An Agent Infra Founder with $25M ARR Says: We're at 6:05 AM on Day One of the Agent Era | A Conversation with William Du of Xiaosu Technology

What Opportunities Has the AI Agent Boom Created for Upstream and Downstream Industries?

👦🏻 Interviewers: Koji & Ronghui

🥷 Editor: Starry

🧑‍🎨 Layout: NCon

What opportunities has the AI Agent boom created for upstream and downstream industries?

In this episode, we invited William Du, co-founder & CEO of Xiaosu Technology, to share how his company — an AI Infra provider behind Agents — seized the moment as Agents took off: with Xiaosu Intelligent Search, designed specifically for Agents, and the one-stop model aggregation platform SkyRouter.ai, the company pushed its overall ARR past $25 million.

As "pick-and-shovel" players in the Agent space, William told us we're only at 6:05 AM on day one of the Agent era — the sun has just risen, with a long road ahead and massive room for growth. Beyond Agents, William also shared his career choices, his transition from secondary market investor to entrepreneur, and key lessons learned. We hope this helps listeners understand Agents and the AI Infra behind them.

PS: Xiaosu Technology has prepared a gift for Crossing listeners. Click the "Read More" link at the end to register and receive 2,000 free intelligent search API calls or an equivalent professional services package.

Listen on WeChat:

Listen on Xiaoyuzhou:

The Infrastructure Provider Behind AI Agents: How to Reach $25 Million ARR

👦🏻 Koji

Hello everyone. This week's Crossing guest is William Du, CEO of Xiaosu Technology. Xiaosu chose the Agent Infra path because they realized that even today's leading agent vendors and products likely have user bases less than 1% of top mobile internet-era apps. That means at least 100x growth potential in the agent market.

William and I go way back — when I founded Jiepang, he was one of the most active and proactive interns on the team. Through Crossing, I've reconnected with many old friends, and it's wonderful to see everyone thriving in their fields. I'm delighted William could join us to introduce Xiaosu Technology, which I consider a force to be reckoned with in this wave of AI entrepreneurship. This episode is also recorded as a video podcast, available on Xiaohongshu, Bilibili, and WeChat Channels — follow us there.

Let's start with today's rapid-fire round.

👦🏻 Koji

William, how old are you?

👦🏻 William Du

Born in 1990, so 35.

👦🏻 Koji

Where did you study?

👦🏻 William Du

Tsinghua University, double major in Aerospace Engineering and Economics.

👦🏻 Koji

Your MBTI and zodiac sign?

👦🏻 William Du

INTJ, Aries.

👦🏻 Koji

One sentence about your current company and product?

👦🏻 William Du

Xiaosu Technology is a one-stop Agent Infra platform. Our core product is Xiaosu Intelligent Search, designed for AI agents, and we also run Skyrouter.ai, a global LLM API aggregation platform.

👦🏻 Koji

Current funding status?

👦🏻 William Du

Series A is closing. Amount and valuation will be in the official announcement. We expect to open Series B by year-end.

👦🏻 Koji

What about revenue and profit — anything you can share?

👦🏻 William Du

ARR exceeded $25 million in June, and we're break-even on a PNL basis.

👦🏻 Koji

Current team size?

👦🏻 William Du

Over 100 people, roughly 70% in R&D, the rest in marketing and sales.

👦🏻 Koji

What were you doing before entrepreneurship? Tell us about that.

👦🏻 William Du

Before this, I was a hedge fund portfolio manager and family office CIO, and the first employee at HSG's secondary market fund.

👦🏻 Koji

So William has extensive finance experience, and his undergrad was in aerospace — quite a跨界 journey. What opportunity did you see when you decided to found Xiaosu Technology?

👦🏻 William Du

Before entrepreneurship, I spent over a decade at Hillhouse and HSG's hedge funds and family offices, riding the waves of China's internet sector, watching cycles of startups and public companies. Choosing entrepreneurship was actually inevitable — working at a family office or HSG's secondary fund is similarly entrepreneurial, just in a different domain.

My career choices look scattered: consulting first, then Baidu, then secondary market investing, family office, now entrepreneurship. But my approach is dual-wheel driven: both investing and entrepreneurship are based on views of the future. Between reality and goals lie interest and rationality. For me, interest drives, rationality guides — these two wheels power my actions.

In Q3 last year, I officially joined Xiaosu full-time. The company already had some overseas infrastructure business, but hadn't found an AI business model yet — AI revenue was minimal. I believed that with our infrastructure capabilities and the founding team's business experience, we could find product-market fit, so I joined without hesitation. Now overall ARR has exceeded $25 million, growing very fast.

👩🏻 Ronghui

Can you explain in simpler terms what you do, for listeners who aren't familiar?

