Who will be the next to stumble?

Recently, Lingyiwanwu, one of the "Six Tiger Cubs of Big Models," has announced a strategic shift. According to founder Kai-Fu Lee, Lingyiwanwu will henceforth concentrate on small-parameter, moderately-sized industry models. Its pre-training algorithm team will join Alibaba's Tongyi team, while the Infra team will integrate with Alibaba Cloud...

The signal is unmistakable: When the game evolves into a protracted battle requiring sustained capital investment, the scales naturally begin to tip.

But that's not all. Recent rumors suggest that Hailuo AI, under MiniMax, is facing a lawsuit from iQIYI for copyright infringement. Although the damages claimed are not substantial, behind this lies a formidable wall of copyright costs accumulated from images and videos, crashing down on all big model players.

It's certainly adding insult to injury.

In fact, it's not just big model startups feeling the pinch; even some major tech companies seem to be under pressure now.

Take Baidu, for example. Facing the life-or-death crisis of JiYue, why doesn't it show any intention of rescuing it? Behind this, it's hard to deny that there isn't a reason related to AI and big models consuming too much capital and energy, making it difficult for Baidu to manage.

Then there's Iflytek. Due to its aggressive investment in big models, its already unimpressive profit performance turned into a state of significant loss in the first three quarters of 2024, with a non-deductible net loss of 468 million yuan.

Companies like SenseTime, once known as one of the "Four Little Dragons of AI," are facing the same dilemma...

And compared to Baidu and Alibaba, which have stable fundamentals, SenseTime and Iflytek have indeed spread their business quite widely. But in reality, each of these sectors is highly competitive, and their core profitability is not prominent.

This has led to a polarized reputation for many second- and third-tier big model investments, including SenseTime: Optimistic investors see them as the next BYD or CATL before their explosion; pessimistic ones view them as a batch of severely ailing patients waiting for treatment...

AI Big Models: Visible 'Money-burning,' Intangible Profitability

Speaking of AI big models 'burning money,' Iflytek and SenseTime certainly have a story to tell.

In the first three quarters of last year, Iflytek's revenue was 14.85 billion yuan, an increase of 17.73% year-on-year; gross profit was 6.007 billion yuan, an increase of 18.15% year-on-year, demonstrating a very capable overall performance.

Unfortunately, the corresponding net profit attributable to shareholders immediately turned into a loss of 343.7 million yuan, swinging from profit to loss compared to the same period last year; non-deductible net profit was a loss of 468 million yuan, with the scale of losses further expanding.

In this regard, Iflytek did not hide anything. Both the third-quarter and half-year reports clearly stated that the company's excessive investment in the research and development and layout of the Spark big model had an impact on short-term performance.

According to Tianyancha APP, Iflytek's R&D investment in the first half of 2024 was 2.191 billion yuan, an increase of 32.23% year-on-year. Among them, the total investment related to big models exceeded 1.3 billion yuan...

A similar imbalance between input and output also occurred at SenseTime.

Against the backdrop of cumulative losses of nearly 30 billion yuan over the past three years, in the first half of 2024, SenseTime achieved revenue of 1.74 billion yuan, but the corresponding R&D expenditure alone amounted to 1.892 billion yuan. Coupled with the inevitable sales costs and administrative expenses, it's no wonder that SenseTime's net profit for the same period was a loss of 2.457 billion yuan.

But even so, why do companies like Iflytek knowingly head into the tiger's den?

In the words of Iflytek Chairman Liu Qingfeng and Vice President and Secretary of the Board Jiang Tao, "Currently, we are still in a critical historical opportunity period for AI layout, and we must persist and not be lenient in the source and application landing of models. We don't want to miss this major (AI) revolutionary opportunity.""

It's actually not hard to understand. The AI market has almost been recognized as a major boon for future global economic growth, so no amount of technological investment is too much. And after experiencing the era of early-stage losses and "throwing money" to gain scale among internet companies, followed by later-stage "easy wins," the current losses of big model players, while looking "scary," are worth it as long as they can succeed.

Moreover, in this regard, Iflytek and SenseTime have indeed achieved many accomplishments.

