Written by Wang Huiying

Edited by Ziye

Halfway through 2024, dubbed the "Year of AI Large Model Applications," the industry has undergone rapid changes amidst a cacophony of voices.

The typically lengthy technology development cycle has been significantly accelerated for large models. In just two short years, players have engaged in a "hundred models battle," with capital pouring in like wildfire, amidst exploration and skepticism alike.

As we enter the second half of the year, the winds in the industry have shifted subtly.

Recent data from analysis firm Similarweb indicates a steep drop in monthly visitors to the ChatGPT website. OpenAI's pace of releasing new products has noticeably slowed. NVIDIA's stock price has fluctuated, dipping below its single-day high in the U.S. stock market. Even in China, there are influential figures casting doubts on large models...

On the other hand, domestic "rising stars" of large models remain red-hot. Dark Side of the Moon has been rumored to have received investment from Tencent, with the company's post-investment valuation reaching $3 billion; Zhipu AI has secured a new round of financing amounting to billions of yuan, with a pre-investment valuation of 20 billion yuan; and Fei-Fei Li's first startup, World Labs, has announced a $230 million funding round, backed by NVIDIA.

The industry finds itself in a state of anxiety, caught between ice and fire. Behind this lies the gradual disillusionment with large models by the market and the return of capital to rationality. Only when the hype subsides will truly valuable enterprises or applications emerge.

Regardless of the stage, commercialization is a topic repeatedly mentioned in the large model industry. The particularity of large models lies in their "burn rate" – the financial costs associated with both technical development and operational expenses require substantial financial support, acting as a "tight noose" around large model enterprises.

Especially for startups like Dark Side of the Moon and Zhipu AI, which lack the financial muscle and resources of industry giants, relying solely on financing without the ability to generate revenue internally will inevitably lead to their elimination.

Changes in the tide push the large model industry forward. From the "hundred models battle" to the "application battle," the large model industry is poised to enter a new phase, where finding a unique path to monetization becomes paramount, in addition to developing products and applications.

1. "Star Enterprises" Remain Popular

During the recent opening keynote at the 2024 Yunqi Conference, Alibaba CEO Wu Yongming made an illustrative comparison: Last year, large models' mathematical abilities were on par with those of middle school students, but today, they can win international math Olympiad gold medals, especially in disciplines such as physics, chemistry, and biology, where they are approaching doctoral levels.

Over the past 22 months, the advent of large models has accelerated AI's development pace unprecedentedly. From the battle of a hundred models to the battle of applications, in addition to internet giants entering the large model fray, numerous AI-related startups have emerged.

After several rounds of elimination, China has given birth to numerous star enterprises. From the "New AI Four Dragons" to the "AI Five Tigers" and "AI Six Little Tigers," capital has played a pivotal role in their rise.

According to CBInsights data, generative AI startups globally raised approximately $20.4 billion in funding in 2023, more than five times the $3.6 billion raised in 2022.

The hype continues today. In August, ZeroOne raised a new round of financing worth hundreds of millions of dollars, with a post-money valuation of 10.4 billion yuan. In the same month, Dark Side of the Moon also secured a $300 million funding round. In July, Baichuan AI completed its Series A funding round of 5 billion yuan and launched its Series B funding round with a valuation of 20 billion yuan. In September, Zhipu AI completed a 1 billion yuan funding round, with investors including Zhongguancun Science City...

Among the "AI Six Little Tigers," ZeroOne, Baichuan AI, Zhipu AI, Dark Side of the Moon, and Minimax have all raised over 100 million yuan in financing this year. Another company, Jieyue Xingchen, is rumored to be in the midst of a new funding round with a valuation of $2 billion.

Among them, Dark Side of the Moon and Zhipu AI are the most favored by capital.

Looking back at Dark Side of the Moon's funding journey, it boasts well-known investment institutions and internet giants among its investors.

