This earnings season has created a clear divide in the financial reports of AI companies. While both remain unprofitable, some firms continue to see losses widen, relying on grander narratives to support their valuations; others show narrowing losses alongside improvements in revenue structure and cash flow. Behind this divergence lies a shift in the yardstick investors use to judge AI companies—they no longer ask "how large is your model," but rather "what exactly are customers paying for?" This question points to a deep-seated transformation in the AI industry's business model, as pricing mechanisms move from "selling solutions" to "selling productivity."

For the past decade, "selling solutions" has been the dominant model in the AI industry: project-based, customized, with revenue recognized once per contract. This model has three inherent flaws: revenue is discontinuous, requiring a constant search for new orders; gross margins are repeatedly eroded by customization costs; and scaling often leads to proportional increases in delivery headcount—making the business bigger without necessarily thickening profits. "Selling productivity" follows a different logic: instead of paying for algorithm licenses and hardware deployment, customers pay for the actual output AI generates in real business operations. Revenue thus gains continuity, repeat purchases and upsells follow, and it can even grow in tandem with customer business volume. This shift has roots in industry trends. As foundation models continue to improve, raw technical metrics are becoming commoditized, moving the competitive focus from "model capability" to "deployment capability." The rise of embodied intelligence has accelerated this process—when AI enters physical world operations, customers naturally demand to be billed for the tasks completed. Policy and market signals are also converging. Embodied intelligence has been included in the nation's "15th Five-Year Plan" as a priority future industry, with policy support intensifying; in June, the Ministry of Industry and Information Technology and the State-owned Assets Supervision and Administration Commission launched a special initiative for humanoid robots and embodied intelligence real-scenario training, encouraging model training and real machine data collection in real-world settings; in August, the National Development and Reform Commission clarified the goal of "letting robots iterate technology in real scenarios and form application feedback loops around real demand," while also warning the industry to "prevent blind imitation and herd behavior." The signal is clear: companies that fail to deliver tangible output and offer only concepts will struggle to secure a ticket to the next cycle.

For investors, this change can be read directly from financial statements. If a company is truly charging for "output," three signs typically appear in its reports: rising contract liabilities, indicating customers are willing to prepay for undelivered output; improving gross margins, suggesting a growing share of standardized, replicable revenue; and converging receivables and cash flow, showing growth no longer relies on capital advances. Conversely, "high growth" without any of these three indicators likely still represents the old project-based story.

Applying this framework to this year's interim reports, RECONOVA (07656.HK) provides a case worth unpacking. There are two reasons for choosing this company as a sample. First, timing: it listed on the Hong Kong Stock Exchange on July 8, earning the market label "first visual embodied intelligence stock," and the interim report disclosed on August 26 is its first post-listing results, framed consistently with its prospectus and offering good comparability. Second, structure: its business spans both ends of this transformation—with a fourteen-year-old visual intelligence "solution" segment and a robotics business just entering real-scenario validation, the same report captures both the legacy of the "old model" and the early form of the "new model."

Looking at the overall picture: first-half revenue was approximately RMB 102 million, up 58.7% year-on-year; by segment, smart civil aviation, smart commerce, and smart safety driving revenue grew by 319.1%, 95.5%, and 41.2%, respectively—while the aviation segment's absolute first-half revenue was RMB 2.791 million, mainly driven by the concentrated recognition of a few cross-period projects in H1, as this segment's revenue has historically been confirmed in the second half, especially Q4, so seasonality should be noted. On the profit side, net loss narrowed by 6.7% year-on-year; notably, this occurred against a backdrop of R&D spending up 51.2% year-on-year, with R&D expenses accounting for 48.8% of revenue. In other words, the narrowing loss was not achieved by cutting investment.

Now cross-referencing those three indicators. First, contract liabilities rose from RMB 2.237 million at the end of 2025 to RMB 11.634 million, an increase of roughly 420%. For a company that started with solution-based business, higher customer prepayments signal improved order visibility and suggest some clients are paying in advance for "future output" rather than "delivered projects." Second, gross margin improved from 21.9% to 26.2%. The 4.3-percentage-point gain is not dramatic, but the direction matters: under a project-based model, margins typically fluctuate with customization level, so sustained improvement generally points to a rising share of standardized, reusable products in revenue. Third, working capital is converging. Trade receivables and notes fell 13.0% from year-end, and operating cash outflow narrowed from RMB 80.093 million to RMB 32.814 million—while growing, both advance funding and cash burn are decreasing, the opposite of the typical "burning cash for growth" path. All three indicators are met. Of course, numbers alone are insufficient for conclusions; the business structure behind them must also be examined.

