Mid-game Struggle of AI Computing Power between China and the US: The Dispute between Openness and Closure

Recently, Google's TPU, riding on the wave of the comeback momentum of Gemini 3, has significantly broadened its incremental prospects. Meta is considering spending billions of dollars on it, and institutions have raised their TPU production forecast by 67% to 5 million units. Based on the full - chain closed - loop of "chips - optical switching networks - large models - cloud services", Google's intelligent computing system has returned to the forefront echelon of the AI track, marking a further step in the American - style closed and monopolistic route.

Meanwhile, open - source models represented by DeepSeek are closely following. At the beginning of the month, DeepSeek V3.2 and its long - thinking enhanced version were released. The former tied with ChatGPT in performance tests, while the latter directly targeted the top - tier closed - source model Gemini. This also indicates that China's open - source and open - ended route is gradually gaining momentum, and the domestic intelligent computing system shows good ecological synergy potential at the application layer.

So far, the game between China and the US in the AI industry has reached the middle stage, and the confrontation pattern between "open collaboration" and "closed monopoly" has become increasingly clear. Especially in the layout of the intelligent computing ecosystem, the two camps may be brewing a peak competition of systematic capabilities.

01 From Gemini 3 to TPU v7, the Closed - loop of Hardware and Software Integration Reaches the Pinnacle

Undoubtedly, the sudden popularity of Google's TPU is largely due to the verification of the model capabilities of Gemini 3. As an ASIC chip specifically designed for Google's TensorFlow framework, TPU laid the foundation for its full - stack closed - loop with its hardware - software integration design. At the same time, it captured the external user market when making high - level breakthroughs in upper - layer applications and was even once regarded as the strongest alternative to NVIDIA's GPUs.

The so - called "hardware - software integration" means that the hardware design fully serves the needs of upper - layer software and algorithms. For example, the training and inference processes of Gemini 3 are highly compatible with TPU clusters, and this customized and dedicated model also shows extremely high value in terms of power consumption and energy efficiency. The power consumption of TPU v5e is only 20% - 30% of that of NVIDIA H100, and the performance per watt of TPU v7 has doubled compared with its previous products.

Currently, Google has formed a closed and efficient cycle through the vertical integration of "chips + models + frameworks + cloud services". On the one hand, it has greatly improved its own AI R & D and application development efficiency. On the other hand, it has carved out its own territory under the mainstream NV system and won the dominance in another intelligent computing track. Meta's intention to purchase TPU has pushed the popularity of this system to a peak.

Some industry insiders point out that from Apple to Google, the American - style vertical and closed - loop approach has almost reached the extreme, showing the ubiquitous monopoly desire of technology giants at the industrial chain level to consolidate and expand their interest maps. However, from the perspective of ecological development, the closed - loop model lacks the spirit of long - termism and can easily lead to the loss of innovation vitality in the upstream and downstream of the industry and form a highly centralized pattern of a single entity.

In addition, from the perspective of TPU's application scenarios, the closed - loop of hardware and software integration is clearly a game exclusive to giants. An analyst said that Google's cluster design and "software black box" require users to re - configure a whole set of heterogeneous infrastructure. Without the need for training trillion - parameter models, one simply cannot fill the systolic array of TPU, and the saved electricity costs may not even offset the migration costs.

At the same time, since the TPU technology route is extremely closed and incompatible with the mainstream development environment, users also need a professional engineering team to operate its XLA compiler and reconstruct the underlying code. That is to say, only enterprises of the level of Google and Meta are eligible to switch to the TPU route, and only when the computing power scale reaches a certain level can the energy - efficiency advantages of customized products be realized.

It cannot be denied that leading enterprises such as Google have achieved rapid single - point breakthroughs in local tracks through vertical integration and self - built closed - loops, and at the same time created a prosperous scene of numerous American technology giants. However, against the background of the China - US AI game, the American - style closed and monopolistic route has completed the track occupation in advance by virtue of its first - mover advantage, and passive follow - up catch - up is difficult to meet the development needs of China's intelligent computing industry.

