# Where China's AI Capital Flows: Apps Get Deals, Giants Get Cash

> In H1 2026, China saw 307.7B RMB across 1,203 AI deals. While apps dominated deal volume, over half the capital still went to foundation model giants.

Source: https://www.starbox.work/en/insights/ai-news/2026-07-20-142/ (Starbox AI Insights, updated 2026-07-20)

Three numbers defined China's AI primary market in the first half of 2026: 307.7 billion RMB in total funding, 1,203 individual deals, and 93 billion RMB captured by just three foundation model makers—DeepSeek, StepFun, and Moonshot AI. 

Those three companies alone absorbed nearly 30% of all capital invested across the market.

Look deeper into the data, and an interesting divergence emerges. By deal count, the vast majority of those 1,203 rounds went to AI application and agent startups. Yet by capital volume, more than half of all funding went straight back into the foundation model layer. 

For the first time, deal volume and capital allocation pointed in completely different directions. At the World Artificial Intelligence Conference (WAIC) in Shanghai on July 20, investors openly addressed this split.

## Volume Chases Applications, Capital Backs Foundational Models

Wen Zhi, General Manager of the Shanghai State-Owned Capital Investment Guiding Fund, noted at the conference that 2026 marks the critical inflection point from models to practical applications. His fund of funds manages 22.5 billion RMB, has deployed over 8 billion RMB, and connects nearly 300 underlying projects across 20 sub-funds. Its operating logic has shifted from simply writing checks to building an ecosystem, using an industry partner model to link capital, technology, talent, and real-world deployment scenarios.

Exhibition data from the event backs this shift: over 60% of companies exhibiting in the foundation model and generative AI sections carried primary labels around AI agents and applications, rather than core models.

Yet capital figures from the same period tell another story. The foundation model layer still pulled in 159.853 billion RMB in H1 2026—over half of all invested capital. DeepSeek alone raised a 51 billion RMB Series A round. 

The transaction volume is in applications, but the heavy money remains anchored to a small cohort of foundation model providers.

## The Questions Investors Ask Have Changed

Zhao Tiancheng, CEO of Om AI, observed that investor focus is pivoting away from language models in isolation toward models integrated with the physical world. Zheng Xuanle, founder of Lightsource Capital, framed this transition as three waves of AI re-industrialization: first digital labor, then scientific discovery, and finally physical labor.

Zhu Min, Tech Innovation Director at Shanghai State-Owned Capital Investment, pointed to an accelerating timeline: the path from founding to IPO for an AI company has compressed to roughly five years. His view is straightforward: the companies that integrate AI most effectively will produce the category leaders.

These shifts change how investors evaluate pitches. The conversation has moved away from parameter counts and total capital raised. Instead, investors now ask: What is your inference cost? When can you deliver at production scale? How long until you can go public?

A practical example on display at WAIC was UIH Intelligent's glasses-free 3D interactive display system. The platform completes multi-organ 3D reconstructions in minutes to assist with surgical planning. Demonstrations like this secure investor interest not because of the specific base model powering them, but because the software can be deployed inside an operating room today.

> What the data reveals is not capital abandoning the model layer, but a classic barbell strategy. On one end, investors place massive capital behind a handful of model giants, betting they will anchor the foundational infrastructure. On the other end, they distribute smaller, more frequent checks across application-layer startups, betting on who can turn model capabilities into scalable, shippable products. Taking positions on both sides is how venture capital hedges against not knowing which base model will ultimately win.

## Enterprises Do Not Need to Hedge Like Venture Capitalists

A barbell allocation makes sense for investors managing asymmetrical risk. When you cannot predict the winning model, spreading your bets limits the downside of choosing wrong.

Enterprises do not have that mandate, nor should they adopt that playbook. A mid-sized business cannot back 300 projects to hedge against choosing the wrong model provider.

Across hundreds of client deployments, we have found that the way to eliminate model dependency is not spreading bets across multiple vendors. It is making the underlying models completely interchangeable. 

Starbox AI built multi-model routing into its three-tier architecture from day one specifically to avoid vendor lock-in. Workloads automatically route to whichever model is most cost-effective for simple tasks, and switch dynamically to higher-reasoning models when complexity demands it. Organizations do not need to guess which foundation model will lead next year, because the architecture hedges that risk in real time.

The asset that actually warrants significant enterprise investment is not any single third-party model. It is the proprietary workflow design and industry knowledge organized inside a structured, four-dimensional knowledge framework—assets that retain their full value regardless of how quickly underlying models evolve.

Venture capital's barbell approach is an admission that even professional investors cannot pick the ultimate model winner. For business leaders, the takeaway is simpler: remove that gamble from your operational roadmap entirely. 

Review your current AI initiatives: are you spending budget betting on a specific model, or are you building infrastructure that remains valuable no matter who wins?