnTechPost, and Humanoids Daily all point to the same 2025 picture: Deep Robotics generated RMB 337 million of revenue, with embodied robots contributing more than 95% of core business income, and product mix led by the industrial Jueying X series, the lower-priced Lite line, and the newly launched Lynx M20 wheel-legged robot. The company’s monetization surface therefore appears hardware-first and enterprise-led, with components and services present but economically secondary to robot-body sales. That matters because it means underwriting starts from shipment volume, ASP, and gross margin by product family rather than from recurring SaaS metrics. Pricing is visible only in partial form. The clearest public signals are 2025 prospectus ASPs of roughly RMB 287.5 thousand for the X series, RMB 27.7 thousand for Lite, RMB 197.6 thousand for Lynx, and RMB 823 thousand for the single DR humanoid sale. Management and third-party coverage also frame Deep Robotics as materially cheaper than Boston Dynamics and some Korean peers, but those are positioning claims rather than realized contract economics. The record still does not disclose discount schedules, software attach rates, maintenance pricing, or what portion of project revenue sits in services versus robot hardware.[CI001, CI003, CI004, CI005, CI006, CI007]

4.2 GTM motion is solution-led around hazardous inspection, with strong ROI proxies but large unit-economics gaps

The reviewed sources consistently describe Deep Robotics as an enterprise solutions seller, not a consumer robot vendor. Official pages emphasize industrial inspection, emergency response, and hazardous-environment deployment; management interviews point to power inspection as the wedge because China has roughly 20,000 substations with repetitive, rules-based inspection workflows; and overseas proof points such as Singapore Power show the same pattern of customized deployment into mission-critical infrastructure. This is financially important because the company appears to win business through operational reliability, environmental resilience, and deployment support rather than through low-friction self-serve checkout. Public ROI signals are stronger than for many robotics startups but still incomplete. Official case studies claim 98% inspection accuracy, sub-35-minute full-station patrols versus more than an hour manually, more than 50% workload reduction, and about 70% operating-cost reduction in one substation setting. Singapore Power expects more than 480 annual labor hours saved. Li Chao’s interview adds a useful commercial test: a product is merely usable when a customer buys once, but “good-to-use” when it leads to 100-, 300-, or 500-unit repeat purchases. Even so, the public record still does not give CAC, payback, churn, NRR, service attachment, warranty burden, or installation margin, so the revenue-quality case is directionally positive rather than fully underwritten.[CI014, CI016, CI017, CI018, CI019, CI020]

4.3 2025 profitability is real, but humanoid expansion and manufacturing scale still make financing dependency central

Deep Robotics enters this chapter with better public financial evidence than most private robotics companies. The 2025 filing summaries indicate first-time profitability, 52.83% gross margin, and a meaningful overseas revenue contribution. Recent private financing also looks substantial: roughly RMB 500 million in July 2025, more than RMB 500 million in December 2025, and a separate pre-IPO round of hundreds of millions of yuan reported in early 2026. The company remains private as of the run date, but its STAR Market application has been accepted and the planned IPO would raise about RMB 2.503 billion for embodied-AI models, new robot bodies, industrialization, and base construction. The underwriting issue is not whether Deep Robotics can raise capital; it is whether the current margin profile is durable without fresh capital and without a favorable product mix. Public coverage repeatedly ties the next phase of growth to humanoid R&D, expanded manufacturing, and continued talent hiring. Prospectus-based reporting says R&D already consumed RMB 84.3 million, or about 25% of revenue, in 2025, and the company itself warned that rapid technological iteration could cause profitability volatility. Because cash on hand, burn, debt, and runway are still undisclosed, the safer conclusion is that 2025 profit demonstrates viability in industrial quadrupeds but does not eliminate forward financing dependence.[CI002, CI008, CI009, CI010, CI026, CI027]

4.4 Financial verdict: commercially credible industrial robot business, but still not a complete diligence-ready case

The positive case is unusually concrete for a private robotics company. Deep Robotics now has prospectus-backed revenue, product-level ASP disclosures, improving gross margins, first-time profitability, visible overseas revenue, and customer ROI stories in real infrastructure settings. Those are not trivial signals: they separate the company from robotics vendors that still rely almost entirely on demos, research sales, or narrative financing. The negative case is that the remaining blind spots are exactly the ones that decide whether a profitable 2025 can compound. Public sources do not disclose cash balances, monthly burn, debt facilities, working-capital strain, backlog quality, customer concentration by revenue, realized discounting, or service and warranty economics. Adverse market sources also warn that the humanoid race is becoming a war-chest contest, while management itself highlights price pressure and talent poaching. The chapter verdict is therefore balanced: Deep Robotics looks like a real industrial-hardware business with defensible demand today, but forward underwriting still hinges on private evidence about liquidity, concentration, and whether margin survives the shift from quadrupeds to more capital-heavy humanoid programs.[CI002, CI015, CI037, CI040, CI041, CI042]

