Agora delivered a fourth consecutive quarter of GAAP profitability in Q3 2025, with revenue up 12% year-over-year to $35.4 million and net income of $2.7 million (a 7.8% margin). The core real-time engagement business rebounded across both the overseas Agora division (+15.9%) and China's Shengwang (+8.4%), with dollar-based net retention improving for a fourth straight quarter. Gross margin dipped slightly to 66% as the company leaned into conversational AI, launching Engine 2.0 and Studio and hosting a record 3,000-plus attendee conference. Management guided Q4 revenue to $37-38 million and reaffirmed a target of full-year GAAP operating profit in 2026, while conversational AI remained in early, largely pre-revenue adoption.
Thanks, Operator. Welcome everyone to our earnings call. I'll first reveal our operating results from the past quarter. We're pleased to report our fourth consecutive quarter of GAAP profitability in Q3, supported by double-digit revenue growth and expanding margins. Total revenue in Q3 reached $35.4 million, up 12% year-over-year. Our GAAP net profit for the quarter was $2.7 million, with a GAAP net margin of 7.8%. We expect our revenue and net profit to continue growing on a quarter-over-quarter basis in Q4. As you can see, our core real-time engagement path business is rebounding strongly and is on track to deliver its first full-year revenue growth since the pandemic, providing a stable, profitable foundation for us. At the same time, we are significantly increasing our investment in conversational AI. Voice-based human-machine interaction is not new, yet most conversational AI solutions today still disappoint users. Why?
Because building voice agents that can converse naturally with a human is just hard. Just a few months ago, Greylock Partners, a leading venture capital firm, published a blog post titled, "Voice Agents: Easy to Use, Hard to Build." They nailed the core challenge. Behind the simplicity users expect lies immense complexity: system obstruction, real-time audio processing, latency management, and compliance requirements. Consider the issue of background noise and multiple speakers. Just two of the many technical challenges. In a real-world setting, like a busy home, office, or car, clean audio is the exception, not the norm. A voice agent must accurately isolate a user's voice from overlapping speech and ambient sounds. Without this, transcription becomes unreliable. Intent is misunderstood, and the agent's reasoning falters, undermining the whole interaction. Furthermore, as Andrew Tapassi has pointed out, there is often a significant gap between a working demo and a production-ready product.
Conversational AI is no exception. For instance, in our discussion with customers and prospects, many have expressed frustration with the reliability and scalability of current solutions, especially when users are distributed across geographies or when concurrent usage is high. Our investment in conversational AI is specifically aimed at addressing these challenges. Recently, we launched our conversational AI Engine 2.0. It integrates over a decade of advanced audio research and development, including AI-powered noise suppression, acoustic echo cancellation, a proprietary audio codec, and adaptation across thousands of device types to ensure that AI hears and speaks with consistent clarity. In addition, the engine also tackles core interaction challenges: selective attention, turn-taking, interruption handling, emotion detection, and natural conversational flow. In short, we're not just providing the transmission pipeline for voice and video. We're building the behavioral intelligence that powers truly responsive, human-like conversational AI agents.
To help developers build voice agents more easily, we announced our conversational AI Studio at our recent Convo AI and RTE Conference in late October, which allows developers to create, configure, and deploy voice agents through a zero-code interface. Complementing this, our conversational AI Benchmark and Orchestration Platform allows developers to evaluate, mix and match, and optimize both our proprietary and third-party modules so they can identify the best-performing combination for their specific use case. Our Open-Source TEN Framework designed for building voice agents continues to gain traction in the developers' community. Recognized for its high concurrency architecture and deep cross-platform integration, it has been adopted by multiple cloud providers and major enterprises for their agent orchestration platforms. All these products are backed by our global distributed real-time inference cloud.
