In 2023, two waves converged on the artificial intelligence boom.
The first wave was flashy, filled with glamor. These people can be easily spotted at high-profile events like the World Artificial Intelligence Conference (WAIC): graduates from prestigious universities, contributors to top-tier journals, with overseas experience. Their conversations frequently revolved around grand topics like artificial general intelligence (AGI) surpassing OpenAI, or the next “killer app.” When asked if AI could be profitable, they would say, “It’s too early to tell.”
But to find the second wave, you needed to venture away from Beijing, Shanghai, and Shenzhen—into less prominent places.
Take Yiwu, for example, a county-level city famous for its small commodity market, hardly associated with high-tech innovation. The winds of AI didn’t sweep into some investment committee meeting room or plush office building here. Instead, walk into a fancy restaurant, and you’d likely see young locals with tablets and wine glasses, demonstrating AI-generated product images for shopkeepers in a private room.
According to Tianyancha, the number of AI startups in Beijing and Shanghai declined on a percentage basis between 2023 and 2024, while it increased in non-first-tier cities such as Shanxi, Hebei, Henan, and Hubei.
In Yiwu, terms like AGI or killer app don’t impress clients. The only metric that matters is whether it cuts costs and makes money. One multichannel network (MCN) operator in Yiwu told 36Kr how she once cut off an engineer pitching an AI agent program: “Young man, just tell me how much I’ll save each month on virtual hosts.”
When Feng Yawei brought his GPT-based office assistant to Yiwu from Wuhan, he knew his background wouldn’t stand out in cities like Beijing, Shanghai, or Shenzhen. He was a nuclear energy major, not trained in computer science. His previous internet venture had folded due to commercialization challenges. But Yiwu businesspeople didn’t care about that—they cared about whether Feng’s AI product could efficiently manage the complexity of live streaming operations.
200 business introduction booklets Feng brought from Wuhan were snapped up in under an hour. He hurried to a print shop to get hundreds more. In down-to-earth markets, if AI proves useful, the returns are palpable.
Two men entered Xia Nan’s (pseudonym) office, lugging large bags filled with stacks of pink cash, a muffled thud as they hit the desk. Xia Nan opened one bag, quickly counting the stacks—around 40 or 50 of them. The cash-wielding visitors were local bosses from an MCN in Fujian. They were showing their sincerity, wanting to buy Xia’s AI face-swapping technology.
In mid-2023, short dramas targeting overseas audiences were booming, and these bosses desperately needed affordable, efficient AI tech to replace Chinese actors’ faces with ones more appealing to foreign viewers. This grassroots counter scenario became Xia Nan’s reality. “I’ve never seen so much cash in my life,” he said. Just two years ago, he was struggling, returning to his hometown in Fujian from Hong Kong after six years of repeated failures.
The beauty of the lower-tier markets lies precisely in their acceptance of AI entrepreneurs—regardless of credentials or business models.
Before donning the label of “AI entrepreneur,” people like Xia Nan came from all walks of life: recent graduates struggling to find work, laid-off big tech employees, or even TikTok streamers and Taobao merchants.
LangChain. AI agents. Most of these entrepreneurs didn’t grasp the technical jargon, but they excelled at getting things done and weren’t afraid to get their hands dirty.
Qu Kai, founder of financial advisory firm 42 Capital, told 36Kr that this group’s defining trait was that, while big companies and star teams were still debating possibilities, they were already making money.
Seizing new opportunities
In corners unseen by elites, everything demands that AI entrepreneurs break the inertia of traditional startup approaches. “You can’t be too academic or theoretical doing AI business here,” a financial advisor told 36Kr.
Instead of crafting business plans, many grassroots entrepreneurs excel at doing business the straightforward way. They sell not just tech but also emotions and human connections. For instance, a couple from Wuhan, who launched an AI self-study room in Meishan, Sichuan, told 36Kr that, within six months they had mostly mastered the local dialect and even learned to play mahjong.
“Not every startup needs to be the next ByteDance,” Kai-Fu Lee once said. “You might turn a profit in the first three months. Of course, it may not be scalable, nor become a public company, but it’s still entrepreneurship.”
In Yiwu, rather than fearing competition from big tech or celebrity entrepreneurs, Feng feared the local business people doing AI. Locals were more sensitive to subtle changes. A Yiwu merchant mentioned to 36Kr that, as early as mid-2022, when generative AI went viral, heirs of local shops and factories had already begun working on model APIs. “Locals understand local needs better.”
Ordinary AI entrepreneurs often spot opportunities not through detailed market research but by being attuned to and quickly addressing the immediate needs of those around them.
After eight years in Fujian, Xia Nan had worked on mobile games and virtual live streamers, each venture ending prematurely due to the pandemic or changing regulations. Then in early 2023, opportunity knocked. During a dinner, an old MCN client mentioned wanting to use AI for face-swapping in overseas short dramas. Xia Nan got to work immediately, taking just weeks to pull together some open-source models and repurpose a digital human platform left over from previous business into an AI face-swapping tool.
