Youdao Abandons the Hardware Red Ocean: Zhou Feng's AI Track-Switching Experiment

This fall's back-to-school season has brought a more brutal promotional battle in the education hardware market than in previous years. The classic Xueersi X5 model dropped from 4,899 yuan to 3,799 yuan on e-commerce platforms, a 22% cut; the Xiaoyuan X6 and iFlytek S30 also rolled out steep discounts. NetEase Youdao, which once carved out a place for itself with dictionary pens and learning tablets, looks quiet amid the clamor.

Youdao frames this quiet as a deliberate strategic retreat. But judging by the financial data, the consecutive declines in hardware revenue are hard to explain entirely as "deliberate." Second-quarter net revenue was 1.47 billion yuan, up only 3.5% year on year; net profit attributable to shareholders swung from a loss of 17.8 million yuan in the same period last year to a profit of 73.8 million yuan, while operating profit reached 111.5 million yuan, nearly 2.9 times the year-earlier figure; overall gross margin rose from 43.0% to 48.9%. The profit improvement is a fact, and so is the near-stagnation in revenue. What Youdao needs to prove to the capital markets is not that "hardware was bad, so we stopped," but that "with hardware gone, can the new engine catch us?"

The scale and boundaries of the hardware pullback

Youdao's retrenchment in smart hardware is indeed substantial. Smart device revenue fell 42.6% year on year in the first quarter; in the second quarter it dropped a further 31.5% to 86.8 million yuan, with gross margin falling from 41.5% a year earlier to 32.8%, which the company attributed to rising material costs. Against the industry backdrop, this is not an entirely isolated choice.

Data from RUNTO Technology shows that in the first half of 2026, China's learning tablet sales across all channels totaled 2.287 million units, down 14.2% year on year; sales revenue was 7.72 billion yuan, down 10.2% year on year. The decline accelerated in the third quarter: sales of 1.25 million units, a year-on-year plunge of 30.8%; revenue of 4.5 billion yuan, down 17.1% year on year. After large-model capabilities became widely accessible, hardware product experiences became markedly homogenized, leaving leading players to compete for share only by cutting prices. Costs, meanwhile, moved in the opposite direction: AI computing demand pushed up memory prices, with one industry source saying memory prices at one point rose seven- to tenfold; memory chips can account for as much as about 56% of a learning tablet's BOM, while the main selling price band is locked at 2,000–2,999 yuan, so the rational response for manufacturers is to cut capacity rather than raise prices.

The gross margin gap is clear: learning services had a gross margin of 65.5% in the second quarter, versus 32.8% for hardware. That explains the direction of the contraction, but it cannot by itself prove that the timing and pace of the contraction were a deliberate choice rather than the result of competitive squeeze. A more accurate framing may be: the hardware red ocean made "giving up the obsession with shipment volume" an easy calculation to make, but the boundary between "deliberate retreat" and "forced contraction" is not so clear in the company's narrative. One detail worth noting is that in the second quarter Youdao still launched the Dictionary Pen X8 series and said it took the top spot in sales on JD.com and Tmall during 618; the new AI hardware OpenPods earbuds are also in the works. Hardware has not been entirely abandoned: what is being scaled back is the tablet-type product line with low margins and heavy reliance on shipments, while niche categories with relatively controllable gross margins and stronger brand barriers are being retained. This looks more like a selective fallback than a wholesale exit.

After cutting hardware, Youdao is staking its growth on AI subscriptions. First-quarter AI subscription sales exceeded 100 million yuan for the first time, up more than 70% year on year; second-quarter sales were about 100 million yuan, up more than 20% year on year. Note that "more than 20%" is a floor statement; actual growth may be close to 20%, a marked drop from the over-70% growth in the first quarter. In vertical scenarios, Youdao has already captured the first wave of high-intent users: Xueshucha sales rose 200% year on year in the first quarter, AI simultaneous interpretation sales rose more than 100% year on year in the first quarter, and Hi Echo sales doubled year on year in the second quarter. But the user base in these scenarios is itself limited; even as paid penetration rises, a ceiling on absolute scale emerges. Getting AI subscriptions from "breaking 100 million" to "doubling again" will be hard to achieve through linear growth in vertical scenarios.

Breaking through that ceiling requires an application with broader mass adoption that can generate sustained dependence. That is the backdrop for bringing LobsterAI to the fore. As China's first 100% open-source desktop-level office Agent, LobsterAI launched in February 2026, surpassed 1 million users in the second quarter, and ranks among China's top four office Agents; next-day retention of active users exceeds 80%, and per-user daily Token consumption exceeds 20 million. It dispenses with seat fees, charges by actual usage, and remains open-source without locking users into an ecosystem.

The risks of this model are equally obvious. Open source lowers the barrier to use, and also lowers the barrier to paying; Token-based billing sounds flexible, but whether users pay for every conversation depends on whether the office Agent truly becomes a daily habit. Next-day retention of 80% shows that "those who came did not leave immediately," but retention does not equal payment, and high-frequency use does not equal sustained payment. How much subscription revenue can be converted from 1 million users is not publicly disclosed. The competitive environment is more realistic still: ByteDance's Doubao Work, Tencent, Alibaba and Baidu have all entered the office Agent space, a track hardly less crowded than hardware. LobsterAI's open-source and security narrative can help Youdao win some developers and young users early on, but when big tech companies push products through ecosystem bundling, "no lock-in" and "pay-as-you-go" alone are not enough to constitute a reason to pay. At a stage when office Agent habits have yet to form, user scale itself does not equal subscription revenue.

