Tencent bought Baidu's chips

In the past two decades, the Chinese Internet has been doing one thing: putting as many capabilities as possible into a single super company.

Tencent started with social media, then ventured into payment, gaming, and cloud computing; Alibaba began with e - commerce and incidentally got involved in logistics, finance, and the entertainment industry; Baidu focused on search and simultaneously expanded into mapping, autonomous driving, and large - model development.

The logic of that era was to claim territory. Whoever had the most entry points and users was the winner.

However, when several seemingly isolated news items are put together recently, the trend has completely changed.

Baidu plans to spin off Kunlunxin for a Hong Kong listing, with a target valuation of about $50 billion.

What does this mean? Currently, Baidu's total market value is only about $36 billion. The subsidiary is nearly 40% more valuable than the parent company.

Almost at the same time, T-Head Semiconductor under Alibaba also reported plans for an independent listing.

What's even more thought - provoking is a report from The Information showing that: Tencent has become an important customer of Kunlunxin.

Meanwhile, according to reports, Kunlunxin put forward an extremely tough condition during its roadshow: Want to subscribe for IPO stocks? First, promise to purchase chips worth 3 to 7 times the subscription amount. "Be a customer before becoming a shareholder."

Many people see this as a simple capital game. But in my view, these events all point to a change in the underlying logic of the Chinese Internet.

In the past, large tech companies built moats through closure and monopoly; today, they must rely on openness and division of labor to scale up.

01 Chips have turned from money - burning black holes into money - printing machines

Why did large tech companies make chips in the past?

The answer is simple: To save money.

Baidu needed to train search algorithms and large models, and self - developed chips were cheaper than buying from NVIDIA; Alibaba needed to support its huge cloud - computing infrastructure, and making its own chips could reduce hardware costs.

In the past, the chip department within the group was under R & D management and was a pure cost center. It was responsible for spending money but did not directly generate revenue. It was only natural for it to be self - sufficient within the group.

Today, the accounts have been recalculated.

The emergence of Agents has made the entire industry face a harsh reality: the most terrifying consumption in AI lies not in large - model R & D but in high - frequency inference.

Every AI response, every Agent task execution, and every code generation is devouring Tokens. At the bottom of Tokens are GPUs, inference chips, networks, and data centers.

When the user volume on the application side crosses the explosion point, API calls turn into real - money revenue. Thus, a previously unremarkable hardware R & D department within the company suddenly has an independent and attractive business model: The chips themselves are a highly profitable business.

According to public information, Kunlunxin's P800 has completed large - scale verification. Since 2025, it has delivered multiple clusters of ten - thousand - card systems and completed the training of Wenxin 5.1 on a fully domestic cluster. Its customer list has quickly expanded from Baidu's self - use to China Mobile, Geely, China Southern Power Grid, China Merchants Bank, and Tencent.

When chips change from a money - spending department to a money - making business, spin - off and independence have changed from a financial means to a strategic necessity.

02 Tencent's purchase of Baidu's chips is more important than the listing itself

In The Information's report, Tencent's appearance is the most dramatic detail.

In the past two decades, Chinese Internet giants have had little interaction at the infrastructure level. Alibaba's cloud services would never be sold to Tencent, and Tencent's technology would never use Baidu's underlying infrastructure. Each company has been reinventing the wheel in the most expensive way.

Today, Tencent has started purchasing Baidu's Kunlunxin chips. This marks the first time that the Chinese AI industry has begun to form a truly mature division of labor.

The infrastructure in the AI era is too expensive. Chip R & D has a long cycle, requires heavy capital investment, and has high talent barriers. If large companies only use the chips they develop within their own systems, the economies of scale will never be achieved, and the cost per chip cannot be spread.

Finally, they find that trying to make everything on their own results in making nothing well.

Therefore, the industry has started to show signs of change. Tencent's purchase of Kunlunxin chips means that leading players are starting to accept one thing: the best AI ecosystem does not require all components to be self - contained.

This is similar to the relationship between Apple and Samsung in the mobile phone industry. They compete fiercely in the terminal market, but the most core OLED screens of the iPhone still rely on Samsung's factories.

A competitor's purchase is the highest - level endorsement. When Tencent is willing to allocate part of its computing power base to Kunlunxin, it shows that domestic AI chips have passed the most rigorous real - world tests, and even competitors find them sufficient and safe.

03 The capital market is starting to re - price computing power

Kunlunxin was established in 2011. Why is it accelerating its listing process in 2026?