👦🏻 William Du

What we do is relatively less sexy and intuitive than ToC AI products, because we're the service providers hiding behind these AI agents.

Simply put, two things:

First, when we saw connected search, AI intelligent search, Deep Research and similar research interaction capabilities emerge, we realized that beyond model capabilities, people need real-time information and data. In the previous era, real-time information and data was embodied primarily through search engines, but in the AI era, search takes new forms.

For twenty years, humans mainly searched by inputting queries, scanning the most eye-catching, interesting links in returned results, clicking to read content, often opening multiple links to filter what they wanted. But as more people search through agents, search behavior becomes machine-to-machine, no longer human-to-machine.

In this process, we saw structural opportunity. For the past two decades, search business models and formats were user-facing — concerned with content presentation, various cards, focused on click-through rates for top results as core metrics of recommendation accuracy and capturing user attention. The business model was mainly selling targeted ads, because user queries expressed their immediate needs.

So we judged that as AI agents rise, ToC search engines that originally held over 90% market share would see ToB's share grow from single digits to near 80-90% of information access entry points. Additionally, for general agents like Manus and Kuaishou's Skywork, implementing each task requires breaking down queries and conducting multi-round searches for each. Even based on current search volumes, the shift from human to agent search will trigger an explosion in search frequency. Meanwhile, B-end share will rise dramatically. These two factors combined led us to set "search APIs designed for machines or agents" as our primary strategic direction — that's the origin of Xiaosu Intelligent Search.

Second, while serving leading agent clients, we found that agents need not just information but one-stop infrastructure services. I remember Koji mentioning that new-era AI agent companies are often one-person, three-person, five-person teams that might grow into unicorns, but won't invest heavily in managing Infra. So beyond intelligent search APIs, we provide the one-stop LLM aggregation platform Skyrouter.ai, letting these small teams use leading models from day one without geographic or usage limits, to achieve their agent product goals.

That's roughly our product matrix.

👦🏻 Koji

How did you come up with Xiaosu Intelligent Search? What's your advantage?

👦🏻 William Du

The average perception of search APIs is probably still stuck on the old model of repackaging consumer search into a B2B API. But our journey to building intelligent search went through three distinct phases before arriving at today's Xiaosu Intelligent Search:

Phase one, we didn't jump straight into search itself. Instead, we spotted a capability gap among leading LLM companies offering AI search and web-connected search. Legacy search APIs returned 10-20 links with brief snippets, but those snippets were neither comprehensive nor long enough to meet the demands of large models or intelligent search. From September last year through January this year, our first product was full-web full-text extraction, designed to work alongside search APIs — initially Bing's. After an agent called the Bing search API, our service could rapidly and accurately capture full articles and multimodal content for the client's LLM to process further.

That was phase one: full-text retrieval and crawling services to complement search APIs.

Phase two, after DeepSeek R1 exploded in February, we discovered that Bing had abruptly stopped providing API services to certain Chinese model vendors and major internet companies. We concluded that leading global search providers, Bing included, would cease serving China's AI industry. This directly drove our second product: self-developed search.

We built a hundred-billion-scale index and developed the complete pipeline from recall to ranking, creating an intelligent search product that could directly replace the Bing search API. This call was validated by late May, when Bing announced it would retire its search API globally on August 11, after which it could only be used bound to cloud or agent services — confirming our strategic judgment three to four months ahead of time.

Phase three was the critical stage that established the comprehensiveness of our intelligent search capabilities. In May-June this year, as Manus and other agents exploded, we discovered through serving leading agent clients that agents demanded far richer search capabilities than before:

• Not just full-text capability
• For example, when using Manus or Kunlun Tech to create presentations or product documents, they needed text search, image search, video search — multimodal search capabilities that legacy search APIs lacked
• Agent products going global from day one needed multilingual search beyond Chinese: English, Spanish, Portuguese, and more
• Customization capabilities, such as content extraction after URL clicks, ad filtering, core content capture, and so on

The new generation of agents demands richer search capabilities, and currently only Xiaosu Intelligent Search possesses these, because we're genuinely serving these leading AI agent clients. Other competitors in the market, both overseas and domestic, mostly still provide traditional search APIs — essentially text-search-text returns, simple in form. We don't really consider them competitors.

👦🏻 Koji

You mentioned seeing this opportunity in February and only then building out an R&D team to do search. It sounds like you produced a sufficiently usable API in a very short time. How did you pull that off? That sounds quite difficult.

👦🏻 William Du

I think we're essentially standing on the shoulders of giants to continue the work. After making this strategic decision, beyond self-development, our other leg was advancing through M&A. There were teams and products in the market doing intelligent search or AI search that had already done substantial work but later shifted strategic direction. So we acquired a mature in-house search team and product as the foundation of our offering, then continued building on top of it with full-text, multilingual, and additional capabilities.