Over the past several months, we've expanded this infrastructure to cover key regions across North America, South America, Europe, and Asia, ensuring consistent latency, reliability, and performance even under high concurrency and varying network conditions. Early adoptions from customers around the world have been encouraging, and our pipeline of use cases and prospects continues to grow as we head into next quarter. Our recent Convo AI and RTE Conference attracted more than 3,000 on-site attendees, a record for us, and made it the largest gathering focused on conversational AI technology globally. Our customers and developers are deploying our conversational AI solutions to build voice agents for outbound marketing, inbound customer service, tutoring, and many other applications. Power manufacturers are also integrating our technology into smart toys, enabling voice-powered companionship and learning experiences.
In conclusion, the convergence of advanced AI models and robust real-time infrastructure is unlocking a new era of possibilities. Backed by proven scalability, deep technology expertise, and a forward-looking product suite, we're well-positioned to empower this next chapter, enabling truly human-like, reliable, and scalable voice agents. With that, let me turn things over to Jingbo, who will reveal our financial results.
Thank you, Tony. Hello, everyone. Let me start by first reviewing financial results for the third quarter of 2025, and then I will discuss outlook for the fourth quarter. Total revenues for the third quarter reached $35.4 million, up 12% year-over-year, representing a third consecutive quarter of double-digit organic growth. If we look at the two business divisions, Agora revenues reached $18.2 million in Q3, representing 15.9% year-over-year growth and flat quarter-over-quarter. The strong year-over-year growth reflects our successful market penetration and growing adoption in verticals such as live shopping. Shenghua revenues reached RMB 122.4 million in Q3, up 8.4% year-over-year and 6% sequentially, driven by continued business expansion and adoption in key verticals such as social, entertainment, and IoT. Dollar-based net retention rate is 108% for Agora and 90% for Shenghua, marking the fourth consecutive quarter of improvement for both businesses.
Gross margin for the third quarter was 66%, slightly decreased 0.7% year-over-year and 0.8% sequentially. Moving on to expenses, R&D expenses were $13.8 million in Q3, decreased 52.8% year-over-year. R&D expenses represented 39.1% of total revenues in the quarter, compared to 92.7% in Q3 last year. Sales and marketing expenses were $6.5 million in Q3, decreased 5.6% year-over-year. Sales and marketing expenses represented 18.3% of total revenues in the quarter, compared to 21.7% in Q3 last year. G&A expenses were $5 million in Q3, decreased 48.4% year-over-year. G&A expenses represented 14.1% of total revenues in the quarter, compared to 30.8% in Q3 last year. Moving on to the bottom line, we delivered net income of $2.7 million in Q3, representing a 7.8% net income margin. This result represents a significant improvement from last year and marks our fourth consecutive quarter of GAAP profitability.
Based on our current business momentum and visibility into the fourth quarter, we expect net income to grow sequentially compared to Q3. Now turning to cash flow, operating cash flow was $0.7 million in Q3, compared to negative $4.6 million last year. Moving on to balance sheet, we ended Q3 with $374.3 million in cash, cash equivalents, bank deposits, and financial products issued by banks. Net cash outflow in the quarter was mainly due to share repurchase of $4.8 million. In the third quarter, we repurchased 5.2 million ordinary shares or 1.3 million ADS, representing 1.4% of our outstanding shares at the beginning of the quarter. Since our board approved the share repurchase program in February 2022, we have repurchased $132.1 million worth of shares through September 30, 2025.
The share repurchase program demonstrates our dedication to returning value to our shareholders, balanced with our ability to continue investing in strategic growth opportunities. Now turning to guidance for the fourth quarter of 2025, we currently expect total revenues to be between $37 million and $38 million, compared to $34.5 million in the fourth quarter last year, representing year-over-year growth rate of 7.2% -10.1%. This outlook reflects our current and preliminary views on the market and operational conditions, which are subject to change. In closing, I would like to express my gratitude to our outstanding teams in Agora and Shenghua. Our sustained double-digit revenue growth and profit expansion are a direct reflection of your hard work and strategic focus. To our shareholders, thank you for your continued trust. We remain focused on executing our roadmap to build a durable, market-leading company at the forefront of AI innovation.
Thank you all for joining today's call. Let's open it up for questions.