Breaking it down, there are three notable aspects to RECONOVA's approach. First, in the "selling solutions" era, it turned solutions into reusable legacy assets. Founded in 2012, the company has spent fourteen years deeply cultivating aviation, commercial spaces, and safe driving scenarios: by 2025 revenue, it ranks first in China's civil aviation visual intelligence product market with an 8.7% share, its products cover over 60% of airports with ten-million-plus annual passenger flow, and it has served over 1.5 billion passenger trips cumulatively; in smart commerce, it serves over 100 enterprise clients, including 20 of the top 40 Chinese commercial real estate developers; and in smart safe driving, over 600,000 commercial vehicles carry its intelligent terminals, covering more than 7 billion kilometers cumulatively. This base of existing customers and installed devices forms the foundation of recurring operating revenue—exactly the structure that serves as the prerequisite for shifting from "selling solutions" to "selling productivity."

Second, in extending toward "hands-on" operations, it chose measurable production metrics over demo effects. Its luggage transfer robot, "Xiao Yi," has entered a real flight support environment at a ten-million-passenger hub airport in East China, conducting routine POC operations. RECONOVA did not pursue one-step full replacement but created a human-machine collaboration model: standard luggage is handled by robots, while long-tail items like soft bags, irregular pieces, and damaged luggage are handled by staff. Public testing data shows "Xiao Yi" processes a single piece of luggage in under 18 seconds, with 99.9% loading accuracy and 7×24 continuous operation—cycle time, accuracy, and continuous working hours are precisely the units on which future "output-based billing" will depend. As Founder and Chairman Zhan Donghui put it: "In the past, we were making the 'eyes' and 'brain.' Now we are moving forward, toward the 'hands.'"

Third, R&D investment remains focused on a reusable capability foundation. In April this year, the company iteratively launched the VTFLA end-to-end embodied large model, which adds tactile and force feedback to the traditional VLA framework, with a core design philosophy of realizing over 80% of intelligence on the edge and adopting a fast-slow collaborative dual-system design to decouple and coordinate task planning from real-time control—"seeing" a box does not mean knowing how much force to use to grasp it steadily, and such capabilities can only be pressure-tested in real operations: scenarios generate data, data feeds the model, and R&D spending compounds accordingly. It should be noted that RECONOVA's 48.8% R&D ratio is not scattered across areas, but concentrated under the principle of "extreme focus, extreme concentration" on a productivity foundation that can be reused across scenarios. At the 2026 World Robot Conference in August, the company fully presented a luggage transfer solution comprising "Xiao Yi," a wheeled dual-arm humanoid robot, and an intelligent transfer robot. According to its roadmap, the wheeled dual-arm robot is expected to launch in 2027 and commercialize in 2028. The overall solution's operational scope will expand from fixed-position luggage handling to mobile loading/unloading beneath aircraft, and further into warehousing, industrial automation, and other fields. The company has also signed a strategic cooperation agreement with Shanghai Industrial Technology to jointly build a Hong Kong embodied intelligence super accelerator, driving industrialization and international expansion; overseas, it has already launched POC trials, with "Xiao Yi" directly targeting European and Japanese markets, avoiding data compliance risks through hardware product form, and exploring new pathways for Chinese AI companies going global.

Of course, this case study is far from conclusive. RECONOVA remains in a strategic loss phase, with distance between POC and scale revenue; aviation revenue recognition is concentrated in H2, so the full-year picture will only become clear in Q4; and robotics business currently contributes limited direct revenue. The long-term value realization of embodied intelligence will require continuous validation in scenario after scenario. But the shift from "selling solutions" to "selling productivity" has already begun in how value is priced. For AI companies, the true watershed lies not in how grand a story they tell, but in whether their financial statements make the transition first—the trajectory of contract liabilities, gross margins, and cash flow will reveal, earlier than any grand narrative, who is preparing for this leap.