The Second Half for AI Startups: Where Is the Capital Headed?

Author | Wang Hanyu

Editor | Zhang Fan

The 2026 venture capital market is witnessing an unprecedented "AI Great Divergence". According to Crunchbase data, global start-up financing in the first half of this year reached approximately 510 billion US dollars, exceeding the total amount for the whole of last year; in the first quarter alone, around 80% of venture capital flowed to AI companies. OpenAI completed a single round of financing of 122 billion US dollars, pushing its valuation to 852 billion US dollars, while Anthropic overtook it with a valuation of 965 billion US dollars, becoming the world's most valuable AI start-up.

Beneath the frenzy, the differentiation is equally drastic: leading model companies still boast astonishing financing capabilities, while a large number of vertical applications have users but struggle to turn a profit.

The market is no longer competing on model parameters, but on application implementation. The narrative of the inflection point of the "second half of the AI era" is replacing the previous computing power arms race — for start-ups still at the table, the rules of survival have completely changed.

As global capital flows to high-growth opportunities, where is the next destination? From November 2 to 3, Hong Kong FinTech Week × StartmeupHK 2026 will be held, drawing the market's attention once again to this city that connects mainland innovation and global capital. Before the opening of the tech week, 36Kr had in-depth exchanges with Ji Wei, Founding and Managing Partner of China Growth Capital, and Lui Kam Yung, Acting Head of StartmeupHK of InvestHK, trying to answer a question: In the second half of the AI start-up era, what is the real judging criterion?

Having competed on model parameters, now competing on the "data flywheel"

Ji Wei, who has reviewed countless projects, first put forward a general analytical framework: AI investment can be divided into three layers: infrastructure layer, model layer and application layer. The infrastructure layer covers GPUs, high-speed optical interconnection, advanced packaging, storage systems, even power supply and thermal management, as well as system-level links such as inference scheduling, heterogeneous computing power adaptation, data governance and security.

In her view, the external questioning of "excessive investment in computing power" is more of a rhythm problem than a direction problem — there is a natural cycle mismatch between infrastructure investment and revenue generation. Just like around the year 2000, submarine optical cables and switches thrived before internet companies became profitable, and Google did not even have revenue at that time.

What truly determines how far a segmented track can go is the "data flywheel" she has mentioned several times: at what cost data is acquired, in what form it enters the model, and whether it can form positive feedback in a short time to feed back and improve model capabilities.

This is also why programming is currently the AI application with the largest revenue scale — the right or wrong of code is clear at a glance, and the positive feedback is extremely strong; while video generation has real demand, it is much more difficult to improve the model because the feedback is hard to standardize. Just as OpenAI shut down Sora and Anthropic has never dabbled in video, their judgment is exactly based on this point.

Market verification came very quickly. Cursor's annualized revenue rose from breaking 1 billion US dollars to about 2.6 billion US dollars within a year, and GitHub Copilot has more than 4.7 million paying users. In Ji Wei's view, the ceiling of any AI application track ultimately depends on the total market size and user attention, rather than the dazzling display of a single technology.

From the demand side, Lui Kam Yung supplemented another layer of criteria: whether it can solve real, high-frequency and high-value pain points, whether the customers are clear and willing to pay, and whether technical barriers can be continuously transformed into cost advantages — people who have experienced the baptism of the internet bubble know that all the companies that survived after the bubble burst have firmly established their business models first.

Embodied Intelligence: the most popular track, in the most "early" stage

In Ji Wei's eyes, embodied intelligence is still the most active field in AI applications, and also the most "early" field. The frequency of model release and evaluation is still continuing, indicating that the smartest minds are still pouring into this field, but the "ChatGPT moment" belonging to it has not yet arrived — her conservative expected timeline is 5-10 years, but she firmly believes this moment will come.

A viewpoint that has been clarified repeatedly is that embodied intelligence is not equal to humanoid robots. It can be bionic forms on the seabed, autonomous aerial vehicles at low altitude, or flexible robots in space, performing work in places inaccessible to humans.

China Growth Capital has made layouts in all these directions, including low-altitude autonomous flying robots, and China's first flexible robot that completed verification in space; its invested embodied intelligence world model company — Independent Variable Robotics — is also one of the active world model teams in Hong Kong at present.

Lui Kam Yung also believes that embodied intelligence is a systematic project of "model + hardware + data + scenario feedback", and all links have not been truly integrated: human engineers still need to stand by to guarantee the quality of screw tightening, and the stability of logistics handling is far from meeting the standards. Currently, the scenarios that can be implemented are mostly relatively restricted scenarios such as patrol and inspection.

Market forecasts also confirm this "early-stage" consensus. Morgan Stanley predicts that the market size of China's humanoid robot industry will be about 2 billion US dollars this year, and may reach 15 billion US dollars by 2030; different institutions have significantly different calculation calibers for generalized embodied intelligence, ranging from hundreds of billions to trillions of US dollars.

