Original · Unique Research · 2026-04-02 · Shanghai
Editor's note: This historical roundtable preserves the source overview and full edited transcript, including repeated statements. Quantitative, product, market and competitor comparisons are source/speaker claims, not independently verified current facts. The source's 1998 founding date, 27-year history and moderator's 25-year reference are retained without recalculation; 400 nodes and 4400-plus facilities are distinct source descriptions, not silently reconciled. The overview locates a competitive comparison in North America while the moderator's corresponding question names Japan; both remain. Wendy's 100-person hiring/layoff example is secondhand, and the labor-law characterization is not legal advice. Race- and gender-based hiring generalizations in Zheng's remarks are preserved as his views, not evidence-based findings or editorial endorsement. Local processing does not itself establish legal compliance, and the reported ten-thousand-fold inference-cost forecast attributed to Jensen Huang has not been checked against a primary transcript. The 10,000/100,000 monthly cost example specifies no currency. “BC from the same source” retains the speaker's label without inventing a formal definition. Sixteen source images still require content review.
Unique Awards
Technology Strategy and Building Defenses in Global AI Competition
Four Companies at Four Stages of Globalization Arrived at an Answer More “Unglamorous” Than Most People Imagine
“
The four answers were really about the same thing: speed and relationships, time and depth.
This was a trends roundtable at Unique Awards · Hangzhou AI WEEK, titled “Technology Strategy and Building Defenses in Global AI Competition.”
The four guests onstage represented four completely different stages of globalization:
Zheng Han, CTO of InfronAI. Their SaaS company was founded in '24 and provides a global AI-model-routing API, primarily for the North American To B market. Its target enterprise customers spend US$300,000–500,000 a month. Sales and BD teams are based in the Bay Area, while R&D headquarters are in Hangzhou.
Wendy, co-founder of Star Shine. The team builds AI emotional-companionship products focused on Japan. Having completed development and testing, it was preparing for an official Japanese launch in April, with South Korea next. It was the smallest and earliest-stage team among the four.
Gao Zhou, CEO of Zhongju Intelligence, based in Hong Kong and Fuzhou. Having shifted from the Infra middleware layer to applications, the company now emphasizes two products: a customer-service and Marketing Agent for the WeCom ecosystem, and an overseas SaaS platform for Shopify sellers.
Li Wentao, Akamai's Asia-Pacific director of cloud architects. The source describes Akamai as a global company founded at MIT in 1998, one of the principal inventors of CDN, with more than 4400 data-center facilities across over 130 countries and carrying 20% of global internet traffic. Of the four, it had the deepest globalization and longest operating history.
The moderator was Qi Wei, founder of the growth Studio Rockbase, which helps Chinese technology and AI companies grow overseas.
The four companies ranged from just before launch to 27 years of operation—an enormous span. Yet by the end, they converged on the same answer, one far more “unglamorous” than most people imagine.
Why Not North America? A Counterintuitive Choice
Startups going overseas have a default script: on Day 1, Go to North America. The reasons are strong purchasing power, high-quality users and the earliest breakout of AI products.
Star Shine did not follow that script.
Wendy recalled that whether to enter Europe and America was “the most intensely debated topic” internally. The team ultimately passed over North America and chose Japan for three reasons.
First, find a less crowded space within developed markets. Companies such as Character.AI had already broken out in North American AI companionship. Mature doll-like AI companions also existed in hardware. How much advantage would a small team have there? Wendy's judgment was that competition in Europe and America was too dense and their own core advantages somewhat weaker.
Second, in the source's assessment, Japan had the lowest AI penetration among developed markets. How low? Moderator Qi Wei offered an on-the-spot anecdote: “I visited Japan this month for exhibitions and AI events. The products felt much like those I saw 10 years ago. 90% were To B products, centered on document organization and company Workflow construction.” The author argues that Japan's local AI supply is no longer at the same Level as development in China and the United States—and that this is the real window of opportunity.
