Listing | 妙投APP
Author | Dong Bizheng
头图 | 动画《机器人总动员》
Since the release of ChatGPT, the world has set off a feast of AI large models.
Lei Jun once said, "Standing in the wind, pigs can fly."
If what Lei Jun said is not a lie, then iFLYTEK should think carefully about why it did not fly.
On April 22, iFLYTEK has released its 2023 annual report, and its performance is lower than expected. In 2023, the company will achieve an operating income of 19.650 billion yuan, a year-on-year increase of 4.41%, lower than the company's original expected revenue of more than 20 billion yuan, and in 2023, the company will achieve a net profit attributable to the parent company of 657 million yuan, a year-on-year increase of 17.12%.
It is worth noting that in 2023, the company is expected to achieve a non-net profit of 118 million yuan, a year-on-year decrease of 71.74%. At the same time, the company also disclosed a quarterly report that the net profit attributable to the parent company in Q1 of 2024 will be -300 million, a year-on-year decrease of 418.99%.
In 2023, when AI broke out, why didn't iFLYTEK enter the fast lane of rapid growth?
First, the old country is still there
First, let's start with the business model of AI.
As a tool to reduce costs and increase efficiency, AI is more integrated with various industries to empower various industries. For example: AI + education, AI + healthcare, smart city (AI + city management), etc.
Therefore, what kind of industry AI companies choose as the landing scenario also determines their competitive landscape and growth ceiling.
From the perspective of AI's three elements, computing power, algorithms, and data, the accumulation of industry-specific data by AI companies and the algorithms formed by know-how of the industry are the key factors that can be successfully implemented in the industry.
According to the annual report disclosed by iFLYTEK, in recent years, the company has built seven core businesses around five major industry fields, including smart education, smart healthcare, smart city, smart automobile, smart finance, operator, platform and consumer business, relying on the "platform + track" strategy.
From the perspective of the layout of the track, the company has achieved obvious results in education, urban management, operators and other businesses.
Before the outbreak of AI large models, the company's smart education, smart city, operator business, open platform and consumer business accounted for 34.11%, 23.62%, 11.14% and 24.66% of the company's revenue in 2022.
The reason for this can be summarized in two aspects: first-mover advantage and customer type:
On the one hand, iFLYTEK has laid out smart education, smart city tracks, and operator tracks earlier, and has accumulated relevant industry data and industry understanding capabilities, such as: according to incomplete statistics, iFLYTEK has been involved in education, operators, and urban management in 2004, 2008, and 2017 respectively;
On the other hand, education and urban management are led by the government, and the G-end business itself has a natural entry threshold, backed by the Hefei government, the Chinese Academy of Sciences and other resources of iFLYTEK is easier to get orders, while telecom operators belong to the "monopoly" of central enterprises, there are certain access requirements for suppliers, and China Mobile is the largest shareholder of iFLYTEK.
As the basis of the company's revenue, the company's smart education, smart city, and operator businesses will decline in 2023.
Due to the impact of the economic environment and fiscal expenditures on the G-end business, the revenue proportion of the company's two major businesses of smart education and smart city in 2023 will decrease to a certain extent compared with 2022.
However, the proportion of the company's operator business revenue remained relatively stable. According to the company's 2023 annual report, the revenue of the company's carrier business in 2023 will decline by 0.05% year-on-year.
It can be seen that the explosion of the AI model and the release of the Spark model have not brought much room for growth to the company's smart education (G-end), smart city, and operator businesses.
So, here we can't help but have a question, what kind of growth does the iFLYTEK Xinghuo model bring to the company?
Second, the profit effect of the large model is average
Since the official release of the Spark model by iFLYTEK in May 2023, the monetization methods of the iFLYTEK Spark model have been divided into three ways: empowering existing businesses, API authorization payment, and enterprise customization model, but only the C-end hardware has obvious effects.
1. The price of existing C-end hardware products has increased
The Spark model empowers C-end hardware products, which can bring a significant increase in product prices.
Taking the learning machine as an example, the current price of the iFLYTEK LUMIE10pro high-end version equipped with the Xinghuo large model reaches 9899 yuan, the price of T20 is about 7999 yuan, the price of T20Pro is about 8999 yuan, and the price of the AI learning machine without a large model is about 2000~3000 yuan.
In addition, the sales of iFLYTEK's AI hardware (AI learning, AI office, and AI health) increased by 125% year-on-year during the "618" period, and the sales increased by 126% year-on-year during the Double 11 cycle.
