With DeepSeek's meteoric rise to global prominence, the AI industry finds itself at a pivotal juncture, prompting a reassessment of its economic dynamics.

Traditionally, AI was perceived as a costly endeavor, accessible only to those with deep pockets. Yet, the DeepSeek R1 model boasts a training cost of just $5.5 million, a fraction (90%-95%) of OpenAI's flagship GPT-4, without compromising on performance. This revelation has Wall Street investors reeling, realizing they've overestimated the barriers to entry.

DeepSeek has shattered the mystique surrounding AI, compelling investors worldwide to reevaluate its commercial viability. Cost-effectiveness and profitability have emerged as key criteria for assessing AI applications.

If DeepSeek excels in cost-performance, the subsequent wave of AI products is poised to shine in profitability. In 2024, global AI products amassed a significant number of paid users: Microsoft Copilot with 55.74 million, Baidu Wenku with 40 million, Canva with 22 million, ChatGPT with 11 million, Claude with 4.17 million, Notion with 3.33 million, Perplexity with 1 million, Cursor with 660,000, and Character with 536,000.

Amid DeepSeek's global acclaim, China's major internet firms faced scrutiny. Critics questioned whether established tech giants, epitomized by BAT, could be outpaced by a startup. Yet, Baidu Wenku ranks among the top in terms of paid users globally. In an era prioritizing AI commercialization, what does this signify?

Have tech giants like Baidu truly been usurped?

Why is AI now 'talking about money'?

DeepSeek's impact transcends technical advancements, extending into the economic realm.

Over the past year, AI enterprises have grappled with escalating costs, even industry leaders like OpenAI. Data indicates that OpenAI incurred losses of approximately $5 billion in 2024, projected to surge to $14 billion by 2026, primarily due to the high costs of training large language models.

Wang Yu, founder of WuWenXinQiong, calculated the magnitude of computing power costs. Assuming GPT-4 Turbo serves 1 billion active users daily, the annual computing power cost could exceed RMB 200 billion, excluding model training investments.

Tech giants have also invested heavily in AI. From January to August 2024, Microsoft, Meta, Google, and Amazon collectively invested $125 billion in AI data centers, a mere fraction of the total expenditure.

Domestic large models face rising cost pressures too. According to the China Academy of Information and Communications Technology, large models' computing power costs have surged tenfold in a year. Competitive pressures intensify as OpenAI enters the Chinese market via Microsoft Azure, leveraging 'technological advantages + mature ecosystems' to crowd out domestic models. A domestic large model's customer churn rate spiked by 30% after GPT-4 opened its API.

The soaring operational costs of large AI models impede not only their development but also the overall AI industry's ecological growth. Notably, many small and medium-sized AI enterprises rely on large models via API interfaces. High training costs for large models inflate invocation costs for these enterprises, dampening their enthusiasm for AI and hindering ecosystem development.

Last year, some domestic large models attempted to attract customers by slashing prices, igniting a price war. ByteDance's AI large model, Doubao, announced a pricing strategy 99.3% cheaper than the industry average. Alibaba significantly reduced prices for its Tongyi Qianwen model, with the API input price of Qwen-Long dropping by 97% and the output price by 90%. Baidu made ERNIE Bot's two entry-level products, ERNIE Speed and ERNIE Lite, free, while Tencent Cloud adjusted the price of its Hunyuan-lite model to be completely free.

Leading manufacturers' price reductions for large models resemble network operators lowering data fees, alleviating the computational cost burden on development enterprises. However, price wars are double-edged swords. Startups in the large model industry face a more competitive environment, shrinking their market share.

Yet, this challenging situation is easing. DeepSeek's popularity is catalyzing a new transformation in the relationship between large models and computing power. Previously, large model manufacturers adhered to the adage 'miracles happen with great effort,' focusing on building larger models and deploying more computing power. DeepSeek has overturned this belief, demonstrating that exceptional AI large models can be achieved at low costs. This lowers the industry threshold, enabling previously marginalized players to enter the market and helping large companies escape low-quality price wars, facilitating easier commercialization.

Currently, major tech companies like Baidu, Alibaba, and Tencent have launched DeepSeek models. Alibaba Cloud's PAI Model Gallery supports one-click deployment of DeepSeek-V3 and DeepSeek-R1 on the cloud. Baidu Intelligent Cloud's Qianfan platform lists DeepSeek-R1 and DeepSeek-V3 models and offers ultra-low-price plans. Tencent Cloud announced support for one-click deployment of the DeepSeek-R1 model on its high-performance application service HAI.

Over the years, we've witnessed numerous technological commercial failures. It's not that these technologies lack potential but rather the absence of a viable business model that hinders progress. Now is the time for AI enterprises to focus on monetization and promote commercialization.

On the AI frontier, how do major tech companies pursue commercialization?

Any technology applied in business must generate revenue. Non-commercializable AI is unethical.

As more large models become open-source, the AI technology gap between major tech companies and startups narrows. As DeepSeek's founder Liang Wenfeng notes, establishing a technological edge over competitors in a short period is challenging for both. With OpenAI paving the way and based on public papers and codes, both can swiftly develop their large language models, presenting opportunities for both.

Many of these opportunities lie in commercialization.

For major internet companies, nearly all products warrant reimagining in the AI era. This process continually unveils vast business opportunities. In 2023, Li Yanhong expressed that among Baidu's restructured businesses, Baidu Wenku's progress was most satisfying. Last September, Baidu integrated Baidu Wenku and Baidu Netdisk. According to the latest data from Baidu AI DAY 2025, Baidu Wenku's AI features surpassed 90 million monthly active users, with over 40 million paid users, ranking second globally after Microsoft Copilot. Moreover, Baidu Wenku's paid rate surged by 60% year-on-year, primarily driven by AI features.