Third, Japan offered a sufficiently long entry window. Low AI penetration meant, in Wendy's view, that big companies would not rush in immediately. “That gave us a relatively long period to enter the Japanese market and favorable timing.”
The logic sounds clear. But there is a crucial issue: entering a market and surviving there are entirely different things.
Japan's Real Barrier to Entry: Flyers Work Better Than Paid Acquisition
The roundtable's shared view was that Japan is easy to defend but hard to enter. Once users accept a product, they do not readily switch, but they are also extremely cautious about new products.
How, then, do you break in? Star Shine's current approach sounds a little out of place at first.
“We go offline, to anime-culture shops in places such as Akihabara, and work with stores to hand out our promotional flyers. We also find university students to distribute advertisements for us.”
Wendy said she had not expected much from such an “old-fashioned” approach. Yet users converted through traditional advertising had the highest retention on their platform, she reported.
The conclusion is counterintuitive but has an internal logic. In the author's account, Japanese users are not accustomed to being bombarded by online traffic, and the large-scale paid-acquisition approach used in North America is difficult to make work in Japan. Users who encounter a product through real offline settings and community relationships, by contrast, arrive with clearer motivations and are more willing to stay.
Qi Wei added a line worth noting separately: “Japan's offline market can rival online traffic. Anyone interested in Japan should try more offline channels.”
Hiring is another trap in Japan. Wendy had not experienced this herself, but had heard about many peers. One team “hired 100 people in Japan in an aggressive six-month push, only to find their efficiency and technical iteration far below the Chinese team's.” When the founder later wanted layoffs, the account says, Japan's extremely strict labor law made regular employees almost impossible to dismiss. The decision nearly brought the company to collapse, according to this secondhand account.
Wendy's advice was to prepare for at least three years. The first-year goal should be to achieve a modest positive cash flow as quickly as possible, rather than take a heavily cash-burning route.
“BC from the Same Source”: Renting a Silicon Valley Villa to Win To B Contracts
InfronAI took a different route: directly into North America, serving To B customers.
Zheng Han said they met the West Coast team every morning at 7. Qi Wei added a jab: “If you don't have meetings at 12 midnight or 7 in the morning, you hardly dare call yourself a global team.”
Taking To B infrastructure overseas is fundamentally a different battle from taking To C applications overseas.
A consumer switching shopping App faces a very low decision cost, perhaps just a matter of personal habit. A company switching API providers, however, typically takes one to two months for R&D evaluation, finance approval and the full decision chain, according to this account. Each To B switch therefore carries high friction and cost; once a deal closes, customer stickiness is relatively stronger.
Their Three Main Approaches to Winning Bay Area Customers
First, Hacker House. They rent a large Airbnb villa and hold events every evening—drinks and Workshop sessions. When meeting people for BD over coffee in the Bay Area, many say, “I've heard of you; I've been to your Workshop.” That greatly reduces the effort needed to build a relationship.
Second, mutual-friend networks. North American users care deeply about Connection, Zheng said: “Many first ask whether we have mutual friends.” Their advantage, as described in the source, is that their partners include a YC Portfolio connection: shared YC alumni ties and the same VC contacts quickly create a foundation of trust.
Third, 24-hour response coverage. Zheng's comparison was that OpenRouter could not provide 24-hour service. InfronAI has teams in Hangzhou and the Bay Area with opposite working hours. When a North American customer encounters a problem, Hangzhou is in its workday and can take over immediately.
Zheng gave an example: “Recently, a Japanese audio model required a Japanese domestic credit card to purchase. The customer couldn't get one. We sorted it out that afternoon and connected the model for them. That was a Surprise for the user.”
As these “Surprise” moments recur, customers begin to Share their corporate strategies with the team. “We build infrastructure and must be deeply connected to our users' businesses. That recognition between people and teams, together with 24-hour service, is our moat.”