With the blessing of the Spark model, the revenue proportion of the company's C-end business has been increased. For example, in the education business, in 2022, the G-end will account for 40%, the B-end will account for 30%, and the C-end will account for 30%, and in 2023, the G-end will account for 31%, the B-end will account for 28%, and the C-end will account for 41%.
It can be seen that C-end hardware products equipped with large models can bring incremental growth to the company's performance by increasing the overall sales price. In 2023, the company's smart hardware revenue will be 1.617 billion yuan, a year-on-year increase of 22.35%.
2. The price of API call service is reduced
In addition to empowering existing products, the company has also launched a charging scheme for developers to call the API interface of the Xinghuo large model. The API interface is like a bridge that allows a software program to request information or services from the Spark model.
In addition, after the release of the iFLYTEK Xinghuo cognitive model, the number of developers has increased significantly. After the release of the Spark model, the number of developers increased by 850,000 in two months. The number of developers increased from 4.974 million in June 2023 to 5.785 million in December 2023.
However, in terms of the revenue of Xinghuo large model API calls, the company did not disclose relevant information.
However, it is worth noting that contrary to the price increase of C-end products, the price of iFLYTEK's API call service has been reduced.
At present, the company has launched different service packages for Spark V1.5 and V2.0 respectively, with prices ranging from 1,800-90,000 yuan/year (V1.5) and 3,600-180,000 yuan/year (V2.0).
At present, according to the official website of iFLYTEK, the company has launched different service packages for Xinghuo V1.5, V3.0 and V3.5, with prices ranging from 750~60,000 yuan/year (V1.5) and 1,500~120,000 yuan/year (V3.0/V3.5).
(Data source: iFLYTEK official website)
After the iteration of the Xinghuo large model, the performance has improved, but the price of calling the API has decreased. Could it be that the API call service of the Xinghuo large model is not as expected?
The reason for this is that, on the one hand, the basic version of iFLYTEK Xinghuo's API pricing is too high, and on the other hand, the competition for large models is fierce, and GPT-4's increase in volume and price reduction has also put pressure on domestic large models such as iFLYTEK.
API calls are denominated based on tokens. As of the beginning of October 2023, the API pricing displayed on the official websites of some major large models is as follows (unit: RMB/1,000 tokens):
| Price comparison of major model APIs | Standard Edition | Enhanced Edition |
|---|---|---|
| Ali Tongyi asked | turbo:0.012元/千token | plus:0.14元/千token |
| Tencent Mixed Element | The standard price is 0.01 yuan/1,000 tokens | Premium pricing: 0.10 yuan/1,000 tokens |
| Baidu Wenxin said | turbo:0.008元/千token | 正常:0.12元/千token |
| Xunfei Xinghuo | Version 1.5 Spark model: 0.018 yuan/1,000 tokens | Version 2.0 Spark model: 0.036 yuan/1,000 tokens |
| Tsinghua Zhipu (GLM) | Lite:0.008元/千token | Pro:0.01元/千token |
| ChatGPT | GPT-3.5:0.015美元/千token | GPT-4.0:0.06美元/千token |
(Data source: official websites of major models)
Among the standard versions of major models, iFLYTEK Xinghuo V1.5 is the most expensive, significantly higher than its competitors.
In addition, with the continuous iteration of large models, GPT has also begun to reduce prices. OpenAI upgraded GPT-4 to GPT-4 Turbo, but the fee for the Turbo version has dropped significantly. Compared with the original version of GPT-4, the Turbo version charges $0.01 for 1,000 tokens, which is three times cheaper, and $0.03 for the output of 1,000 tokens, which is twice as cheap.
Therefore, it is reasonable for iFLYTEK Xinghuo's API pricing to decline.
3. The customized large model of enterprises has limited contribution to revenue growth
Compared with the C-end market, the B-end market is the force of the large model.
It is generally believed in the industry that a model adjusted by industry expertise will perform better work on the basis of a large model than a general basic model that has not been optimized.
For example, Zhu Xiaohu, an investor who "quarreled" over the AI large model startup, and Fu Sheng, CEO of Cheetah Mobile, both agreed that the results of many vertical scenarios optimized by open source small models are far better than those of general large models, and the importance of vertical scenarios and data far exceeds the large model itself.
At present, iFLYTEK and leading enterprises in the industry have jointly released 12 industry models, including finance, automobiles, operators, industry, housing and construction, and law.