4.5 Exhibits

5.1 The lineup is real, segmented by workflow, and the brief's "L20" label does not match public product surfaces

The public product surface is now broad enough to underwrite Deep Robotics as a platform company, not just a one-model quadruped vendor. Official pages and late-2025 releases show four distinct lanes: the X series for industrial quadruped inspection, the Lynx series for wheel-legged operation in tighter or rougher terrain, the Lite3 family for education and developer use, and the DR humanoid line for industrial embodied-AI expansion. Within that stack, X20 remains the lighter industrial quadruped, X30 the more rugged flagship, Lynx M20 the heavier industrial wheel-legged robot, Lynx S10 the smaller wheel-legged robot for narrow spaces, Lite3 the open research platform, DR01 the expo-stage humanoid demonstrator, and DR02 the first product page that looks like a commercial humanoid offer. One briefing assumption does not survive source review: no reviewed official page, FAQ, GitHub surface, or third-party coverage names an "L20" model. The recurring names are X20 and Lynx M20, so the most defensible interpretation is that the brief meant one of those. The workflow split is also meaningful. Utilities and tunnel operators buy X-series or Lynx robots for patrol and defect detection; emergency teams buy terrain access plus gas, thermal, and imaging payload carriage; researchers buy Lite3 because it exposes ROS, SDK, and sim-to-real tooling. That segmentation makes the portfolio easier to map to actual budgets and deployment motions than a generic "robot dog" story.[CE001, CE002, CE003, CE004, CE005, CE006]

5.2 The core architecture is a full-stack perception-to-actuation system, with the most concrete public proof in the developer stack

Deep Robotics' most credible technology claim is not any single speed or payload number; it is the coherence of the stack the company has exposed across product pages, GitHub repositories, and later 2025 announcements. The corporate description and Series C release use the same "perception-decision-action" framing, while DeepVLA 1.0 adds a more explicit software layer: open-vocabulary semantic point clouds, large language model command parsing, and a topological road network for multi-floor door-to-door navigation. On the hardware side, public pages expose a consistent pattern of multimodal sensing, onboard compute, modular payload mounting, and ruggedized locomotion. Lynx M20 is the clearest example because its official page publishes dual industrial processors, dual 96-line LiDARs, dual wide-angle cameras, Ethernet, hot-swap batteries, and OTA upgrades in one SKU. The developer surface is stronger than for many private robotics companies. DeepRoboticsLab publishes robot models in MJCF/URDF/USD, IsaacLab-based reinforcement-learning environments for Lite3 and M20, a ROS2 deployment SDK, and CAN-level joint control for J60 modules. Lite3 documentation on the U.S. site goes further by describing ROS1/ROS2 support, Motion SDK access, LiDAR SLAM examples, Jetson-based perception hosts, and sim-to-real reinforcement-learning workflows. That matters for diligence because it turns the product from a closed appliance into a platform that customers, universities, and integrators can extend. The caveat is scope: the open ecosystem is deepest for Lite3 and newer platforms, while X20/X30/DR02 remain much more marketing-page-visible than SDK-visible.[CE011, CE012, CE013, CE014, CE015, CE016]

5.3 Deployment proof is strongest in utility inspection, but trust and support controls lag the locomotion story

The product is easiest to believe where the workflow is narrow, repetitive, and hazardous. Utility and tunnel inspection fit that pattern, and the reviewed cases show Deep Robotics repeatedly selling around terrain access, data collection, and reduced human exposure. The substation evidence is unusually specific for a private robotics vendor: official pages claim nearly thirty China Southern Power Grid substations, cost reduction of roughly 70 percent, efficiency gains above 50 percent, greater than 95 percent inspection-detection accuracy, and 96.5 percent recognition accuracy in a Ningxia desert deployment. Singapore Power's SPock deployment gives a second, more international workflow proof point: X30 patrols a 40 km power tunnel, detects leaks and thermal anomalies, and triggered a wider-scale deployment plan. Trust and quality signals are real but uneven. Hardware ruggedization is concrete, with IP66 or IP67 ratings across the industrial line and DR02's modular quick-detach design explicitly positioned as an O&M feature. At the same time, public compliance evidence is thin beyond ingress ratings and marketing claims of reliability. The support and privacy pages create real diligence flags: after-sales terms are stated for mainland China only, warranty exclusions are broad for harsh environments and unauthorized modification, and the privacy policy allows collection of logs, time, location, device configuration, and IP data with disclosure to authorities when legally required. In other words, the locomotion stack looks farther along than the public trust, security, and global support package.[CE027, CE028, CE029, CE030, CE031, CE032]