Overseas Customers Can Actually Be Easier to Serve Than Domestic Ones
Zhongju Intelligence builds customer-service and sales Agents for Shopify, serving two groups: Chinese cross-border e-commerce teams and local North American sellers on Shopify.
Most people would assume foreign customers are more troublesome to serve. Gao Zhou's answer was the opposite.
“Serving domestic customers actually brings more challenges. Their requirements seem to keep changing.” In his account, local North American customers accept a standardized product: they configure their own Agent, knowledge base and SKU through an easy end-to-end process, perhaps requiring only one local Support person.
“Domestic customers have many customization requests. They may not be genuine needs, but they want the feature.”
Qi Wei could not resist: “It seems we Chinese are rather hard to deal with.”
Gao also said those domestic users had unexpected value: “They reveal pain points we hadn't previously considered, which we can turn into standardized features.” By the time those features reach overseas customers, the product has been refined.
On global compliance, Gao was direct: “The underlying issue isn't dealing with tax or legal compliance. Those are relatively standardized; find professionals and it's OK. My main Focus is the product: does it really solve the user's problem?”
That sounds simple, but it is a starting point often overlooked in discussions of overseas expansion.
Can 27 Years of Infrastructure Become a Moat in the AI Era?
Akamai's story is unusual. Its question is not “how to go overseas,” but “how to respond to change with a stable foundation”—using a globally distributed platform accumulated over 27 years to help others expand abroad.
Li Wentao said Akamai had built local network capabilities, carrier interconnections and a foundation for compliant operations in over 130 countries, with more than 4400 facilities worldwide, deep partnerships with 1200 carriers and 20% of global internet traffic. In his account, leading internationally expanding companies in virtually every sector were customers.
Its new move in the AI era was deploying NVIDIA's latest Blackwell GPU computing power at the internet's edge.
“The edge platform is one network hop and ten milliseconds away from 90% of users worldwide.”
Why does that matter? For highly real-time businesses—emotional interaction, connected cars or real-time gaming—compute needs to be closer to users. In Li's account, processing local data nearby both supports compliance and improves user experience and personalization.
On cost, Li identified a trap founders often overlook: he characterized cloud providers' pricing as exponential. Going from 0 to 1 feels easy, perhaps with Credit subsidies, but costs can rise sharply as scale increases. “CTO and technical teams should think about decoupling from cloud providers on Day 1. Don't overdepend on one provider. Make applications multi-cloud-friendly and embrace open source.”
He also relayed a forecast attributed to Jensen Huang at GTC: “Future AI inference costs will be ten thousand times those of training.” How to build scalable inference architecture that substantially lowers Token costs in long-term operations is something to consider from Day 1.
The Nature of a Moat: Four Answers Pointing to the Same Thing
Toward the end, Qi Wei asked the roundtable's central question: what is the moat you believe competitors would find hardest to replicate?
Gao Zhou was most direct and most disruptive: “In the AI era, I don't think anything has a moat, especially a technical moat.”
He has worked in quantitative trading for almost ten years, centered on quantitative strategies. The strategy itself is not the moat, he argued; it is “the dynamic updating process as the strategy interacts with the market.” The same applies to AI products. The product is not the endpoint: the continuing combination of product, market, data, iteration and the latest technology is the moat. Moreover, “a moat is always dynamic. What is a moat now may not be one in three or six months.”
Zheng illustrated this with a phenomenon he had seen: “Whenever a new large model appears, traffic in your back end falls off a cliff. The supposed Prompt Engineering and engineering moats people used to discuss disappear the instant the new model arrives.”
His response is to spend 60% of each day on customers—not coding or thinking up features, but on customers. If responses are fast and deep enough, customers share their strategic secrets, bringing the business into something more durable than a technical moat: a relationship moat.
Wendy's answer was time. Surviving a year in Japan and completing the business cycle is itself a barrier. “Many startup teams die before a year is up. That doesn't fit the Japanese market.” Product strength is the entry requirement, but sustained market cultivation and deep user insight generate real organic growth and high user stickiness.
