From Hedge Fund to AI Startup Frontlines: A Conversation with Du Zhiheng on Rebuilding the "Information Superhighway" for Agents | Yunqi Capital Attent!on Podcast

The "last mile" of search engines is being rebuilt.

In May this year, Microsoft announced it would shut down the Bing Search API. This didn't just mean external developers and AI application builders lost a stable, high-quality search data source — it also meant that in the era of rising Agents, a critical pathway for "connecting to and sensing the world" was severed.

But the ecological void left by giant contraction is being filled by emerging forces. AI startup Cloudsway (小宿科技), founded at the end of 2022, is dedicated to building a search engine API and model aggregation platform for AI Agents.

Recently, we spoke with William Du, CEO of Cloudsway, and Chen Yu, Partner at Yunqi Capital, about what it means to build a search engine for Agents, the strategic considerations behind the evolution of API ecosystems, and the entrepreneurial journey and industry observations amid the Agent wave.

Scan the QR code above or follow "Attent!on" on Xiaoyuzhou to listen to this episode

Guests

William Du, CEO of Cloudsway

Cloudsway.AI is an Agent infrastructure service provider, offering one-stop real-time search, model aggregation, and compute services for Agents. Its model aggregation platform Skyrouter.ai has connected with over 100 mainstream global models. Cloudsway's intelligent search API handles hundreds of millions of monthly calls, with business resources deployed across more than 80 countries and regions, successfully helping hundreds of enterprises achieve AI transformation.

Chen Yu, Partner at Yunqi Capital

Deeply engaged in frontier technology sectors including AI and embodied intelligence. Representative investments include MiniMax (large model unicorn), DeepRoute.ai (leading intelligent driving company), and PingCAP (open-source database unicorn).

Linda, Managing Director at Yunqi Capital (Host)

Timeline

Part 1: Getting to Know the Guests

01:05 From internet to hedge funds to AI entrepreneurship: the constants and variables in "unconventional" choices

Part 2: Opportunities Behind the Search Engine Ecosystem Shift

08:59 The signals and opportunities behind Bing Search API's discontinuation

15:22 What is Cloudsway doing? Why choose the dual product of "search API + model aggregation"?

22:20 How big is the market for Agent search engines? How do the technology and product differ from conventional search engines?

28:16 Cloudsway's user base and competitive advantages

31:42 Several business models for "Agent services"

Part 3: Heart-to-Heart: On AI Entrepreneurship, ToB Entrepreneurship, and Globalization

34:07 Advice for AI entrepreneurs: passion-driven vs. rational analysis — both engines matter

35:41 Global ToB entrepreneurship experience: why is the China market a "grinding field," not a "profit field"?

40:21 Cloudsway looks forward to building the Agent world together with more developers and entrepreneurs!

Below is an excerpt from this podcast episode (edited for clarity)

From Investor to Entrepreneur: A Role Transformation

Linda (Host):

Your life experience is indeed quite rich. As a STEM guy, crossing over into secondary investment and then entrepreneurship — what's the internal logic behind these seemingly unconventional choices?

William Du (CEO, Cloudsway):

My experience has been quite diverse. I joined Baidu in 2013, working in the PMO team for Mingyuan in mobile cloud — probably the first PMO team in China's internet industry. Back then my main responsibilities were Baidu Mobile Assistant and Mobile Baidu strategy, plus some external investments. I was among the earlier wave of people transitioning from consulting and investment banking into the internet sector, and also among the first to work on mobile search projects. Our team was quite interesting — basically everyone came from investment banking or consulting, and today we're all quite active starting businesses in various fields. After leaving Baidu, my first move was actually to chat with Chen Yu. We hit it off quite well. But in the end, due to personality considerations, I went into secondary markets and spent the next ten years there.

After leaving Baidu, I joined Hillhouse, responsible for global gaming and entertainment investments, then helped found HSG's hedge fund. During this period we made some investments in infrastructure companies and also worked with cornerstone investors. This was when I started observing and entering the infrastructure field, and through the infrastructure industry began observing how AI was changing the world.

Linda (Host):

Chen Yu, do you still remember your first impression of William more than ten years ago?

Chen Yu (Partner, Yunqi Capital):

I remember it was at a café in Huamao. I thought William was quite sharp — you could tell he was smart from the first conversation.

William is still sharp today. But after more than a decade of experience, he's become much more grounded in temperament. Of course, our topics of conversation are different now — basically everything revolves around AI.

Linda (Host):

William, how does it feel to still be old friends with Chen Yu despite your different identities over these ten years?

William Du (CEO, Cloudsway):

Yeah, we've never lost touch. It's just that before, as a secondary market investor, I would often seek Chen Yu's opinions on various questions — how to view NVIDIA, how to view inference versus training compute, all kinds of questions. Chen Yu would always explain them to me, which was very helpful. This time, Chen Yu's encouragement was also tremendous when I decided to start a business. When I really began thinking about this venture, I actually asked for Chen Yu's opinion — around Q3 last year when I started preparing. At the time, Chen Yu encouraged me: "As long as you go for it, combined with the infrastructure capabilities you've accumulated, Cloudsway will definitely find its scene in AI. If you find the scene, we'll definitely invest in you." So in September last year, I officially joined the company to go all-in on entrepreneurship. Back then the company had some basic overseas cloud-native business, but hadn't found its scene in AI yet. AI revenue was basically approaching zero — almost nothing.

