In 2025, the global AI battlefield ignites with unprecedented fervor. Microsoft invests $80 billion to bolster the Azure computational fortress, while Amazon allocates $100 billion to target the Bedrock ecosystem. These colossal moves are destined to be etched in history.

According to Bloomberg Intelligence's latest data, global tech giants' cumulative AI-related capital expenditures will surpass $320 billion for the first time in 2025, marking a staggering 152% increase from 2023. In China, AI server procurement has sustained a year-on-year growth rate exceeding 50% for three consecutive quarters, with Huawei's Ascend 910B delivering over 20,000 units in a single month.

Clearly, AI armaments have become the focal point of this race. China's tech giants are not to be outdone: Alibaba invests heavily in AI infrastructure with $380 billion, Tencent's annual capital expenditure exceeds $100 billion, and Huawei's Ascend chip annual production capacity soars by 300%. This "trinity" arms race, comprising large models, agents, and computational power infrastructure, is reshaping the technology industry landscape with unprecedented intensity.

In this "new cold war of the intelligent era," computational power serves as the nuclear weapon. NVIDIA CEO Jen-Hsun Huang warns that "the next generation of AI requires 100 times more computational power," putting Chinese manufacturers under the spotlight to lead this fierce technological race. Behind the AI arms race lies a transformative wave reshaping human commerce and the noble mission of leveraging technology for the benefit of all.

This battle is crucial.

Vertical Large Models Seek the Next Ignition Point

Change happens swiftly. Vertical large models have long been heralded as the future trend, and indeed, that future is upon us. In the first quarter of 2025, the deployment of vertical industry large models in China surged by 317% year-on-year. The daily average number of calls to AI quality inspection models in the manufacturing industry exceeded 8 billion, and financial risk control models prevented over $20 billion in suspicious transactions in a single month. These numbers underscore a strategic shift by cloud vendors towards vertical tracks. As competition in general large models falls into "parameter involution," industry-specific models have emerged as the sexiest growth pole for AI commercialization, boasting gross margins of 45-60%.

Large companies have a keen sense of direction. The 2024 financial reports reveal that while the growth rate of general cloud services for Alibaba Cloud, Tencent Cloud, and Huawei Cloud has slowed, the growth rate of industry intelligent clouds has outpaced the overall market.

After a decade of AI development, when it became apparent that general large models couldn't address industry pain points, attention naturally turned to vertical large models. Conversely, the large-scale penetration of vertical large models into industries is inextricably linked to the arrival of technological inflection points. Computational power, multimodality, strong reasoning, open sourcing, accurate data, agents, and deep applications have become crucial trends in the development of vertical large models. Among them, two trends—one fast and one slow—are particularly noteworthy.

The "fast" trend refers to computational power. Models have always craved speed, and with the era of computational power explosion upon us, signs of declining computational power costs have emerged. DeepSeek, with its low-cost training of $5.57 million, nearly overturned the underlying logic of large model operations, shocking the industry. As the US continues to invest heavily in data center construction, chip procurement, and network establishment, it is invaluable for China's large models to collectively focus on reducing training costs.

However, while DeepSeek's emergence is commendable, it is also tinged with narrow-minded emotion. China still faces a supply bottleneck for high-end chips. Despite companies like Huawei, Suiyuan Technology, Moore Threads, Hygon, and Biren designing domestic chips comparable to NVIDIA's A100 single card, the road to mass production is long and arduous, posing significant risks for AI development.

In addition to the rapid pace of computational power, vertical large models are also pursuing "slowness." Currently, slow thinking has become a standard feature of models. Inspired by the DeepSeek effect, the evolution speed of large models globally has accelerated: OpenAI launched GPT-4.5 and the reasoning model o3; Anthropic unveiled the hybrid reasoning model Claude 3.7; and Google expedited the launch of multiple reasoning models.

Meanwhile, Tencent Hunyuan released the strong reasoning model T1, capable of responding in seconds, garnering considerable attention.

The continuous improvement of reasoning model capabilities, characterized by slowness, accelerates the arrival of the inflection point for model applications. Every "slow down" in thinking accumulates energy for a true intelligent leap. Moreover, once the evolution of reasoning models breaks through the bottleneck, models may possess other human characteristics such as self-exploration and reflective verification. If that day comes, it's not impossible for AI to attain consciousness.

Thus, the most fascinating aspect of AI emerges: everyone awaits the historic moment, yet no one can predict where the explosion point will be.

All Are Heavy Investments, but Path Disputes Have Emerged

The pie is growing. According to a report by the China Academy of Information and Communications Technology, China's cloud computing scale has grown at a compound annual rate of 38% from 2021 to 2023, and it's projected that China's cloud computing market size will exceed 2.1 trillion yuan by 2027.

How to divide this expanding pie has become the primary task for large companies, yet each company's path differs. For instance, Alibaba's approach is to create miracles through strength, announcing an investment of at least 380 billion yuan over the next three years to build cloud computing and AI infrastructure, focusing on supporting the application and deployment of large models, exceeding the total of the past decade. Such a massive investment stems from Alibaba foreseeing the dawn of AI.

