Youdao Abandons the Red Ocean of Hardware: Zhou Feng's AI Track-Shifting Experiment
This fall's back-to-school season has seen a more brutal promotion war in the education hardware market than in previous years. The classic Xueersi X5 model has been cut from 4,899 yuan to 3,799 yuan on e-commerce platforms, marking a 22% drop; the Xiaoyuan X6 and iFlytek S30 have also rolled out substantial discounts. NetEase Youdao, which once secured a solid market position with its dictionary pens and learning tablets, has stayed noticeably quiet amid all this hustle and bustle.
Youdao frames this quietness as a proactive strategic retreat. However, based on financial data, the consecutive decline in hardware revenue can hardly be fully explained by "proactivity". Its net revenue in the second quarter hit 1.47 billion yuan, posting a meager 3.5% year-on-year increase; its attributable net profit turned around to 73.8 million yuan in profit from a loss of 17.8 million yuan in the same period last year, with operating profit reaching 111.5 million yuan, surging nearly 2.9 times year on year; its overall gross margin rose from 43.0% to 48.9%. While profit improvement is an undeniable fact, so is the near-stagnation of revenue growth. What Youdao needs to prove to the capital market is not that "we stopped making hardware because hardware business is not good", but that "after scaling back hardware business, can the new growth engine pick up the slack".
The Scale and Boundary of Hardware Business Contraction
Youdao has indeed made a large-scale cut in its smart hardware business. Smart device revenue in the first quarter dropped 42.6% year on year; in the second quarter, it further fell 31.5% to 86.8 million yuan, with the gross margin declining from 41.5% in the same period last year to 32.8%, which the company attributed to rising material costs. Against the industry backdrop, this is not an entirely isolated choice.
Data from Luo Tu Technology shows that in the first half of 2026, the total channel sales volume of learning tablets in China reached 2.287 million units, down 14.2% year on year; sales revenue hit 7.72 billion yuan, down 10.2% year on year. The decline accelerated in the third quarter: sales volume stood at 1.25 million units, plummeting 30.8% year on year; sales revenue reached 4.5 billion yuan, down 17.1% year on year. After large model capabilities became widely accessible, obvious homogenization emerged in the experience of hardware products, and leading players could only seize market share by cutting prices. However, the cost side moved in the opposite direction: AI computing power demand pushed up storage prices, and some manufacturer insiders revealed that memory prices once rose 7 to 10 times; storage chips account for up to about 56% of the BOM cost of learning tablets, while the mainstream selling price range is locked at 2000-2999 yuan. The rational response of manufacturers is to reduce storage capacity rather than raise prices.
The gross margin gap is clear: the gross margin of learning services in the second quarter was 65.5%, while that of hardware was 32.8%. This explains the direction of contraction, but cannot independently prove that the timing and pace of contraction are active choices rather than results of competitive squeeze. A more accurate statement may be that the red ocean of hardware makes "letting go of the obsession with shipment volume" an easy account to calculate, but the boundary between "proactive retreat" and "forced contraction" is not so clear in the company's narrative. A noteworthy detail is that Youdao still launched the dictionary pen X8 series in the second quarter, and claimed that it topped the sales revenue list on JD.com and Tmall during the 618 shopping festival; the brand new AI hardware OpenPods earphones are also under development. Hardware business has not been completely abandoned: what has been scaled back are low-margin, high-shipment-dependent tablet products, while sub-categories with relatively controllable gross margin and stronger brand barriers are retained. This is more like a selective defensive retreat rather than a one-size-fits-all exit.
AI Subscription Growth Slows Down, LobsterAI Faces User Conversion Test
After cutting back hardware business, Youdao has bet its growth chips on AI subscriptions. The sales revenue of AI subscription services exceeded 100 million yuan for the first time in the first quarter, growing by more than 70% year on year; the sales revenue in the second quarter was about 100 million yuan, growing by more than 20% year on year. It is worth noting that "more than 20%" is a lower limit expression, and the actual growth rate may be close to 20%, forming a clear gap with the growth rate of over 70% in the first quarter. In vertical scenarios, Youdao has already reaped the dividends of the first wave of high-willingness users: Academic Chacha's sales revenue in the first quarter increased by 200% year on year, AI Simultaneous Interpretation's sales revenue in the first quarter increased by more than 100% year on year, and Hi Echo's sales revenue in the second quarter doubled year on year. However, the user base of these scenarios is limited in itself. Although the payment penetration rate is increasing, the ceiling of absolute scale has also emerged. It is difficult for AI subscriptions to grow from "breaking 100 million yuan" to "doubling again" relying on linear growth in vertical scenarios.
To break through this ceiling, a more popular application that can generate continuous user dependence is needed. This is the background for LobsterAI to be pushed to the forefront. As China's first 100% open-source desktop office Agent, LobsterAI was launched in February 2026. Its number of users exceeded 1 million in the second quarter, making it one of the top four office Agent products in China; the next-day retention rate of active users exceeded 80%, and the per capita daily Token consumption exceeded 20 million. It abandons seat fees, charges based on actual usage, and remains open-source without locking the ecosystem.
The risks of this model are equally obvious. Open source lowers the threshold for use, and also lowers the threshold for payment. Token billing sounds flexible, but whether users will pay for every conversation depends on whether the office Agent really becomes a daily habit. The 80% next-day retention rate indicates that "users who come will not leave immediately", but retention does not equal payment, and high-frequency use does not equal continuous payment. There is no public data on how much subscription revenue can be converted from 1 million users. The more realistic competitive environment is that ByteDance's "Doubao Work", Tencent, Alibaba and Baidu have all entered the office Agent track, which is no looser than the hardware track. LobsterAI's open source and security narrative can help Youdao gain some developers and young users in the early stage, but when large manufacturers push products through ecosystem bundling, "no locking" and "pay-as-you-go" alone are not enough to constitute a reason for users to pay. At the stage when office Agent habits have not yet been formed, user scale itself does not equal subscription revenue.
