AI Education 2026: Duolingo's 23% DAU Growth, Chegg's 99% Collapse, and Who Captures Value in the Chain

Three distinct "AI lines" matter when reading edtech financials: direct AI subscription revenue, AI-embedded hardware revenue, and AI-driven cost savings. Duolingo is the textbook case of the third: its Q2 2026 print (Aug 5) showed per-Video-Call inference cost falling from $0.30 to under $0.01 via open-source models, underpinning a raise of FY adjusted-EBITDA guidance to 26.5%; revenue was $298.5M (+18%) with 58.7M DAUs (+23%) and a 100M-DAU-by-2028 ambition. Yet the stock sits ~76% below its May 2025 high above $544, and the April 2025 "AI-first" memo backlash forced a public walk-back. At the other end, Chegg — its paid homework bank commoditized by free frontier models — closed at $0.96 on Jul 24, 2026, down ~99% from peak after two rounds of layoffs. Chinese listcos monetize the second line: TAL's content-solutions (device-led) share rose from 16.3% (FY23) to 31.8% (FY25) of revenue; Youdao's AI subscription sales topped RMB100M in a single quarter (+70% YoY) in Q1 2026. Mixing these three lines is the sector's biggest reporting trap.

The July 2021 "double reduction" closed the capital path for academic tutoring but not household willingness to pay for personalized learning. Five years on, three compliant channels have emerged: hardware sales (learning devices sit outside tutoring regulation and legally embed AI-tutor capability), non-academic content and practice scenarios, and in-school G/B-side services. Hardware leads: IDC counts 1.54M learning tablets shipped in Q2 2025 (+44.6% YoY), with iFlytek first by sales value that quarter and its device revenue doubling in H1; AVC's January 2025 share data shows a TAL (28.7%) / Zuoyebang (24.6%) / iFlytek (16.4%) top three, and Yuanfudao's Xiaoyuan practice device passed 1M cumulative units by April 2025. Content compliance tightened in parallel: the Aug 2023 GenAI Interim Measures require filing, and the MOE's two guidelines of May 12, 2025 established a school "whitelist" — banning unsupervised open-ended generation for primary students and copying AI output as answers. For minors-facing products, pedagogy guardrails are the entry ticket — compliance is not a cost line but a structural moat for domestic players.

In 2025–26 the growth engine shifted from consumer downloads to institutional procurement. The landmark contract is California State University: signed with OpenAI in January 2025 at ~$17M for year one across 22 campuses and ~460k students and staff — the largest US higher-ed AI deployment — then renewed in 2026 at $13M/year for three years, cutting per-seat cost from ~$37 to ~$28 and revealing a volume-discount pricing curve. A 2025 systemwide survey found 95% of responding students had used AI tools (84% ChatGPT), yet 82% worry about job security and some faculty petitioned against renewal. Anthropic's Claude for Education (launched April 2025 with a Socratic Learning Mode) spans Northeastern (50k across 13 campuses), LSE, Syracuse, Dartmouth, Columbia and Stanford, plus a Feb 2026 CodePath partnership reaching 20k+ community-college and HBCU students. Nation-scale procurement began too: Estonia's AI Leap serves 20k high-schoolers and 3k teachers from Sept 1, 2025, expanding by 38k students in 2026. Meanwhile free tiers compress B-side pricing — ChatGPT for Teachers free until June 2027; Gemini for Education bundled at no cost, with free training for all 6M US educators — and South Korea's AI-textbook reversal (over 1.2 trillion won spent before losing legal textbook status) stands as procurement's biggest cautionary tale.

Outcome evidence split sharply in 2025–26 — and the dividing line is consistent: pedagogy guardrails. On the positive side, a World Bank RCT in Nigeria (≈800 students, six weeks of teacher-supervised GPT-4 after-school English) found +0.31 SD overall (+0.23 SD in English), cost-effectiveness equal to 1.5–2 years of business-as-usual schooling, ranking among the strongest rigorously evaluated education interventions; a Harvard RCT (Kestin et al., Scientific Reports 2025, n=194) found students learned more than twice as much in less time with a purpose-built AI tutor. On the negative side, MIT Media Lab's "Your Brain on ChatGPT" preprint (2025, n=54) found the weakest EEG connectivity in the LLM group and coined "cognitive debt" — though small samples limit generalization; a Penn-led RCT in Turkish high schools (2024) found unguarded GPT-4 practice help boosted practice scores but left unassisted exam scores ~17% below control, an effect largely eliminated by a guardrailed "GPT Tutor". Products and regulators are converging on the same answer: OpenAI's Study Mode (Jul 29, 2025), Anthropic's Learning Mode, and China's whitelist plus no-copying rules all wrap generation in guided scaffolds. These findings remain contested, vary by sample and context, and constitute no educational advice.