The answer actually lies not with Baidu but with the shift in the capital market's targets.

Five years ago, AI chips were just R & D investments in large companies' financial reports. Without a large enough commercial throughput, the capital market would never give high valuations to hardware companies.

Today, the situation has completely changed. NVIDIA, Samsung, and SK have all broken market - value records, setting a new benchmark for the global capital market: In the AI era, the biggest winners are often those who sell the shovels.

Especially with the explosion of Agents and multi - modal applications this year, the demand for inference has increased exponentially. The entire investment community has started to recalculate the value of AI companies. In the past, people talked about parameters and scores; today, they only calculate Token costs, inference efficiency, and data - center utilization.

For the first time, AI infrastructure has an extremely clear and viable commercial return model.

Kunlunxin didn't just start thinking about listing today. It's only now that the capital market has finally understood and is willing to give domestic AI chips a high enough price.

In fact, this has become a collective capital journey. T-Head Semiconductor under Alibaba has launched an independent listing; Cambricon has completed capital verification on the A - share market; domestic GPU companies such as Biren, Moore Threads, and Muxi have all gone public, and Suiyuan is also entering a new IPO rhythm.

In the past decade or so, the most difficult task for domestic chip companies was to prove whether they could make chips and whether anyone would use them.

Today, that stage has passed. What they need to prove now is something else: Who can become the domestic chip base for China's AI era, apart from NVIDIA.

04 The certainty war among global giants

Taking a global perspective, you'll find that everyone is doing the same thing.

Recently, OpenAI and Broadcom jointly released their first self - developed inference chip, Jalapeño. It took only 9 months from design to tape - out, and it is planned for commercial deployment by the end of 2026, targeting gigawatt - level data centers. (Extended reading: GPT designs GPT)

Why does OpenAI want to make its own chips? Essentially, the call volume from monthly active users is too large, and every call costs real money. Even if self - developed chips can improve performance per watt by just 20%, it can save billions of dollars a year. More importantly, they understand that their fate cannot be tied to a single supplier like NVIDIA.

Looking at other giants, Google's TPU has reached its 8th generation and is the most mature self - developed chip system globally; Amazon has Trainium and Graviton; Microsoft has Maia; Meta has MTIA.

All the global leading AI players have reached out to the most fundamental hardware.

The reason is straightforward. Inference cost is the largest single expenditure for AI companies. Whoever can reduce hardware costs can truly establish a viable business model. Moreover, when you control both the large - model architecture and the chip architecture, the software - hardware collaborative optimization you can achieve is a barrier that buying external general - purpose GPUs can never match.

Kunlunxin's IPO is just a signal, announcing that China's AI infrastructure is officially moving from the backyards of large companies to the public market.

05 The dimension of competition has completely shifted down

In the past two years, everyone has been discussing GPT, Claude, and DeepSeek. People habitually think that the parameters and scores of large models are the whole of AI competition.

But now, models are becoming more like operating systems. They are crucial, but they are no longer the only variable determining victory or defeat.

What really determines how much money an AI company can earn and how long it can survive are those more fundamental and boring hard indicators:

Who can minimize the cost of a single Token? Who has the most efficient inference cluster? Who can have a continuous and undisturbed computing - power supply?

In 2023, it was a war of models, and people were competing on who was smarter; in 2024, it was a war of applications, and people were competing on who was more useful.

By 2025 - 2026, the war has reached the most fundamental level. Global AI competition has officially entered a war of attrition in infrastructure.

06 Words beyond the page:

In the past, we were always used to viewing Chinese Internet giants from the perspective of empires and monopolies.

They were like black holes devouring everything, extending their tentacles to every corner and trying to complete all business closures within their own walls. It was a twenty - year period when large companies became larger and the ecosystem became more closed.

However, the spin - offs of Kunlunxin and T-Head Semiconductor, as well as Tencent's orders, have put an end to that era.

Behind this is not the decline of Baidu or Alibaba, but an inevitable fate: The AI industry chain is so huge that no single super company can swallow the entire infrastructure alone.

Models are iterating every day, and applications are reshuffled every three months, but the underlying chips, networks, and data centers, once built, will be the industrial foundation for a decade.

Today, it may seem like Baidu and Alibaba are "breaking up their families," but in fact, the Chinese Internet is completing a twenty - year decoupling.

In the Internet era, large companies became larger; in the AI era, large companies are becoming smaller.

They are starting to release their capabilities to form a larger industry.

This article is from the WeChat official account "Beyond the Page," written by Huahua and published by 36Kr with authorization.