Next, he posted a video using the face-swapping tech to his social media, offering it for free to a few old clients as a trial. Days later, news of Xia offering AI face-swapping services spread quickly in Fujian’s MCN circles. Xia received over a dozen friend requests each day. “Demand for short dramas grew several fold. Prices weren’t high per order, but demand surged.”
Moreover, what enabled these entrepreneurs to earn their first bucket of gold during the AI boom was often their own needs.
Zhang Luyu, who dropped out in junior high school, registered his AI company Dify in Suzhou, a city somewhat peripheral in the AI boom, intentionally avoiding elite clusters like Beijing, Shanghai, and Shenzhen. “The consensus among Beijing investors and star founders is too strong—every day, they discuss the same things. Isn’t entrepreneurship about doing something different?”
For him, there are two best ways to find a direction: first, evaluate if your past skills match an unmet need; second, solve a problem you face yourself. “I can’t empathize with others’ needs. It’s like opening a restaurant—someone tells you they want hotpot, but you don’t like it. You can’t really run a good hotpot place,” he told 36Kr.
But could one’s own needs represent a vast market? Zhang had no definite answer. He simply felt that if he followed conventional startup logic, calculating the market size, he’d have to anticipate all kinds of competitors: “Some big name enters, or some big company swoops in—in that case, why bother starting?”
On November 31, 2022, the night ChatGPT launched, Zhang stayed up fiddling with it. But soon, he realized that developing new applications using ChatGPT’s API would be challenging—not just engineering-wise, but it would require months of prompt testing. Yet, there were no tools available to assist developers. Like Newton being struck by an apple, an idea struck him: “Isn’t this a business opportunity?”
Going where there’s more money
In April 2023, a month before Dify’s launch, Zhang met with two or three investors, testing the waters. Judging by their lukewarm interest, he knew funding wasn’t in the cards.
“This round of funding—unless you have a superstar background—is entirely different from the mobile internet boom,” he told 36Kr. Product data and commercialization milestones, once requirements for later funding rounds, had become prerequisites for the first round.
In an interview, Lee said that current funding environments are the worst in a decade. Qu mentioned that, in the capital winter, investors prefer stability and proven founders—the so-called “elites.”
The reality of financing pressures grassroots founders to think about commercialization from day one. This is why the AI ventures now boasting impressive revenue figures often belong to these ordinary individuals.
Feng’s team cobbled together ConnectAI Technology from open-source communities, raking in tens of thousands monthly; Xia Nan’s revenue enabled him to upgrade to a more upscale office. Stories of monthly income over RMB 100,000 (USD 14,000) or annual net profits in the millions float abound on apps like Xiaohongshu and Jike, with many using AI to write or design.
Meanwhile, “most AI companies in the market today have an annual recurring revenue (ARR) lower than the daily wage of a domestic worker at RMB 400 (USD 56),” Qu said.
For elite founders, deciding whether to target the domestic or international market—where to raise funds, which users to serve, where to register the company—is a cautious decision. For grassroots entrepreneurs, there’s only one correct answer: go where the money is.
Back in 2018, when he was building a developer collaboration product called Feie.work, Zhang realized software-as-a-service solutions couldn’t survive in China. Despite good reviews, the income was a cold splash of reality—”far from my ideal.” With that bitter lesson learned, Dify chose to go global. At the outset, they developed three language versions: Chinese (simplified and traditional), English, and Japanese. With no one in the team fluent in Japanese or English, they resorted to Google translations, word by word.
To Zhang, the US and Japan are must-have markets for SaaS products: “The US is a given, and Japan’s SaaS market is uniquely reachable for Chinese, with strong customer loyalty and a high concentration of global companies.”
Initially, Dify had no overseas personnel, nor any promotional funds. Zhang set low expectations for overseas sales: “I didn’t expect much from the numbers.” But to his surprise, an article about Dify’s team became a hit on Note, Japan’s equivalent to GitHub.
Spurred by this exposure, overseas deals poured in within two months of Dify’s launch. During that period, cheers often erupted in the office each time an email requesting cooperation came through—and it would typically turn out to be from yet another Fortune 500 company.
However, even with growing buzz, they had to carefully deliberate on pricing: “Charging too soon would stifle growth, too late and the business model wouldn’t work—users would assume it was always free.”
In terms of monetization, the one common understanding between elite and grassroots founders is that charging starts with user traction. After reaching over 100,000 global installations in seven months, Zhang felt it was time to monetize. “The market has recognized your product, you understand your customers’ price points—now it’s time to set the price.”
ConnectAI’s pricing also came from customer demand. Its smart office assistant, launched in May 2023, received nearly 100 trial requests within its first month. Within a week of those trials, customer repurchase rates approached 100%.
Revenue came several months sooner than the co-founders anticipated, so much so that ConnectAI’s website didn’t even have a pricing plan. The company remodeled its fees on a similar tool, publishing the rates online. Then, after gifting several months of services to early customers for feedback, the pricing was gradually fine-tuned.
