September 23, 2026, (Inside AI) — In a closed-door roundtable held in Beijing on September 16, a group of early-stage investors and founders gathered to dissect the first-year survival playbook for AI startups. The event, titled "The First Year of AI Startups," was co-hosted by TechNode, Asia Capital Exchange (ACE), Lighthouse Capital, and BEYOND. The discussion focused on a counterintuitive reality: the most effective fundraising strategy often involves not chasing investors at all.
The roundtable featured Qiao Jianlong, editor of AI Insider's Demo Club, who moderated. Panelists included Zheng Jinliang, co-founder of Bensu Intelligence; Song Haocong, founder and CEO of Aurano; Zhou Junting, CEO of Lingqi Xiyuan; and Li Zhengwei, managing director at Lighthouse Capital and co-head of its 3i incubation business. Their conversation revealed a stark divide in the funding market and offered concrete tactics for founders navigating the earliest, most fragile phase of company building.
Chasing Investors Can Backfire
Li Zhengwei, who rejoined Lighthouse Capital earlier this year after a stint founding and incubating companies, delivered a blunt assessment of fundraising dynamics. He noted that investors who proactively reach out to a startup are at least twice as likely to invest compared to those introduced by a third party. Furthermore, an introduction from a trusted contact is at least twice as effective as a cold approach by the founder. This creates a paradox: the harder a founder pushes for meetings, the less likely those meetings are to yield capital.
Qiao Jianlong illustrated the point with a personal anecdote. He once introduced a well-qualified former classmate to an investor friend. The investor's private feedback was dismissive, suggesting that startups who come knocking are often just there to make up the numbers. Li's advice is to secure a credible champion who understands the business and can make thoughtful introductions to three to five high-value investors. Founders lacking such a champion should consider working with a capable financial adviser.
Read: AI startup Mantic raises $25 million for superhuman forecasting
This trust deficit also explains why first rounds typically come from friends and family. However, that goodwill has a shelf life. Qiao warned that the first six months rely heavily on pre-existing credibility. Without a convincing product or meaningful progress after that window, raising a subsequent round becomes significantly harder. Even a well-connected supporter can only risk their reputation so many times. Li emphasized that founders must first know what they are building and who can build it with them. Only then does fundraising become a smoother process.
Pick a Direction Too Small to Notice
Before fundraising, founders must choose a direction. Song Haocong, who previously worked at Spark Education and foundation model company 01.AI, observed that since the arrival of GPT-3.5 in 2022, human-computer interaction has been confined to a text box. He founded Aurano to find a more natural way to interact with AI at the application layer. To avoid competing in a crowded space, Song uses a "three-step" approach. Looking one step beyond the obvious reduces competitors. Looking three steps ahead requires testing more assumptions but can reveal a niche that seems small today and becomes vital later.
"From day one, I look for a direction that is too small for others to care about today, but could grow as the conditions around it change," Song said.
Zheng Jinliang, a doctoral student at Tsinghua University, founded Bensu Intelligence in July after three to four years of embodied AI research at the university's Institute for AI Industry Research. His company focuses on the "brain" of embodied systems rather than robot hardware. He believes the field's prevailing approach may be nearing a ceiling. Instead of starting with better robots and more data, his team asks what problem a model should solve and works backward to the data and hardware required.
Zhou Junting, an undergraduate at Peking University's Yuanpei College, chose a third path: AI for Science. Lingqi Xiyuan aims to build infrastructure for an AI lab that improves through its own work. His proposed system has two layers: a "Science Agent" to move from hypothesis to experimental protocol, and a "Physical Agent" to execute that protocol in the real world. Zhou argues that for AI to contribute to scientific discovery, it must enter the full research loop, including physical experiments.
These three companies pursue distinct opportunities. What they share is a refusal to compete solely by improving what everyone else is already building.
Hire for Conviction, Not Credentials
Once a direction is set, founders need people who can pursue it. Song looks for conviction. With new models and headlines arriving weekly, he said, people who change course with every development can lose sight of the team's goal. He also values a research mindset. Aurano keeps textbooks close at hand, returning to books and papers when the team runs out of ideas.
Read: As Trump and Xi Meet, Investors Play Both Sides of AI Divide
Zhou assembled his team before seeking outside funding. For a student founder, a campus can be a good place to find people who already believe in the same goal. He is drawn to people with a distinctive way of thinking. "I don't want a team made up only of people who are excellent in a generic way," he said. High grades matter less to him than an ability to form an independent judgment.
Zheng's first seven team members had already worked together in a laboratory on embodied AI research. They knew one another's strengths and how to divide the work. When meeting potential recruits, they rely on their technical ideas and research results to attract them. People who have spent time at the frontier of embodied AI, he said, can tell whether a team understands the hard questions.
Li framed hiring from an investor's perspective. In the internet and mobile internet eras, founders were judged chiefly on their understanding of user demand and their ability to execute. Those qualities still matter, but he now pays particular attention to technical judgment and business sense. A team must be able to turn a research result into a product and a company. Otherwise, it risks building an impressive laboratory that never becomes a business.
Competition with large technology companies is difficult to avoid in a startup's first year. For application-layer founders, the immediate threat comes from foundation model companies. Those companies already interact with users, and each improvement to their models can absorb features that once supported a standalone product. Song said investors ask him about this frequently. His answer returns to choosing a direction that looks too small to justify a large company's attention today, while showing signs of future demand. Big companies often need a quicker or larger return to approve a project internally.
"I need to find an opportunity where I can build what looks like a lightweight product now," he said, "but one that can later connect to capabilities many other companies are racing to develop. Others may already be building those capabilities. The entry point is what remains open."
Zhou faces a similar question in AI for Science, where companies such as Google DeepMind are also active. He does not assume they must be rivals. Better foundation models, more accessible interfaces, and a market that understands the technology could all help a smaller company. His team's immediate task is to make the loop between computational work and real experiments function inside a laboratory.
Zheng sees embodied AI as too large a field for any one company to address in full. Since 2023, embodied AI has advanced in data collection, hardware, and model training, he said. But the basic approach to learning has not changed as much. A small team cannot match a large company's resources; it may, however, be better placed to test a new approach in a compact, complete system.
What would count as getting through the first year? Li said the companies that falter almost immediately often lack a committed core team. Some founders chase whichever area is attracting attention. Others start a company because colleagues have done so, or because they see a chance to turn an existing resource into quick money. A team that genuinely believes in a difficult, unfashionable direction can be more resilient.
Lighthouse Capital's 3i incubator has set a goal of seeing half of its projects reach Series A within two years. Li acknowledged that this may be higher than the success rate many early-stage investors actually achieve. To him, a Series A round can indicate that a company has tested whether its product meets a real market need. Few startups go all the way from identifying a direction to building a product and raising that round without setbacks.
Qiao closed on a more optimistic note. The mobile internet era often followed a winner-takes-all pattern, but AI and hard technology have more links in the chain, more applications, and more specialized niches. No company, including a foundation model provider, can do everything. In the first year of an AI startup, the scarcest resource may not be money or model capability, but the ability to make others believe in what the team can do. The aim is to have investors seek you out, talented people want to join, users actively choose your product, and opportunities emerge while big companies have yet to take notice of your niche.