In June 2023, Dark Side of the Moon secured its first angel funding round of over $200 million, with investors including ZhenFund and Sequoia China, valuing the company at $300 million at the time. In July of the same year, Dark Side of the Moon received its Series A funding from investors including Meituan Longzhu and BlueRun Ventures. In February 2024, Dark Side of the Moon raised over $1 billion in Series A+ funding, with investors including Sequoia China, Xiaohongshu, Alibaba, and existing shareholders.

Notably, this February funding round represents the largest single-round funding received by a Chinese large model startup since the emergence of ChatGPT, propelling Dark Side of the Moon's valuation to $2.5 billion.

Capital's favor has driven Dark Side of the Moon's valuation upwards. After the August investment, its post-money valuation reached $3.3 billion (21 billion yuan), leading the "AI Six Little Tigers."

On the other hand, Zhipu AI and Baichuan AI have also joined the "20 billion yuan club."

In early September, Zhipu AI completed a new round of financing worth billions of yuan, with a pre-investment valuation of 20 billion yuan. This round was led by Zhongguancun Science City.

According to Qichacha, Zhipu AI has completed 11 funding rounds, with investors including Beijing AI Industry Investment Fund, Social Security Fund Zhongguancun Independent Innovation Fund, Lightspeed China Partners, Meituan, Ant Group, Alibaba, Tencent, Xiaomi, Kingsoft, Shunwei Capital, Sequoia China, and Hillhouse Capital, among others.

In July, Baichuan AI completed its Series A2 funding round of 5 billion yuan, with a post-money valuation of 20 billion yuan. Prior to this, Baichuan AI announced its Series A1 funding in October 2023, disclosing investors including Alibaba, Tencent, Xiaomi, and other technology giants, as well as top investment institutions.

Within two months, three unicorns with valuations of 20 billion yuan emerged in China's large model industry. Behind capital's preference for star enterprises lies some food for thought.

One notable point is that while the large model hype has been raging for nearly two years, with capital chasing leading unicorns, the industry as a whole has remained relatively cautious.

On the one hand, large models are "difficult." The path from developing large models to deploying them in real-world applications requires not only financial and human resources but also robust technology. On the other hand, large models are "expensive," with training costs reaching tens of millions of yuan and a yet unclear commercialization path, making them a challenging endeavor that capital is hesitant to pursue blindly.

More crucially, "20 billion yuan" is often seen as a watershed for startups. To join the 20 billion yuan club, capital demands higher returns, necessitating that enterprises find a self-sustaining path to profitability.

In other words, for enterprises valued at 20 billion yuan, capital has given them their moment in the spotlight, and they must deliver satisfactory results – make money and move fast.

2. Balancing Spending and Earning: "Rising Stars" Must Learn to Monetize

'What we're doing is challenging and requires significant financial and resource support,' said Zhipu AI CEO Zhang Peng in an interview. In the current economic climate, AI investments are substantial, and results often fall short of expectations, leading to considerable pressure and anxiety.

Indeed, under the pressure of capital's urgent return cycle, financing is only the first step for unicorns to secure a ticket in the large model race. Learning to make money is their compulsory course.

Moving beyond last year's "hundred models battle," the large model industry has shifted towards application deployment, i.e., commercialization. Similar to last year's debate on technical approaches, this year's commercialization path remains a focal point of industry discussions.

For instance, at this year's Zhiyuan Conference, ZeroOne founder Li Kaifu stated, 'ZeroOne is committed to To C business and will not engage in unprofitable To B operations.'

Academician Zhang Yaqin of the Chinese Academy of Engineering, on the other hand, believes that in the embodied intelligence stage, To B applications may be implemented faster than To C, stating, 'At this stage, the real money-making opportunities for large models lie in B-end infrastructure, including chips, hardware, and servers.'

Essentially, this represents a debate between the B-end and C-end paths for large model commercialization. One side argues that B-end applications are relatively clear, cover a wide range of industries, and can quickly achieve multi-scenario applications, while C-end competition is fierce, and it takes a long time to develop a popular application. The other side believes that intensified price wars in the industry have compressed profits for B-end large models, while C-end can see revenue faster.

Under this thinking, domestic startup large model vendors initially pursued two main commercialization strategies: those focused on C-end businesses, such as Dark Side of the Moon, Baichuan AI, and ZeroOne; and those pursuing both B-end and C-end markets, represented by Zhipu AI and MiniMax.