Li described a multiplication of scale and time: 27 years of global infrastructure, every carrier interconnection, every local compliance-certification system and every problem encountered by a customer are things competitors cannot reproduce quickly.
The four answers were really about the same thing: speed and relationships, time and depth. Not one person said technology.
A Final Word
In a sense, this roundtable's conclusion runs against what most people imagine “taking technology overseas” means.
They picture AI expansion as better models, cooler products and faster technical iteration—followed by an attractive growth curve.
This conversation says: before entering Japan, ask whether you can last three years; for North America, get the Hacker House going first; in To B, spend 60% of working time on customer calls; in infrastructure, think today about decoupling from cloud providers.
Gao said something worth recording separately: “We no longer think about changing the world. We first tackle a small pain point in one industry.”
Moving from “disrupting the world” to “tackling one small pain point” may be the real first step in taking AI overseas.
Can you last three years in the market you are working on now?
Live Conversation Transcript
The following is an edited transcript of the roundtable, preserving the guests' original expressions and the atmosphere of the discussion.
Qi Wei: Thank you to all four guests for joining this roundtable. As I briefly introduced earlier, Rockbase is a growth studio helping Chinese technology and AI companies expand overseas. The four guests here are at different stages of globalization. Let me briefly introduce the companies before we start: Star Shine works on AI emotional companionship in Japan and the Japanese and Korean markets; InfronAI provides global AI-model routing; Zhongju Intelligence is moving from enterprise services in China to overseas markets; and Akamai is a global company which, in this introduction, has 400 nodes worldwide and annual revenue of several billion US dollars. These four companies differ greatly in stage and scale. Our shared topic is how to build moats in global AI competition. First, please introduce yourselves and your companies in one sentence, and tell us how far you have gone in globalization. Let's start in this order.
Zheng Han: Hello, I'm Zheng Han, Andrew. Our SaaS company was founded in '24. We provide an API: whether developers need large language models, multimodal models or a search API, our platform offers one-stop access. We can deliver global routing stability of 99.99% and discounts of up to 30% off the original price. Compared with competitors such as the North American company OpenRouter—we have spoken with its CTO in the Bay Area—OpenRouter mainly targets developers and long-tail consumer users. We focus more on To B customers spending US$300,000 to US$500,000 monthly. Our main customer base is North America; sales and BD are currently based there, while R&D headquarters are in Hangzhou. That is roughly where we stand.
Wendy: Hello, I'm Wendy. First, thank you to Unique Research and Mr. Wu for inviting me to Hangzhou to exchange ideas with industry peers. Compared with the other guests, we are a very small startup, still quite small in scale, but happy to share here. Our product focuses on AI emotional companionship, primarily for Japan. We have completed some product development and testing and expect an official Japanese launch in April. We may next extend to South Korea, so our first step mainly concerns Japan and Korea. That is our current stage.
Qi Wei: Thank you. Next, Zhongju Intelligence.
Gao Zhou: Good afternoon. Let me introduce our company. We are mainly based in Hong Kong and Fuzhou. In '23 to '24 we mostly worked on Infra, developing middleware similar to Dify and Coze. In '25 we shifted our focus to customer-service and Marketing Agents. We now have two products in the market. One provides customer-service and Marketing Agents for the WeCom ecosystem. The other serves overseas markets, especially Shopify: a SaaS platform offering customer-service and sales Agents to Shopify sellers. So, apart from large models themselves, we have worked across middleware and applications. That is roughly our situation. Thank you.
Qi Wei: Next, let's hear from Akamai's director of architects.
Li Wentao: Hello, I'm Li Wentao, and I lead Akamai's Asia-Pacific cloud-computing architect team. We are headquartered in Boston and were founded at MIT in 1998. Our work is to use a global platform to address internet performance, security and reliability, helping businesses reach the world and overseas markets faster. In the AI era, we offer global AI compute. Through close cooperation with NVIDIA, we provide the latest Blackwell GPU capacity to help Chinese companies expand abroad faster and take AI applications and Agents overseas with greater speed, stronger compliance and higher ROI. Thank you.