Since you asked, I think it's worth explaining where the competitive moat of a search product actually lies, because it's an enormous engineering challenge. This manifests in three main areas:

First, the moat isn't money or resources — it's the team and talent. Over the past decade, virtually no new technical talent has been cultivated in search, because search is an old, traditional product form. In the early internet era, Baidu, 360, and other traditional search companies trained many excellent people, but after the explosion of Toutiao and TikTok, most young technical talent graduating first thought of building recommendation engines rather than search. So search talent in China is an extremely limited, tight-knit circle. Startups or external teams can't just instantly assemble a 20-30 person R&D team. You have to find core people first, who have influence and pull within that circle, to then "recruit pyramid-style" and bring related talent in — only then can you build a team. If a small company claims to have built a native search engine with just 5-10 people, and no one on the team has led specific modules at a traditional search company, that's simply unrealistic. It can only be a wrapper product. That's the first major moat.

The second moat is the human, financial, and compute resources you mentioned, invested over 6-9 month cycles to produce a first version. We skipped this phase through our acquisition approach, and based on core talent, completed our current product.

The third moat is that ToC products don't need to serve major clients — with users come feedback, positive or negative. But doing search for agents is different. You can't directly obtain user feedback, because users aren't directly using your search results, and additional feature requirements only come through client feedback. If your clients on day one are just small developers, you can't get effective guidance. But leading agent clients like Manus, Kunlun Tech, and Shenyan Technology constantly make demands — multimodal capabilities, full-text extraction, Markdown output — and these are the key factors driving continuous product improvement. That's the third competitive moat.

👩🏻 Ronghui

The strategic choices you mentioned earlier are quite interesting too. One was anticipating Microsoft's likely moves, the other was anticipating the agent explosion. Could you talk about what clues led you to predict these things?

👦🏻 William Du

Microsoft's adjustments to Bing played a critical role in our thinking. If Microsoft hadn't stopped API services to Chinese LLM vendors, we merely would have felt that geopolitical competition might affect future business models and availability. But Microsoft's decisive, absolute shutdown of many search API interfaces made us realize this could happen quickly, immediately, right now.

We also did our homework and believe they had internal considerations outsiders wouldn't know about:

The first consideration is that historically, providing ToB APIs was part of the ecosystem — giving APIs to Yahoo, browsers, or major vertical clients, with Microsoft standing behind the product, not viewing these products as challenging its search ecosystem or portal status.

But over the past year, as products like Perplexity and some Chinese AI search offerings emerged, Microsoft discovered that search APIs were actually helping emerging AI industry search competitors obtain the best infrastructure at very low cost, thereby threatening its portal position. This is also a major battleground we're seeing in China. If in the future users interact with various agents (chatbots, etc.) to make decisions and search, today's portal apps will lose their portal status and may be reduced to APIs behind these agents.

For example, when a user wants a Tokyo travel guide, the agent will call our search API to get relevant information, and the high-quality content comes from Xiaohongshu, Trip.com Group, Mafengwo, and others — but they will no longer directly face user demand, merely serving as information sources. This is what giants don't want to see, and what Microsoft doesn't want to see. Hence they shut down search APIs to slow competitors like Perplexity from challenging their portal position.

The second consideration: if search APIs hold such strategic value for obtaining real-time information within agents, why wouldn't Microsoft integrate them with other products? Some major clients have already received Microsoft's suggestion: to continue using the Bing search API, they need to migrate cloud services to Microsoft Cloud; or instead of providing a native API, bind it to Copilot or large models and output analysis results — this may become their main next strategy. Meanwhile, price increases are inevitable. Over the past two years, they've already raised prices substantially twice, and now search APIs combined with model capabilities have risen to $45, compared to just $5 two years ago.

👦🏻 Koji

I'd like to interject here — Google doesn't offer a similar API, right?

👦🏻 William Du

Right, strictly speaking, Google doesn't offer a ToB Search API. They have Custom Search for small-to-medium developers, but with very low usage limits that can't be used in B2B commercial scenarios. Globally, there were roughly 20 billion searches per day, with only 2%-3% happening in B2B, and the vast majority of that share was captured by the Bing API.

Our judgment is that of these 20 billion daily searches, 80%-90% will shift from C-side search to agent search. Agents will break queries apart and search across multiple rounds, causing search call volume to grow by an order of magnitude of 10x. That's why we're firmly committed to investing in multilingual, multi-capability agent search.

👩🏻 Ronghui

Do you think providing one-stop services for agents will become a market that similar companies all rush to capture this year?

👦🏻 William Du

I think this year is a very good competitive opportunity.

Currently among overseas competitors, after Bing's retirement, only 1-2 companies are doing similar things. One transformed from a privacy browser and now provides API services in the AI era. The other is a Chinese team backed by US VCs, but without major clients, serving developers as their starting point.

