Chasing iFLYTEK

07/01 2024 400

"China must establish an independently controllable general large model foundation, which is the key to achieving general artificial intelligence and aligning with international standards."

Recently, API (Application Programming Interface) developers have gradually received "disconnection" notification emails from OpenAI. It is reported that starting from July 9, 2024, OpenAI will begin to block API services from unsupported countries and regions, including mainland China and Hong Kong, China.

The "disconnection" notification has stirred up market nerves, adding new variables to the competitive landscape of the domestic large model industry. The hardest hit are undoubtedly the "shell" developer teams that once relied on OpenAI technology.

However, some people have seen opportunities. While OpenAI blocks API access in China, the domestic large model market has launched a wave of beachhead operations, introducing their respective "migration" plans at a lightning-fast speed.

On June 26, iFLYTEK announced the launch of the "iFLYTEK Spark API Universal Benefit Policy," including the Spark Lite version being permanently free, the Spark Pro version offering a free trial for 1 month, the Spark Max version giving away 100 million tokens for free, the first publicly released Spark 4.0 Ultra version offering buy-one-get-one-free on the first order, and providing technical expert 1V1 support services.

In fact, the iFLYTEK Spark large model has another "game-changing" advantage. Starting from the V3.5 version, all models are trained based on "Feixing No. 1," China's first purely domestic computing power foundation, representing genuine "Made in China." It is reported that "Feixing No. 1" is an independently controllable large model computing power foundation platform based on the Ascend ecosystem, jointly developed by iFLYTEK and Huawei.

In the end, this is a race for domestic large models to catch up with internationally leading general large models.

Large Model Competition

With the advent of ChatGPT in November 2022 as a representative, artificial intelligence has undergone a directional change, shifting from discriminative AI to generative AI. The latter's underlying technology is the large model, which has rapidly sparked a global competition in large model technology.

Entering 2023, there has been a wave of hype surrounding pan-AI concepts in China. Since the beginning of 2023, iFLYTEK's share price has soared to a historical peak of 80.68 yuan per share, and its market value has risen to 170 billion yuan.

Grasping the opportunity, iFLYTEK has gradually shifted its R&D focus towards AI in 2023, even announcing its all-in on large models. In May 2023, iFLYTEK officially launched the "Spark Cognitive Large Model" and continues to update and iterate it.

On June 27, 2024, Spark has been upgraded to version V4.0. Liu Qingfeng stated at the press conference that Spark V4.0 has surpassed GPT-4 Turbo in terms of text generation, language understanding, knowledge question answering, logical reasoning, and mathematical abilities.

However, while the "national team" is striving to catch up, GPT is also constantly updating and progressing. Recently, Mira Murati, OpenAI's chief technology officer, revealed that GPT-5 may be released by the end of 2025 or early 2026, reaching a "doctorate-level" intelligence in specific tasks.

It can be predicted that after the launch of GPT-5, the industry will usher in another significant leap, and the pursuit of domestic large models will continue.

Liu Qingfeng also mentioned in the upgrade press conference that once GPT-5 is released, the gap will be narrowed to about one year, as they have more computing power, data, and larger models. "But we are confident in continuing to catch up quickly."

What is the Source of Tragedy?

The game of the brave will continue, but on the eve of the Spark V4.0 upgrade press conference, iFLYTEK was shrouded in a wave of public opinion.

Recently, news of an employee of iFLYTEK dying suddenly from overtime work has spread on social media. According to online messages, the deceased employee was a senior tester at iFLYTEK who suddenly felt unwell at home after work. The family called 120 and sent the employee to the hospital, but the rescue was unsuccessful. Later, according to an internal notification from iFLYTEK, the news of the employee's death was confirmed.

In fact, this is the second employee of iFLYTEK to die suddenly from a heart attack in the past year. In August 2023, media learned from the PR department of iFLYTEK that an employee suddenly felt unwell at home, and after being sent to the hospital by family members, was diagnosed with a heart attack, ultimately resulting in death despite rescue efforts.

