Source | Bohu Finance (bohuFN)

Editor's Note: As 2024 draws to a close, we reflect on a year filled with change and turmoil. E-commerce giants strategized to capture market share, showcasing their top talent. More electric vehicles were sold, yet fewer brands emerged. Pangdonglai refrained from expanding beyond Henan but planted seeds nationwide. Overseas, TEMU expanded rapidly yet faced growing regulatory pressures.

Our 'Reflecting on 2024' series documents, analyzes, and summarizes the most representative companies and trends across various industries. This inaugural article focuses on ByteDance's rise in the AI industry.

Recognized as an innovative tech company, ByteDance initially appeared slow in AI initiatives. However, once its direction was set, it emerged as the most aggressive player.

In the current large model race, processing 500 billion tokens daily is remarkable, signifying a leading position. In natural language processing, a token represents the smallest text unit, be it a word, punctuation, or subword.

Processing 500 billion tokens daily involves handling vast data, akin to millions of long articles or tens to hundreds of millions of social media posts.

Recently, MiniMax, a Chinese large model startup, claimed to process over 3 trillion tokens daily, shocking the industry. However, as competition intensifies, token volume is no longer the sole metric; focus has shifted to application implementation.

Known as the 'king of competition,' ByteDance, despite entering the large model scene less than two years ago, has left its mark.

In November, Doubao App ranked second globally and first domestically in monthly active users (MAU) for AI large models, with 59.98 million MAUs, second only to OpenAI's ChatGPT. Its overseas version, Cici, ranked 22nd with 12.67 million MAUs.

In February 2024, Doubao had just 1.73 million MAUs.

Doubao has become a significant player among top domestic large models, popular and promising commercially. This positively impacted Doubao-related stocks, making them a leading growth option in the AI sector within a month.

From a low profile in the first half to full commitment in the second, ByteDance's large model empire continues to expand. How did this 'latecomer' achieve such a turnaround?

01 From 'Latecomer' to 'Game-Changer'

On August 18, 2023, ByteDance renamed its AI chat product Grace to 'Doubao.' Amid the 'heavyweight releases' in the hundred-model war, this seemingly minor news ushered in a new era for ByteDance's large models.

Compared to others, ByteDance's entry was delayed. At the 2024 annual meeting, CEO Liang Rubo acknowledged the company's inferior technical sensitivity, noting GPT discussions only began in 2023, while industry leaders founded their startups between 2018 and 2021.

In 2019, Baidu led with Wenxin, marking a milestone. Alibaba's DAMO Academy and Tencent's Research Institute followed, accelerating in-house large model development. In 2021, Alibaba released the world's first 10-trillion-parameter multimodal large model, while Tencent launched multiple models with 100 billion to 1 trillion parameters, laying a solid foundation.

By 2023, domestic large models boomed, with companies like Alibaba Cloud, Tencent, 360, Huawei, iFLYTEK, SenseTime, Baichuan, and Zhipu AI releasing their own models.

As the industry matured, ByteDance entered late.

ByteDance didn't underestimate AI. It developed its first recommendation engine in 2012 and an AI Lab in 2016. However, core AI Lab departures slowed exploration. ChatGPT's emergence highlighted ByteDance's lag, prompting Liang Rubo's reflection at the 2024 annual meeting.

To catch up, ByteDance adjusted its strategy, recruiting AI talent, establishing a dedicated AI department, integrating resources, and implementing a horse racing mechanism to accelerate Doubao's development.

In March, Doubao's downloads and MAUs soared, continuing through November. While ChatGPT leads with over 300 million MAUs, Doubao's rise surpasses competitors. Domestic rivals like Kimi, Wenxiaoyan, and Tongyi Qianwen lag behind, with Doubao surpassing their combined MAUs.

From C-end users to B-end industries, Doubao boasts large usage and diverse applications. According to Volcano Engine President Tan Dai, Doubao processes 120 billion tokens and generates 30 million images daily.

In over a year, ByteDance achieved results envied by AI entrepreneurs. Despite being a late starter, it transformed from 'latecomer' to 'game-changer.'

02 Growth Strategy: Competing for Users, Prices, and Computing Power

ChatGPT reignited AI imagination, with large models taking center stage. Over two years, a fierce 'hundred-model war' ensued, focusing on technology and product iteration speed and commercial implementation efficiency.

Domestic tech giants like Baidu, Alibaba, Tencent, and Huawei, along with AI-focused vendors like iFLYTEK, SenseTime, and Megvii, and startups like Zhipuhuahuangzhang, Baichuan Intelligence, and Daguan Data entered the fray. Vertical industry companies also developed large models based on their technology and data.

Universities and research institutions actively participated, releasing about a quarter of large models. From major companies to startups and academia, all entered the competition, making it fiercely competitive.

After the hundred-model war, nearly 'one in ten' survived. How did ByteDance, a 'latecomer,' achieve its turnaround?

It competed for C-end users, B-end prices, and computing power. Doubao's 'saturation attack' strategy outpaced competitors in every field.

Since the year's beginning, Doubao spent over 1 billion yuan on C-end advertising, capturing users' attention on social media, search engines, and short video platforms.

