In recent months, the spotlight in the tech industry has increasingly focused on artificial intelligence (AI), especially with remarkable developments surrounding large language modelsOn February 17, Baidu Group experienced a significant decline of 6.94% in its stock price on the Hong Kong market, raising eyebrows among investors and industry observers alikeOnce seen as a leader in AI innovation, Baidu now finds itself at a crossroadsThe launch of its large language model, Wenxin Yiyan, aimed at challenging OpenAI's ChatGPT, has left many wondering if it can truly compete on the global stageDespite Baidu's history of anticipating technological trends, it seems that merely being ahead in strategy is no longer sufficient to maintain a competitive advantage.

It's not that Baidu isn't trying hard enough.

The end of 2022 and the beginning of 2023 marked a pivotal moment in AI technology, notably driven by Microsoft's OpenAI and its groundbreaking language modelThis moment proved that AI plays a crucial role in economic relations and decision-making processes globallyIt was during this time that Baidu, a giant in the Chinese internet industry, seemed to trail behind, much to the dismay of its supporters who had long championed its innovative capabilities.

At the start of 2023, Baidu unveiled Wenxin Yiyan 3.0, boasting a parameter scale exceeding 500 billion, which led to an optimistic forecast among investorsBy September, this model upgraded to version 3.5, showcasing a remarkable 35% reduction in training costsThis dynamic led to a 25% monthly increase in Baidu's stock price, displaying regained confidence in its AI venturesMorgan Stanley's 2024 AI industry report emphasized that Baidu's large model achieved an 18% market share in China, surpassing Alibaba's cloud services at 16%. By the same year, Wenxin Yiyan 4.0 launched with an extraordinary leap to 20 trillion parameters, rivaling the performance of OpenAI’s GPT-4 in several analysts’ assessments.

For Baidu, 2024 is poised to become the year of full commercialization of its large AI models

Advertisements

The quarterly financial reports revealed that Wenxin 4.0 Turbo effectively responded to increasing market demands, propelling Baidu's AI-related business revenues up by 14% year-on-yearSubsequently, Baidu's revenue from AI sectors steadily increased, with the popular Baidu Wenku AI application showing an impressive monthly active user base of over 90 million by December 2024, second only to ChatGPT globallyFurthermore, Baidu Wenku's paid user numbers exceeded 40 million, with a notable rise in the monetization rate.

When looking forward to 2025, Baidu's ambition in the AI sector still harbors many potential avenuesRecently announced was the decision for Wenxin Yiyan to go fully free starting April 1, while the 4.5 series of models will be completely open-sourcedDespite the anticipated stagnation in its core search engine advertising sector by the fourth quarter of 2024, analysts speculate that Baidu's cloud computing business, driven by its AI operations, is likely to sustain robust growthThis suggests a steady upward trend in overall revenue for Baidu, despite increasing market pressures.

However, lingering questions remain: why have these positive developments not translated into groundbreaking technological advancements or paradigm shifts within the industry?

As the Lunar New Year of 2025 approached, a lesser-known company, DeepSeek, launched its R1 model, generating global buzz almost overnightThis revolutionary model, which enabled the production of high-performance large language models at lower costs, utilized an open-source approach that resulted in a sharp decline in stock prices for dominant AI hardware suppliers like NvidiaIn the wake of this release, Wall Street reportedly lost $1 trillion in value on the first trading day.

Following this, Alibaba's Qwen 2.5-Max model emerged as another focal point in media discussions, aligning closely with DeepSeek's approach

Advertisements

Reports confirm that Apple chose Alibaba's AI model to develop AI functionalities for its Chinese end-usersAs of February, Alibaba's Hong Kong stock continued its upward trajectory, capturing market attention away from Baidu.

Throughout the period from 2023 to January 2025, major players in China's internet sector generally mirrored Baidu's strategiesThe focus on foundational models became evident, with firms striving to replicate OpenAI's advancements while launching competitive offeringsConcurrently, businesses sought ways to transform these advancements into revenue streams through various subscription and upgrade models.