5.4 The roadmap is credible from quadrupeds into embodied navigation and humanoids, but commercialization proof thins quickly outside core inspection

The 2025 roadmap on the official surface shows real shipping progress rather than a vague future promise. Deep Robotics launched Lynx M20 in April, expanded utility and rescue case studies through mid-2025, unveiled DR02 in October as an industrial humanoid, and then used the December Series C announcement to tie products, DeepVLA 1.0, and an embodied-AI pilot production base into one scaling narrative. The same announcement also says new capital will fund more R&D, more talent, and continued quadruped plus humanoid development, which implies the company does not view humanoids as a side project. The main technical risk is not whether the robots can move; it is whether Deep Robotics can turn a strong locomotion and perception base into durable, supportable, and governable production systems across many geographies and product families. DR02 is still early in that respect: public evidence centers on product pages and expo demonstrations rather than named production deployments or public SDKs. DeepVLA 1.0 is strategically important, but the reviewed corpus still shows it mainly as a company-claimed enabling layer, not a publicly benchmarked field system. And while the company is clearly thinking about clustered robot-dog, drone, and fixed-sensor operations, the open-source and public documentation are materially stronger for Lite3 and M20 than for the entire commercial fleet. The technology thesis is therefore positive, but diligence should underwrite supportability, data governance, and deployment repeatability at least as hard as raw motion capability.[CE017, CE031, CE035, CE036, CE037, CE038]

5.5 Exhibits

6.1 Customer segments are real, but the evidence quality differs sharply by vertical

Deep Robotics is not selling to one monolithic buyer class. The public record supports at least four practical customer buckets. First, utilities and grid operators buy quadrupeds and wheel-legged robots for tunnel, substation, cable-gallery, and hazardous-inspection work; this is where the company has the most concrete named proof and the clearest operating outcomes. Second, public-safety and security buyers use robots for patrol, fire-hazard detection, perimeter monitoring, and incident response, with named references including Xining Public Security Bureau, Yiwu International Trade City, and Singapore's Certis industrial park. Third, emergency-response organizations and fire brigades appear in drills and scenario validation, which proves product-market fit for harsh environments but does not necessarily prove recurring revenue. Fourth, universities and research labs use Lite3 and related open platforms as education and R&D tools. The key diligence distinction is not whether these segments exist, but which are evidenced by independently corroborated operational deployment versus company-curated examples, partner-like showcases, or logo-and-reach marketing.[CU001, CU002, CU010, CU014, CU021, CU032]

6.2 Named customer proof is strongest in utilities, with Singapore Power as the anchor case

The best-supported customer proof is Singapore Power Group. Across the official SPock case study and multiple independent trade articles, the same core story holds: X30 patrols sections of SP Group's underground power transmission tunnel network, performs thermal and visual inspection tasks, required site-specific localization work, and is expected to save roughly 480 manual inspection hours per year if rollout expands. China Southern Power Grid is the next-most-important named account because Deep Robotics says its robots now serve nearly 30 substations with meaningful cost and efficiency improvement, but that scale is still disclosed through company materials rather than customer filings. A broader State Grid relationship is plausible and increasingly important, especially given 2026 utility procurement plans and Jiemian's report of a prior framework agreement, yet those signals still sit one level below SP Group in proof quality. Outside utilities, named deployments exist in security, conservation, warehousing, and Saudi outdoor inspection, but most of them remain official-only proofs with less disclosed commercial depth.[CU003, CU004, CU005, CU006, CU007, CU008]

