News 2024-09-29

Kimi's Commercial Dilemma

The resurgence of fierce competition in the Chinese internet sector has manifested in the realm of AI model applications, reviving a phenomenon that many thought had faded. Over the past decade, any industry that faced disruption and reinvention via the Internet was typically accompanied by audacious capital expenditures. These ‘money-burning battles’ have raged across various sectors, from group buying to food delivery, e-commerce to travel, and ride-hailing to the shared economy. Tech giants and venture capitalists have consistently engaged in a relentless cycle of financial investment and strategic maneuvering.

However, following the dramatic conclusion of community group buying wars, major internet companies began to shift their focus, adopting a more meticulous and sustained approach to growth during challenging macroeconomic cycles. Some nascent trends, like the metaverse and Web3, failed to gain substantial traction. The era of unrestrained spending seemed to recede, replaced by a more cautious approach.

Yet, the landscape has dramatically changed with the emergence of AI. Since the beginning of 2023, a new phase of competition has emerged among what are referred to as the ‘Five Tigers of Large Models’ in China—Zhipu AI, The Dark Side of the Moon, Baichuan Intelligence, MiniMax, and Zero One Everything—alongside industry giants like ByteDance and Tencent. These entities are embroiled in a battle for user acquisition across various platforms, marking a return to aggressive financial strategies akin to the past.

Advertisement

A startup founder, entering the realm of large models in the wake of ChatGPT’s meteoric rise, remarked to Techknow that last year, industry peers were primarily focused on product development and bridging the gap in technological prowess with foreign counterparts. However, fast forward to this year, and the conversation has pivoted to the headache of user growth and acquisition costs.

According to the "2024 China Mobile Internet Half-Year Report," as of June this year, the monthly active users of AIGC (AI Generated Content) apps in China reached 61.7 million, representing a staggering growth of 653% year-on-year. This influx of users has inevitably driven up customer acquisition costs, resulting in a highly competitive landscape.

Leading applications in AI models, such as Kimi from The Dark Side of the Moon and Doubao from ByteDance, are increasingly turning this groundbreaking technological revolution into a fierce financial contest. Early in 2023, industry insiders revealed that Kimi had an average customer acquisition cost of 12-13 yuan per person, with daily acquisition costs reaching a minimum of 200,000 yuan. Reports later indicated that Kimi was spending at least 30 yuan to acquire a single registered user through advertising on Bilibili, underscoring the hefty financial stakes involved.

As the sudden calm of the internet sector is shattered, the internet battle rekindles with the new titans, like The Dark Side of the Moon, striving to compete against established heavyweights such as ByteDance. How they navigate this landscape remains a critical question.

Launching a novel business model demands significant upfront investments, usually with minimal immediate returns, to cultivate the market. Internet companies are able to undertake these large-scale burn rates largely due to constant backing from capital sources, which are enticed by the distinctive 'winner-takes-all' characteristic inherent in the internet industry.

For example, Meituan triumphed in the “thousand group war,” monopolizing merchant resources and establishing itself as the leader in food delivery. Similarly, Didi leveraged capital tactics to acquire and eliminate competition, ensuring a monopoly in ride-hailing. However, the bonuses and subsidies that users enjoyed during these high-stakes battles ultimately faded away after the dust settled.

The traditional model of cash-burning in the internet sector secured users and market positioning or paved the way for future financing. However, the motivations driving the aggressive spending seen in AI model applications now extend beyond just user growth; they encompass a broader strategic vision.

Kimi’s soaring user traffic not only allows The Dark Side of the Moon to stand tall among investors but also provides invaluable fresh data for the training of core models. As algorithms undergo internal optimization and upgrades, the need for high-quality data can only be satisfied through external sources. While some have suggested using AI-generated data for training purposes, research published in high-profile journals like Nature warns that allowing models to self-train on automated data could lead to significant degradation of performance within a few generations, resulting in outputs that are nonsensical.

Perhaps due to the acute need for precise data, Kimi has opted to focus on students and young professionals as its primary demographic on Bilibili, a key platform for user acquisition. Official statistics from Bilibili reveal that over 80 million users engage with AI-related content monthly, with average daily views for such material skyrocketing more than 80% in the past year, while active participation from AI content creators surged over 60%.

Kimi's monopolistic approach to advertising on Bilibili has left competitors like Zhipu AI, MiniMax, and Kunlun Wanwei scrambling to keep pace amid soaring unit costs. A content creator who regularly accepts sponsorships from platforms commented, “I’ve never seen a client as financially generous as The Dark Side of the Moon, aside from Pinduoduo.”

The combination of substantial financing and high-cost marketing strategies has created a virtuous cycle for The Dark Side of the Moon. As reported by third-party platform aicpb.com in July, Kimi's monthly traffic reached 24.56 million, substantially putting it ahead of its closest competitor, Baidu's Wenxin Yiyan. Subsequent financing rounds pushed The Dark Side of the Moon's post-investment valuation to a staggering $3.3 billion (approximately 21 billion yuan), cementing its lead in the domestic AI model startup race.

However, the immense financial outlay for traffic acquisition often cannot be offset by the revenue generated from new user subscriptions. According to analytics platform Zpeidia, Kimi’s smart assistant saw 24.05 million visits in July, but only 3.99 million were unique visitors, translating to an average of merely six app openings per user. Even a small number of free uses by new users incurs significant background computational costs.

In a similar vein to other cash-burning initiatives, Meituan and Didi could eliminate subsidies after users formed new consumption habits, but currently, the active marketing efforts of The Dark Side of the Moon fail to retain enough sticky users, signaling that this unprofitable model cannot be sustained indefinitely.