Nevertheless, the previously dominant internet firms now face an unprecedented change—the Internet's paradigm shiftThis transformation evidently impacts their approach to AI, which could define their fates as they navigate this evolving landscape.

The Paradigm Revolution of Chinese Internet Enterprises

A personal observation made during a trip to Japan highlighted that most app and internet services there are operated by American companiesThis phenomenon extends to various digital domains such as social media, e-commerce, job hunting, and news portalsDespite the significant success achieved by the Chinese internet sector during its earlier stages, exemplifying the rise of well-established, locally-adapted applications, the tide is beginning to change.

In summary, the ultimate battle for Chinese internet enterprises will be fought in the realm of AICentral to attaining a competitive advantage in AI lies the concept of a "paradigm revolution." Historically, Chinese internet companies have successfully followed American tech trends and monetization strategies, recreating products that resonate with local consumers while maximizing user engagement and revenue.

However, although this strategy has previously proven effective, it also harbors significant pitfalls.

Technologically, many Chinese internet firms opted for a follower strategy, initially replicating American innovations before slowly introducing adjustments based on their understanding of domestic user preferences and internet conditions.

Yet, in the AI era, many leading American enterprises find that their strategic decisions may not yield successful outcomes

Advertisements

A number of these firms adhere to a theory known as "AI scaling law," which posits that only through investing substantial computational power and utilizing vast data sets can AI model performance be significantly enhancedThis reliance on significant financial investment in computational infrastructure explains the staggering stock price surges for companies like Nvidia and Broadcom, which supply AI chips.

In contrast, lesser-known companies like DeepSeek in Hangzhou have taken an alternative path, championing a "small model + big data" approachUtilizing techniques like the MoE architecture and multi-head potential attention (MLA), they've significantly reduced computational demands, curtailing per-query inference costs.

This disruptive innovation arises from DeepSeek's years of commitment to technology without chasing quick profits or uncritically mimicking American counterpartsAs traditional firms continue to follow OpenAI's blueprint in this way, they risk facing financial burden due to exorbitant infrastructure investments that yield little return.

On the commercial side, conventional internet firms prioritize trafficUpon reaching peak user levels, they either opt to invest heavily to eliminate competitors or begin monetization efforts to turn user engagement into revenue.

Nonetheless, this approach may not suffice in the AI era.

Reflecting on January 2025, how many people had heard of DeepSeek prior to its model launch? How many enterprises had adopted its open-source offerings? Yet a mere few weeks post-launch, DeepSeek has secured a dominant position in AI trafficThis surge can be attributed not only to its technological advantages but also its open-source business model.

The emergence of open-source models has fundamentally transformed the current commercial landscape for large models

DeepSeek-R1 matches the performance of premier models while maintaining costs significantly below industry averages due to its sophisticated and cost-effective technology.

According to DocsBot, uploading 1 million tokens to DeepSeek-R1 costs only 55 cents, while generating 1 million tokens incurs a cost of $2.19. For contrast, OpenAI's o1 outputs priced at an astonishing $60 for the same volume.

Musk's xAI model Grok offers services at comparatively lower rates than o1—$12 to $15 per 1 million tokens—but still far exceeds DeepSeek-R1's pricing.

Last year, Baidu set its pricing for Wenxin Yiyan 4.0 Turbo at approximately 0.03 yuan per 1,000 tokens for inputs and 0.06 yuan for outputsBased on a 3:1 input-output ratio, this marked a staggering 70% reduction compared to the general versions.

However, Baidu is not the first to announce free, open-source AI models.

Embracing the open-source model will undoubtedly draw in more users and developers, rapidly accumulating traffic and fostering a user-scale effect, thereby approaching a near "monopoly" in profitable modelsThe rationale is straightforward: a well-functioning product used by the masses will likely generate sufficient revenue through advertising or related services.

Furthermore, disparities in establishing open-source ecosystems could lead to strategic dividesA vibrant open-source environment with community developers will promote rapid growth of model visibility on platforms like GitHub, contrasting with companies adhering to paid models, who risk missing the essential window for building developer ecosystems.

Conclusion

By 2025, global investors will inevitably focus their attention on the Chinese market

Advertisements

Advertisements