6.3 Research and education adoption is real, but logos and reach metrics should not be overread

Deep Robotics also has a second customer motion that looks very different from industrial inspection. Lite3 and related open platforms are positioned for universities, labs, and developer teams that value modularity, SDK access, sensor openness, and lower price points. Official and third-party sources repeatedly cite Tsinghua, Peking University, University of Edinburgh, and University College London, which is enough to treat academic adoption as real. References to Tencent and Alibaba are weaker: they show platform interest or R&D usage, not the kind of production-deployment detail that would let an investor infer durable enterprise revenue. The same caution applies to the company's large top-line reach claims. Statements like 100+ clients in 26 Chinese provinces, 44 countries and regions, and 600+ industry scenarios are useful directional adoption signals, but they are not the same as a run-date paying-customer count, and they do not tell us how many accounts are active, repeat buyers, or revenue-significant. This is the chapter's main line between verified named proof and marketing breadth.[CU014, CU015, CU016, CU017, CU030, CU033]

6.4 Customer durability looks promising in utilities, but retention, concentration, and policy risk remain underdisclosed

The customer verdict is constructive but not fully underwritten. The positive read is that several deployments are clearly beyond demo-stage theater: SP Group operationalized SPock inside a mission-critical tunnel workflow, China Southern Power Grid reportedly expanded across many substations, and public-safety/security scenarios show that the robots can handle more than static inspection. The negative read is that Deep Robotics still does not disclose the metrics that convert usage proof into revenue-quality proof. There is no public NRR, GRR, churn, renewal, cohort, or top-customer revenue breakdown. International customer durability also depends on localization and support intensity: the Singapore case required months of adaptation, and the public mainland after-sales terms do not provide a clean picture of overseas SLA strength. Finally, U.S. lawmakers' 2026 push to limit Chinese-made quadruped and humanoid robots in federal settings is not a current Deep Robotics customer loss, but it is a real category-level blocker for future public-sector expansion in the United States.[CU022, CU024, CU025, CU026, CU027, CU028]

6.5 Exhibits

7.1 Regulatory and legal risk is now a real market-access constraint, not just a theoretical geopolitical overhang

Deep Robotics' most serious external risk is not a single export ban already aimed at the company; it is the widening lattice of U.S. procurement, supply-chain, and technology-control rules that can narrow addressable markets before a direct designation ever arrives. The FY2025 NDAA is especially relevant because it covers the exact components and system types Deep markets: LiDAR-dependent inspection systems and unmanned ground vehicles from covered foreign countries. At the same time, the June 2026 1260H notice shows that a close quadruped peer, Unitree, has already crossed from abstract scrutiny into formal U.S. military-company designation. Deep's own positioning compounds the exposure. Its official materials repeatedly target substations, tunnels, gas detection, emergency response, and public-safety workflows, which are precisely the kinds of environments where public buyers, defense-adjacent contractors, and critical-infrastructure operators apply elevated supplier screening. The practical implication is that Deep should be underwritten as a company whose best overseas opportunities are civilian industrial accounts willing to localize around Chinese hardware, not as a vendor with frictionless access to Western public-sector or defense budgets. A second legal-commercial constraint is certification. ANYbotics now advertises ATEX/IECEx Zone 1 readiness with Yokogawa-backed channel support into energy and hazardous industries, while the retained Deep public pack shows optional explosion-proof sensing but not a system-level certificate. That does not disqualify Deep from utilities and tunnels, where it already has strong proof, but it does create a clear boundary around the most demanding hazardous-process environments. The right diligence stance is therefore to assume regulation and certification will shape product-market fit, channel design, and valuation simultaneously rather than treating them as downstream compliance tasks.[CR001, CR002, CR003, CR004, CR005, CR006]

7.2 Operational proof is credible, but it is concentrated in customized utility deployments that depend on partners, trust, and field execution

Deep has stronger real-world operational evidence than many robotics startups. The public record shows a technically capable inspection platform: X30 is ruggedized for harsh environments, the substation case claims meaningful accuracy and labor savings, and the Singapore SPock deployment moved beyond demo language into a live tunnel patrol workflow with measurable time savings. That matters because it proves the company can solve expensive inspection problems in real infrastructure rather than simply shipping research hardware. However, the same evidence also reveals the operational risk profile. Deep's best references are customized utility and substation deployments that require route mapping, localization, software refinement, and tight workflow integration. The foreign flagship in Singapore required months of safety assessment and local tuning, which is exactly what an investor wants to see from a serious deployment but also exactly what makes expansion service-heavy and harder to scale than pure hardware revenue. Public proof is also concentrated. State-linked and utility-style accounts dominate the named reference set, and even the strongest overseas proof is still a partner-enabled pilot rather than a broad cohort of repeat deployments across multiple private operators. That concentration is amplified by dependency risk. Overseas delivery already relies on partners such as EGP, while high-trust hazardous inspection channels are being locked down by competitors with stronger certifications and industrial-software alliances. A further operational complication is cyber trust: the sector now faces customer skepticism about Chinese-origin robots in data-sensitive sites, especially after peer hardware-security concerns surfaced. Put simply, Deep's current moat is deployment know-how in harsh utility environments, but that moat is only as strong as its service bench, partner quality, and ability to clear customer trust screens account by account.[CR012, CR013, CR014, CR015, CR016, CR017]