The frequent occurrence of similar tragedies has sparked heated discussions among netizens. Some iFLYTEK employees have expressed that "iFLYTEK is very intense, overtime work is a culture, and if the company says there is something wrong at midnight, you have to go and work overtime until early morning."

In recent years, the death of big company employees due to overwork has not been an isolated incident, reflecting issues such as companies increasing task targets and working hours for employees to enhance efficiency.

Apart from excessive overtime, another hot topic in the internet circle in recent years is layoffs.

iFLYTEK is no exception. As of the end of 2023, iFLYTEK had a total of 14,356 employees, a decrease of 690 employees compared to the end of 2022. Among them, 283 sales personnel, 373 technical personnel, 5 financial personnel, 96 administrative personnel, and 9 management personnel were reduced.

All accidents have their inevitable causes behind them. Layoffs, business cuts, and welfare reductions are aimed at cost reduction, while overtime, additional tasks, and increased targets are aimed at efficiency enhancement. "Tightening the belt" has become a corporate consensus regardless of industry or stage.

And for iFLYTEK, which is in the "golden mine" of the AI race, this phenomenon will only be more apparent. With the continuous upgrading and iteration of its large model products, performance pressure is now prominent.

Worst Net Profit Performance

In 2021, iFLYTEK proposed a revenue target of "achieving 100 billion yuan in 2025." This determination mainly stemmed from the fact that iFLYTEK had achieved revenue growth of over 25% for 10 consecutive years, a performance that is among the best in the entire A-share listed companies.

However, after entering 2022, iFLYTEK's high growth came to an abrupt halt, with revenue growth almost stagnating in the past two years. In 2022, iFLYTEK's revenue was 18.82 billion yuan, with a sharp drop in growth to 2.77%. Although there was a rebound in 2023, it was still a single-digit growth, increasing by 4.41% year-on-year to 19.65 billion yuan, achieving less than 20% of the 100 billion yuan revenue target.

In the first quarter of 2024, iFLYTEK regained some dignity, with revenue growing by 26.27% year-on-year. However, compared to the growth needed to achieve the 100 billion yuan revenue target in 2025, a 26% growth rate is clearly not enough.

When iFLYTEK proposed its 2025 revenue target of 100 billion yuan in 2021, it also set specific targets for its various businesses.

Among them, the education segment is expected to achieve revenue of 30 billion yuan, the medical and elderly care segment is expected to achieve 20 billion yuan, the open platform and consumer segment is expected to achieve 20 billion yuan, and the three directions of automotive business, operators, and smart cities are each expected to achieve 10 billion yuan.

However, combining the 2023 financial report data, iFLYTEK's education segment achieved revenue of 5.852 billion yuan, the open platform and consumer segment achieved revenue of 6.185 billion yuan, operator-related businesses achieved revenue of 2.096 billion yuan, the smart city segment achieved revenue of 3.933 billion yuan, the smart automotive segment achieved revenue of 696 million yuan, and the smart medical segment achieved revenue of 540 million yuan, all significantly falling short of the target values.

Apart from the "deceleration" in revenue, iFLYTEK's profitability is also concerning.

After a year-on-year decline of 63.94% in net profit attributable to shareholders in 2022 to 561 million yuan, iFLYTEK's net profit attributable to shareholders in 2023 increased by 17.12% year-on-year to 660 million yuan, seemingly showing a significant rebound, but mainly relying on government subsidies for "embellishment."

In 2023, iFLYTEK's government subsidies included in current profits and losses were 404 million yuan, accounting for 61.21% of the net profit for the period. Excluding non-recurring gains and losses, iFLYTEK's net profit after deducting non-recurring gains and losses decreased by 71.74% year-on-year to 118 million yuan, the lowest level in the past decade.

This situation continued into the first quarter of 2024, with the amount of government subsidies for the period being 400 million yuan, a decrease of 59.57% compared to the same period in 2023. The reduction in government subsidies further expanded iFLYTEK's losses for the period. In the first quarter of 2024, iFLYTEK suffered a net loss of 300 million yuan, a significant expansion compared to the 57.9 million yuan loss in the same period last year, also the worst first-quarter performance since iFLYTEK went public 15 years ago.