Navigating the B2C (ToC) landscape is proving to be a challenging endeavor for large model applications, as profit models rely heavily on subscription fees. Companies like OpenAI, the parent of the globally leading ChatGPT, boast over 7.7 million subscribers worldwide, achieving $1.9 billion in consumer revenue. However, this pales in comparison to the reported annual operating costs of $8.5 billion, illustrating a significant shortfall.

The Dark Side of the Moon previously introduced a paid feature called “Cheer for Kimi,” allowing users to purchase gifts ranging from 5.2 to 399 yuan for priority usage during peak times. This trial of user payment willingness appears more as a survey than a well-considered profit strategy.

Is it feasible to utilize business-to-business (ToB) models to cover the losses in ToC? There is considerable division within the industry on this topic. For instance, Li Kaifu, founder of Zero One Everything, has publicly asserted that “Zero One Everything is committed to ToC business and will not engage in unprofitable ToB business,” whereas Zhipu AI's CEO Zhang Peng has consistently argued that the willingness to pay in the business-to-business space is significantly higher compared to consumer markets.

Li Kaifu's concerns are not unfounded; while the ToB business model for large models is clearer in terms of profitability and client revenue willingness, the sector is currently embroiled in a pricing war initiated by powerful industry players.

Reports suggest that executives from Alibaba Cloud and Baidu Smart Cloud disclosed that, as of May, domestic margins for AI modeling computational power were above 60%, aligning closely with international benchmarks. However, subsequent price reductions in May led to a decline below the break-even point.

Entities that jumped into the large model pricing wars, such as Alibaba, Baidu, ByteDance, and Tencent, are often leveraging profits from their cloud operations as a buffer. Just as OpenAI may be losing money, the surge in demand for Microsoft’s cloud services due to rising rental needs has significantly boosted its performance. As reported by Sina Technology, part of the $1 billion venture capital raised previously by The Dark Side of the Moon included significant backing from Alibaba, with $600 million of this earmarked for consuming Alibaba Cloud services, showcasing a strategic interdependence.

Regardless, venturing into the ToB space, despite its potential downsides, is a necessary step for model startups. Zhipu AI has adopted its CEO’s philosophy by focusing on ToB while simultaneously exploring ToC, while MiniMax has poised itself in both B and C sectors, utilizing Kimi’s popularity in the consumer market to launch an enterprise-grade API in August.

A parallel scenario unfolded in the autonomous driving sector. Similar to The Dark Side of the Moon's strategy of spending on data acquisition through marketing, driving companies also require rapid deployment of software solutions to gather operational data on real-world road conditions, enhancing their AI driving capabilities.

While automakers with both complete vehicles and intelligent driving solutions can rely on hardware sales to fund their software investments, software-only companies must adopt the ToB approach, partnering with vehicle manufacturers to gain operational rights and potentially earn service fees.

The challenges faced by small and medium-sized startups in the presence of dominant players are a hard reality across industries.

During the era of significant capital expenditure in the internet sector, founders often faced one nagging question during venture pitches: “What if BAT (Baidu, Alibaba, Tencent) copies you?” This challenging inquiry resonates painfully in the current landscape of large models.

Celebrating a year since its inception, The Dark Side of the Moon announced a remarkable upgrade for Kimi, enhancing its long-text capabilities from 200,000 to 2 million words. Data from SimilarWeb indicated that Kimi's leading long-text abilities resulted in a peak daily active user count of 346,000, in turn causing a 45% week-on-week growth in user activity.

Once a unique feature is showcased, it becomes inevitable that competitors will replicate it. Within a month of Kimi's upgrades, Alibaba’s product Tongyi Qianwen and 360 AI Browser announced plans to support long texts of 10 million and 5 million words, respectively. Following suit, Baidu unveiled an upgrade for Wenxin Yiyan, enhancing its long-text capabilities to span 200,000 to 500,000 words.

With the incorporation of long-text tools by larger corporations, Doubao has emerged as a formidable rival to Kimi in the traffic acquisition battle. According to industry insiders, Douyin (the Chinese version of TikTok) temporarily 'banned' all other large model product advertisements to provide an exclusive promotional buffer for Doubao. After successfully instilling consumer loyalty among Douyin users, Doubao has launched a counter-offensive into Kimi's territory by purchasing keyword advertisement space for “Kimi” on Bilibili. This titanic clash has forced competitors like Zhipu to explore platforms such as Zhihu and Weibo for additional outreach.

Though The Dark Side of the Moon maintains some competitive edge against deep-pocketed giants, the historical content ecosystems crafted by firms like ByteDance and Tencent present formidable barriers that startups can struggle to overcome.

For instance, responses to similar queries from different large model applications may exhibit varied depth and quality; Baidu’s Wenxin relies heavily on content sourced from its proprietary channels like Baijiahao, while Tencent's Yuanbao is primarily fed by WeChat-based sources. Conversely, responses from startups like Kimi are often based on official documents and third-party media. Such differences in sourcing inevitably impact the relevance and accuracy of generated content in addressing user inquiries. OpenAI has already invested significantly in acquiring data from popular community platforms, an expense that will likely befall startups like The Dark Side of the Moon soon enough.

Although the accumulation of capital fosters technological advancement and overall industry progression—a broadly accepted notion—companies responsible for the cash outflows bear the obligation to prudently allocate funds into areas that yield long-term value. While avoiding immediate answers to the “What if BAT copies you?” dilemma may not be detrimental in the short term, a failure to devise strategies for long-term survival could transition concerns toward the more somber question of “How to be acquired by BAT?”

Say something