Artificial intelligence is attracting an increasingly large share of venture capital, but the funding remains concentrated among a handful of companies. According to Crunchbase, global startup funding reached approximately USD 510 billion in the first half of 2026, surpassing the USD 440 billion invested throughout 2025. AI companies alone accounted for about 80% of global venture funding in the first quarter.
OpenAI secured USD 122 billion in committed capital in a single round at a post-money valuation of USD 852 billion in March. Anthropic subsequently raised USD 65 billion at a valuation of USD 965 billion in May, overtaking OpenAI.
Yet while leading AI model developers continue to attract substantial investment, many businesses built around specialized applications have struggled to translate user adoption into profitability.
As attention shifts from model capabilities and computing infrastructure to practical applications, startups face a different set of tests.
Ahead of the main conference at Hong Kong FinTech Week x StartmeupHK 2026, 36Kr spoke with Tracy Ji, founding managing partner of Meridian Capital (also known as Huaying Capital), and Chin Yung Lu, acting head of startups at Invest Hong Kong. The pair shared how investors generally assess AI businesses, how founders should manage limited cash reserves, and what Hong Kong can offer companies expanding internationally.
Why useful feedback matters
Ji divides AI investment into three layers: infrastructure, models, and applications. Infrastructure encompasses GPUs, high-speed optical interconnects, advanced packaging, memory, power supply, and thermal management. It also includes software systems for inference scheduling, heterogeneous computing, data governance, and security.
She views concerns about excessive investment in computing capacity primarily as a matter of timing. Infrastructure spending and the revenue it eventually supports operate on different cycles. Around 2000, for example, investment in subsea fiber optic cables and network switches surged before many internet businesses became profitable.
For Ji, a more important test of an AI application’s prospects is its ability to establish a feedback loop: how much data costs to acquire, how it enters a model, and whether feedback can improve the model quickly.
Coding illustrates this advantage: whether a piece of code works is relatively straightforward to determine, creating a clear feedback mechanism for improving AI-generated output. It is therefore perhaps unsurprising that coding has emerged as one of the largest, if not the largest, AI application categories by revenue.
Video generation presents a different challenge. Although demand exists, feedback is less standardized, making it harder to evaluate and improve model performance. Ji sees this distinction as relevant to the strategic choices of AI developers, pointing to OpenAI’s discontinuation of Sora and Anthropic’s decision not to pursue video generation.
The commercial growth of AI coding tools illustrates the opportunity. According to Forbes, Cursor’s annualized revenue surpassed USD 4 billion in June, up from USD 2 billion in February. The company was subsequently acquired by SpaceX in August. Microsoft, meanwhile, reported more than 4.7 million paid GitHub Copilot subscribers in January.
In Ji’s view, an AI application’s growth potential ultimately depends on the size of its addressable market and the attention users are willing to devote to it, rather than any isolated technical achievement.
Lu added that demand presents another critical test. Assess that by asking these questions:
- Does the product solve a real, frequent, and high-value problem?
- Are its customers identifiable and willing to pay?
- Can its technical advantages translate into sustained cost reductions?
He pointed to the dot-com bust as a reminder that businesses need viable commercial models to survive periods of market correction.
Embodied intelligence remains at an early stage
Ji considers embodied intelligence one of the most active AI application fields, but also among the least mature. Frequent model releases and evaluations suggest continued research activity, yet the sector has not reached a breakthrough comparable to the emergence of ChatGPT.
Her conservative estimate is that such a breakthrough could take another five to ten years, although she remains confident it will happen.
Embodied intelligence also extends beyond humanoid robots. Applications could include underwater machines modeled on living organisms, autonomous aircraft operating at low altitudes, and flexible robots in space, performing tasks in environments that are difficult or dangerous for people to access.
Meridian Capital has invested across these areas, including autonomous flying robots and a company that Ji said was the first in China to validate a flexible robot in space. Its portfolio also includes X Square Robot, an embodied AI model developer active in Hong Kong.
Lu views embodied intelligence as a system that must integrate models, hardware, data, and feedback from real-world use. These components do not yet work together reliably enough for many complex tasks.
Robots fastening screws may still require engineers to intervene, for example, while logistics robots have yet to achieve consistent performance across different operating conditions. More practical deployments currently involve constrained tasks such as patrols and inspections.
According to Morgan Stanley, China’s humanoid robot market could reach approximately USD 2 billion this year, and USD 15 billion by 2030.