In China, competitors are mainly large companies that previously had ToC search capabilities and will also provide search API services. But we feel we're not entirely in competition with them right now — these large companies' primary goal is maintaining C-side portal status; B2B is just a side business. And doing optimization and feature completeness for agent clients will become increasingly important, because a simple Web Search API can no longer satisfy general or vertical agent needs.

So our current competitive position is quite good. One, large companies won't go all-in on this this year. Two, we're already serving leading AI agents across verticals — currently over half of China's top AI-native applications have become our clients. So as we're meeting client requirements, we're continuously driving product progress and iteration. That's a very good state to be in.

👩🏻 Ronghui

What you're describing fits the classic early-stage startup playbook: catching the strategic window when large companies haven't fully covered their side businesses, where they haven't yet invested that much in resources. But that's also the challenge — how do you do this better than the large companies?

👦🏻 William Du

On foundational services, we need to match the giants — that's our requirement for the search team. On relevance, timeliness, authority, latency, availability, and other content and search quality metrics, we need to reach parity with Bing. That already puts us ahead of most major companies' search API quality, and we're even surpassing them on some metrics.

Beyond that, our competitive advantage comes from our understanding of agent needs. The peripheral search capabilities demanded by leading agents are important drivers of our product improvement and refinement. If competitors aren't serving these top agents, they can't conjure up search capabilities that surpass a basic web search API. This is our key path to gradually building moats — by continuously improving and enriching search functions critical to agents, we maintain sustained leadership.

👦🏻 Koji

These agent clients you're working with — their search requirements sound fairly straightforward: comprehensive coverage, real-time results, authoritative sources. Are there any agent-driven search needs that people wouldn't expect?

👦🏻 William Du

I think that's a great question.

The first agent need to address is globalization. The vast majority of Chinese entrepreneurs today don't want to be limited to the China market — they're treating the global market as their AI agent playground. To go global from day one, you need multilingual capabilities, especially for the top dozen or so languages. Doing business in South America requires Spanish and Portuguese; in Eastern Europe you need Russian and Arabic; everywhere else you need English at minimum. Without multilingual capabilities, these emerging agents won't even consider your service.

Second is whether your data sources and data residency are compliant. In many developed markets, the requirements here are stringent — data localization is mandatory. We have a major advantage on this front because Xiaosu's affiliate company is actually the world's second-largest CDN provider, with 2,800 available nodes globally. These nodes don't just have traditional compute and robust network connectivity — they can also host GPU compute. So we can help clients meet compliance requirements very effectively.

I think these two factors are the primary considerations when choosing a search API under the globalization wave.

Beyond that, on functionality — things like text-to-image search or image-to-image search, whether these can be conveniently solved via API; when crawling full text, whether you can distinguish real content from ads and related recommendations; whether the crawled formats meet clients' needs across different scenarios. These are capability gaps to fill, but the prerequisites are multilingual support and data compliance.

👩🏻 Ronghui

Beyond what you just mentioned, having worked with so many agent companies, have you noticed any other latent needs?

👦🏻 William Du

I think latent needs will gradually emerge as agents satisfy more product scenarios. But in reality, we can't "define needs ourselves" — our needs come entirely from what clients demand as they're achieving PMF and refining their products and processes. So I'm not at all worried about search functionality improvements. I think today is year one of agents — Koji proposed this earlier, and I completely agree — but I'd add that we're actually only at 6:05 AM on day one of year one. The sun has just risen.

If you compare leading native applications to the mature mobile internet era, looking at DAU, we're at less than 1% of true internet prosperity. Leading general-purpose agents have DAUs in the hundreds of thousands; Tier 2 or vertical agents have DAUs in the tens of thousands. Compared to the mobile internet boom, we're not even at 1% — more like one-hundredth or one-thousandth of that state.

So from a search functionality perspective, this will definitely continue evolving. Just serve the leading agents well and keep pace with them. This is also why we're building one-stop agent infrastructure. We've found that many clients — even well-known companies that have been on Koji's show and received funding from top VCs — might have only 10–20 people total, with just one person actually responsible for infrastructure.

Take my good friend and our client Mindverse: their main product is personal digital avatars or AI avatars, the product is mature and iterating fast, but they have just one person handling AI infrastructure. Through us, they can one-stop call various large model APIs — they don't even need to manage billing themselves. We provide daily breakdowns of all model and search usage and costs, which helps them enormously. And when a particular large model goes down, we automatically route model calls to other APIs that are up and can meet the need — clients don't feel any impact from the outage.

This is a typical scenario: from day one, clients can focus on product and go-to-market, outsourcing backend services and infrastructure needs to us. This significantly boosts their efficiency and probability of success.