7.3 The financial model has crossed into profitability, but the next stage mixes price pressure, capital intensity, and an unusually crowded 2026 financing window

The positive financial fact pattern is real. Multiple prospectus-based summaries agree that Deep reached roughly RMB 337 million of 2025 revenue, delivered gross margin above 52%, and turned profitable for the first time. That is materially better than the public evidence available for many private robotics names. The underwriting problem is what comes next. Every source that makes the business look more investable also makes it look more capital intensive. The company is now simultaneously pushing the Lynx family, international field deployments, and the DR02 humanoid program, while prospectus-linked coverage says the next financing step is meant to fund embodied-AI algorithms, large models, new robot bodies, and expanded manufacturing. In other words, today's profitable quadruped business is being asked to finance tomorrow's broader robotics platform. Competition makes that balancing act harder. Unitree remains the category's price and volume reference point, with better headline hardware benchmarks and much deeper unit scale. SemiAnalysis and other retained sources suggest Deep still prices at a premium to Unitree while lacking Unitree's shipment base and consumer or academic flywheel. At the same time, the China robotics IPO queue is becoming crowded: peers are also pushing toward public or quasi-public liquidity events, and industry commentary keeps returning to the same unresolved question of mass-production economics. The result is a risk stack that looks more like a hardware growth company than a pure software story: margins can compress, working capital can expand, and the valuation case can unwind quickly if the IPO window slips or if humanoid spend outpaces industrial cash generation.[CR024, CR025, CR026, CR027, CR028, CR029]

7.4 People and execution risk sits in organizational breadth: the thesis breaks if Deep cannot add compliance, service, and capital discipline as fast as it adds products

Deep's execution challenge is broader than product engineering. Its own company materials present a full-stack robotics organization spanning core components, quadrupeds, humanoids, manufacturing, sales, and service. That breadth is a strategic ambition, but it is also a people and control problem because the company must scale compliance, field support, certification work, partner management, and capital allocation at the same time. The strongest current evidence still sits in industrial quadrupeds for utilities and harsh inspection routes. The farther Deep moves from that lane into international channel building, hazardous-industry certification, and humanoid commercialization, the more the bottleneck shifts from technical possibility to management bandwidth. The most useful investor framing is therefore trigger-based. Some risks are manageable if they remain bounded: a service-heavy overseas rollout is acceptable if customer diversification improves; humanoid expansion is acceptable if quadruped margins stay healthy and the financing path remains open; national-origin screening is tolerable if Deep remains focused on markets where on-prem deployment and civilian workflows are enough. But the thesis breaks if these mitigations do not mature. The kill criteria are not exotic. They are concrete events: no certification path, no diversification beyond power utilities, visible margin erosion under price pressure, or IPO and financing delays while R&D intensity stays elevated. Until management proves that the organization can industrialize controls and channels as quickly as it industrializes robot bodies, execution risk should be treated as one of the load-bearing valuation constraints.[CR017, CR030, CR031, CR035, CR036, CR037]

7.5 Exhibits

8.1 Deep Robotics is deep into the IPO process, but it remains private and the valuation record is still noisy

Deep Robotics is no longer a concept-stage robotics startup looking for its first market-clearing event. By the run date, the company had completed IPO tutoring, had its STAR Market application accepted, and had prospectus summaries circulating through multiple Chinese and English-language outlets. That matters because it gives investors much more operating evidence than most private robotics companies can offer. At the same time, none of those steps means the company is already public. As of 2026-06-21, Deep Robotics still sits in a pre-listing state: private, filing in motion, and dependent on final review, pricing, and market-window conditions before liquidity is real. The valuation trail is therefore mixed rather than clean. Caixin described the company at roughly RMB 8 billion in December 2025, Humanoids Daily later framed the business at more than RMB 10 billion, and Pandaily described a potential post-listing valuation around $3.5 billion. Those are not equivalent reference points. Some are current-private or pre-IPO marks, some are media summaries of filing-era value, and some are headline expectations for a completed public transaction. The prudent underwriting stance is to separate what is observable from what is aspirational: the company is clearly IPO-bound, but the fully supported valuation range is materially narrower than the most promotional headlines imply.[CV001, CV002, CV003, CV004, CV005, CV040]

