By 2025, the global AI battlefield will be ablaze with intense competition. Microsoft has invested $80 billion to bolster its Azure computing power, while Amazon has deployed $100 billion into the Bedrock ecosystem—both moves are historic and monumental.

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

Clearly, AI investment has become a buzzword in this tech race. China's tech giants cannot afford to lag behind. Alibaba has invested RMB 380 billion in AI infrastructure, Tencent's annual capital expenditure has exceeded RMB 100 billion, and Huawei's Ascend chip annual production capacity has soared by 300%! This "trinity" arms race, encompassing large models, agents, and computing power infrastructure, is reshaping the technology industry landscape with unprecedented force.

In this era's "new cold war," computing power is the nuclear weapon. When NVIDIA CEO Jen-Hsun Huang warns that "the next generation of AI requires 100 times more computing power," all eyes are on which Chinese vendor will break through first in this fierce technological race. Behind AI investments lies a wave of change reshaping human commerce and a sacred mission to practice inclusive technology.

This battle must be won.

Vertical Large Models Seek New Breakthroughs

Change happens swiftly. Vertical large models have always been heralded as the future trend, and indeed, the future is here. In the first quarter of 2025, the deployment of vertical industry large models in China surged by 317% year-on-year. The daily average invocation of AI quality inspection models in the manufacturing industry exceeded 8 billion times, while financial risk control models prevented suspicious transactions worth over RMB 20 billion in a single month. Behind these numbers lies a strategic shift by cloud vendors towards vertical tracks. As competition for general large models turns into a "parameter arms race," industry-specific models are emerging as the sexiest growth pole for AI commercialization, with gross profit margins of 45-60%.

Big tech companies have a keen sense of where the wind blows. The 2024 financial reports reveal that while the growth of general cloud services at Alibaba Cloud, Tencent Cloud, and Huawei Cloud has slowed down, the growth of industry-specific intelligent clouds has outpaced the overall market.

After a decade of AI development, people have realized that general large models cannot solve industry pain points, naturally turning to vertical large models. Conversely, the large-scale penetration of vertical large models into industries is inseparable from the arrival of technological inflection points. Computing power, multi-modality, strong inference, open source, accurate data, agents, deep applications, etc., have all become important trends in the development of vertical large models. Among them, two trends deserve special attention: one fast and one slow.

Fast refers to computing power. Models have always craved speed, and with the advent of the era of computing power explosion, there are already signs of declining computing power costs. DeepSeek, with its low-cost training of $5.57 million, almost overturned the underlying logic of large model operations, which was indeed shocking. As the United States continues to invest heavily in data center construction, chip purchases, and network setup, it is particularly valuable for China's large models to collectively focus on reducing training costs.

However, caution is warranted. While DeepSeek's emergence is commendable, it is also tinged with narrow-minded sentiment. Currently, China still faces challenges in the supply of high-end chips. Even though companies like Huawei, Suiyuan Technology, Moore Threads, Hygon, and Biren have designed domestic chips matching the performance of NVIDIA's A100 single card, the road to mass production is long and fraught with difficulties, posing significant risks for AI development.

In addition to the rapid pace of computing power, vertical large models are also pursuing slowness. Currently, one major trend is that slow thinking has become a standard feature of models. Stimulated by the DeepSeek effect, the evolution speed of large models at home and abroad has increased again: OpenAI launched GPT-4.5 and the inference model o3; Anthropic introduced the hybrid inference model Claude 3.7; and Google hastily rolled out multiple inference models.

Meanwhile, Tencent Hunyuan released the strong inference model T1, which can respond in seconds, attracting considerable attention.

The continuously improving capabilities of inference models, characterized by slowness, are accelerating 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 inference models breaks through the bottleneck, models may possess other human characteristics such as self-exploration and reflective verification. By then, it is conceivable for AI to possess consciousness.

Thus, the most fascinating aspect of AI emerges: everyone is waiting for the moment that will be recorded in history to arrive, but no one can predict where the explosion point will be.