Outside the "small courtyard with high walls", how to give full play to the advantages of the national system and unite all forces to break down the walls and build roads has become the key to narrowing the gap between the Chinese and American AI systems.

02 Multi - heterogeneous Ecological Collaboration, the Open Path Leads to the Next Checkpoint

Compared with the American - style oligopoly model, China's intelligent computing industry is gradually decoupling based on a multi - heterogeneous system and reshaping an open - ended ecosystem. From top - level design to industrial implementation, "open - source and open + collaborative innovation" has become the consensus of the entire domestic hardware and software stack.

At the policy level, the "Action Plan for the High - quality Development of Computing Power Infrastructure" proposes to build a computing power Internet with reasonable layout, ubiquitous connection, and flexible and efficient operation, enhance the integration ability of heterogeneous computing power and networks, and realize the cross - domain scheduling and orchestration of multi - heterogeneous computing power. Moreover, relevant departments have repeatedly emphasized the encouragement of all parties to innovatively explore the construction and operation models of intelligent computing centers and the multi - party collaborative cooperation mechanism.

Extending to the AI application layer, the "Opinions on Deeply Implementing the 'Artificial Intelligence +' Action" also requires deepening high - level opening up in the field of artificial intelligence and promoting the open - source availability of technologies... It is not difficult to see that the country has provided a completely different Chinese solution in the fields of artificial intelligence and intelligent computing - instead of blindly chasing the closed - loop approach in the closed route, it is necessary to seek a staggered catch - up in an open pattern.

In fact, the top - level design is completely based on the actual needs of the industry. Under the US technology blockade, China's intelligent computing industry mainly faces two major challenges: the bottleneck of single - card computing power performance and high computing power costs. In addition to continuous efforts in core technology fields such as chips, models, and basic software, the currently more effective way is to develop larger - scale, more diverse, and more efficient intelligent computing clusters to break through the AI computing power bottleneck.

Industry research results show that there are no less than 100 domestic computing power clusters claiming to have a scale of thousands of cards, but most of them use heterogeneous chips. It is imaginable that if different hardware systems are mutually closed, the standard interfaces are not unified, and the software stacks are incompatible, it will be difficult to effectively integrate and utilize intelligent computing resources, let alone meet the application requirements of large - scale parameter models.

According to the mainstream industry view, domestic AI computing power is characterized by diversification and fragmentation, and at the same time has considerable scale advantages. The top priority is not to push forward a single technical route independently, but rather to quickly break through the "technology walls" and "ecological walls" as soon as possible, realize open cross - layer collaboration in the industrial chain, truly unleash the potential of the overall computing power ecosystem, and move from single - point breakthroughs to integrated innovation.

Specifically, the so - called open route aims to promote collaborative innovation in the industrial ecosystem based on an open computing architecture. For example, by formulating unified interface specifications, enterprises in the upstream and downstream of the industrial chain, such as chip manufacturers, computing system providers, and large - model developers, can be linked to jointly participate in ecosystem construction, reduce repetitive R & D and adaptation investments, and share the benefits of technological breakthroughs and collaborative innovation.

At the same time, as the collaboration standards in the open architecture tend to be unified, commercialized hardware and software technologies can be further developed to replace customized and proprietary systems, thereby reducing the application costs of computing products and achieving computing power inclusiveness covering the entire industrial stack.

Obviously, under the Chinese - style open system, domestic AI computing power is breaking through the generalization and popularization dilemma of Google's TPU, widely connecting the intelligent computing ecosystem with various developers and users, and finally forming systematic collaborative combat capabilities to more flexibly and efficiently empower the implementation of "Artificial Intelligence +". By then, the China - US AI game will also move beyond single - card competition and single - model comparison, and comprehensively enter the ultimate confrontation of ecosystem capabilities.

The cover image and accompanying pictures of this article are copyrighted by their respective owners. If the copyright holders believe that their works are not suitable for public viewing or should not be used free of charge, please contact us in time, and our platform will make corrections immediately.