👦🏻 Koji

Actually, when ChatGPT first emerged, there was also an AI infrastructure wave, with many companies doing what Skyrouter.ai does today. I remember one company called Martian — besides API aggregation, they told people they could do "optimal routing," judging which large model to dispatch to based on the task for the best answer. But later it seemed people didn't really need this automatic optimization — they ultimately preferred to choose the most stable model with consistently best performance.

Now you're also doing this kind of service, and domestically SiliconFlow seems quite formidable too. When clients ask, how would you describe your advantages and differences compared to them?

👦🏻 William Du

There are actually two major judgments behind this, which also responds to that company you just mentioned — we followed them at the time too, but they were slightly early. When we started this last year, we had two main judgments:

First, we believe Chinese and Americans will be the main players in the global AI industry. Beyond their home markets, everywhere else is a playground for entrepreneurs from both countries. The first step is finding ways to well serve Chinese agents and entrepreneurs going global.

Second, we judged that 2024's compute structure would still be training-dominant over inference, but from 2025 this would invert — inference demand would significantly exceed training demand, and inference compute would mainly manifest as APIs rather than direct compute leasing. This determined that we'd build a product like Skyrouter.ai, rather than something like Fireworks or SiliconFlow's product form. Because we believe that among open-source and commercial models, most clients will choose the best model for their current scenario, and commercial models will long-term outperform open-source models in some domains. Since these clients need to use multiple APIs from commercial models, "one-stop service" is what many entrepreneurs need from day one.

We actually have some cooperation with SiliconFlow, but there are two significant differences:

Their product sales mainly focus on self-deployed open-source models like DeepSeek. Their core competitiveness lies in optimizing open-source model utilization on compute — getting more tokens from the same compute. Their technical strength is in inference optimization.

Skyrouter.ai's competitiveness comes mainly from three aspects:

• Sufficient nodes and distributed compute resources globally;
• Platform stability and automatic routing capabilities (model capability is determined by the model itself; we focus on platform-layer stability and intelligent scheduling);
• Resource operation capabilities and discount advantages, stemming from our close cooperation with multiple cloud platforms.

AI Infra Industry Analysis

👩🏻 Ronghui

As we discussed, after ChatGPT emerged, AI Infra did have a wave of entrepreneurial enthusiasm. I remember reading a Sequoia report — people originally expected 2023 to be the year of AI application explosion, but later statistics showed most funding flowed to AI Infra, leading to a proliferation of companies in this space.

Could we ask William, as an industry insider, to share your overall observations on the AI Infra sector? In your view, what major changes have occurred in AI Infra over these past two and a half years? And at this point in time, how are AI Infra companies evolving?

👦🏻 William Du

I think the biggest change is everyone shifting from "grabbing GPUs" to "grabbing data." The earliest wave of infrastructure vendors, including CoreWeave and Lambda Labs, were really competing at the IaaS layer — providing foundational services from data centers to GPU cloudification for large companies, enabling them to train models quickly. This was phase one, where the competitive point was compute resources — everyone was competing on who could provide compute more efficiently and at greater scale.

From last year to this year, we've observed more agent vendors emerging. Their biggest difference from large companies is that they don't just need IaaS — they want infrastructure that can provide full PaaS-layer services. The PaaS layer mainly contains several components:

• Model APIs themselves
• Real-time data and information that models can't cover
• Tool calling and multi-agent collaboration capabilities

PaaS-layer competition suits us better. The previous generation of PaaS had many product forms, with one major category being CDN, and Xiaosu's team has accumulated substantial technology and capabilities in this area — capabilities that transfer directly to today's AI PaaS scenarios. So what we're currently providing, Skyrouter.ai and Xiaosu Intelligent Search, are both PaaS-layer products. At the same time, we're very clear that we don't do ToC, don't do SaaS, don't compete with clients for business — we let clients avoid managing underlying IaaS, maximizing their efficiency.

The biggest challenge right now isn't insufficient infrastructure capability, but that clients — the entire agent industry — are still in the process of finding PMF. I hope clients find PMF quickly; once they do, usage will immediately grow. This is also a good industry observation point. Just like when we did CDN in the past — whoever's traffic grew, their usage grew.

So our expectation for the entire industry is: leading agent vendors can find PMF quickly and effectively, so that demands on infrastructure capabilities and needs will also continue rising rapidly.

👦🏻 Koji

Do you see any signs of that now?

👦🏻 William Du

I think people have mentioned many scenarios. Among the scenarios that have truly worked out, coding is a very clear one. And coding itself also calls search. My understanding is: whatever the scenario, as long as AI is assisting or replacing humans, people's work habits will definitely be searching while writing code, searching while making PPTs, searching while stuffing in images. So the key scenario we're seeing is helping users improve efficiency in this "search while working" pattern.