For Hong Kong, Lu sees opportunities over the next three to five years in AI-enabled manufacturing, scientific research, green energy, and enterprise applications across finance, healthcare, and logistics.
He pointed to the IP accumulated by Hong Kong’s research universities. AI could shorten the trial-and-error process involved in scientific discovery by connecting computational experiments with laboratory testing, reducing human error and improving the likelihood of useful results.
Closer links between Hong Kong’s research institutions and the Greater Bay Area’s manufacturing ecosystem could also provide the supply chains and industrial environments needed to develop and deploy embodied AI systems.
Plan cash flow before it runs out
Asked what smaller AI startups should do when their cash reserves begin to run low, Lu recommended assessing cash flow early, prioritizing businesses that generate positive cash flow, and cutting nonessential spending.
Founders should also communicate candidly with investors before funding shortages become critical. Government grants, where applicable, should be incorporated into financial planning early rather than pursued only when companies urgently need capital.
Ji considers cash flow management a basic responsibility for founders:
“If you have only one to six months of cash left and are still asking how to spend it, you probably need to rethink your planning.”
Business milestones should align with financing milestones, she said. A company might progress from government research funding and early institutional support for commercialization to market-based investors, followed by strategic industry investors capable of providing both orders and capital.
Each stage requires advance planning, rather than spending a year developing a product before beginning another search for funding.
Ji also views mergers and acquisitions as proactive growth decisions rather than measures reserved for companies in financial distress. A transaction should create a combined business worth more than its constituent parts.
She cited a company that originally developed automated cutting and polishing systems for lab-grown diamonds. Drawing on its technical expertise, the business subsequently expanded into thermal management materials for stacked semiconductor chips, establishing a new revenue opportunity.
With AI adoption still at an early stage across many industries, Ji believes companies possessing technologies that are difficult to replicate may have opportunities to apply their capabilities in different markets.
Start overseas expansion with the customer
Ahead of the conference, Ji noted that she was scheduled to join two other speakers at Hong Kong FinTech Week x StartmeupHK 2026 for a discussion on global capital trends and emerging investment opportunities.
“It’s actually quite similar to what we’re discussing today,” she said.
Lu laughed: “We’ve already given everyone a preview.”
The conversation then turned to mainland Chinese companies expanding internationally.
Ji pointed to Fairino, a collaborative robot manufacturer backed by Meridian Capital. Fairino ranked first among Chinese collaborative robot brands in both total sales volume and overseas shipments in 2025.
For Ji, its expansion illustrates the importance of combining global ambitions with an understanding of local customers.
Companies often begin overseas expansion by establishing a legal entity, attending trade shows, and signing distributors before identifying customer demand. Ji recommends reversing that sequence.
What do prospective customers need? What are their budgets? Who makes purchasing decisions? What alternatives are available? How will their data be handled, and who will provide after-sales service?
These questions should guide market entry before companies commit substantial resources.
Ji also recommends addressing data compliance, IP, export controls, and security reviews during product development rather than after entering a new market. Companies should establish a viable business in one initial market before expanding into adjacent ones, instead of launching across multiple markets simultaneously.
Lu emphasized that local knowledge matters as much as speed.
International expansion is not simply a repeat of moving from one Chinese province to another. Markets differ in culture, regulation, and historical context. Respecting local rules and building local sales and delivery teams can be more important than technological advantages alone.
Hong Kong and the wider Greater Bay Area could serve complementary functions in this process. Companies can locate R&D, supply chains, and application development within the region’s interconnected manufacturing ecosystem, while using Hong Kong for international financing, regional headquarters, compliance structures, and talent recruitment.
According to the interviewees, Hong Kong’s common law system and experience with international arbitration can also provide overseas customers with greater confidence when entering commercial agreements with Hong Kong-based entities.
Over the past year or so, Meridian Capital has been building its presence in Hong Kong, applying for licenses and preparing an HKD-denominated fund. Its investment plans encompass founders worldwide, frontier technologies, and the commercialization of research from local universities. Ji has also observed more venture capital firms strengthening their presence in the city.
For AI startups, Hong Kong’s role is extending beyond its traditional position as a destination for stock market listings. Its combination of financial infrastructure, research capabilities, and connections to mainland China’s manufacturing ecosystem could make it a base for international expansion.
Lu summed up his advice to companies considering overseas growth with a coastal metaphor: before heading overseas, take a look by the sea, in Hong Kong.
KrASIA features translated and adapted content that was originally published by 36Kr. This article was written by Wang Hanyu for 36Kr.