Bilibili pours 1 billion RMB into AI development, what is it aiming for?

Hardly had Bilibili figured out how to make sustainable profits when it prepared to stake its entire annual earnings again.

In 2025, the company achieved full-year GAAP profitability for the first time, booking a net profit of 1.19 billion yuan. Several months later, the management announced that it would add roughly 1 billion yuan in extra AI-related capital expenditure in 2026.

This figure is almost equivalent to Bilibili's total net profit of the previous year. But in the current AI race, 1 billion yuan is far from a competitive amount.

ByteDance was exposed earlier this year to have further raised its AI infrastructure budget to over 200 billion yuan, and secured a $29.6 billion loan from nearly 30 banks in September, with a large share of the funds continuing to flow into chips and data centers. Tencent's capital expenditure in the second quarter reached 52.8 billion yuan, up 176% year on year, marking the first time its free cash flow turned negative in two decades.

Tech giants are spending hundreds of billions of yuan to buy their next ticket for technological advancement, while what Bilibili puts forward is only a tiny fraction of what they spend in a single quarter.

It is worth noting that Bilibili does not intend to join this most capital-intensive battle. Chen Rui, CEO of Bilibili, drew a clear boundary on the Q2 earnings call: "We will not participate in many segments along the AI industrial chain." The funds will only be invested in three directions: video understanding, video distribution and video creation.

For the AI industry, this sum of money is too small to secure technological leadership; for Bilibili, it is large enough to eat up most of the annual profit of this newly profitable company.

Bilibili neither wants to join the arms race, nor can it stay completely outside the arms race.

Thus the question arises: why does a content platform that does not build large models still have to spend so much money on AI? Is this sum of money paying for growth, or just buying the qualification of "not falling behind"?

01. What can 1 billion yuan buy?

1 billion yuan is not a large sum in the entire AI industry, but for Bilibili itself, it represents a notable upgrade in investment scale.

In 2023, Bilibili's capital expenditure on fixed assets such as equipment was about 180 million yuan, which rose to roughly 480 million yuan in 2024 and around 500 million yuan in 2025. By 2026, the planned extra AI-related capital expenditure alone will reach about 1 billion yuan, an increment twice as much as the full-year equipment investment in 2025. By the end of the second quarter, 70% to 80% of this budget has been implemented, mainly for purchasing servers and computing power.

This also means that Bilibili's AI investment in this round is mainly to supplement AI infrastructure for its existing businesses, enabling large-scale AI implementation in search, recommendation, content understanding and creation tools.

This is the direction Bilibili set as early as 2023.

In July of that year, at the peak of the "Hundred Models War", internet companies were competing for general foundation models, comparing parameters, training scales and model capabilities, and then launching the models as independent external products. Bilibili also released its self-developed bilibili index large model, but its goal was not to participate in this competition, but to make the model understand Bilibili's own videos and users, to serve internal scenarios such as search, content moderation and video summarization.

Image source / IndexTTS2 demonstration platform

At the Q2 earnings call one month later, Chen Rui made this route clearer: Bilibili will not participate in the "Hundred Models War". He took content moderation as an example -- if large models can replace a large amount of repetitive human labor, they already have practical value for Bilibili.

In the three years that followed, Bilibili's AI route has barely deviated from this direction. From the perspective of Wu Jiexi, founder of Haoye Technology, this is more of a realistic choice than active restraint. "It's already too late to start developing (video models) now, and Bilibili does not have the corresponding genes."

She introduced that training a video generation model from scratch and pushing it to commercial use does not only rely on capital. Models such as Keling and Seedance are not products of short-term investment, but are backed by long-term accumulation of teams, data and engineering systems. Even today, video generation models are still iterating rapidly on a cycle of several months. For latecomers, increasing the budget can buy computing power and talents, but it is difficult to make up for the already formed technological and engineering gap in a short time.

For Bilibili, a more cost-effective approach is to let upstream companies bear the most expensive model competition, and integrate their capabilities into its familiar video business after the technologies mature and the calling cost drops.

In the early stage, AI was mainly applied in the background for moderation, recommendation and intelligent subtitles, before gradually being presented to users and creators. In 2024, Bilibili launched the digital avatar tool "Bcut Studio"; its 2025 "AI Original Voice Translation" further integrated translation, speech synthesis, subtitle processing and lip sync simulation into one workflow, allowing Chinese videos to enter other language markets at lower costs.