For Hong Kong, AI and intelligent manufacturing, AI for Science, green energy management, and high-value enterprise-level applications in finance, healthcare, logistics and other sectors are the most promising directions in the next 3-5 years.

Lui Kam Yung specifically mentioned that Hong Kong's research universities have accumulated a large amount of scientific research IP. The significance of AI intervening in laboratories is to greatly shorten the "Edison-style trial and error" process — from dry experiments to wet experiments, it reduces human errors and improves the success rate, which may give birth to a number of high-value breakthrough achievements. The manufacturing ecosystem in the Greater Bay Area is accelerating integration with Hong Kong's scientific research resources, which also provides a more realistic implementation soil for the supply chain of embodied intelligence.

Cash flow becomes a required course for founders

When 36Kr asked what small and medium-sized AI start-ups that are running out of cash should do, Lui Kam Yung's suggestion was very straightforward: conduct a calm cash flow analysis as early as possible, focus on businesses that can bring positive cash flow, and cut non-core expenses; when funds are about to run out, communicate frankly with strategic investors instead of waiting for the crisis to break out before seeking help. Government subsidies should also be included in the plan as early as possible, instead of applying for them when financing is urgently needed.

Ji Wei also mentioned that cash flow management is a required course for founders. "If you only have 1-6 months of cash left in your pocket and are still asking how to spend it, you mostly need to reflect on your planning ability."

In her view, business milestones must be advanced in tandem with financing milestones — from government scientific research funds, achievement transformation support from early institutions, to market-oriented institutions, and then to industrial capital that can bring orders and funds, every step must be planned in advance, instead of working hard for a whole year before looking for financing.

She also believes that mergers and acquisitions should not be regarded as crisis management, but as an active choice in the development process, and only the real "1+1>2" effect is worth pursuing.

Combined with past investment cases, Ji Wei specially mentioned a real transformation sample: a company that originally engaged in automatic cutting and polishing of lab-grown diamonds, relying on its accumulated core technology, turned to heat dissipation materials required for chip stacking, and found its second growth curve in the semiconductor field. In her view, the application of AI in all walks of life is still in the early stage, and companies with core technical barriers are not late to try again in another track.

How to get a good score on the "going global" test

Talking about the upcoming Hong Kong FinTech Week × StartmeupHK 2026, Ji Wei revealed that she will have a dialogue with two other guests at the tech week, with the theme "New Trends of Global Capital: Decoding the Next Golden Investment Track". "It's actually quite similar to what we talked about today." Lui Kam Yung smiled and added: "You have already spoiled the content today."

Standing in the Hong Kong market, the topic naturally falls on the required question of mainland enterprises going global.

Ji Wei specially mentioned Fao Robotics, a collaborative robot enterprise invested by China Growth Capital. This company ranked first in total sales of collaborative robots in China and first in overseas shipments in 2025. In Ji Wei's view, its development perfectly fits the phrase "Think Global, Act Local".

Ji Wei believes that the traditional going-global path is usually to set up overseas entities first, participate in exhibitions, sign agents, and then find customer demands; but the truly effective path may start from customers, figuring out what the other party needs, how much their budget is, who makes procurement decisions, what alternative solutions there are, how data is processed, and who is responsible for after-sales services.

Therefore, a series of plans such as data compliance, intellectual property rights, export control and security review should be advanced to the product design stage, instead of making up lessons after the market is fully laid out. She also does not recommend expanding globally right from the start, but suggests running through the operation in a benchmark market first, and then replicating the experience to adjacent markets.

Lui Kam Yung added that apart from being fast when expanding business, it is more important to be precise and down-to-earth. When enterprises go global, they should try to avoid directly copying the "expanding across provinces" model used in the mainland to overseas markets — every city has different culture, regulations and even historical backgrounds. Respecting local rules, establishing local sales and delivery teams, and operating businesses in a down-to-earth manner are often more important than technology itself.

In this going-global narrative, Hong Kong's unique role is also being re-understood. R&D, supply chains and application scenarios can be placed in the Greater Bay Area within the one-hour living circle, while international financing, regional headquarters, compliance structures and talent introduction are settled in Hong Kong. Relying on the common law system and international arbitration experience, overseas customers are more willing to sign contracts with Hong Kong entities; Hong Kong is also helping enterprises transform mainland innovation capabilities into long-term trusted business relationships for international customers.

Over the past year or more, China Growth Capital has regarded Hong Kong as a base for "re-entrepreneurship": applying for licenses, preparing for a Hong Kong dollar fund, whose future investment will target outstanding global entrepreneurs and forward-looking technological breakthroughs, and deeply participate in the scientific research transformation of local universities in Hong Kong. At the same time, according to Ji Wei's observation, more and more venture capital peers are also strengthening their layout in Hong Kong.

Entering the second half of the AI era, Hong Kong, the two-way gateway connecting mainland innovation and global capital, is no longer just the listing terminal for large enterprises, but is becoming a going-global bridgehead for more AI start-ups. Just as Lui Kam Yung's advice to enterprises going global: before you go sailing, come to the seaside and take a look first.

Follow us for more updates