Qi Wei: Let me add something—Wentao is being modest. I only just learned that Akamai's Founder was one of the principal inventors of CDN. It is an impressive company. I work in growth, so globalization and localization are unavoidable terms for us. We often say AI companies Go to North America on Day 1 because purchasing power is strong and users are high quality. But companies such as Star Shine may find more potential in Japan and Korea, where a deeper combination of product and culture can create more opportunity. So my first question is for Wendy: why choose Japan as the first stop for AI emotional companionship? We know that in Japan, both software and hardware companionship are highly competitive. In software there are companies such as Character.AI, and in hardware many excellent doll-like AI companions. Why Japan, then, and what is your biggest problem there now?
Wendy: This was an important question that kept troubling our decision-making. Why Japan? Many peers went to the United States. Why not the Middle East or Southeast Asia? It was our most intensely debated internal topic. There were three main reasons. First, we wanted a market with growth potential that could sustain growth over the long term. Our assessment favored Europe, America and Japan: developed economies with mature markets and more standardized compliance. For a startup, they look expensive, but the cost of mistakes can actually be lowest. Second, purchasing power and the overall user ecosystem are very good, whether B or C, and completely different from China. Considering long-term development and overall demand, we judged Japan to have the greatest potential outside Europe and America. Why rule those out? As mentioned, many peers and the leading applications of the AIGC generation are concentrated in the United States. A small team competing with them in Europe and America might have weaker core advantages. Third, consider the market window. When we founded the company in '24, we compared similar products in Europe and America, Japan and other Asian markets. We judged Japan to have the lowest AI penetration, with product models and forms still quite basic. That gave us a relatively long entry window and favorable timing. If there is room to survive long-term, we might focus more on Japan. We considered our team's overall competitive strengths and the space the market offers—including areas big companies might not enter. Only there do we have opportunities to develop. On balance, we chose Japan first.
Qi Wei: Thank you. I visited Japan this month and strongly felt that its supply of AI products was relatively weak. At exhibitions and AI events, the products seemed much like those I saw 10 years ago. They appeared to carry the Agent label, but were not particularly AI-native. For example, 90% of the exhibition products were To B, perhaps still centered on document organization and company Workflow construction. Japan's local supply and AI development in China and the United States were no longer at the same Level, in my impression. That creates a substantial opportunity for Chinese and American entrepreneurs. But Japan is a market we describe as easy to defend and hard to enter: once Japanese users accept a product, they do not switch readily because of loyalty, yet they are cautious about trying new products and services. At your current stage, have you tried growth approaches that offer a better way into the local Japanese market?
Wendy: We have studied Japanese emotional, software and social products extensively. Growth there requires long-term cultivation; results do not come quickly. In Europe and America, heavy paid acquisition might bring many users in a few days, but that model is difficult to make work in Japan. You need the ability to keep working deeply in the market. Some of our growth methods may seem very old-fashioned. Yet these traditional methods can have surprising effects in Japan. I was just discussing this with Qi: how do we promote? Sometimes through local, on-the-ground marketing. We go to anime-culture shops in places such as Akihabara and partner with stores to distribute flyers. We also ask university students to distribute advertisements. We found that users converted through this traditional advertising actually have the highest retention on our platform.
Qi Wei: Yes, Japan's offline market can rival online traffic. Anyone interested in Japan should try more offline channels. Thank you, Wendy. My next question is for Akamai's director of architects, Li Wentao. As a veteran with more than 20 years of globalization experience, looking back over the past 25 years, what was the most important step? In today's AI-inference arena, how does your globalization strategy fundamentally differ from when you were building CDN?