The second scenario is office software. Take our client Kunlun Tiangong, for example — it's a typical product trying to use AI to replace Office. Whether writing documents or making PPTs, it simulates real human workflows: constantly calling search, guiding path planning, finding images, organizing content, and completing output. The process looks a lot like a high schooler making a PPT. As model capabilities improve, it will eventually reach the point where working professionals can directly deliver usable PPTs. Throughout this process, both model and search call volumes will grow rapidly.

Beyond coding and office work, several vertical scenarios have already found decent product-market fit:

• AI ad placement
• AI travel route planning with automatic booking
• AI information retrieval (for instance, our client Shenyan Technology DeepLang, founded by an alumnus of mine, does customized content推送 similar to an "AI agent version of Toutiao," requiring real-time抓取 of the latest full Chinese and English materials)

All of these show AI rewriting and optimizing traditional internet products, though not yet to the point of complete disruption.

👩🏻 Ronghui

So what changes do you think will happen in the AI Infra market landscape, short-term and long-term?

👦🏻 William Du

I don't think there's a definitive answer to this yet. But I believe the most important thing in the industry is this: AI-native applications need to achieve a qualitative leap in user scale and retention. To get there, two driving forces matter:

• Continuous improvement in model capabilities
• Continuous improvement in founder capabilities (right now, the best founders are basically all in this space. Once foundational capabilities keep improving, more scenarios will find better PMF within six months to a year)

As an infrastructure provider, what I need to do is ensure platform availability and provide one-stop, cost-effective services for Chinese agent companies going global — removing barriers in language, market, and culture. I'm very confident this day will come soon, and that Chinese founders will make up an enormous share of it. Just like in the mobile internet era, when China was streets ahead of the United States on the application layer. The AI agent era won't be fundamentally different.

👩🏻 Ronghui

What about long-term?

👦🏻 William Du

Long-term, the same logic extends: the importance of compute competition will decline, because everyone's compute will reach a baseline level. Including domestic compute — there won't be a significant gap with NVIDIA on inference. Inference will become the primary compute usage scenario, while training's share will keep shrinking.

What we're observing: data will become more important, especially real-time, dynamic data. Model training has a 3–6 month cycle. If you want qualitative change, the evolution speed won't be particularly fast, so you need real-time data to work with it. The compliance of this data, whether its sources can remain available to Chinese vendors, and how vertical-domain data (finance, news, healthcare, etc.) can be obtained and better integrated with agents — these will all be key directions for data-driven AI Infra going forward.

👦🏻 Koji

Can you introduce which search providers the big model companies use? Like OpenAI, Claude (Anthropic Claude), or domestic ones like Moonshot AI, MiniMax — do they use someone else's API search, or build their own?

👦🏻 William Du

Top model companies will definitely build their own search, but not all of it. Because search has very strong query long-tail effects. Head manufacturers typically build their own core query indexes — say, the top 100 million or 300 million queries, which might cover 90% of total query volume. The ROI on building this portion themselves is very high.

But the remaining 10% of long-tail queries are scattered across billions of entries. This portion typically gets outsourced to third parties. This is a very significant opportunity for us. We're already working with or testing with some top model companies. For niche languages or niche queries, they will definitely use third-party calls, because the ROI of building it themselves is too low.

👦🏻 Koji

Sounds like if you only do the long tail, the server and algorithm compute requirements are both very high. Can you still make money?

👦🏻 William Du

That's the other side of the coin. The first side is that top model companies will definitely build core indexes themselves. But the second side is: I believe AI agent proliferation will be highly personalized and diversified.

Top model companies won't monopolize the agent market. Over the next 3–5 years, it will definitely be a blooming array of agents that take center stage, because personalization and verticalization needs are so strong. These agents are our real target customers. While model companies have high demands on us but low margins, the capabilities we build serving them will help us better serve the large volume of general and vertical non-model-company agents. That's actually the main part of our business — it's just that this ecosystem is still in the very early morning of day one.

Personal Career Choices and Entrepreneurship Lessons

👦🏻 Koji

Can we talk a bit about career choices? I feel like friends around me who ended up in secondary markets all seem to think that's the final destination for their careers. But I didn't expect that after achieving some success in secondary markets early on, you'd return to the trenches of entrepreneurship. What was the thinking there?

👦🏻 William Du

Actually, the endpoint in secondary markets is often starting your own fund, and starting a fund is also a form of entrepreneurship. For me, it was comparing the ROI of starting a fund versus doing AI entrepreneurship — which is higher? Which am I more passionate about?

I was actually deeply influenced as an undergraduate. That was my first internship, working with Koji at Jiepang, during sophomore year. I was profoundly influenced by Koji back then — I found entrepreneurship fascinating, and that it could create new things. Because investing is more about resource allocation, while entrepreneurship is about manufacturing new productive forces and resources.