Up to this year, AI has begun to expand from auxiliary tools to two core links: understanding content on one end, and generating content on the other.

Bilibili is dominated by medium and long videos. A video of dozens of minutes may contain characters, viewpoints, emotions and complex narrative structures at the same time. Traditional recommendation algorithms rely more on titles, tags, clicks and user behaviors, but can hardly truly understand what a video is about, let alone why users like it. Chen Rui mentioned on the Q2 earnings call that the new generation of models can gain deeper insight into video content and user intentions.

According to Bilibili's vision, if this goal is achieved, it will not only reduce the number of moderators or subtitle production processes, but may even reshape search and recommendation -- the platform can understand hundreds of millions of its videos more accurately, and match them to the right users.

The other end is content generation. In the past, an animation or a video with complex production usually required collaborative work of multiple people; now, a single creator can use AI to complete work that used to need a small team. There have already been cases on Bilibili where individuals used AI to produce animations that gained tens of millions of views.

In July this year, Zeng Ailing, who previously worked at Tencent Hunyuan and Anuttacon, joined Bilibili to lead the AI video generation business, reporting directly to Chen Rui. This means AI video business has become a priority directly followed up by the founder.

Therefore, for Bilibili, in the past it could test AI at low cost by integrating models into several marginal links; now when AI begins to enter the core chains of content understanding, distribution and production, Bilibili must allocate real computing power, servers and organizational resources for it.

02. 1 billion yuan can buy capabilities, but not moats

This route has already brought some quantifiable benefits to Bilibili.

In Q1 2026, Bilibili's advertising revenue reached 2.59 billion yuan, up 30% year on year; the figure rose to 3.13 billion yuan in Q2, up 28% year on year. In the second quarter, the comprehensive ad click and conversion indicator CTCVR increased by 19% year on year, and search ad revenue doubled. The management attributed part of the improvement to the enhanced AI capabilities in understanding content and user intentions.

AI companies themselves are also contributing to the revenue growth. In the first quarter, the advertising budget of the AI industry on Bilibili increased by 170% year on year, and maintained a doubling growth rate in the second quarter.

The same trend appears in the content sector. In Q2, Bilibili's average daily upload volume increased by 28% year on year, and the number of creators with more than 1,000 followers grew by 30%; an animation creation campaign launched in May this year gained more than 180 million views in three months.

From this perspective, Bilibili has indeed reaped the dividends of "not building large models": it does not need to bear the most expensive technological competition, and can share the efficiency improvement and new demands brought by model advancement.

But the problem also arises here.

Bilibili can purchase better video understanding capabilities, and so can Douyin, Kuaishou and Xiaohongshu. AI can improve Bilibili's ad matching efficiency, and will also improve the efficiency of other platforms at the same time. When models are increasingly becoming a general infrastructure, they will first raise the overall efficiency of the whole industry, rather than bring exclusive advantages to a single company.

In other words, Bilibili can spend 1 billion yuan to "keep up" with the trend, but it is difficult to achieve "leading" position relying solely on this sum of money.

Image source / pexels

AI is not only changing the internal efficiency of platforms, but also reshaping the competitive environment where Bilibili operates.

The first change lies in content supply. In the past, video was a content format with high production costs. Live-action shooting, animation, dubbing and editing all required time and labor, which naturally limited the supply speed. After the emergence of AI short dramas, manhua dramas and generative videos, the originally expensive and low-frequency video content has become cheaper and easier to produce in batches.

Hongguo is a direct example. Data from QuestMobile shows that as of July 2026, Hongguo Short Video's daily active users reached 168 million, exceeding the sum of the DAUs of iQiyi, Tencent Video, Youku and Mango TV. In February this year, Hongguo's average daily usage time per user hit 125 minutes; by contrast, Bilibili's average daily usage time, which hit a record high in Q1 this year, was 119 minutes.

Hongguo's growth cannot be simply attributed to AI. Its free model, ByteDance's traffic system and the IP supply from Fanqie Novel are still its more direct advantages. But AI is further amplifying the competitiveness of this model: as video production costs drop and supply speeds up, Hongguo can continuously expand its content pool at lower costs, and push new content to users relying on its existing traffic and IP system.

