The domestic AI computing power industry is moving past the initial phase of hardware catch-up.

In the past few years, the market's focus has consistently been on chip parameters, peak computing power, and wafer fabrication processes. A number of domestic AI chips have been deployed, with smart computing center projects launching one after another across regions, and domestic computing clusters have entered a cycle of large-scale delivery. However, after many projects were implemented, the industry discovered a harsh reality: the availability and installation of hardware do not guarantee stable operation of large models.

The real bottleneck has shifted to the software stack, development tools, and industrial ecosystem. While the hardware gap continues to narrow, the shortcomings in the software ecosystem have become the core obstacle to large-scale commercialization of domestic computing power.

Accelerated Hardware Delivery: Cluster Deployment Shifts from Pilot Projects to Mass Rollout

Three years ago, most domestic computing clusters were confined to laboratories and demonstration projects, primarily used for technical validation. Today, the situation has changed significantly.

Several domestic AI chip manufacturers can now stably deliver single cards and entire systems, supporting the construction of clusters with hundreds or even thousands of cards. Local smart computing projects and private deployment orders from government and enterprises continue to materialize. Domestic computing power is no longer just a concept on PowerPoint slides but now has the foundational conditions for large-scale delivery. Hardware metrics such as single-card computing power and memory bandwidth are narrowing the gap with mainstream overseas products, and in some scenarios, they can already handle large-model inference and small-to-medium model training tasks.

However, after the deployment of numerous projects, new issues have emerged. For clusters of the same scale, hardware installation takes only a few months, but model adaptation, operator debugging, and performance optimization often require much longer cycles. Many enterprises that have purchased domestic computing clusters find it easy to power on the hardware but face significant engineer-hours to fully migrate mainstream large models, requiring extensive adaptation and modification.

Hardware is a tangible asset that can be procured and stacked; however, the adaptation capabilities of the software stack cannot be delivered in a one-time package. As hardware shortcomings are gradually addressed, the main battleground of industry competition has quietly shifted.

Software Ecosystem Becomes the Core Bottleneck: Migration Costs Hinder Large-Scale Adoption

The value of a computing cluster lies not in the hardware itself but in its ability to smoothly run various large models and industry applications. A complete computing software stack includes compilers, operator libraries, drivers, distributed frameworks, debugging tools, and the upper-layer model ecosystem. Overseas computing ecosystems, after years of development, allow most open-source large models and industry tools to run out-of-the-box.

In contrast, the biggest pain point in the domestic computing ecosystem is migration costs. Many open-source models are not natively compatible with domestic chips, requiring extensive redevelopment and optimization of operators. Enterprises looking to migrate models to domestic clusters need dedicated algorithm teams for adaptation, resulting in high labor costs and long debugging cycles. Once a model version is updated, the adaptation process must be repeated.

Ecosystem deficiencies create a chain reaction. Developers are more accustomed to mature frameworks and lack motivation to migrate; fewer developers mean fewer upper-layer applications; insufficient applications reduce enterprise procurement willingness, which in turn limits chip manufacturers' financial investment in iterating the software stack, forming a negative cycle.

This does not mean there has been no progress in domestic software stacks. In recent years, domestic vendors have continuously invested in compiler and operator library development, collaborated with large model enterprises on joint adaptation, and established industry benchmark tests. However, ecosystem development is a long-term endeavor that cannot be accomplished through short-term investment by a single enterprise. While hardware can quickly catch up, ecosystem cultivation requires years of sustained effort.

Conclusion

In the first half of the domestic computing power race, the focus was on chip hardware, addressing the question of "availability." In the second half, the focus shifts to the software ecosystem, addressing the question of "usability."

Hardware gaps can be quickly narrowed through sustained R&D, but the software stack, developer ecosystem, and model adaptation systems require collaborative efforts across the entire industry chain. For domestic computing clusters to truly transition from demonstration projects to large-scale commercialization, the focus cannot remain solely on hardware parameter competition.

Future competition will no longer be about the performance of individual chips but about the integration of the entire hardware-software ecosystem. The hardware foundation is now in place, but the long-term battle of ecosystem development has only just begun.

Source: Touzizhe Wang (Investors Network)