Following the unveiling of OpenAI and Google’s new product last week focused on innovative features and capabilities, ByteDance held its own product launch conference. Unlike its American counterparts, the highlight of ByteDance’s conference was the exceptionally low prices of its artificial intelligence models.
The price of ByteDance’s large Doubao language models is significantly lower than the industry standard. For example, the inference entry price for the Doubao General Model Pro-32k version is 0.0008 RMB per thousand tokens, while similar models on the market cost about 0.12 RMB per thousand tokens, or 150 times higher than that of Doubao. In other words, Doubao’s price is 99% cheaper than the market price.
“With just one yuan, you can buy 1.25 million tokens of our main Doubao model, which is roughly equivalent to 2 million Chinese characters, or about three volumes of ‘Romance of the Three Kingdoms’,” ByteDance introduced during his event.
In the age of AI, Chinese tech companies like ByteDance are using a familiar weapon – lower prices – to compete. But will it work this time?
A different strategy or a choice of no choice?
Large language models (LLMs) are still in the early stages of development and, according to some industry experts, the crucial goal is to establish technological superiority and maintain a sustainable lead over competitors. They argue that companies that continually iterate their products to impress users will achieve economies of scale, eventually allowing them to lower their prices. However, they also argue that relying solely on low-cost products to drive technological progress may not be the winning strategy in the era of generative AI.
Price competition also risks losing ground to competitors willing to take short-term losses to gain users. For example, Baidu significantly improved the performance of its Ernie model by 105 times and reduced inference costs to just 1% of the original in April. This allowed customers who previously performed 10,000 queries per day to now perform 1,000,000 at the same cost.
For ByteDance, the high costs of AI models pose an obstacle to its progress in China. Tan Dai, president of ByteDance’s Volcano Engine, said this year marks a significant breakthrough in LLM applications, but the cost of innovation is prohibitive. Only by reducing the cost of trial and error can we achieve wider adoption of generative AI. This is the main reason why ByteDance chose to lower the prices of its AI models.
However, the focus on low prices already highlights ByteDance’s delicate position: a company that initially rose thanks to its impressive recommendation engines is now forced to compete on price rather than innovation to ensure its future. This marks a significant change in fortunes for a company that had previously miraculously ridden the wave of AI technology.
The rise of ByteDance: driven by innovation and AI
More than a decade ago, ByteDance captured the transition of Internet content distribution from search engines to machine learning-based recommendations. Leveraging its pioneering recommendation algorithms, the company launched two flagship products: Toutiao and Douyin (and later its international version, TikTok).
Toutiao, launched in 2012, revolutionized the distribution of textual and image content on the Internet thanks to its innovative recommendation algorithm. By outperforming traditional news portal customers who rely on centralized, manually curated content, Toutiao has firmly established ByteDance as a dominant force in China’s internet landscape.
In 2016, Douyin emerged, bringing recommendation algorithms to the realm of short videos. Riding the wave of booming video content, Douyin has quickly become a major hub for internet traffic. This exponential growth has propelled ByteDance’s traffic, user base and business value to unprecedented heights, solidifying its position as one of China’s most valuable technology companies.
However, the advent of generative AI innovations has shifted ByteDance from an innovator to a target for disruptive innovations.
ByteDance’s Pursuit of GenAI: Will the App-Centric Approach Pay Off?
While many leading Large Language Model (LLM) startups emerged between 2018 and 2021, ByteDance entered the generative AI arena relatively late, beginning to consider GPT during its technology review of the first semester 2023.
In its pursuit of innovations in generative AI, ByteDance is adopting a diversified strategy reminiscent of its application-centric days. Positioned as an “applications factory,” ByteDance casts its net wide by incubating a diverse range of GenAI applications.
ByteDance now has a robust product portfolio. At its recent conference, ByteDance introduced its new Doubao LLM family, including two general models and several specialized models for use cases such as role-playing, speech recognition, and text-to-image generation.
The company has also rolled out more than ten products, including AI dialog assistant Doubao, AI bot development platform Kouzi, and animated AI chat bot Hualu. ByteDance has enhanced its video editing app Jianying and desktop app Feishu with GenAI features.
Recent metrics indicate promising user engagement: Doubao’s daily active users (DAU) reached around 3 million in March this year, while Kouzi reached one million DAU. Gauth, competing in overseas markets, saw its DAU peak at over two million after integrating AI features.
Additionally, Zhu Jun, Vice President of Product and Strategy of ByteDance, revealed that more than 8 million intelligent agents have been created on Doubao, with 26 million monthly active users.
ByteDance’s AI strategy: the verdict is still in
Despite ByteDance’s impressive track record with products like TikTok, its foray into generative AI (GenAI) has not seen the same level of success. While TikTok has accumulated more than 2 billion downloads worldwide in two years, ByteDance’s AI offerings have yet to demonstrate their viral growth potential. Even compared to emerging startups like Moonshot AI’s Kimi chatbot, ByteDance’s GenAI products lag behind.
At the event, ByteDance attributed its significant price reductions to both capacity and necessity. Upgrading from single-machine inference methods to distributed inference methods improves calculation efficiency, allowing ByteDance to reduce its prices.
While the move could benefit China’s GenAI sector, it raises questions about ByteDance’s reliance on price rather than innovation. The effectiveness of ByteDance’s GenAI strategy remains to be seen, leaving the final verdict undecided.