Under large model reconstruction, Baidu Wenku transformed from an office tool into a 'one-stop AI content acquisition and creation platform,' introducing hundreds of AI capabilities, including smart PPT, smart writing, and AI web search, covering search, creation, and editing. For instance, the smart writing feature aids users in generating logically clear reports, summaries, or plans through keywords, allowing outline, text modification, and writing expansion anytime, seamlessly integrating into users' work and study creation processes. Thus, Baidu Wenku holds a clear lead in China's smart PPT market share.

One of Baidu Wenku's direct competitors is Kingsoft Office's WPS AI.

Launched in November 2023, WPS AI targets three strategic areas: AIGC (content creation), Copilot (intelligent assistant), and Insight (knowledge insight). Kingsoft Office's third-quarter 2024 financial report reveals that WPS AI 2.0 drove its domestic individual subscription business to achieve RMB 762 million in revenue, a 17.24% year-on-year increase. During the reporting period, WPS AI 2.0 added functions like AI writing assistants and AI reading assistants, previously deploying in content creation and smart assistant scenarios.

In major tech companies' AI commercialization endeavors, results are as crucial as the process. When asked about the large model competitive landscape, Liang Wenfeng acknowledged that major tech companies have advantages but must swiftly apply them to persist, as they need tangible outcomes.

Alibaba is also infusing new life into its established businesses. Last year, Alibaba spun off its AI application 'Tongyi' from Alibaba Cloud and merged it into Alibaba's Intelligent Information Business Group. In January 2025, Quark, also part of the Intelligent Information Business Group, unveiled a new slogan: 'The AI All-rounder for 200 Million People,' clarifying its AI positioning. Alibaba is clearly strengthening its AI C-end applications layout. Currently, Quark, Tongyi APP, and Tmall Genie represent layouts in productivity tools, chatbots, and AI hardware.

In AI's C-end direction, Baidu and Alibaba share similar visions, adjusting organizational structures to connect and integrate businesses while activating previously high-quality products with AI.

Tencent's AI commercialization strategy can also be described as 'old trees sprouting new shoots.' At Tencent's 2025 internal employee conference, Pony Ma mentioned that Tencent would continue investing in computing power reserves, encouraging all BGs to embrace large model productization and landing scenarios. Products including WeChat, QQ, input methods, and browsers will launch AI agents, while games, WeChat Reading, and Tencent Video will explore more AI applications based on Hunyuan.

After competing on parameters, long texts, agents, and prices, domestic major tech companies now share a common front in AI – restructuring existing businesses and driving the commercialization process.

The Game Between Startups and Major Tech Companies

As major tech companies accelerate AI commercialization, AI startups face commercialization pressures.

Even leading startups with robust technology grapple with balancing costs and benefits. Kimi, the intelligent assistant from startup Dark Side of the Moon, has emerged as one of the hottest AI applications recently. Its first reasoning-enhanced model, k0-math, rivals or even surpasses GPT-4, developed in just two months.

According to APPGrowing data, since March 2024, Kimi's advertising investment has reached tens of millions of yuan monthly, peaking at RMB 220 million in October and RMB 200 million in November. However, Kimi's advertising effectiveness fell short of expectations. QuestMobile data shows that AI native applications like Kimi have a next-day retention rate of around 30%, lower than the industry standard of 50%-60%. Dark Side of the Moon must balance advertising investment, organic growth, and commercialization. Thus, last year, Kimi introduced a tipping function, exploring accelerated commercialization.

Liang Wenfeng candidly admitted in an interview that existing vertical scenarios are not in startups' hands, and this stage is unfriendly to them. Major tech companies have more resources in vertical scenarios and have already established a foothold, giving them a first-mover advantage. Post-AI transformation, many of their existing vertical businesses exhibit better user activity and paid conversion than AI native applications. This means startups must urgently seek differentiated application scenarios, with their survival space potentially further compressed.

Nevertheless, in commercialization, AI startups have advantages over major tech companies.

Liang Wenfeng noted that major tech companies' models are tied to their platforms or ecosystems, whereas DeepSeek is completely free. Simply put, major tech companies develop models based on their respective businesses, creating certain limitations. For instance, some cloud vendors' previous demands were scattered. It wasn't until 2022 that autonomous driving began demanding machine rental for training and the ability to pay, prompting cloud vendors to build infrastructure. Major tech companies rarely engage purely in research and training; they are often driven by business needs.

Startups like DeepSeek, starting from scratch, possess greater development potential.

In commercialization, AI startups and major tech companies are not in a 'challenger vs. gatekeeper' relationship but rather construct differentiated value networks leveraging their unique ecological positions. Their core advantage lies not in resource scale but in edge innovation efficiency and value reconstruction capability.

Conclusion

The commercialization of large AI models and native applications is akin to a grueling double marathon that traverses both the technological and market landscapes.

In the immediate term, enterprises must strike a delicate balance between scenario-specific focus and cost management. DeepSeek's rising popularity has undoubtedly exerted pressure on tech giants such as Baidu. Nevertheless, from a long-term perspective, successful commercialization hinges on groundbreaking advancements in foundational technologies and robust ecosystem synergy. Companies like Baidu are far from being out of the race; the game is still very much in play.

The true champions will be those players who adeptly translate technological superiority into distinctive commercial worth and construct a comprehensive, closed-loop industrial ecosystem.