For example, in the 2024 half-year report, iFLYTEK Spark was still the only big model technology achievement based on domestic computing power training that could be downloaded by the public to date. In the first three quarters, the number of new developers reached 1.777 million, and the download volume of the "iFLYTEK Spark" APP on the Android platform exceeded 190 million times.

Moreover, and more importantly, Iflytek's solid technical strength has made the "iFLYTEK Spark" big model the preferred choice for central and state-owned enterprises. According to the latest "China Big Model Bidding Project Monitoring Report (2024)" released by third-party agency Smart Hyperparameters, in 2024, Iflytek disclosed a bid-winning amount of 847.808 million yuan, ranking first and leading Baidu and Alibaba Cloud by a wide margin, becoming the bidding champion of the year 2024.

On one side is recognition from central and state-owned enterprises, and on the other side is the successful landing of toB and toG commercialization bids. This gradually clearer path to technology monetization has also become a key basis for many investors to be optimistic about Iflytek.

But then again, these strategic monetization possibilities would be even better if they weren't accumulated in "accounts receivable."

In recent years, the scale of Iflytek's "notes receivable and accounts receivable" has been very large.

For example, in the first half of last year, its notes receivable and accounts receivable reached 13.65 billion yuan, even higher than the revenue scale of 9.325 billion yuan for the same period. By the first three quarters, this indicator had further increased to 14.51 billion yuan, accounting for 36.16% of total assets.

And from the half-year report, Iflytek has about 3.067 billion yuan in accounts receivable with an aging of more than 2 years.

In this regard, SenseTime and Iflytek are almost identical. During the same period, SenseTime's trade receivables also amounted to 3.534 billion yuan, more than twice the corresponding revenue scale...

Fortunately, however, Iflytek has stated on investor interaction platforms that these accounts receivable mainly come from high-quality customers such as governments, financial institutions, and large and medium-sized B-end enterprises. The quality of customers is good, and the safety of accounts receivable is high. In recent years, the actual bad debt write-offs have accounted for less than 0.1% of accounts receivable.

In other words, major customers do not or rarely default on payments.

Although this is good news, on the one hand, from the current loss situation, since Iflytek cannot temporarily recover more accounts receivable, neither its cash flow nor its capital reserves look good. As of the first three quarters of last year, Iflytek's net cash flow from operating activities was -821 million yuan, and its monetary funds were only 2.751 billion yuan, which undoubtedly increased its market risk resistance pressure to a certain extent.

On the other hand, the longer the account period of accounts receivable, the greater the possibility of bad debts and the more uncertainty of potential negative impacts.

Looking at troubled brother SenseTime, in 2023, the proportion of loss provisions for SenseTime's trade receivables from 6 months to 1 year was 27.9%, 1 to 2 years was 43%, 2 to 3 years was 58%, 3 to 4 years was 75%, and over 4 years was 100%.

What was the total loss provision for trade receivables for the whole year? 3.52 billion yuan. Equivalent to 103% of 2023 revenue.

The two companies have similar toB and toG business routes, the same high input-profit losses, and high accounts receivable. This has caused the capital market to have some doubts about the commercialization prospects of AI big models and has led to a conservative judgment on the investment valuation of companies like Iflytek...

Commercialization 'Involution': Startups Pave the Way, Major Companies Overtake?

In fact, the difficulties of the big model market are not only reflected in internal operational cost pressures but also in external aspects such as commercial monetization applications and technological market competition.

Especially since the second half of 2024, due to the poor landing of AI commercialization applications, which has led to a tightening of financing channels for big models, this year has been regarded by many investors as a year of financial testing and commercialization elimination for China's big models.

But to be honest, the current performance of big model commercialization is indeed lacking.

For example, in terms of C-end software, against the backdrop of major companies frantically investing in traffic to acquire customers, according to third-party monitoring data from QuestMobile, over the past year, the average daily usage frequency of products such as Doubao, Kimi, and Wenxiaoyan has hovered between 4 and 5 times, and the average usage time per user has also remained between 5 and 10 minutes, with no significant growth shown.