8.2 Commercial proof is real, but comparable multiples still argue for entry discipline rather than momentum pricing

The strongest argument for paying a meaningful premium over a generic industrial-hardware multiple is that Deep Robotics has crossed several thresholds that many robotics peers have not. Prospectus-derived reporting now shows RMB 337 million of 2025 revenue, RMB 28.68 million of net profit, 52.83% gross margin, heavy embodied-robot mix, and visible industrial deployment proof in power inspection and Singapore utility tunnels. Those are serious signals, not demo-theater. But valuation has to be framed against what other robotics marks already buy. Unitree is the closest private comparator because it shares a China base, legged-robot heritage, and IPO path; yet Unitree crossed RMB 1 billion of revenue and was still discussed at roughly RMB 12 billion private and a much more aggressive ~$7 billion IPO target. UBTECH provides a live Chinese public benchmark, with 2025 turnover of about RMB 2.0 billion and a multi-billion-dollar market value, while Boston Dynamics and Skild AI show how much optionality public and strategic investors can assign to category leaders or software-first platforms. Deep Robotics sits below those ceilings on scale and breadth. It has one sold humanoid unit in 2025, meaningful quadruped and Lynx traction, and a strong industrial wedge, but it also faces the same price pressure, geopolitical discounting, and humanoid-capex burden highlighted by adverse market sources. That mix supports a premium to undifferentiated hardware, but not a blank check.[CV006, CV007, CV008, CV009, CV010, CV011]

8.3 The evidence supports a track recommendation: buyable near disciplined private marks, speculative at aggressive IPO headlines

The key valuation question is not whether Deep Robotics is a good company in the abstract. It is whether the currently observable price corridor leaves room for attractive risk-adjusted returns. On trailing 2025 revenue of roughly $49.5-49.6 million, the December 2025 Caixin-style RMB 8 billion mark already implies around 22x sales. A north-of-RMB-10-billion interpretation pushes the company toward roughly 28x sales, effectively Unitree-style territory despite materially smaller scale and much less humanoid proof. A $3.5 billion post-listing headline would imply roughly 70x trailing sales, which is closer to software-style or speculative-humanoid-optionality pricing than to evidence-supported industrial robotics. That does not make the company uninvestable. It means valuation needs scenario discipline. In a bear case, hardware multiple compression, IPO delay, or margin normalization could drag fair value well below recent private headlines. In a base case, continued industrial growth plus credible but still early humanoid optionality can justify a low- to mid-$1 billions range. In a bull case, Deep Robotics would need not just continued quadruped execution, but sustained 2026 growth, stable high gross margins, and real evidence that humanoids can become a repeatable business rather than a narrative extension. That is why the recommendation is track rather than buy. There is enough proof to stay interested, but not enough to underwrite the most aggressive marks without demanding a sharper entry price or deeper diligence.[CV040, CV041, CV042, CV043, CV044, CV045]

8.4 Exit readiness is better than at most private peers, but public-market readiness is still not the same thing as valuation readiness

Deep Robotics looks more exit-ready than many private robotics names because it has already completed tutoring, has an accepted STAR Market filing, has disclosed 2025 profitability through prospectus summaries, and has a capital-use plan that fits its next-stage roadmap. That is materially better than relying on venture decks or unpriced rumor rounds. Even so, an IPO pathway is not the same as a fully diligenced investment case. BigGo’s prospectus summary is especially important here: it says investor valuation-adjustment rights can reappear if the IPO process stalls, which means preference and governance terms may still distort common-equity economics. Other unresolved issues are equally practical. The public record still does not show exact current post-money, preference stack, liquidation protections, customer concentration by revenue, repeat-order cohorts, warranty burden, service attach, or whether 2025’s margin profile survives a broader humanoid push. Regulatory and geopolitical friction also remain material. US procurement and export-control regimes continued tightening into 2026, while global competition intensified through crowded robotics IPO calendars and heavily funded AI-software challengers. The final investment question is therefore operational: can a buyer obtain enough private diligence to confirm that Deep Robotics is not only IPO-capable, but also worth paying a premium for before the market sets a public price? Until that answer is clearer, the right posture is to monitor actively, define kill triggers, and insist on evidence that closes the cap-table and quality-of-revenue gaps.[CV032, CV033, CV034, CV035, CV048, CV049]

8.5 Exhibits