Li Wentao: Akamai is an inventor of CDN. We were founded to solve waiting on the internet. We wanted people to watch high-definition video, shop and interact with websites and APP services smoothly. Global operations were therefore inherent from the start. Over the past 27 years, we have extended infrastructure to more than 130 countries, with over 4400 facilities. We have deep partnerships with 1200 carriers worldwide. In major countries, that has built local network capabilities, carrier interconnections and the ability to operate stably and compliantly over the long term. Those 27 years have produced substantial experience. Chinese companies target developed markets such as Europe, America, Japan, Korea and Australia, but we also cover Belt and Road markets, including the Middle East and Central Asia. Our highly global platform can help you reach users worldwide at once. Beyond the platform, we place great importance on people. In major countries, local teams support customers' local businesses and operations, even offering information-security and compliance advisory assistance. The platform plus our people's experience and local operational knowledge are important differentiators.
Qi Wei: Thank you. Our approaches to globalization and localization really differ. Star Shine described a To C approach, whereas Akamai is wholly a To B platform supported by local teams. Globalization can go very deep: in Japan, for example, a local spokesperson may be necessary. A Chinese or American person promoting a product may be less effective than a Japanese endorsement. There are also lightweight approaches: without local employees, you can operate a Global business from China by establishing local channels, growth and promotion. My next question is for Zhongju Intelligence's CEO Gao Zhou. Enterprise services inherently have strong local characteristics—tax rules and contracts differ by country. You mentioned Shopify as your main global focus, with a large North American share. Compared with To C services, how do enterprise-service products take that first step globally? For example, tax requirements differ across countries. You are now in Japan and North America; will you choose other markets next?
Gao Zhou: OK, thank you. For us, the important issue is not really facing tax or legal compliance, because we resolved those when establishing the company in this field. They are relatively standardized. Find professionals to help and it is OK. The underlying logic is to make the product good first: what problem does it truly solve for users? For overseas compliance, find the right professionals. That is therefore not my main Focus. My focus is whether the product genuinely helps users solve their problems.
Qi Wei: OK, that is central. Your customers are mainly Shopify users, perhaps mostly Chinese cross-border e-commerce businesses going overseas. Their pain point is indeed technical services for cross-border e-commerce rather than country differences. Could you briefly compare serving those locally based cross-border teams with Local teams, such as American e-commerce businesses?
Gao Zhou: Their needs are actually the same because they serve similar audiences. The Local store Owner may be American and the domestic owner Chinese, but their customers are similar, so their functional requirements are alike. When serving the groups, though, domestic customers honestly pose more challenges because their requirements keep changing. Overseas, ours is a highly standardized product: users can easily configure the Agent, knowledge base and SKU and complete the whole chain themselves. Domestic customers request many customizations, even if a feature is not a genuine need. In China, we often need on-the-ground Sales staff to serve them. For North American customers, we may need just one local Support person. Once we resolve registration or store-linking Bugs, there are not many other requests. The upside is that domestic users bring pain points we had never considered, which we can turn into standard features. Each group has advantages and disadvantages, but we will serve both well.
Qi Wei: OK, it seems we Chinese can be quite difficult! So far, we have discussed applications and services going overseas. InfronAI is different: you provide the global AI-model pipeline layer. You differ from OpenRouter, which focuses more on independent developers, whereas you already do substantial local To B business. Zheng mentioned getting up at 7 to meet the West Coast team; I said that without meetings at 12 midnight or 7 in the morning, you hardly dare call yourself a global team. In AI Infra services, what differs most from application-layer globalization? How does North America differ from Europe, Japan or Korea, including technology and service?
Zheng Han: Let me break that down. First, we do To B business. At home or abroad, To B differs from To C. Second, ours is an infrastructure-oriented product. A user switching a shopping App faces low decision costs—personal habit or culture. A company changing an API supplier, however, usually takes one to two months. To B decisions involve R&D, finance and the overall decision process. We call our North American approach “BC from the same source.” First, we have our own Bay Area Office and something called Hacker House: we rent a large Airbnb villa and host nightly events, drinks and Workshop sessions. When we do BD and meet people for Coffee, many have heard of us or attended our Workshop. That is useful in To B BD. Second, North American users care about Connection and often begin by asking about mutual friends. Our card is a YC Portfolio connection among our partners, which quickly brings people closer: shared YC alumni ties and the same VC contacts, plus the service our product provides. In my comparison, OpenRouter cannot offer 24-hour service, but Hangzhou and the Bay Area have opposite time zones, letting our two teams provide 24-hour technical responses to North American users. Those are our three main approaches to winning customers in the Bay Area.