In my senior year, I did my first startup — a campus dating website called "Ting Yuehui." We promoted it at Beijing universities in 2011 and accumulated about 50,000 users. We raised some investment at the time, but shut it down when the core founders graduated. We saw Miliao, WeChat, and Momo rising. We were early in doing mobile web plus app, but also realized the scenario would get eaten by these head apps. As a user myself, I wouldn't use my own webpage to arrange meetups anymore. We also saw our capability ceiling. As CEO, I didn't know what resources were available to call upon, or whether our product choices were correct, or how to validate choices. This was the prequel to my first startup ending without conclusion.

Later, working at Hillhouse, the TMT team I was on did both secondary and primary market projects. Before the primary market team was fully spun out, we also looked at quite a few primary deals. I mainly covered entertainment — gaming, media — and looked at global markets. I believed secondary and primary markets weren't separate. For example, HUYA Inc. spinning out from YY was a very successful case I participated in. HUYA entered new markets with parent company capabilities — this is similar to our current relationship with Xiaosu and sibling companies. We're coming out with global distributed infrastructure resources to build infrastructure products and services for the AI era.

When I saw this opportunity, I was incredibly excited — because we had foundational capabilities, could ride the AI wave, and could work with an excellent founding team. With these conditions in place, I was still full of passion for entrepreneurship. I didn't think too long — just a day or two before deciding to go full-time.

👩🏻 Ronghui

Did your life change dramatically?

👦🏻 William Du

I think the change was earth-shattering. As a fund manager, most decisions were relatively short-term — looking at one month, three months, or six months P&L. You needed to judge whether a company, especially a mature company, had the right strategy, what product numbers were changing, how to guess the company's future direction. Most things were out of your control.

The uncertainty in entrepreneurship isn't about whether you can do it, but what results you can get from doing it. My agency and certainty in execution are very high — as long as I judge clearly, I can invest resources and team to do it. Success isn't fully controllable, but beyond thinking strategy, you need to build teams and execute, with massive granular work in between. There's friction between people and between tasks — this process is both draining and interesting. Building something from zero is incredibly dopamine-stimulating for me. Every day there's adrenaline and dopamine flowing.

When investing, you only get that hair-standing-on-end, goosebump-inducing tension during earnings season. After starting a company, you frequently get that aha moment when landing a big client or doing a major product iteration. You encounter this in investing too, but very rarely — in entrepreneurship, it happens constantly.

👩🏻 Ronghui

And the shift in posture when talking to projects as an investor versus talking to customers as a founder must be quite large too.

👦🏻 William Du

Ronghui is absolutely right. After starting a company, I realized how shallow I was as an investor. When investing, you can only listen to CEOs or CFOs talk about what they're doing — you're miles away from actual business collaboration. Though relatively objective, you still carry a certain arrogance, always feeling others' business models aren't good enough, wondering why CEOs and teams don't adjust according to ideal strategy. But after entrepreneurship, I'm filled with reverence for business.

Businesses that survive in the market are already one in ten. Making it to public company, even if secondary markets think the business model is poor or the founding team isn't strong enough, is incredibly rare. I first need to become a qualified CEO, and I'm still some distance from that — let alone a great CEO. This gives me strong humility.

👩🏻 Ronghui

Earlier we talked about strategically correct choices from a results perspective — like Microsoft, the timing of agent explosion. You mentioned entrepreneurship requires massive strategic judgment, though you did strategic judgment in investing too, but the feeling of bearing consequences is different. Can you share what worries keep you up at night as a founder?

👦🏻 William Du

I think the biggest difference in doing wrong things is bearing different consequences. When investing, especially in secondary markets, your probability of being right or wrong is roughly 60% — if you exceed 51% correctness, that's already quite good. Bearing results usually means stopping losses and moving to the next decision.

But entrepreneurship is completely different. If a strategic decision is correct, it may guide the company toward long-term success. If wrong, it produces massive consequences — including wasted manpower and resources, and how to handle the mistaken business. My personality is relatively mild, so wrong decisions generate massive conflict — layoffs, cutting businesses and budgets, requiring massive communication and coordination, draining energy.

For a CEO, energy is the most precious resource. It needs to be allocated across strategic thinking, team building, execution, and other things. One wrong decision may consume massive energy for a period on cleanup execution — this is very real pain. So I believe in thinking more, testing small-scale first, then gradually amplifying and validating.

These days, I actually have fewer sleepless nights. When I was investing, I had to track multiple markets — China, Japan, and Korea during the day, Europe in the evening, and the United States at night. Not being able to fall asleep was common and almost rhythmic, especially during earnings season when I couldn't go to bed early. I'd need to watch the pre-market and after-hours moves of companies reporting earnings and make trades. That brought tension and excitement. For example, if 10% of the portfolio was in a company reporting the next day, and not all earnings outcomes are clearly predictable, I'd be unable to sleep. This cycle repeated every quarter. Good fund managers can reduce that proportion, but they can't eliminate it entirely.