Search API Industry Shift: What Does Bing Search's "Supply Cut" Mean?

Linda (Host):

We mentioned Bing Search's API shutdown at the beginning — this is actually quite a significant event. Zhihang, based on your overall understanding of Bing Search API and Microsoft's AI layout, how do you view the function of APIs?

William Du (CEO, Cloudsway):

For the vast majority of users, what we use and see are C-end search. The biggest is Google, the biggest in China is Baidu — this is what people use search engines for in daily life, obtaining real-time and comprehensive data and information. Many mobile internet and SaaS products also use such real-time data, so in the past era, the mainstream product form was search APIs.

Bing's search API was one of the most mainstream products globally. So its shutdown has had fairly obvious impact on everyone in the market who needs real-time data and search data. But recently, for Microsoft past and present, the value of Bing has changed significantly. Because in the past, Bing's ToB API was just an extension of its C-end product — packaging the C-end product into an API to make some easy money and serve customers. Beyond enterprise clients, many well-known search products actually called Bing in the backend — Yahoo, DuckDuckGo, these search products all called Bing — because Microsoft only treated them as regional or vertical supplements, not strategic competitors.

But in the AI era, search APIs have taken on greater strategic significance. Next-generation search products like Perplexity, and various deep research AI products domestically and internationally, are actually competing with Bing and Google for C-end entry points, forming direct challenges to each company's search traffic. From Microsoft's perspective, its goals are actually quite clear. First, don't let cheaply available APIs create more competitors for yourself. Second, in my own understanding, combine data capabilities with model capabilities to maximize its own advantages.

Linda (Host):

Chen Yu, from your perspective, how do you understand Microsoft's choice logic?

Chen Yu (Partner, Yunqi Capital):

Let me explain from Google's angle. Google also used to provide search APIs, but later felt this might give competitors an advantage. For example, if you provide search APIs to other AI search engines like Perplexity, it's essentially handing over the entry point to others while only serving as infrastructure — very unfavorable to Google's business model, since a large part of Google's revenue comes from search advertising. Only by becoming the entry point yourself can users see and click on your ads. If you provide APIs out, it's not worth it, because API revenue is certainly far less than advertising revenue. Microsoft probably has similar considerations.

Linda (Host):

For other similar API suppliers, will there be other impacts? William, what's your take?

William Du (CEO, Cloudsway):

It's definitely good news for similar API suppliers. Because Microsoft was the largest global supplier of B-end search APIs, and its alternatives are actually very few. As Chen Yu just mentioned, B-end search monetization efficiency has been far lower than C-end search. If calculated by CPM, C-end search monetization efficiency is 5-10x that of B-end, and doesn't require such heavy sales. So there are very few suppliers truly doing B-end search APIs. Looking overseas, one search API was originally a Web3 token-issuing privacy browser project that did search because it had its own browser. It has recently received a lot of traffic on the B-end. Among Chinese overseas-connected search and AI agent clients, we currently believe Cloudsway's intelligent search is the preferred choice. Because among domestic suppliers, we are one of the few that simultaneously possess the capability to provide real-time search data and content APIs in Chinese, English, and even Spanish, Portuguese, and more than 30 other languages. This is good news for everyone.

What Is Cloudsway Doing? Search API + Model API Aggregation

Linda (Host):

Zhihang, walk us through Cloudsway's business systematically.

William Du (CEO, Cloudsway):

Cloudsway is an AI company founded in Singapore at the end of 2022. Early on we were constantly thinking: what is something that fits Cloudsway's DNA? If you look at it today, the answer is very clear — we are a product platform serving Agents. We currently mainly provide two products: one is the search API for Agents that we've discussed a lot; the other is a model API aggregation platform. These are our current businesses. But the better question is: why do we do these two things? The origin was around mid-to-late last year, when we actually had three judgments, and now they're being verified one by one — all of which might be hot takes.

First judgment: Chinese and Americans will become the main protagonists in the AI world, while all other regions and countries will be the playground for these two groups. Each side will defend its own battlefield while competing comprehensively in all other battlefields.

Second judgment: we believed inference demand would rapidly exceed training demand, and within inference compute demand, 90% or even 95%+ would manifest in API call form, because leading foundation models would rapidly converge. This led us to decide not to build our own compute, but to figure out how to leverage existing leading API capabilities.

Third judgment: we believed Agents would become the main product form after large models. Beyond the large model capabilities inside, real-time data would become an important need, and a large portion of real-time data could be met through search APIs. Of course some portion couldn't be met by search APIs because they weren't indexed or were in closed ecosystems — this portion might be addressed through methods like AI browsers. But we believed perhaps more than half of real-time data could be met through search APIs, which determined our focus on Agent-facing search.

Linda (Host):

What's the difference between search engines for Agents and search engines for humans? I'd like to cue Chen Yu on this.