Qi Wei: Thank you. That is a great Topic—How to build in Silicon Valley—and gives us a useful template. We have discussed global differences and how to expand. Next: how do our products build moats in global competition? Competitors are not only other Chinese companies going overseas, but companies from everywhere. Please each take two or three minutes to describe the moat you believe rivals would find hardest to copy.
Wendy: Since we are the youngest team, I'll begin. We have not yet promoted widely in Japan; I can only share what we have learned while exploring. Entry barriers are high, and it is hard to acquire users through heavy paid advertising. Getting users to stay organically is a test. Product strength is the entry requirement, but the core moat in Japan is the ability to keep iterating and cultivate the market over the long term. Many startup teams die within a year; that does not fit Japan. The first-year goal might be to achieve a modest positive cash flow as quickly as possible, rather than burn heavily. Only after completing the business cycle can you put down local roots. My advice for companies entering Japan: prepare for at least three years. Product and technology are requirements, but sustained cultivation and deep user insight matter most. They generate future organic growth and high retention.
Qi Wei: An excellent answer. From what I have learned about Japan over the past two years, it is much harder than imagined. Weak supply does not mean your product can instantly take the market. Patience matters enormously. Next, Gao Zhou of Zhongju Intelligence: what moat is hardest to surpass?
Gao Zhou: I have long believed that in the AI era, nothing has a moat, especially technology. I have worked in quantitative trading for almost ten years, centered on strategies. Everyone uses strategies, but the true moat is their dynamic updating as they interact with the market. Is the product itself the moat? Not necessarily. It is the accumulated data, iteration and integration with the latest technology as the product is repeatedly refined with the market. When we began Infra, we also wanted a moat, but Dify and others were already emerging. Looking back, moats are always dynamic. Today's moat may be gone in three or six months. The moat must therefore be the mechanism formed through continuous product-market refinement.
Qi Wei: Indeed, technology iterates so fast. When underlying models and capabilities change, something built two or three months ago may be overturned. Teams must keep iterating through the interaction of technology and markets. Zheng Han, CTO of InfronAI: in AI middleware or basic services, beyond sales and marketing, what technical or service differentiation and moats can you build?
Zheng Han: In AI, understanding and product iteration change weekly. A new large model comes out—Claude 3, for example—and other models immediately lose traffic. Users embrace new technology with almost no loyalty. Product, technology and UI aesthetics are therefore basics; without them, you have no chance even to discuss it. We believe our moat is spending 60% of each day on customers. When they have a problem or want a new model, they come to us first. Recently, a Japanese audio model could only be purchased with a Japanese domestic credit card. The customer could not get one; we resolved it that afternoon and connected the model. That was a Surprise for them. As this keeps happening, our connection deepens. They Share corporate strategy with us. We build infrastructure and must be deeply connected to users' businesses. Recognition between people and teams, plus 24-hour service, is our moat, I think.
Qi Wei: Let me ask: in this era, is there no shortcut other than doing work more intensively and deeply?
Zheng Han: Things really do change too quickly. Another example: when we talk with ByteDance and model providers, everyone strongly feels that whenever a new large model appears, back-end traffic falls off a cliff. The supposed Prompt Engineering and engineering moats vanish the moment a new model arrives. We can only embrace the market faster and more deeply and stay close to customers. That is how you dynamically maintain PMF; otherwise, PMF disappears quickly.
Qi Wei: Thank you. The previous three speakers discussed building local moats or deepening service. Akamai has 27 years of history and many advantages in its platform, channels and customer relationships. As new players and technologies keep emerging, how do you maintain those moats?