Substantial Investments with Distinct Pathways

The pie is growing. A report by the China Academy of Information and Communications Technology shows that from 2021 to 2023, China's cloud computing market grew at a compound annual rate of 38%, and it is expected that by 2027, the scale of China's cloud computing market will exceed RMB 2.1 trillion.

How to carve up this expanding pie has become the primary task for big tech companies, but each company's approach differs.

Take Alibaba as an example. It is going all in, announcing an investment of at least RMB 380 billion over the next three years to build cloud computing and AI infrastructure, with a focus on supporting the application and deployment of large models, surpassing the total investment over the past decade. The reason for such a massive investment is that Alibaba foresees the dawn of a new era.

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

Another example is Huawei, which is innovating in infrastructure technology. At the recently concluded Huawei Cloud Eco-Conference 2025, Huawei unveiled the CloudMatrix 384 super-node. This server, composed of 384 Ascend computing cards, not only pushes the computing power of a single cluster 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 allows 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 rise. Through integration with the HarmonyOS system, Huawei AI has achieved breakthroughs in the intelligent automobile field. Its "end-edge-cloud" collaborative business model is more diversified, stable, and commercially imaginative compared to other giants.

Baidu's approach is more application-centric. Baidu's 2024 financial report revealed that the daily average invocation of the Wenxin large model reached 1.65 billion times, a 33-fold increase from 50 million times at the end of 2023. Among its achievements, the cooperation with the State Grid in the energy sector to predict grid load through AI and improve power dispatch efficiency, and its application in the medical field where lightweight large models enhance the accuracy of CT image diagnosis, helping top tertiary hospitals achieve large-scale deployment of AI-assisted diagnosis, are all commercial successes based on applications. Behind this lies Baidu's trend judgment: AI can completely restructure product logic and business models.

In this tide, no one dares to slow down. Tencent's AI-related capital expenditures in 2024 have exceeded RMB 76.7 billion, a year-on-year increase of 221%, hitting a record high. ByteDance's AI capital expenditures also reached RMB 80 billion. 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 shown more rapid effects.

Behind the collective increase in AI investment by the five major players, although their pathways differ, they all converge on the same goal. This huge and attractive AI market pie, as long as there is an angle to enter, players will naturally seize the opportunity.

A Crucial Year for AI Hardware Expansion: An Ecosystem Battle Looms

IDC states that AI is fully integrated into cloud terminal products and services, and 2025 will be the year when internal and external definitions and consensus AI cloud terminal products and market standards are established within the industry.

The combination of cloud computing and local computing has sparked another explosion in the terminal hardware market. Devices such as personal computers (PCs), clients, tablets, and mobile phones are experiencing increased sales when equipped with 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%.

Simultaneously, IDC predicts that the overall cloud terminal market will continue to maintain high growth in 2025, with an expected growth rate exceeding 16%; by 2028, the scale 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 demand and AI rewriting business logic, a crucial point 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 insights.

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

Looking at the layout of big tech 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 has gone even further, with the Hunyuan large model being opened up to small and medium-sized developers, resulting in 50,000 teams developing small games based on it, 30% of which have monthly revenues exceeding RMB 1 million. The practice of the "platform-developer" win-win model shows that benign interaction is not a difficult goal to achieve.

AI is too grand, and single-point advantages are quickly submerged in market competition. Therefore, the more leading players there are, the more they are focusing on ecosystem development and binding upstream and downstream partners. This is the only way for them to become international giants, with each containing the other.

AI Agents Accelerate Commercialization

How Many Hurdles Remain for Enterprise-Level AI to "Go from Virtual to Real" and Achieve Grounding?

Evolution of Large Models: Where Will AI Industrial Landing Roll Up?

Li Wei from the China Academy of Information and Communications Technology: How Can Cloud Computing and AI Accelerate Integration to Usher in a New Era of Intelligent Computing?

[Original from Tech Cloud Report]

Please indicate "Tech Cloud Report" and attach the link to this article when reprinting