Reviewing previous financial reports, it can be seen that iFLYTEK has relied on government subsidies for a long time. There are still doubts in the market about whether iFLYTEK is a "subsidy-dependent enterprise."

Combining previous data, since 2016, iFLYTEK's government subsidies have accounted for a high proportion of net profit. From 2016 to 2022, iFLYTEK's government subsidies included in current profits and losses were 128 million yuan, 77 million yuan, 276 million yuan, 412 million yuan, 426 million yuan, 439 million yuan, and 473 million yuan, accounting for 180.28%, 104.16%, 50.83%, 50.24%, 31.23%, 28.21%, and 84.31% of the net profit for the respective periods.

It can be seen that even with external subsidies for "blood transfusion," as iFLYTEK continues to increase investment in R&D, the company's inability to generate its own profits may be difficult to hide.

In its first-quarter report for 2024, iFLYTEK stated that the losses were mainly due to the company's active seizing of the historical opportunity of general artificial intelligence and its firm investment in general artificial intelligence cognitive large models. In the first quarter of 2024, the company invested approximately 300 million yuan in large model research and development, as well as in achieving autonomy and controllability of core technologies and the industrial chain, and in expanding the industrial application of large models.

Liu Qingfeng said that 2023 was the "Spark Year," with increased R&D investment and less market investment, while 2024 is the "Spark Prairie Fire," with increased R&D investment and also increased market investment. Clearly, continuous capital injection remains the main theme of iFLYTEK's 2024 strategy.

"In the future, the company's investment in basic large models may account for 30% of R&D investment, and the proportion of investment in applications will increase. The scale of R&D investment around basic large models will still grow with the increase in the scale of industrial implementation."

What's Next?

As the saying goes, "The golden nuggets are revealed after the waves wash away the sand." After enduring the brutal baptism of the "Hundred Model War" in 2023, domestic general large models entered the second half in 2024. Nowadays, only those with solid strength can remain in the competition.

Apart from Spark, domestic large models such as Wenxin Yiyan, Tongyi Qianwen, Tencent Hunyuan, Kimi, and Doubao are also constantly catching up, possessing the ability to compete with ChatGPT.

However, while there have been significant breakthroughs in model capabilities, computing power challenges, and other issues, there is still a significant gap in the commercialization of domestic large models.

In 2024, large models began to truly enter the first year of commercialization, and several leading domestic technology companies have begun to accelerate the deployment of large model applications in various fields such as finance, automobiles, operators, and government affairs.

Liu Qingfeng once revealed that iFLYTEK's Spark large model has achieved leading implementation in various industries such as automobiles, home appliances, and robots, becoming the first brand in these fields.

Liu Qingfeng has planned three commercialization paths for Spark. The first is to empower C-end hardware, the second is to achieve implementation in scenarios such as education, healthcare, and automobiles, which involves B-end service empowerment and large model customization, and the third is to sell API services through the iFLYTEK open platform.

Overall, with the iterative updates of Spark, the effectiveness of large models is becoming apparent. However, continuous large investments have made the company's profitability a challenge, and it still relies on external "blood transfusions" for relief.

More urgently, iFLYTEK is facing financial strain.

In the first quarter of 2024, the net cash flow from operating activities was -1.384 billion yuan, compared to -1.668 billion yuan in the same period last year. As of the end of the first quarter of 2024, the company's monetary funds were 2.678 billion yuan, while non-current liabilities maturing within one year reached 1.24 billion yuan. Looking back at 2021-2023, iFLYTEK's monetary funds were 5.87 billion yuan, 4.346 billion yuan, and 3.563 billion yuan, respectively, reflecting a significant burn rate.

Wind data shows that iFLYTEK has conducted several targeted share issuances, raising approximately 10.48 billion yuan. The last issuance occurred in 2021. As the entire industry enters the first year of commercialization, iFLYTEK is bound to face a "tough battle." So, what is the next step?

Solemnly declare: the copyright of this article belongs to the original author. The reprinted article is only for the purpose of spreading more information. If the author's information is marked incorrectly, please contact us immediately to modify or delete it. Thank you.