Chen Yu (Partner, Yunqi Capital):

This is a great question. Everyone knows Agents substitute for humans in executing tasks, but problems arise during web browsing — often when you visit websites, they ask you to click images to verify you're human. For many so-called Agent browsers, or for Agents generally, this is a painful problem. So I believe the industry needs to define a standard for how Agents can declare themselves as Agents when interacting with web pages, while maintaining human-like operations yet providing programmatic convenience for Agents. When humans interact with browsers and web pages, there's a lot of extraneous stuff — fancy visual effects to ensure there's a real person in front, or additional prompts through voice or other methods. But none of this is needed for Agents.

For Agents, the hope is to obtain the information needed to execute operations in the simplest, most efficient way. And if the information is structured and unambiguous, that's ideal. So I believe this entire mechanism of interacting with web pages needs to be redesigned.

Linda (Host):

Then why does Cloudsway also do a model API aggregation platform?

William Du (CEO, Cloudsway):

We actually approach this from the customer perspective. Because our judgment is that the manifestation of inference demand — the vast majority — won't be customers buying GPUs or renting GPUs to deploy models themselves, but rather calling through APIs. When serving these AI customers, we found that basically every AI customer is calling large models, but cost and efficiency are fairly significant challenges.

If you look at general-purpose Agents today, their spending on models far exceeds their spending on search. This is the logic of how our own DNA can meet customer needs, because Cloudsway's DNA is actually as a distributed infrastructure vendor, born from distributed infrastructure. So for API aggregation platforms, we naturally have this capability. We have distributed infrastructure, channel resources with various global cloud providers, and resource operations capabilities — through this we can meet the needs of the AI customers we serve. So we quickly launched this service, and customers have been very enthusiastic in their demand for it.

Cloudsway's Market Position and Competitive Advantages

Linda (Host):

Curious to ask William — who is using your products now?

William Du (CEO, Cloudsway):

Our typical customer groups — we actually look forward to a flourishing AI Agent ecosystem. Currently there are already some categories with commercialization or considerable scale. For example, first is connected search: some well-known large model platforms' overseas connected search uses our English search interface.

For example, some general-purpose Agents — recently there's been a very hot general-purpose Agent whose search and model interfaces both call through us. Beyond this, there are many vertical Agents, such as e-commerce, Web3 gaming — we have benchmark customers in these areas. For example, as Chen Yu mentioned, Yunqi's portfolio company Creao, an Agent generation platform, is also one of our typical customers. We're all working hard to provide the best service.

Linda (Host):

What are Cloudsway Search's competitive advantages?

William Du (CEO, Cloudsway):

I think this can be divided into two parts. First, since starting the business, most of my time has been spent without the leisure to look at others — every day I'm thinking about how to make my own products better and better meet customer needs. So I actually don't have much energy to pay attention to competitors. If I look at my own selling points or what stands out from customer service interactions, there may be three aspects where others don't have or don't do as well as us.

One aspect I can think of is multilingual capability. Because the vast majority of domestic search APIs only provide Chinese search services. We should be one of the few that can provide Spanish, Portuguese, Russian. Looking from a VC investment perspective like Yunqi's, the Agent developers they invest in, like Creao, are basically global from day one. They not only need global go-to-market but also global real-time information data services, so from this angle I think we've gained some first-mover advantage.

Second is ensuring high quality — providing a search product with high relevance, real-time capability, and authority, while guaranteeing high concurrency and high SLA on the resource side. Beyond this, we can additionally provide full-text crawling capability on top of traditional search ToB APIs. This need was mentioned earlier — when many Agents read information, short summaries can't meet customer needs. They actually need to crawl the full text, for example processing the top 20 results together. Because of our accumulation in edge cloud and distributed infrastructure, our full-text capability should be the strongest on the market. Third is probably cost-effectiveness — after Bing's price increase, the price is quite outrageous. Our price is only one-third of Bing's.

How to Build the "Service Industry" for Agentic AI?

Linda (Host):

With the explosion of Agent applications, serving Agents and building Agents has become a fairly active entrepreneurial track. In your view, what are some viable business models for serving Agents?

William Du (CEO, Cloudsway):

Our observation is quite simple. I'm actually now focused on doing just these two things: providing Agents with easily accessible, low-cost model aggregation platforms — essentially wholesale-to-retail, but with considerable value for customers. Second is what we believe to be search APIs — of course we've also observed other aspects beyond real-time information acquisition. Essentially, from this angle, what we've observed is the perspective of real-time data and information that models themselves cannot provide, and within this we've chosen search APIs. Of course other potentially large perspectives exist, including AI browsers and other information acquisition methods. We believe real-time data will become an important and commercially viable domain.

Linda (Host):

Chen Yu, from your perspective, what other viable business models do you see for serving Agents?

Chen Yu (Partner, Yunqi Capital):

Let me add two more. First, a good Agent, especially a general-purpose one, needs many tools to support it. So providing additional tools for Agents may be a potential business model. Second, Creao, which we interviewed in our previous episode, is actually an Agent development platform that can help developers quickly develop Agents. We believe this is also a way of serving Agents.

For the full content of this episode, subscribe to the "Attent!on" channel on Xiaoyuzhou app~