Ad cost cuts pay off, but cyclical dependence remains

During the transition period when hardware is receding and AI subscriptions are still ramping up, the foundation supporting Youdao's return to profitability is the advertising business. In the second quarter, online marketing services revenue was 584.4 million yuan, down 7.7% year-on-year, which the company explained as a deliberate contraction to focus on high-ROI business; gross margin rose from 25.8% in the same period last year to 28.7%. AI plays the role of an efficiency multiplier here: the iMagicBox platform integrates more than 2,000 globally localized digital human avatars, reducing the cost of producing ad creatives by about 70% compared with manual work, while click-through rates have also improved. In an environment where advertisers increasingly value ROI, efficiency gains directly improve profit margins.

But the soft spot of the advertising business remains unchanged: it is highly dependent on macroeconomic conditions and advertisers' budget cycles. The 7.7% year-on-year revenue decline in the second quarter itself shows that "deliberately contracting low-margin business" and "weakening demand" may exist simultaneously; 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 considerable extent on a cyclical business rather than a new engine. Treating advertising as the absolute backstop for profit means that Youdao's earnings quality will fluctuate with the broader environment. This is not a perfect second curve, but more like a source of cash flow during a transition period. Its value lies in buying time for AI subscriptions to ramp up, but it does not itself solve the question 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 choice. He believes that the model only completes "thinking," while the Harness is responsible for making thinking understandable, collaborative, reproducible and able to run over the long term, including context management, tool invocation, memory, state persistence, evaluation, loop control and permission governance; in a complex Agent, the model may account for only 20% of the value, while the peripheral engineering system accounts for 80%. He further laid out a seven-point methodology: design for next-generation models, prioritize high-intelligence products, make tokens worth spending, treat context engineering as the main task, treat tools as the product interface shown to the model, drive development with evaluation, and start with a single Agent by default.

This framework holds up in engineering terms, and it is also a strategic narrative: it concentrates resources on the application layer and engineering layer, avoiding the capability race in general-purpose large models. Youdao did not stubbornly pursue trillion-parameter general models; instead, it upgraded "Ziyue 4" into vertical-domain multimodal capabilities, strengthening visual-mathematical and complex reasoning, and then packaged it into application matrices such as LobsterAI. Youdao's bet is that in strong scenarios such as education, office and advertising, the capability gap between general models will eventually be closed, and the barrier lies in who can package models into reliable, reproducible workflows.

But the judgment that "the Harness accounts for 80% of the value" is still a proposition awaiting paid validation, not a verified fact. If general model companies also make the Harness layer good enough, or if highly standardized Harness components emerge in the open-source ecosystem, Youdao's relative advantage accumulated at the engineering layer could be compressed. Whether scenario data is truly irreplaceable, and whether engineering capability is enough to constitute a reason for users to pay, depends on whether LobsterAI can, under the premise of being free and open source, move enough users from "usable" to "indispensable." That leap has not yet happened.

Two unverified assumptions behind the valuation re-rating

Youdao is currently in the awkward period of switching between old and new growth drivers: the old engine, hardware, is decelerating, while the new engine, AI subscriptions, is accelerating, so total revenue growth has stalled, at only 3.5% in Q2. The healthy books are a fact: operating profit has been positive for eight consecutive quarters, cash and short-term investments totaled 849.3 million yuan as of the end of June, and net operating cash inflow in the quarter was 334.2 million yuan. In a capital winter, being able to generate its own cash flow is itself a trump card.

The capital market is revaluing Youdao using SOTP: the advertising business provides the profit floor, and AI subscriptions determine valuation elasticity. But this valuation framework relies on two assumptions that have not been fully verified. First, whether the advertising business can continue to contribute stable profits—second-quarter revenue fell 7.7% year-on-year, and the gross margin improvement came more from structural optimization and AI cost reduction than from revenue expansion, so the "stability" of the profit floor needs to be discounted. Second, whether LobsterAI can convert free open-source users into paying subscribers—1 million users, 80% next-day retention and average daily token consumption of 20 million per person prove depth of usage, not willingness to pay. Before payment conversion data emerges, the valuation elasticity of AI subscriptions is more option value than cash flow value.

From "selling hardware" to "selling tokens," the direction is already clear. But changing tracks cannot be completed just by announcing a strategy; it depends on organizational inertia, user mindset and, most practically, how many people are willing, under a free and open-source model, to pay real money for every conversation. The decisiveness of the hardware contraction deserves recognition, the cash flow backstopped by advertising is worth cherishing, and the Harness narrative is also internally consistent in engineering terms. But taken together, these conditions still only give Youdao a ticket to enter the track switch, not a ticket to the destination. Before it can eat the next piece of cake, Youdao must first keep its balance on the tightrope between profitability and growth—and on this tightrope, it has not yet reached the other side.

This article comes from the WeChat public account "Edu Guide," author: Edu Guide, published by 36Kr with authorization.