Regarding C-end large model applications, none is more well-known than Dark Side of the Moon's Kimi. Launched in October 2023, Kimi became an instant hit with its exceptional long-text capabilities. Subsequently, Dark Side of the Moon enhanced Kimi's long-text abilities tenfold and rapidly iterated and optimized the product.

Data shows that from December 2023 to February 2024, Kimi's monthly active users were 508,300, 1,128,500, and 2,984,600, respectively. Notably, in February 2024, the number of users was nearly six times that of December 2023.

While Kimi's popularity is undeniable, it's also true that while the C-end market is closer to consumers and yields faster returns, many players have entered, and there's still a long way to go before a true super app emerges. Everyone has an opportunity to break through, and no one can afford to slack off.

Currently, the revenue model for large models in the C-end market is relatively limited, with other charging models still under exploration besides subscription fees. For example, Kimi previously introduced a paid option called "Refuel Kimi," with prices ranging from 5.2 yuan to 399 yuan, similar to a "tipping" model, to explore new commercialization avenues.

On the other hand, vendors like Zhipu AI, which have made faster progress in B-end business commercialization, have focused on building a large model ecosystem.

Since its inception, Zhipu AI has set its sights on catching up with OpenAI. To date, Zhipu AI has developed a comprehensive suite of model products benchmarked against OpenAI, including Zhipu Qingyan for AI efficiency enhancement, CodeGeeX for efficient code generation, CogVLM for multimodal understanding, and CogView for text-to-image generation.

Zhang Peng, CEO of Zhipu AI, has repeatedly emphasized that compared to the C-end market, the B-end market has a stronger willingness to pay. Along this path, Zhipu AI has made numerous layouts in the B-end market.

For example, by introducing the concept of "Model-as-a-Service," Zhipu AI encapsulates large models into an open platform, providing APIs for developers and enterprises to call and charge based on usage. Addressing the data security needs of medium and large enterprises, Zhipu AI offers cloud-based private deployment solutions to help users establish dedicated model zones in the cloud.

Regardless of the path chosen, startups' goal is to make money, but the challenges are the same. C-end commercialization faces issues of low user retention and high customer acquisition costs, while B-end commercialization grapples with industry price wars, putting considerable pressure on startups.

In such cases, pursuing both B-end and C-end markets is a strategy some vendors are considering.

In August this year, Dark Side of the Moon launched its Kimi Enterprise API, continuing to expand in the B-end market. Compared to general-purpose models catering to C-end needs, enterprise-grade model inference APIs offer higher levels of data security and concurrency to support complex workflows and large-scale data processing within enterprises.

Meanwhile, Zhipu AI has also begun exploring the development and deployment of C-end businesses.

In July, Zhipu AI officially launched its video generation model Qingying, capable of generating a 6-second video in just 30 seconds. In August, the Zhipu Qingyan app added video call functionality.

Another unicorn, MiniMax, also pursues a dual strategy in its product offerings, targeting both C-end and B-end markets. For C-end users, it offers AI chat apps like Glow for role-playing, immersive AI content community products like Xingye, and paper-writing support through Hailuo AI. For B-end users, it has successively released the MoE large language models abab 6 and abab 6.5 and plans to open APIs.

Since its explosion into popularity, the large model industry has reached several critical milestones, with the most important being the evolution from parameters to applications. The standard for a successful large model is shifting from speed to usability and practicality. The industry consensus is that no matter how capable a general-purpose large model company may be, it ultimately relies on commercialization for sustainability.

Currently, almost all startup large model vendors' revenue scales are far from sufficient to support their valuations, and applications are mired in homogenized competition. Learning to make money while spending is the core of their past and future commercialization efforts, as time waits for no one.

3. Competing with Industry Giants: Pressure on Large Model Rising Stars

Amid the large model wave, internet giants and startups stand on the same starting line. If last year was the qualification round for large models at the technical level, this year's application level has reached the finals.