While C-end investment captured users' minds, B-end saw a lethal 'price war.' At the Volcano Engine FORCE conference in May, Doubao Pro pricing was set at 0.8 cents per 1,000 tokens for the 32k model (99.3% lower) and 5 cents for the 128k model (95.8% lower).

For perspective, one yuan can buy 1.25 million Doubao tokens, roughly 2 million Chinese characters, equivalent to three copies of 'Romance of the Three Kingdoms.'

Unlike companies announcing models with various evaluations, ByteDance took a pragmatic approach, aiming for user recognition with affordable prices, especially among enterprises.

Large model competition also involves computing power and talent. ByteDance's stable resource injection played a crucial role.

In 2024, ByteDance invested heavily in AI, with expenditures reaching 80 billion yuan, close to Baidu, Alibaba, and Tencent's combined total (approximately 100 billion yuan). In 2025, expenditures will reach 160 billion yuan, with 90 billion for AI computing power and 70 billion for IDC infrastructure and network equipment.

In AI talent acquisition, founder Zhang Yiming personally recruited from Alibaba Group, Lingyiwanwu, Zhipu, and others, emphasizing 'Artificial General Intelligence' internally.

Due to this, ByteDance stood out in AI application competition, gradually building a unique competitive advantage with Doubao's stability, reliability, and powerful application development capabilities, establishing a strong presence in AI.

03 Beyond Doubao: ByteDance Aims for the Next 'App Factory'

'In the AI era, all products deserve a large model upgrade,' former Alibaba CEO Daniel Zhang said. This is now validated. Many domestic platforms deeply integrate AI into their core businesses, accelerating AI product launches in traditional sectors.

ByteDance is no exception. Doubao is just one aspect of its large model applications. As models mature, departments like Douyin, Volcano Engine, and Giant Engine also explore AI.

To date, ByteDance has launched the Doubao large model family, Volcano Ark, and advanced AI applications and cloud infrastructure products.

The Doubao family expanded to nine product lines, covering general-purpose, role-playing, speech synthesis, voice cloning, text-to-image, speech recognition, vectorization, and function call models, meeting diverse user and enterprise needs.

Volcano Ark focuses on B-end applications, including intelligent outbound calls, digital humans, and data assistants, reducing enterprise costs and technical barriers. It cooperates with leading automotive, mobile, finance, and food & beverage companies like Geely Automobile, Great Wall Motors, OPPO, vivo, Xiaomi, ASUS, China Merchants Bank, and Haidilao.

Empowered by Doubao, AI software and internet applications emerged, including Kouzi (AI agent platform), Xinghui (image generation), Maoxiang (role dialogue), Doubao Aixue (AI education), and Gauthmath (overseas homework search).

From large models to AI development, social networking, and AIGC creation, ByteDance built a comprehensive AI product ecosystem.

In the mobile internet era, ByteDance was the 'App Factory,' creating hit products like Toutiao and Douyin, making it one of China's most profitable internet companies.

In the large model era, ByteDance aims to create another growth miracle – an 'AI Factory,' exploring the possibility of a third super app.

Beyond AI software, ByteDance also integrates large models with hardware. In the Internet of Things era, hardware is the software's carrier and user traffic gateway to the ecosystem. AI hardware development translates software advancements.

At the Volcano Engine FORCE conference in May, ByteDance showcased robotic dogs, learning machines, and robots developed by partners.

During the Mid-Autumn Festival, ByteDance launched Xiangyanbao, an AI companion toy, as a Volcano Engine special gift to customers.

Unlike conventional toys, Xiangyanbao stands out with its integration of the advanced FoloToy large model AI core, Magicbox, showcasing the prowess of the Doubao large model and Kouzi. Users can converse with Xiangyanbao via straightforward commands, and this AI toy will respond accordingly.

Although ByteDance hesitates to label Xiangyanbao as an official product, this attempt to infuse AI into toys marks a novel and groundbreaking initiative.

Notably, ByteDance is no stranger to hardware exploration.

From acquiring the Nut mobile phone team and certain patent rights from Smartisan Technology in 2018, leading to the launch of Nut mobile phones, TNT displays, speakers, and other peripheral products, to focusing on educational hardware in 2020 with the introduction of the 'Dili Education' brand and products like smart study lamps, educational tablets, and electronic dictionaries, and ultimately acquiring PICO, a leading domestic VR vendor, in 2021.

However, amidst fierce competition in the hardware market and the impact of the 'double reduction' policy, ByteDance's achievements have been modest. Currently, only the intelligent study lamp remains listed on the Dili Education official website. Meanwhile, PICO has undergone several rounds of layoffs over the past year and now maintains a small hardware team.

As AI computing increasingly permeates various industries, a transformative shift may be imminent. Nevertheless, ByteDance's ambitions in the AI era are evident – from the Doubao large model to the creation of an AI application pipeline spanning software to hardware, the company aims to identify a robust growth framework within the burgeoning AI landscape and achieve rapid expansion.

The cover image and illustrations in this article are the property of their respective copyright holders. If any copyright holder believes their work is unsuitable for public viewing or should not be used without charge, please contact us promptly, and we will take immediate corrective action.