From the perspective of industrial resources, the iFLYTEK Xinghuo model can occupy a place in education, law, housing and construction, operators and other tracks, but its competitiveness in new tracks such as automobiles and finance still needs to be observed.
At this stage, the landing of the Spark model in education, law, housing and construction, operators and other tracks is also limited to revenue growth.
The implementation of large-scale models in G-end projects such as law and housing construction is also affected by the economic environment and financial expenditure. In the short term, the large model is expected to land in these G-end tracks, but it remains to be seen whether it can bring better revenue growth to the company.
On the operator track, iFLYTEK claims that iFLYTEK Xinghuo has enabled many applications, such as: iFLYTEK Xinghuo and China Mobile have created a 5G network to C highlight application, which can realize call change minutes.
It's just a transcription of speech, which has always been a feature that iFLYTEK has been developing for a long time.
If iFLYTEK Xinghuo can organize the content of the call into a condensed version of the text minutes, then, why is it the Kimi model that has recently come out of the circle with its long text processing and semantic understanding, instead of iFLYTEK Xinghuo?
In addition, in the context of the company's revenue growth of only 7% in 2023, the proportion of the company's carrier business revenue remains relatively stable. It can be seen that after the Xinghuo model empowers the operator's business, it has not brought immediate results to revenue.
In the new tracks such as finance and automobiles, iFLYTEK, a latecomer, lacks customer resources.
Here, taking the car model as an example, Tesla, BYD, Weilai, Ideal, etc. have developed their own large models, and Hongqi, Dongfeng Nissan, Aiways, Leap, Haima, Nezha, etc., including car companies, have joined Baidu Wenxin Yiyan's circle of friends. According to incomplete statistics, the iFLYTEK Xinghuo model has empowered Chery, GAC, Changan and other car companies.
In 2023, iFLYTEK's smart car business will achieve 695 million, a year-on-year increase of 49.71%, which has a good performance, but has a limited contribution to the company's revenue of nearly 20 billion.
Overall, the positive effect of the Xinghuo model on performance is reflected in the C-end business, smart cars and other businesses, but the contribution to the company's revenue base is still limited.
3. A "big gamble"
It is worth noting that large models also burn money.
As of 2023, iFLYTEK's R&D expenses will be 3.481 billion, a year-on-year increase of 11.89%, accounting for 17.71% of revenue.
In addition, according to iFLYTEK's annual report, the company will increase investment in the research and development of cognitive large models on an independent and controllable platform. This is also the main reason why the company's non-net profit is expected to decrease by 71.74% year-on-year in 2023.
According to the previous analysis, due to the large revenue of iFLYTEK, although the Xinghuo model can have a certain positive impact on the company's C-end business, it is not enough to make the company's revenue to a new level, and the iteration of the Xinghuo model still requires a large R&D investment, so that the company's profitability will continue to be under pressure.
In addition, iterations of large models continue. The company said that iFLYTEK Xinghuo V4.0, which is currently benchmarking GPT-4's current capabilities, is being trained and is expected to be released in June 2024.
However, what people didn't expect was that before GPT-5 was released, Microsoft had quietly started training GPT-6 for OpenAI. This training will use 100,000 H100 GPUs, and the cost of one H100 is about 30,000~40,000 US dollars, and the cost of GPUs alone is as high as 2.1 billion US dollars. Of course, the training of the model can also be done in the form of computing power leasing, but the cost is still expensive.
It can be seen that as long as there is competition and you want to surpass your opponents, there is no end in sight to the investment in large models.
And iFLYTEK's bet on the large model this time is undoubtedly a "big gamble". Under the tuyere, iFLYTEK has come out of the predicament of profitability with a large model, but there is still great uncertainty.
In addition, iFLYTEK's performance in the capital market is also unsatisfactory.
In 2023, when major manufacturers have released large models, the capital market has high expectations for iFLYTEK's large models, and the stock price of iFLYTEK once rose to 81.90 yuan.
With the release of the 2023 annual report and the first quarter of 2024, iFLYTEK's performance has poured cold water on the capital market, and its rolling price-to-earnings ratio has reached 232.41.
Similar to performance, iFLYTEK's valuation can't fly.
Disclaimer: The content of this article is for reference only, the information or opinions expressed in this article do not constitute investment advice, readers are advised to make investment decisions with caution.
This content is the author's independent view and does not represent the position of Tiger Sniff. May not be reproduced without permission, please contact [email protected] for authorization
People who are changing and want to change the world are all in the Tiger Sniff APP