It's difficult to deny that startups relying solely on capital infusions face pressure when competing with industry giants with richer resources and more comprehensive ecosystems.

For startups, the primary challenges lie in the high costs of large model training and the rising costs of customer acquisition for applications.

Even for OpenAI, foreign media have cited unpublished internal financial data claiming that OpenAI will face up to $5 billion in losses this year. Its annual revenue is estimated to be between $3.5 billion and $4.5 billion, but operating costs could reach $8.5 billion, with inference costs accounting for $4 billion alone.

As we enter this year, the large model industry's internal competition has intensified, and with low user retention for large model applications, C-end applications have reached new heights in their efforts to acquire new users.

Taking Kimi as an example, according to AI Emergence, Dark Side of the Moon pays at least 30 yuan for every registered user Kimi acquires through Bilibili. Data from Similarweb shows that in March 2024, following its promotion on Bilibili, Kimi's traffic surged significantly, with a peak growth rate of 402.9%, outpacing Zhipu AI's Zhipu Qingyan and MiniMax's Hailuo AI by an order of magnitude.

From online platforms like Bilibili, Xiaohongshu, and Douyin to offline venues like subways and office buildings, the large model industry has ignited a battle for advertising and marketing. The results have been remarkable, with Similarweb statistics showing that total visits to products from the "AI Five Dragons" (Zhipu AI, MiniMax, Baichuan AI, ZeroOne, and Dark Side of the Moon) have soared by 963% in six months.

But spending hundreds of millions on advertising doesn't necessarily lead to long-term consumer loyalty, and retention remains an unknown for large model vendors.

This year, the issue of making money has become even more challenging. In May, major vendors like ByteDance, Alibaba, Baidu, and Tencent reduced the prices of their flagship model APIs by over 90%, officially entering a price war in the large model market.

Some vendors have chosen to follow suit. Zhipu AI reduced its prices twice in a month; MiniMax quietly launched a promotion offering 100 million tokens for free registration and certification, as well as free TPM expansion; Kimi Open Platform reduced its context cache storage fees by 50%.

Others have chosen to hold firm. Wang Xiaochuan, founder of Baichuan Intelligence, publicly stated that he would not follow the price-cutting trend; Li Kaifu, CEO of Zero One Universe, bluntly said that the crazy price war in the domestic large model market is a lose-lose proposition.

Regardless of whether they choose to lower prices or not, startups face a common challenge: where is the next round of funding coming from?

Behind this lies the collective anxiety of startups. Without users, startups lose the data to train their models and investor enthusiasm, pushing them to the brink of collapse.

Zhu Xiaohu, Managing Director of GSR Ventures, has pointed out that investing in domestic large model companies may not be profitable. He believes that the more embarrassing situation for large model companies is that even if they are willing to spend tens of millions of dollars, they may end up wasting their money if others open-source their models.

In June this year, Goldman Sachs published an article titled "Too Much Invested, Too Little Returned," stating that large companies plan to invest $1 trillion in AI-related areas such as data centers, chips, and power grids over the next few years. However, so far, these investments have only marginally improved developer productivity, without significant other results.

Judging from the current market response, large companies have largely cornered the market, making it difficult for startups to secure funding from them. Coupled with limited investor enthusiasm, startups with insufficient self-sustaining capabilities face challenges in survival.

Drawing on the development path of foreign unicorns, it is more likely that they will be acquired by large companies. However, whether they can sell at a good price and find a good buyer remains uncertain.

Currently, Character.AI has sold to Google at a 50% valuation discount, while Inflection and Adept have been acquired by Microsoft and Amazon, respectively. Meanwhile, Reka AI is still looking for a buyer, Runway has been embroiled in a data deletion scandal, and Stability AI has reported a funding crisis following significant management changes.

Starting from the same starting line does not guarantee reaching the finish line together. Even if large model businesses are loss-making, large companies can still make up for it through other businesses within their ecosystems. In contrast, startups must raise funds and generate revenue simultaneously, leveraging their product strengths to compete with large companies.

While a single dominant player is unlikely to emerge in the short term, a fierce elimination round is inevitable. All players must give their all this year.