news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

In 2000, the second year of iFLYTEK's founding, IDG was interested in investing in iFLYTEK.

However, after listening to his subordinates' introduction, Chairman Xiong Xiaoge said: "It looks lively but it doesn't make money." In the end, the two sides failed to reach an agreement.

Now, more than 20 years have passed, and iFLYTEK has long since grown into a member of the national AI team, but Xiong Xiaoge's evaluation still seems to be coming true.

According to the performance forecast previously released by iFLYTEK, in the first half of this year, its net profit attributable to shareholders of the parent company plummeted by 616.50% to 725.24% year-on-year, turning from a profit of 73.572 million yuan in the same period last year into a loss of 380 million to 460 million yuan.

It is worth mentioning that this is the first time that iFLYTEK has reported a semi-annual loss since its listing in 2008.

Turn profit into loss

In its earnings forecast, iFLYTEK attributed its "sudden" losses in the first half of this year mainly to its greater investment in large models.

"In the first half of 2024, the company's new investment will exceed 650 million yuan in large-scale model research and development, independent control of core technologies and industrial chain, and the implementation and expansion of the large-scale model industry," iFLYTEK wrote.

In the middle of last year, iFlytek officially released the iFlytek Spark large model. For it, this part of the investment is of course critical. "iFlytek's technical strength is its core competitiveness, and the large model can help it go further," said Yu Fenghui, a special researcher at the China Financial Think Tank.

With the understanding that "all industries are worth redoing with big models", iFLYTEK's business span is quite large, from education to medical care, from automobiles to finance, and its presence can be seen in almost every vertical field.

The advantage of doing this is that the probability of success is higher, but the disadvantage is that the cost is also higher - iFLYTEK has to adapt to these fields one by one, which naturally means more money.

It can also be seen from the financial report that after announcing its embrace of the big model at the end of 2022, iFLYTEK's R&D investment reached 3.827 billion yuan last year, a year-on-year increase of 14.36%.

In 2021 and 2022, iFLYTEK's R&D investment increased by 21.50% and 14.28% respectively. Last year, its R&D investment accounted for 19.53% of its revenue, an increase of 1.7% from 2022; and in 2022, this proportion increased by 1.8% from the previous year.

Behind this bet on big models, not only is iFLYTEK's inevitable choice to stay at the forefront of technology, but it also reflects iFLYTEK's urgency to break through bottlenecks.

"iFlytek's speech recognition and synthesis business has fully contributed to the company's performance, but in other AI fields, iFlytek lacks mature applications and profit points," said Pan Helin, a member of the Information and Communication Economy Expert Committee of the Ministry of Industry and Information Technology. Therefore, finding new growth points is a necessary step for iFlytek.

However, the company's sales expenses are growing faster than iFlytek's R&D expenses. As can be seen from the figure below, from 2021 to 2023 and in the first quarter of this year, iFlytek's sales expenses grew 1.22%, 7.56%, 1.38%, and 2.87% higher than its R&D expenses year-on-year, respectively.

existChatGPTIn 2022, the year of its birth, iFLYTEK's sales expenses even surpassed its R&D expenses. In that year and 2023, iFLYTEK's sales expense ratio (16.81%, 18.24%) was also higher than its R&D expense ratio (16.53%, 17.72%).

iFLYTEK is quite frank about this. At the first quarter performance briefing this year, Chairman Liu Qingfeng also introduced that as the business scale expands, the company's cost investment is also increasing. In addition to the relatively large R&D expenses, iFLYTEK has also increased its market investment this year.

Still not good at making money

The question is, what did such a huge investment bring in return?

On the one hand, it is the iteration of technology. In May last year, iFlytek Spark Cognitive Model was officially released, and then it went through three rounds of iterations, rushing all the way to benchmark GPT-3.5.

On the other hand, under iFLYTEK's long-standing "platform + track" strategy, iFLYTEK Spark was able to be put into different test fields.

"These industry application scenarios provide iFLYTEK with rich data resources and scenario verification opportunities, which help to continuously optimize and improve its big models, and also enable its big models to be closer to actual application needs," Jiang Han, a senior researcher at Pangu Think Tank, told China Newsweek.

"iFlytek Spark has become the basic platform for the company's artificial intelligence industrialization, and has gradually begun to play an increasingly important supporting and leading role in the company's various business tracks and open platforms." iFlytek wrote in last year's financial report.

However, China Newsweek noted that iFlytek listed a number of benefits that iFlytek Spark has brought to the company's various businesses, such as the joint release with China Mobile of "Business Shorthand" that allows business people to easily transcribe and summarize calls, the Spark Teacher Assistant that helps teachers prepare lessons more easily, and the government affairs big model that makes "one network for all management" more feasible. However, iFlytek did not mention how much benefit these big models can bring to the company while upgrading the original business.