By the Lunar New Year of 2024, Zhang was pleasantly surprised: Dify had become profitable just eight months after launch, with an 80% gross margin and nearly half of its revenue from overseas.
Zhang remembered how, during Dify’s initial launch month, he flew from Suzhou to Beijing, meeting over 30 investors in a week to fund product development. His business plan back then was modest: a screenshot of a note on an iPhone summarizing who he is, what he’s making, and the progress to date.
He didn’t set high expectations, but figured that, out of 20 pitches, a 5% success rate would be enough. Today, Zhang is the one making the choices, having turned down several term sheets: “Dify is already profitable, so I’m not fixated on raising funds.”
Trends don’t last forever, only entrepreneurs do
In late 2023, the market’s focus shifted from model training to implementation and commercialization. More and more major companies moved into the same markets the grassroots players were in. AI companies at the grassroots level could no longer keep “quietly making money.”
At the end of 2023, Zhang was preparing for Dify’s second funding round. During his pitch, he openly analyzed where Dify might falter: competition from giants, changes in technology, or team limitations. He reckoned, with Dify’s capabilities, competing against a company that can throw in huge resources is improbable.
Before long, that well-resourced competitor appeared. In November 2023, ByteDance launched an AI app development platform called Coze. Zhang opened Coze’s developer panel and was stunned—it was strikingly similar to Dify. Zhang recalled that, because Dify was moving fast when first launching, some words were misspelled in its interface. He said he didn’t expect to find the same typos in Coze—insinuating that Coze may have copied Dify.
Meanwhile, stronger competitors also raised customer expectations of AI. ConnectAI’s stay in Yiwu lasted less than three months. During those months, Meitu, Baidu, Taobao, and Temu—all eyeing Yiwu’s coffers—arrived as well.
Turning a set of complicated live stream operations into an AI assistant took ConnectAI three months. Just as it had a product to deliver, Feng realized the big players had already captured Yiwu. Former clients poured cold water on him: “Why doesn’t your AI help us interview hosts?” “Why can’t it do product image recognition like Baidu’s?”
As competition intensified, prices plunged and profits shrank. For instance, the cost of AI face-swapping in short dramas fell from RMB 100,000 for 100 minutes to mere four-figure RMB sums, courtesy of Baidu and Meitu. Only three months after starting AI face-swapping, Xia realized he had been outplayed by the big players with low prices.
Old customers gradually lost contact. Occasionally, an MCN would inquire, and Xia Nan would offer rock-bottom prices, offering to translate 100 dramas for just RMB 10,000 (USD 1,400). His coping mechanism became late-night snacking. By the end of 2023, AI face-swapping had nearly no demand, and Xia Nan gained several pounds from stress eating. Friends who saw him assumed he was doing well.
“Thought about finding a job?”
“No.” Xia Nan answered without hesitation. Gaming, the metaverse, AI—none of these trends made him enough to retire. “But I’m not giving up. Why should only the bigwigs make money?”
While Coze ventured into freemium options, Zhang never considered lowering Dify’s pricing to compete with ByteDance. “Many customers come for the subsidies, and once those end, they leave,” Zhang told 36Kr. He chose to stick with the original pricing, avoiding subsidies in favor of attracting long-term, committed customers.
The strategy of sacrificing some customer volume for healthier revenue proved effective. At an AI event, Zhang met with members of Coze’s team. A product manager admitted that Coze’s freemium approach had attracted fewer serious users compared to Dify.
The entrepreneurs who once struggled also worked to carve out new paths. After three failed ventures, the only remaining valuable asset for Xia Nan was his server farm’s computing power. As long as technology advances, he realized computing power is a steady need. His new direction: move upstream, selling resources.
He reviewed his assets—over 100 GPUs and cloud service reseller contacts. Believing he can make money off these, renewed excitement fueled him, and Xia Nan transformed into a computing power leasing agent, relaunching his business. Lately, Xia’s social media was active again. Posts had shifted from AI face-swapping videos to RTX 4090 machines that he was looking to lease out.
Feng saw Yiwu as an avenue for him to understand product-market fit. His team quickly regrouped, reaching out to old clients for research, seeking new opportunities. Many clients said that standardized smart office assistants still didn’t meet their unique needs. So, within weeks, the co-founders rolled up their sleeves and launched a platform to develop enterprise AI applications.
For Zhang and his co-founders at Dify, the most urgent task was enrolling in an English course. In May, Dify hosted its first developer meetup in Japan. In June, Zhang visited the epicenter of AI—the US—for the first time. He felt that his team’s growth had to match Dify’s global expansion. “Our team isn’t overseas, and our English still isn’t good enough.”
“How long do you think Dify can last?” 36Kr asked Zhang.
“Until the next tech comes along to disrupt large models.”
“And what will you be doing then?”
“Still starting businesses.”
This article was written by Zhou Xinyu for 36Kr and was originally published by 36Kr.