To the extent that some ByteDance employees refer to Doubao as an "ICU product" – it cannot survive without the "intubation" of investment.

In addition, there are products like "E Tingshuo Middle School" launched by Iflytek for the education sector. In the Apple App Store, "E Tingshuo Middle School" has a rating of 2.9 stars, while in the Android app store, the rating is even lower at only 1.3 stars, with many negative reviews such as "inaccurate recognition, impractical functions" in user evaluations.

Even so, the price for purchasing "E Tingshuo Middle School" for one year ranges from 308 yuan to 460 yuan. At the same time, there are also many voices in public opinion pointing out that during the campus promotion of "E Tingshuo Middle School," although some schools claim that students and parents voluntarily purchase and use the software, in actual execution, there have been cases of excessive assignment of homework related to this software, which seems to be suspected of de facto forced purchases...

As for smart hardware, concepts like AI phones and AI learning machines did gain popularity for a while, but when it came to actual implementation, they did not bring the market more amazing moments like the iPhone 4. This makes it so that even though some smart learning machines and other products are selling well now, such as Iflytek's smart hardware business revenue increasing by 56.61% year-on-year in the first half of last year, it still has not fully ignited the big model commercialization market.

Then looking at B-end applications, which were once a major source of confidence in the investment market. According to the blue paper "AI Development Report (2024)" released by the China Academy of Information and Communications Technology, by 2026, more than 80% of enterprises will use generative AI APIs or deploy generative applications.

But there's that old problem: Big model startups are all eagerly waiting for funds to prevent being kicked out of the game, but when they look at the high levels of accounts receivable at companies like Iflytek and SenseTime, as well as the visible 'money-burning' and intangible profitability, it's truly suffocating...

Of course, the commercialization of big models is somewhat unpredictable. After all, from software to hardware, and from medical care, office work, entertainment to automobiles, industrial manufacturing, and other fields, almost all are already trying to integrate with AI, so market explosions may happen at any moment.

Even so, the future market competition landscape can be roughly predicted. Because the domestic big models that can stay in the game now are basically players with real strength. Even apart from iFLYTEK Spark, ERNIE Bot, Tongyi Qianwen, and Doubao all have certain capabilities to compete with ChatGPT.

So in terms of technological competition, even if there are gaps in a certain field, it's not to the extent that one cannot catch up.

And this will lead to the situation where, whenever anyone discovers a sign of breakthrough in big model commercialization, involution will quickly follow. In other words, many AI fields that seem to be blue ocean markets now can actually turn into highly competitive sectors very quickly.

Just like the price reductions in the big model industry last year, if Doubao reduces by 90%, Alibaba dares to reduce by 97%...

So what will be the result? Small and medium-sized startups will blaze a trail, and then major companies occupying internet traffic entrances will quickly catch up and overtake latercomers by leveraging their capital and technological resource advantages.

After all, many products currently supported by big models are still mainly in the form of apps, so model capabilities are indeed an important driving force for enhancing the experience, but user demand insight, function design, and traffic acquisition are also crucial, and they are even the decisive factors for retaining users.

But this is undoubtedly a natural disadvantage for tech-focused players like the AI Four Dragons, the Six Tiger Cubs of Big Models, and Iflytek.

So some investors bluntly stated: Before the big model track is whittled down to only two or three players, compared to upstream "shovel companies" and downstream application companies, big model startups in the midstream face more challenges than opportunities.

On one hand, high input costs loom large, while on the other, output results and competitiveness are shrouded in significant uncertainty. Consequently, when firms such as Lingyiwanwu and SenseTime struggle to 'regain strength' swiftly enough to keep pace with their rapid cash burn rates, a fresh wave of bankruptcies and consolidations among big models may be on the horizon...

Disclaimer: This article is based on the company's legally mandated disclosures and publicly accessible information for the purpose of commentary. However, the author does not warrant the completeness or timeliness of the information presented. Furthermore, the stock market involves risks, and investors are advised to exercise caution. This article does not constitute investment advice, and investors are solely responsible for making their own decisions regarding investment opportunities.