For Bilibili, the threat is not just that "there are more videos on the market", but that a competitor that already has huge traffic and content supply capabilities can use AI to increase video supply at lower costs and faster speeds, and continue to compete for users' limited daily viewing time.

"I believe no content platform will feel relaxed when seeing Hongguo's surging user data," Wu Jiexi said.

Competitive pressure also appears on the creator side.

AI can help UP owners improve production capacity. For a video that used to take several days or even weeks to finish, part of the production process can now be completed with tools for script generation, image generation, dubbing, translation and editing. For Bilibili, this increases the in-site content supply; for creators, it also reduces the cost of distributing content across multiple platforms.

Therefore, AI brings two simultaneous changes to Bilibili.

It improves the efficiency of recommendation, advertising and creation inside the platform, and also improves the content supply efficiency of the whole industry. The former change makes Bilibili willing to invest, while the latter change leaves Bilibili little choice not to invest.

According to Wu Jiexi, Bilibili's 1 billion yuan investment is more of a defensive project than an active offensive. This sum of money can only help the company "not fall behind", rather than build a new competitive moat.

03. A mandatory admission ticket

How much long-term cost will this defense require, and how much extra revenue can it bring? The answer lies in a more practical calculation: whether the increased cost for maintaining competitiveness can be covered by higher commercialization efficiency and content efficiency.

This calculation is not easy for newly profitable Bilibili.

The 1 billion yuan is mainly capital expenditure, which will not be recorded in the income statement at one time. But according to the statement from Fan Xin, CFO of Bilibili, on the earnings call, the investment in servers, computing power and related R&D is expected to increase the annual R&D expenditure by about 500 million yuan, accounting for more than 40% of the 1.19 billion yuan GAAP net profit in 2025.

On the other hand, by the end of the second quarter, Bilibili held about 24.3 billion yuan in cash, cash equivalents, time deposits and other funds. The 1 billion yuan accounts for less than 5% of this total. It will not threaten the company's cash security, but is enough to affect the profitability that the company has just established.

Fan Xin said the company will cut other operating expenses to offset part of the new costs, and its long-term gross margin target of 40%-45% and operating margin target of 15%-20% remain unchanged.

The capital market has also begun to recalculate Bilibili's AI investment based on this logic.

Earlier this year, UBS, JPMorgan Chase and Citi all listed the ad demand and efficiency improvement brought by AI as new growth drivers for Bilibili. After the release of the Q2 earnings report, Morgan Stanley and Bank of America maintained positive ratings, but lowered their target prices or profit forecasts. Morgan Stanley pointed out that R&D expenditure increased and gross margin faced pressure, while the advertising business also faced macro headwinds; Bank of America also referred to the increased R&D investment and slightly lower gross margin.

The market's concern has shifted from "whether AI is useful" to "whether the incremental value brought by AI can outpace its continuously rising costs".

Image source / pexels

If the annually increasing costs of servers, computing power and R&D can continuously bring higher ad conversion rate, better recommendation efficiency and more high-quality content, the investment will generate returns. But if AI eventually becomes a standard configuration for all platforms, servers and computing power will turn from "investment" to a kind of "tax" -- you have to pay the fee every year just to avoid falling behind.

A long-time investor focusing on internet platforms holds a cautious attitude towards how much incremental value the 1 billion yuan can create.

In his opinion, 1 billion yuan is not enough to help Bilibili gain real technological advantages. For content platforms, what will eventually widen the valuation gap is still user scale, community relations, commercialization efficiency and product capability, rather than whether they have deployed AI. As for why Bilibili still needs to invest, he believes there are also considerations at the capital market level: when AI has become a mandatory question for tech companies, complete absence itself may be regarded as backwardness.

In his opinion, many companies are now doing AI for AI's sake, and the capital market and companies are "cooperating with each other in a performance".

Wu Jiexi's judgment has both similarities and differences with his. Both of them agree that the model itself can hardly become Bilibili's exclusive asset. The real divergence lies in whether the same AI capabilities can amplify the inherent advantages of different platforms after being integrated into different platforms.

According to Wu Jiexi, Bilibili's advantages mainly come from two ends: one end is a large number of long-accumulated C-end content creators, and the other end is highly sticky ACG users. AI may make these two assets more valuable.

Generative AI is turning content production into a business for smaller teams. Many current AI manhua drama teams only have 1 to 5 members, and a small studio can even produce hit works. This