The reason I didn’t mention it might be because it doesn’t look that good.

Last year, iFLYTEK, with the support of big models, made great strides in the fields of education,Smart CityThe revenue of the 2019 and 2020 operators segments declined to varying degrees. Among them, the revenue of education and smart cities, as the two major performance pillars of iFLYTEK, decreased by 8.85% and 11.52% respectively.

"Although the Spark large-scale model launched by iFlytek has received a good market response, it cannot significantly boost the overall performance due to its limited share in the B-end market," said Qu Fang, an investment consultant at Wanlian Securities. "In recent years, iFlytek has taken an increasingly large share in the C-end market, and the large-scale models in the C-end market do not have strong technical barriers, so the competition is more on price, user experience and channels."

In other words, the customized large models provided by iFLYTEK for B-end customers are unlikely to bring significant growth in performance in the short term; and while the C-end market demand can drive revenue growth by raising prices for smart hardware in the short term, it may be difficult to sustain large-scale sales in the long term.

The result last year was that iFlytek's R&D and sales kept up, but its revenue did not. In 2023, iFlytek's R&D expenses and sales expenses increased by 11.89% and 3.27% year-on-year respectively, but its revenue only increased by 4.41% over the previous year.

In 2021, Liu Qingfeng set a goal for iFLYTEK to achieve "one billion users, one hundred billion in revenue, and a ten billion ecosystem" by 2025. However, by the end of 2023, iFLYTEK's revenue had not yet crossed the 20 billion yuan threshold.

This also means that iFlytek, which has already been questioned for being "not good at making money", faces greater profitability challenges.

How difficult is it to make money with big models?

Investors' confidence in iFlytek is also being shaken.

Since the beginning of this year, iFlytek's stock price has fallen by 20.11%. Even though iFlytek forecasted that its revenue and gross profit in the first half of this year are expected to grow by 15% to 20%, and the growth rate of gross profit is higher than the growth rate of revenue, the capital market did not pay for this rare growth when the net profit was negative.

In fact, the commercialization of large models is still uncertain.

At present, there are two main monetization paths explored by big model companies: one is to develop paid applications for C-end users, such as Miaoya Camera, which charges 9.9 yuan, and Wenxin Yiyan, which charges 49.9 yuan for a continuous monthly subscription of professional version; the other is to target B-end users, including providing API interface services, embedded intelligent solutions, customized applications combined with the industry, and cloud services based on big models, for exampleBaichuan IntelligenceThe closed-source large model charges API interface call fees for the B-side, Huawei launched Pangu 3.0 that can match vertical fields such as government affairs/finance/manufacturing/meteorology, and cloud platforms such as Baidu and Alibaba provide enterprises with a series of services such as models, computing power, and development tools.

But as the heat of the Hundred Model Contest gradually faded, a series of problems were exposed.

The first problem is that the technology is not mature enough. "Although domestic large-scale models have made certain progress in technology, there are still many technical problems to be solved, such as model performance and stability," Chen Yixin, director of CIC Consulting, once pointed out to China Newsweek.

Previously, a math problem of primary school difficulty, "Which is bigger, 9.11 or 9.9?", stumped a number of AI models at home and abroad. A media outlet tested 12 large models on this question, and the results were, includingOpenAIofGPT-4oAll 9 major models including the one given above got the answer wrong.

In addition, the degree of understanding of scenarios by large model companies, as well as the problem of homogenization of a large number of large models and corresponding products on the market, have also added more uncertainty to the companies' path to commercialization.

"The pace of AI technology innovation is getting faster and faster, which requires companies to continuously invest in research and development. However, under the existing technology framework, it is difficult for (companies) to launch breakthrough products or services and cannot meet the rapidly changing needs of the market," Zhang Xiaorong, director of the Deep Technology Research Institute, told China News Weekly. "Without seeing obvious advantages, current investment may not necessarily lead to technological breakthroughs and increased market share, thereby driving performance growth."

This is also the reason why Zhang Xiaorong called large-scale model technology "the forbidden area of ​​a few giants" - it is capital-intensive and technology-intensive, with huge investment and very little output, and the industry competition is more intense than other fields, and the risks are extremely high.

Pan Helin believes that although iFLYTEK has a certain amount of technical accumulation, its basic conditions such as large model data, computing power, and algorithms have not formed outstanding barriers. In addition, its AI technology lacks originality and often adopts a follow-up strategy. Therefore, it still has a long way to go to realize future industry application scenarios.

"Technology is only part of success. Whether or not we can really capture the market ultimately depends on whether the product is easy to use and whether the price is reasonable," Yu Fenghui also said.

References:

Time Business Interview, November 14, 2019, Time Weekly

Author: Shi Hanxu