After starting a company, I actually sleep better. Every day I just need to lie in bed and reflect on the gains and losses of recent strategy and execution, make short-term adjustments the next day, and discuss longer-term problems when needed. Now I mostly go to bed before 11 p.m., start work early, and don't open my computer after 9 p.m. — I only do thinking and communication work. Overall, the pressure is more concentrated during the day, and my rest is more regular.

👩🏻 Ronghui

I feel like this is pretty rare.

👦🏻 William Du

I don't think anxiety helps. Maybe many of the founders you've talked to before either came out of big tech as leaders or after running a business unit, or went straight from one startup to another. Their state is still somewhat different from the kind of long-term tension I experienced in secondary markets.

The second thing is that my goals became longer-term. Primary market investing requires a relatively clear long-term view, but in secondary markets, even when you have a long-term view, short-term volatility still has a huge impact on performance. After starting a company, the long-term goals became clear — build the team well, and the rest is daily tactics and execution. This actually makes it easier for me to sleep well.

Although whether a particular client can be won makes a difference, the beauty of ToB business is that even if you don't win them today, if you keep improving the product, understanding client needs, and building better relationships, you still have a chance to win them next time or next month when you talk to them. So I don't feel anxious about this. If a deal doesn't close, or a client isn't using our service for now, I don't get anxious about the result itself. Instead, I look back at what client needs we failed to meet during the testing process, what the product still needs to improve, and whether the client scenario is the highest priority problem to solve right now.

These are questions to think through when you're emotionally calm, not when you're tense and anxious. So I feel like this job suits me quite well.

👦🏻 Koji

Indeed. Looking back, I don't think we've ever had a founder on the show who came back to entrepreneurship after doing secondary markets. So one can imagine that having gone through the secondary markets experience before doing this long-term-oriented thing of entrepreneurship, plus Xiaosu Technology being ToB — which inherently requires being friends with time — your mindset and sense of balance are better compared to the past.

👦🏻 William Du

I don't think anxiety comes to me. Whenever I feel anxious, I think: right now, the biggest AI-native apps only have a few hundred thousand to a million DAU. They are the best people on this path, the best teams, with the most abundant resources. So beyond waiting for them to grow, waiting for the ecosystem to flourish, and simultaneously optimizing my own product and client coverage — anxiety is useless. The clients are already excellent enough. If there are still unsolved problems, those also require time and environmental iteration to resolve. So I can only keep a level head and focus on doing the present things well.

👩🏻 Ronghui

Is there any realization you've iterated toward after being in entrepreneurship for a while, that you wish you had known when you first started?

👦🏻 William Du

I think what I've learned most is that "strategic choices really matter." For example, our evolution from an API-based supporting product to building our own capabilities, then doing M&A during that self-building process, and the evolution of product form. Strategic choice seems abstract, but it actually truly determines a company's long-term fate. This is something I gradually came to appreciate and understand through the process — a very important thing.

Beyond strategy, another important thing I learned is: business itself is hard.

Doing business involves many elements. Beyond product quality and capability, it includes relationships with clients, what stage clients are currently at, and many other aspects. The vast majority of commercial behavior isn't a "0 or 1" judgment, but rather a process of gradually deepening from 0 to 0.1 to 0.2. So don't conclude too quickly. Keep trial-and-erroring in small steps, and understand needs through serving clients. This is also something I learned along the way. From a strategy perspective, we haven't made any major mistakes overall, so there aren't any big lessons. But from an experience perspective, I think these two points are critical.

Another point is to maintain an open mindset continuously, because the world is changing too fast. Compared to when I first entered the industry and joined the internet industry at Baidu, the speed of change is now at least ten times faster. Back then when we were doing the PC-to-mobile transition, company OKRs or KPIs were adjusted on a "year" basis. But our mobile cloud department at the time was adjusting strategy and OKRs on a "quarter" basis — we already thought that was fast.

Now our clients are iterating their product ideas and understanding on a "week" or "day" basis. Many clients, after transformation, see better opportunities or things more suitable for them to do within their own circles of competence. So we ourselves also need to maintain a very open mindset continuously, because this world is changing too fast.

👦🏻 Koji

Alright, thank you William for your time today. We also very much look forward to Xiaosu Technology having new moves in this massive transformation that's iterating and changing "by the day, by the week." If any listeners today are interested in Xiaosu intelligent search or Skyrouter.ai, you're also welcome to reach out to Xiaosu Technology to learn about their products.

Thank you again, William. Bye bye.

👦🏻 William Du

Thank you Koji, thank you Ronghui. Bye bye.

---

References

[1] SkyRouter.ai: http://skyrouter.ai/