Under the China-US AI Gap, To B or Not To B — Here's What Kai-Fu Lee Thinks

By Lili Yu

This is an era when everyone stands at a crossroads — and for this generation of AI entrepreneurs, perhaps more so than anyone. If the internet and mobile internet eras represented a remarkable convergence, a near-perfect duet between Chinese and American founders who had timing, geography, and human talent on their side, then the new AI 2.0 era of large models has placed entrepreneurs in almost the opposite position against a backdrop of US-China competition. From day one, they've had to confront an unprecedented set of divisions: build models or build applications? Go global first or focus on the domestic market? And before a killer app emerges, choose a B2B or B2C path to monetization? All of this has grown especially murky with Trump's return to the presidency.

In many ways, Kai-Fu Lee is among the most qualified people to discuss these questions. As the founder of 01.AI, one of China's "Six Little Tigers" of large model companies, he has remained active in both the Chinese and American tech communities, and 01.AI has continued to explore both domestic and overseas markets, both B2B and B2C models.

On US election day, 01.AI unveiled its "Ruyi" digital human solution for e-commerce livestreaming and office meetings, alongside an AI infrastructure solution. Combined with its existing open platform centered on the Yi series of large models, these form a complete suite of B2B offerings.

The following are edited remarks from Kai-Fu Lee and 01.AI co-founder Qirui Feng at the launch event, as compiled by "Dark Tide Waves":

1. On the pursuit of AGI, many people ask how we can compete with American companies. The answer: using America's playbook, of course we can't beat them — but we can use China's playbook.

Where America is strong: first, its ability to break through and invent. Second, it can raise enormous amounts of money for massive resource investment. They use money to buy time. We use the diligence and hard work of Chinese engineers to drive down training and inference costs.

American large model companies aren't planning to enter China for now, and would face difficulties if they tried. Chinese companies, meanwhile, can operate in China and also find opportunities abroad — Belt and Road countries for B2B, and B2C without geopolitical constraints. Going forward, China and the US will have their own separate markets across industries.

2. We've been talking with many European and American clients. Our sense is that for large model consumer applications, the US is indeed ahead of China. But for B2B applications, China may well end up ahead of America.

American clients, when it comes to large models — forget private deployments — aside from GPT, few even use supervised fine-tuning (SFT). One major reason is that US and European compliance and data requirements create so many constraints that they become less open, which actually slows them down in application and deployment.

On the supply side, only one American large model company is truly doing B2B. OpenAI and others just offer APIs, maybe a dedicated instance at most, but SFT you do yourself. I have a client in the US, DoorDash — America's Meituan — their intelligent customer service saved $50 million this year using large models directly, but they did all the SFT themselves. No one in America does on-site private deployment for clients. So once engineering and delivery capabilities come into play, China has ample ability to compete globally.

3. For Chinese companies doing B2B overseas, the difficulty will be substantial. But B2C overseas should have opportunities. For a startup, several things matter: first, does your domestic team truly understand foreign enterprises? Second, foreign domains change very quickly. Third, China's advantage is having very good large models with strong English capabilities and lower average inference costs — how to leverage this advantage.

4. For 01.AI's B2C, we're building up abroad first, establishing a viable path, then looking for suitable opportunities to continue domestically. ChatGPT ignited all of America; China hasn't developed that atmosphere yet, so we hope to first raise our product capabilities and validate our commercial path in more mature regions with better payment habits, where we're already seeing quite good revenue and growth. We'll definitely want to enter the domestic market later. Our strategy remains focused: B2C first, focused overseas; B2B focused on creating real value, not customization or project-based work.

5. Project-based work means the product itself is slightly non-standard, while customization accounts for the bulk. From SaaS to big data, BI to AI 1.0, to AI 2.0 — China's repeated attempts in the B2B ecosystem have all been about how to transform the products and services clients need into standardized offerings delivered through cloud services, SaaS, or token-and-service models within large models, rather than man-day-based customization. Solving this requires returning to the essence of the business, finding domains deep and broad enough that large models can restructure their application interfaces.

6. Looking at where large models are actually landing in B2B today, the use cases converge on two core scenarios. First: using models combined with private data in a knowledge base to power generation — this is about cost reduction and efficiency. Second: Q&A powered intelligent assistants. Frankly, today's intelligent assistants can only assist as copilots, only answer questions — they can't do autopilot. For enterprises, if content is your product, both scenarios work well.

7. For retail, manufacturing, and similar industries, content generation is only auxiliary; it doesn't truly enter business workflows. Retail is the first industry solution we're presenting in a relatively complete form. We didn't build a knowledge base or HR assistant — we looked at clients' real business pain points: first, growth; second, costs; third, how these traditional industries keep pace with new technology. So we're not focused on back-office HR, finance, or even intelligent customer service — we'll do those too, because the ROI is clear, but that's merely cost reduction. We want to truly understand the store.

8. Before building digital human livestreaming for retail, we worried it might be a red ocean. But what's interesting is: among a hundred digital human vendors, very few focus on retail livestreaming. Why? Because in the past, using PCs controlled by IT staff was convenient enough. But China still has many counter-style shops — just twenty, fifty square meters, with no space to put a PC. So we didn't just use large models; we rebuilt this product from the mobile side.

9. There's also the centralized management platform. With current digital human livestreaming, setting aside how complex it is to start a broadcast, every shop employee has to press buttons ten times just to go live each day. Ten presses times 365 days times ten thousand stores — that's an enormous burden. We've put all digital human livestreaming, including broadcasting, operations, and all management, into a back-end cloud service. We're single-button go-live, zero-interference operations — we don't disturb the staff.

With large models plus mobile plus rebuilt management and service platforms, we found we not only delivered growth value to clients, but turned a red ocean product into a blue ocean, because many hadn't built from the business scenario outward. For us, being six months ahead is enough. Most critically, if we can find such an entry point, we can follow the store's logic to find others in different industries.

10. Six months ago we put it simply: we wouldn't do B2B that loses money. Our B2B solution launch now isn't a strategic change — it's because B2B can now generalize, creating value for users and partners. These things are one and the same with not doing unprofitable B2B; the strategy hasn't changed.

For B2B opportunity: first, multimodal must enter the mainstream and commercialize. Second, the fundamental shift from OS on CPU to intelligence centers with models on GPU brings enormous opportunity — these are core points.

11. AGI remains fairly distant — about seven years, in my personal judgment. But multimodal becoming mainstream, AI search becoming mainstream, AI B2C and B2B becoming mainstream, helping B2B companies make money, being used by more people, mobile apps being rewritten — all of this will happen in the next two years.

What we need to hold firm to: China's engineering capabilities are stronger than many American companies, and our technical talent is more diligent and hardworking. The app era was China's blessing — China's understanding of app methodology, its insight, and the people who can execute exceed America's. The injection of multimodality will also give birth to richer AI-first applications. This is already happening.

Image source | IC Photo