Summary of Key Points
This article examines the strategic transformations of ZeroOneAllThings, founded by Kai-Fu Lee, over the past three years, revealing the realities of China's AI startup scene. Initially, ZeroOneAllThings aimed to create a "WeChat/Douyin" for the AI era—super apps that would compete with GPT-5-level large models. Later, the company shifted its focus to B2B services and abandoned the development of cutting-edge large models. More recently, it has aimed to become like Palantir, providing enterprise-grade AI solutions. Each adjustment was presented as an proactive choice made by "seeing the end of the era ahead." The article satirizes how some AI entrepreneurs use storytelling to mask their forced transformations and reflects the shift from a bubble to a more pragmatic approach in the AI startup landscape.
Breakdown and Interpretation
1. Three Major Course Changes in Three Years: ZeroOneAllThings' Evolution
ZeroOneAllThings' journey over the past three years has been akin to a rollercoaster, with different goals at each stage:
- 2023: Chasing trends, focusing on B2C super apps and large models
After ChatGPT became popular, Kai-Fu Lee launched ZeroOneAllThings with the aura of an AI evangelist. The company's slogan was to create world-class large models and join the global elite, and they open-sourced the Yi-34B model (bilingual in Chinese and English). Within half a year, its valuation exceeded $1 billion (making it a unicorn). They also aimed to develop B2C super apps similar to WeChat and Douyin, considering B2B business too challenging.
- 2024: Shifting to B2B, abandoning large models
Reality quickly struck: B2C users were attracted by competing platforms like DouBao and Kimi, and the API price war reduced profits significantly. Training large models was extremely costly (up to $3 million per iteration). ZeroOneAllThings shifted its focus to B2B, making it the main source of revenue (70%) and abandoned the plan for trillion-parameter models like GPT-5, instead launching cost-effective smaller models (e.g., Yi-Lightning, which costs only 0.99 yuan per million tokens).
- 2026: Rebooting as a "Chinese Palantir"
Now, they aim to become the "Chinese version of Palantir," providing on-site engineers to help companies solve real problems with AI (e.g., improving financial reports). The company's homepage has also changed from promoting "AI super apps" to emphasizing "transforming AI into measurable business outcomes."
Each change felt like a complete shift in direction, yet they were all justified as "pre-emptive moves to see the industry's future."
2. Why the Constant Changes?
Behind each adjustment was a practical survival issue:
- B2C failed: Users had already adapted to free tools like DouBao and Kimi, and ZeroOneAllThings' B2C products (e.g., WanZhi, PopAI) lacked differentiation, leading to slow user growth and losses (AI services required token fees, which increased costs with more users).
- High cost of large models: Training large models required extensive resources (GPUs and funding), beyond what startups like ZeroOneAllThings could afford compared to giants like Alibaba and ByteDance. They realized that the marginal benefits of developing trillion-parameter models were minimal, so they had to give up.
- Urgent need for commercialization: Investors focused on revenue, and the company needed to survive. Although B2B was more demanding (long negotiation periods and customization), it provided immediate financial returns. In 2024, B2B revenue accounted for 70% of ZeroOneAllThings' income, allowing the company to stay afloat.
In essence, the changes were necessary due to practical constraints.
3. Turning Passive Transformations into Proactive Strategies: Kai-Fu Lee's Storytelling Magic
Kai-Fu Lee's strength lies not in technology but in his ability to tell stories that make these adjustments seem like pre-emptive moves:
- 2023 (B2C focus): "B2C is the path for great companies; B2B is too difficult."
- 2024 (B2B shift): "Only by creating real value for businesses can we survive."
- 2025 (abandonment of large models): "Large models are a resource war for giants; startups shouldn't chase excessive parameters."
- 2026 (Palantir comparison): Using the concept from the Tao Te Ching, he presented this shift as a new direction driven by practical needs.
This consistent narrative made it seem like ZeroOneAllThings was always one step ahead of the industry.
4. Comparing with Other AI Companies
Other startups, such as Kimi and Zhipu, continued to pursue large models:
- Kimi: Founded in 2023, released K2 (trillion parameters) in 2025, and K3 (2.8 trillion parameters) in 2026, with open-source models.
- Zhipu: Launched GLM4.5 in 2019 and improved to GLM5.2 in 2026, catching up with international top models in programming capabilities.
This shows that it's not inevitable for startups to give up on large models; it's often a strategic choice (e.g., to reduce costs or accelerate profitability). Kai-Fu Lee has resources (Innovent Factory, connections with big companies, and fame), but he chose a more pragmatic path.
5. Is the Palantir Model Easily Replicable?
Palantir is an American company that helps governments and large enterprises with AI solutions (e.g., anti-terrorism, financial risk management). Its approach emphasizes on-site engineering and customized development, with long sales cycles (6-9 months) and high fees. It took 20 years to become profitable, thanks to accumulated trust from government and corporate clients.
ZeroOneAllThings faces two major challenges in following Palantir's model:
- Customer base: Palantir has a 20-year history with these clients; does ZeroOneAllThings have enough time to build such a customer base?
- Delivery capability: Palantir's engineers understand industry-specific needs (e.g., finance, defense); can ZeroOneAllThings' team quickly adapt to corporate requirements?
Whether this approach will be successful remains uncertain.
Conclusion
ZeroOneAllThings' three-year journey reflects the broader trend in China's AI startups: from chasing trends and conceptualism to a more pragmatic focus on survival. Kai-Fu Lee's storytelling is impressive, but success in AI entrepreneurship ultimately depends on products, revenue, and customer value. Future entrepreneurs would do well to avoid overly definitive narratives and focus on practical solutions.