Advertising Cost Reduction Delivers Results, Cyclical Dependence Remains
In the transition period when hardware business is receding and AI subscriptions are still climbing, the advertising business is the foundation supporting Youdao's profit turnaround in its financial statements. Online marketing service revenue in the second quarter reached 584.4 million yuan, a slight decrease of 7.7% year on year, which the company explained as proactive contraction to focus on high-ROI business; the gross margin rose from 25.8% in the same period last year to 28.7%. AI acts as an efficiency amplifier here: the iMagicBox platform integrates more than 2000 global localized digital human avatars, reducing the production cost of advertising materials by about 70% compared with manual production, and also improving the click-through rate. In an environment where advertisers pay more and more attention to ROI, efficiency improvement has directly boosted profit margins.
However, the weakness of the advertising business remains unchanged: it is highly dependent on macroeconomic prosperity and advertiser budget cycles. The 7.7% year-on-year decline in revenue in the second quarter itself indicates that "proactive contraction of low-margin business" and "weakening demand side" may exist at the same time. The company's narrative tends to emphasize the former, but the impact of the latter cannot be completely ruled out. The current profit improvement relies to a large extent on a cyclical business rather than a new growth engine. Taking advertising as the absolute bottom support for profits means that Youdao's profit quality will fluctuate with the general environment. This is not a perfect second growth curve, but more like a cash flow source for the transition period. Its value lies in buying time for the growth of AI subscriptions, but it itself does not solve the problem of "when the new engine will take over".
Zhou Feng's Agent Formula: Engineering Path and Strategic Narrative
Zhou Feng uses "Agent = Model + Harness" to explain Youdao's path selection. He believes that the model only completes the "thinking" part, while Harness is responsible for making thinking understandable, collaborative, reproducible and operable for a long time, including context management, tool calling, memory, state persistence, evaluation, loop control and permission governance; in a complex Agent, the model may only account for 20% of the value, and the peripheral engineering system accounts for 80%. He further put forward a seven-point methodology: design for the next generation of models, prioritize the development of high-intelligence products, make every token cost-effective, take context engineering as the main task, make tools the product interface for models, drive development with evaluation, and start with a single Agent by default.
This framework is valid in engineering, and it is also a kind of strategic narrative: it concentrates resources on the application layer and engineering layer, avoiding the capability competition of general large models. Youdao did not spare no effort to develop trillion-parameter general models, but upgraded "Ziyue 4" to multi-modal capabilities in vertical fields, strengthened visual mathematical and complex reasoning, and then encapsulated it into application matrices such as LobsterAI. Youdao's bet is that in strong scenarios such as education, office and advertising, the capability gap of general models will eventually be narrowed, and the barrier lies in who can package the model into a reliable and reproducible workflow.
However, the judgment that "Harness accounts for 80% of the value" is still a proposition to be verified by payment, rather than a proven fact. If general model companies also do a good enough job in the Harness layer, or highly standardized Harness components appear in the open source ecosystem, Youdao's relative advantages accumulated in the engineering layer may be compressed. Whether scenario data is truly irreplaceable, and whether engineering capabilities are sufficient to constitute a reason for users to pay, depends on whether LobsterAI can enable enough users to move from "usable" to "indispensable" under the premise of free and open source. This leap has not happened yet.
Two Unverified Assumptions for Valuation Re-rating
Youdao is currently in an awkward period of switching between old and new growth drivers: the old hardware engine is decelerating, while the new AI subscription engine is accelerating, so the total revenue growth rate has stagnated, reaching only 3.5% in Q2. Healthy bookkeeping is a fact: it has achieved positive operating profits for eight consecutive quarters. As of the end of June, its cash and short-term investment balance reached 849.3 million yuan, and the net operating cash inflow in the current quarter was 334.2 million yuan. In the capital winter, the ability to generate blood by itself is a fundamental card in hand.
The capital market is revaluing Youdao with the SOTP (Sum of The Parts) method: the advertising business contributes the profit bottom line, and AI subscriptions determine the valuation flexibility. However, this valuation framework relies on two assumptions that have not been fully verified. First, whether the advertising business can continuously contribute stable profits — the 7.7% year-on-year decline in revenue in the second quarter, and the increase in gross margin came more from structural optimization and AI cost reduction rather than revenue expansion, so the "stability" of the profit bottom line needs to be discounted. Second, whether LobsterAI can convert free open-source users into continuous paying users — 1 million users, 80% next-day retention rate, and per capita daily consumption of 20 million Tokens, these indicators prove the depth of use rather than the willingness to pay. Before the paid conversion data comes out, the valuation flexibility of AI subscriptions is more of an option value rather than a cash flow value.
The direction from "selling hardware" to "selling Tokens" is clear. But track switching cannot be completed just by announcing a strategy, it depends on organizational inertia, user mindset, and the most practical thing: how many people are willing to pay real money for every conversation under the free and open source model. The decisiveness of hardware contraction is worthy of recognition, the cash flow supported by advertising is worthy of cherishing, and the Harness narrative is logically self-consistent in engineering. But all these conditions combined only give Youdao an admission ticket for track switching, not a final destination ticket. Before getting the next piece of the market cake, Youdao must first maintain balance on the tightrope between profitability and growth — and this tightrope has not yet reached the opposite bank.
This article is from the WeChat official account "Edu Guide", written by Edu Guide, and published by 36Kr with authorization.