According to the latest financial report, Alibaba Cloud achieved revenue of 31.742 billion yuan in the fourth quarter of 2024, a year-on-year increase of 13%, with quarterly sequential growth jumping from 7% to 13%, setting a new high in nearly 18 months. Its Tongyi Qianwen series of models drove AI-related revenue to maintain triple-digit growth for six consecutive quarters. The sweeping impact of AI and the explosion of public cloud business underpin Alibaba's heavy investment.

Another example is Huawei, which takes the route of innovating infrastructure technology. At the recently concluded Huawei Cloud Eco-Conference 2025, Huawei unveiled a bombshell: CloudMatrix 384 super-node. Utilizing a server composed of 384 Ascend computing cards, it not only pushes the single-cluster computational power to 300 PFlops but also achieves a single-card decoding throughput of 1920 Tokens/s, directly comparable to NVIDIA's H100.

Simultaneously, its ultra-high-speed interconnection built with 6812 400G optical modules enables data to flow nearly losslessly among the 384 cards, with training efficiency approaching 90% of NVIDIA's single-card performance.

As a result, the value of Ascend AI chips continues to soar. Don't forget that through integration with the HarmonyOS system, Huawei AI has achieved breakthroughs in the smart car field. The "end-edge-cloud" collaborative business model is more diverse, stable, and commercially imaginative compared to other giants.

Baidu's approach is closer to applications. Baidu's 2024 financial report shows that the daily average number of calls to the Wenxin large model reached 1.65 billion, a 33-fold increase from 50 million at the end of 2023. Among its applications, cooperation with State Grid in the energy sector to predict grid load through AI and improve power dispatch efficiency stands out. In the medical field, lightweight large models enhance the accuracy of CT image diagnosis, aiding tertiary hospitals in achieving large-scale deployment of AI-assisted diagnosis. These are commercial fruits borne from Baidu's application-centric approach. Behind this lies Baidu's trend judgment: AI can completely restructure product logic and business models.

Under this tide, no one dares to slow down. Tencent's AI-related capital expenditure in 2024 exceeded 76.7 billion yuan, a year-on-year increase of 221%, setting a new record. ByteDance's AI capital expenditure also reached 80 billion yuan. In comparison, Tencent has more advantages in the AI deployment of diverse sectors such as gaming and entertainment, while ByteDance's AI applications on Douyin e-commerce have a more rapid effect.

Behind the collective surge in AI investments by the five major players, although their paths differ, they all aim for the same goal. This vast and enticing AI market pie is naturally up for grabs for players as long as they have an angle to cut in.

A Crucial Year for AI Hardware Expansion, but Behind It Lies the Battle for Ecosystems

IDC states that AI is fully integrated into cloud terminal products and services, and 2025 will be the year when the industry internally and externally defines consensus AI cloud terminal products and market standards.

The fusion of cloud computing and local computing has sparked another surge in the terminal hardware market. Devices such as personal computers (PCs), clients, tablets, and smartphones are experiencing increased sales as long as they incorporate AI concepts. According to IDC data, the top five in overall cloud terminal market shipments are ZTE, Centerm, Sangfor, Ruijie, and Lenovo, with a combined market share of nearly 59%.

Meanwhile, IDC predicts that the overall cloud terminal market will continue to maintain high growth in 2025, with an expected growth rate of over 16%. By 2028, the size of China's cloud terminal market is expected to exceed 6.15 million units, with a five-year compound annual growth rate of 15.8%.

Why is AI terminal hardware sweeping the market at such a rapid pace? Besides business needs and AI rewriting business logic, a crucial factor is ecosystem development. Microsoft spent over 30 years building a moat under the Windows system; Google spent nearly 20 years making Android invincible. What lies behind this? Tens of millions of system applications, tens of millions of enterprise interconnections, and tens of millions of developer think tanks.

No one can predict how many years faster the building of an AI hardware ecosystem will be compared to the old era, but since the popularity of DeepSeek, one trend has become irreversible: open sourcing. This signifies that competition among enterprises has shifted from technological breakthroughs to ecosystem collaboration capabilities, with openness, sharing, and symbiosis becoming the survival rules of the new era.

Looking at the layouts of large companies, Huawei's HarmonyOS ecosystem has expanded to 7.2 million developers, and the Ascend AI cluster has united 1,200 enterprises to build a complete industrial chain from chips to applications. Alibaba Cloud's partner revenue has increased by 30%, with real money always being a solid foundation for cooperation. Tencent is even more generous, with the Hunyuan large model opening up to small and medium-sized developers, and 50,000 teams developing small games based on it, 30% of which have monthly revenue exceeding one million yuan. The practice of the "platform-developer" win-win model makes it a realistic goal to achieve benign interaction.

AI is ambitious, and single-point advantages are quickly overwhelmed in market competition. Therefore, the more top players there are, the more they are laying out ecosystem development and binding upstream and downstream partners, forming a symbiotic relationship necessary to become international giants.

AI Agent Snaps Its Fingers, and the Commercial Gears Accelerate

Enterprise-level AI is "transforming from virtual to real," but how many hurdles remain before landing?

Evolution of Large Models: Where Will the Landing of the AI Industry Head?

Li Wei from the China Academy of Information and Communications Technology: How does the accelerated integration of cloud computing and AI usher in a new era of intelligent computing?

[Original from Tech Cloud Report]

Please indicate "Tech Cloud Report" and attach this article's link when reprinting.