From narrowing losses to scenario implementation: iFlytek advances in independent control and application expansion

$iFLYTEK(002230.SZ) Recently, iFLYTEK released its semi-annual earnings forecast, with several financial indicators exceeding our previous expectations.

First, the net loss narrowed: the company expects a net loss of 200 million to 280 million yuan in the first half of 2025, a year-on-year reduction of 120 million to 200 million yuan; the non-GAAP net loss was 320 million to 400 million yuan, narrowing by 30% to 50% compared to the same period last year; revenue and gross profit increased by 15% to 20% year-on-year, maintaining a strong growth rate.

Notably, the company's sales collections in the first half totaled approximately 10.3 billion yuan, a year-on-year increase of about 1.3 billion yuan; operating net cash flow increased by over 700 million yuan compared to the same period last year, a rise of about 50%. Given the market conditions the company primarily faces, the significant improvement in collections is highly beneficial for cash flow stability, maintaining steady operations, and enhancing risk resilience.

Behind this earnings forecast is the breakthrough achieved by leading Chinese AI companies through strategic resilience amid the intersection of independent technological advancements and industrial applications. In many people's impressions, aside from the emergence of Deepseek, there seemed to be no particularly impressive industry progress in the Chinese AI sector in the first half of the year. However, in reality, the Chinese AI industry is accelerating its evolution from L2 to L3 at a globally leading pace, giving rise to many noteworthy innovations.

Looking at the development history of generative AI in recent years, there is a clear phase division. In OpenAI's framework, AI applications are categorized into five levels from L1 to L5. L1 represents conversational bots, the initial product form of ChatGPT. With the emergence of Deepseek, although the product form remains conversational, its essence has achieved a leap, reaching the L2 stage, where it can provide complete logic and help solve problems through simple reasoning.

Now, competition in the L1 and L2 stages is nearing its end, and the market's focus will shift to L3-level AI applications, particularly their expansion in enterprise settings. According to publicly available bidding information, large model-related projects in China experienced explosive growth in the first half of 2025: the cumulative number of winning bids reached 1,810, with a total value exceeding 6.4 billion yuan, surpassing the full-year total for 2024, indicating a surge in market demand.

Additionally, full-stack autonomy became another focal point in the Chinese AI industry during the first half of the year.

Starting from February and March of this year, NVIDIA founder Jensen Huang has publicly acknowledged China's technological prowess in AI on multiple occasions, praising Huawei as one of the world's strongest tech companies. Particularly at the Taipei Computer Show in May, he openly stated that the U.S. semiconductor blockade against China is doomed to fail, recognizing the innovation capabilities of Chinese domestic companies. On international platforms, a significant portion of news related to Jensen Huang revolves around this topic.

Amid Jensen Huang's effusive praise, Huawei's Ren Zhengfei, who had not given a public interview in six years, spoke with People's Daily in June, immediately addressing Huawei's Ascend chips. He stated that China's chip technology is still one generation behind the U.S., but China can compensate for this gap with mathematics—leveraging its strong power supply to build larger computing centers to bridge the technological disparity in individual chips.

Domestic computing power adaptation and scenario expansion are two developmental trajectories that few Chinese AI companies can simultaneously fulfill, and iFLYTEK is one of them. Its Spark large model is the only general-purpose large model in China fully trained on Huawei's Ascend 910B chips. Through joint R&D, the training efficiency of the Ascend 910B has improved from initially 30%-50% of NVIDIA's A100 to 85%-95%. In Jensen Huang's view, this already constitutes a computing ecosystem capable of challenging NVIDIA.

The Spark Deep Inference Large Model X1 (70B), released in January this year, received a significant upgrade on April 20. Despite having model parameters an order of magnitude smaller than comparable industry models, it achieves performance on par with top international models (e.g., DeepSeek-R1 671B) while significantly lowering the threshold for enterprise AI privatization and customization.

The realization of "full-stack autonomy" must ultimately be reflected in practical applications. According to the China Large Model Bidding Project Monitoring and Insights Report, the second quarter of 2025 saw 1,295 large model-related winning projects through public channels, with disclosed amounts totaling approximately 3.96 billion yuan. Among these, iFLYTEK surpassed Baidu and Volcano Engine to rank first in both the number of winning bids and total value.

iFLYTEK's application breakthroughs are particularly evident in the central and state-owned enterprise markets. The Kunlun large model, co-developed with China National Petroleum Corporation, improved code development efficiency by 73%, with industry knowledge Q&A adoption rates exceeding 95%. The intelligent bidding assistant developed for China Energy Investment Corporation received praise from the State-owned Assets Supervision and Administration Commission. The "Haineng" AI model, developed in collaboration with CNOOC, has been applied in areas such as bidding and procurement and employee health, with ongoing R&D for specialized scenarios like offshore oilfield production stabilization.

In the TOC business, rapid growth continued. AI hardware based on the "iFLYTEK Spark Large Model" saw a 42% year-on-year sales increase during the "618" shopping festival. Among these, iFLYTEK's AI learning devices ranked first in sales on JD.com and Tmall for the third consecutive year during the "618" period, with semi-annual revenue in 2025 doubling year-on-year.

This is iFLYTEK's "uniqueness"—it is the only company conducting application R&D on a domestic computing platform while achieving scaled commercial applications in both enterprise and consumer markets. The day after the earnings forecast, the company's stock price climbed from an intraday low of -3% to close at 46.98 yuan per share, a 0.36% increase for the day, reflecting capital market recognition of its semi-annual performance.

In the near future, the capital market will once again seek the main driver for further growth, and AI is highly likely to remain that driver. Autonomy and commercial expansion will bring sustained fundamental improvements, further propelling iFLYTEK's growth.$NVIDIA(NVDA.US)