Li Wentao: Akamai's distinction mainly has two parts: global service experience and a global cloud platform. On experience, Akamai carries 20% of global internet traffic. Leading Chinese companies expanding abroad in every industry rely on our platform. We want them to use these best practices to avoid pitfalls. AI changes rapidly, and Time to Market matters; Akamai is a stable platform that can help. On the platform side, we are a public cloud with compute coverage in more than 130 countries. Through our strategic NVIDIA partnership, we deploy GPU compute at the internet edge. The platform is one network hop and ten milliseconds from 90% of global users. We hope Akamai will become the boundary where the physical and AI worlds interact. Highly real-time businesses, such as connected cars, real-time games and emotional interaction, need compute closer to users. Processing local data nearby supports compliance and improves user experience and personalization, in our view. China's AI industry has strong innovation and engineering capabilities. We hope to respond to change with a stable foundation, using distributed global compute and service experience to help companies compete globally faster.
Qi Wei: Thank you. Another Topic: no one globalizes without mistakes. Could each guest share the most costly pitfall you encountered—product decisions, market selection or team building? If you have not had one, a positive case is fine. Please share freely.
Wendy: Why is it ladies first every time? I'll go first. We have not encountered a particularly major pitfall ourselves, but I have heard of many peer teams facing a huge issue in Japan: hiring local people is difficult. It is a major obstacle to genuine localization. In Japan especially, once you hire regular employees, dismissing them is very difficult. One team aggressively hired 100 people in Japan within six months, then found efficiency and technical iteration much weaker than the Chinese team's. The founder's difficult decision to lay people off came at a painful cost and nearly put the company at risk of collapse. Companies considering Japan must think through local team building and collaboration with their China team. They need a local team while also drawing on Chinese engineers' strengths.
Qi Wei: Thank you. Could InfronAI share next?
Zheng Han: Besides business, our biggest challenge in North America now is organization and personnel. First, people are dispersed: we have colleagues in Sydney, New York and Los Angeles, with R&D in Hangzhou. Time differences make meetings hard to coordinate. Second, Remote work makes information transfer inefficient while the market changes quickly. Our hiring now is like “dating.” If we meet a potential partner at a Workshop and have a good conversation, we try working together for three months with Commission. If an outside salesperson in the Bay Area brings customers, we pay up to 10% Commission. People we meet through external events gradually become BD or sales leads this way. We are also wondering whether to hire local white people, because, in my own generalization, Asian women find it easier to start conversations in the Bay Area, whereas men end up talking only about technology. To build deep Connection, I think local people are still needed.
Qi Wei: I've learned something. Next, Gao Zhou of Zhongju Intelligence.
Gao Zhou: We had a small pitfall during our transition. We initially built the Infra layer, but when we saw major Chinese companies doing it, we decided to abandon middleware development. We were very anxious. We kept wanting to use AI to disrupt things, but in vertical sectors we found every industry had giants and rules. So we became more restrained and first looked at how others worked. If something used to cost 10,000 or even 100,000 a month, can AI cut it to half or even one tenth? We no longer think about changing the world. We first tackle a small pain point in an industry. That was a pitfall and challenge in our startup journey.
Qi Wei: Thank you, Gao. Finally, Wentao.
Li Wentao: Two things. First, cost management. Cloud-resource use differs completely from 0 to 1 and from 1 to 10. Many providers offer Credit to help you start quickly, but I characterize all cloud-cost pricing as exponential. CTO and technical teams should therefore think about decoupling from cloud providers on Day 1 and avoid overdependence on one vendor. Applications should be multi-cloud-friendly and embrace open source to support future expansion. Second, many customers focus on model fine-tuning and training, but as I recall Jensen Huang saying at GTC, future AI inference costs will be ten thousand times training costs. For long-term operations, how to make inference more scalable and substantially reduce Token costs is another challenge to consider early.