EN / 中文

Compute Stacking Illusion: Kernel Talent Bottleneck Restricts Domestic Large Model Deployment

by huanghaifeng·November 12, 2026

Have you noticed that recent launch events for domestic large language models follow a similar pattern? Model parameters are larger, capabilities are stronger, and benchmark scores are higher. However, almost no one talks about kernels.

Beneath the noisy competition at the upper layers, an invisible race in kernel development determines how far domestic AI compute can go. Kernels act as translators between hardware and large models, dictating whether AI chips can release their full performance and whether models and AI applications can run with lower costs and higher efficiency.

Every step up in large model capability raises the bar for underlying kernel requirements—the spotlight is on the top layer, while the hard work falls on the bottom. This misalignment defines the current situation of this layer.

Compared to foreign AI ecosystems, "writing kernels" is even more critical for the domestic AI ecosystem because China started later. How can China's AI sector make up for this gap in kernel development? How can we encourage more people to join domestic kernel development?

Recently, the author paid attention to a major kernel development challenge that attracted registrations from 177 universities and 539 teams. After fierce competition, 28 teams advanced to the finals. From the rich outcomes of this competition, we see the answer to how China's AI can address its shortcomings.

01Kernels:The "Translators" of the AI Compute Foundation

To answer this question, one must first understand what a "kernel" actually is.

Kernels are the basic units of computation in deep learning. Matrix multiplication, convolution, attention—these are the most frequent operations in models, which translate into individual kernels when mapped onto chips. They are the foundational blocks upon which all upper-layer capabilities are built.

Here, the author uses a vivid analogy to help readers understand "kernels." In the AI industry, chips have their own "native language," recognizing only machine instructions, while large models speak a complex mathematical language.

Kernels are the "translators" that convert mathematical language into the chip's native language. A good translation hits the mark immediately; a poor translation is convoluted, inefficient, and prone to errors, causing the model to stall or fail to run.

At this point, it becomes easier to understand the role of "kernels" in the compute industry: the quality of kernel optimization directly determines the actual running speed and compute consumption of a model on a chip.

In short, if kernels are well-adapted, the hardware's capacity utilization rate can be truly released; if poorly adapted, even powerful compute cards remain mere paper specifications.

For the domestic compute industry, there is a more practical challenge: cross-chip migration. When vendors need to switch large models from overseas chips to domestic compute hardware, it is not simply a matter of replacing compute cards to get models running on these facilities.

The industry requires rewriting, aligning, and deeply optimizing "kernels." However, this process is not easy and often takes a long time. Clearly, the "kernel" layer is one of the key factors determining whether a model can "truly run."

02The Illusion of Stacking Compute:Why Doesn't Performance Improve Even When Cards Are Sufficient?

Currently, there is a common misconception in the industry: as long as you stack compute cards and build intelligent computing centers, large models will be supported.

The plight of many enterprises in reality poses thought-provoking questions. We often hear complaints from companies that although they purchased intelligent compute services with leading hardware specifications, their own large models failed to achieve expected results. Compute power was increased, but performance did not follow.

Why does this contradiction occur? As suggested in the first part of this article, the problem likely lies at the "kernel" layer. To explain this more clearly, let us use a fitting analogy.

A chip is like an engine, providing horsepower; a large model is like a passenger, presenting business needs; and a kernel is the driver who understands both the hardware and the business. No matter how strong the engine's horsepower, if the driver does not understand the vehicle's characteristics, the car cannot perform to its potential, regardless of its inherent strength.

Breaking it down, when a model fails to achieve expected performance on compute resources, the issue is not insufficient quantity of compute or inadequate data center operations, but rather insufficient adaptation and optimization between kernels, models, and chips.

Compute is about "quantity," while kernel adaptation is about "quality." Merely stacking "quantity" without improving "quality" means that even massive amounts of compute are just inactive equipment; the more you stack, the more waste occurs.

For vendors, the compute purchased from operators must be adapted at the "kernel" level to realize its true value.

This mismatch between chips and models triggers a chain reaction in the industry. It discounts the massive compute investments made by vendors, drives up training and inference costs, and limits business experience.

These phenomena collectively trap the industry in a dilemma: "Compute is available, but it cannot be used effectively." This may be the most realistic awkwardness facing domestic compute today.

03The Kernel Dilemma:The Bottleneck Is Essentially a Talent Gap

Is the difficulty in "kernel" optimization a technical tool issue or a human resource issue? A kernel development competition targeting real-world scenarios provides a valuable observational window.

The competition mentioned is the recently held "China Telecom Star Cup · University AI Kernel Development Challenge," co-hosted by China Telecom Artificial Intelligence Technology (Beijing) Co., Ltd. and Huawei. Its most significant innovation is transforming kernel development capabilities—which were previously difficult to compare horizontally—into a unified, measurable standard, making the talent shortfall at the foundation of domestic compute visible for the first time.

In the past, assessing kernel optimization capabilities was difficult: under different hardware and verification standards, kernel implementation results could not be compared on the same dimension. This competition adopted unified problems, a unified evaluation platform, and a mode of online assessment plus on-site verification to ensure reproducible results.

The competition targeted university students nationwide, with all problems derived from real-world underlying kernel scenarios for large models. It attracted 539 team registrations from universities across the country. After multiple rounds of elimination, 28 teams advanced to the finals.

It is understood that the final round problem was the Sparse FlashAttention backward kernel, which tests the highest level of skill. It is a core kernel in sparse large model training, directly determining the memory usage and training speed of long-context large models, with extremely high optimization difficulty.

The more standardized the competition, the clearer the issues it reflects. What this competition reveals is precisely the baseline of industrial talent stock: there are currently not enough talents willing and able to engage in domestic "kernel" development, yet this also implies huge opportunities.

In other words, the hurdle for China's compute infrastructure lies less in "kernel" technology itself than in the talents who can "write kernels." The author believes there are two main reasons.

First, the barrier to entry for "kernels" is extremely high. To engage in "kernel" development, talents must understand both chip hardware architecture and deep learning model logic, requiring typical interdisciplinary capabilities combining software and hardware.

To put it bluntly, these talents must decipher instruction sets, memory hierarchies, and compute unit divisions in chip manuals, master the mathematical principles and numerical precision of models, and repeatedly weigh trade-offs between "accuracy" and "speed."

A single "kernel" must accommodate multiple numerical precisions simultaneously; a slight oversight can cause model training to diverge. Similarly, unoptimized "kernels" can result in significantly different speeds on the same chip.

As Liu Cong, Vice President of iFLYTEK and Dean of the Research Institute, stated at the 2026 Zhiyuan Conference, training large models on domestic compute is not a simple adaptation issue but a tough battle requiring deep engagement with software and hardware architectures to overcome underlying bottlenecks one by one.

Those who know algorithms may not know chips, and those who know chips may not know models; those proficient in both are rare indeed. This dictates the scarcity of talents who can "write kernels."

Second, talent supply cannot keep up with industrial expansion. On one hand, large models are continuously iterating and intelligent computing centers are being launched on a large scale nationwide; on the other hand, people capable of kernel development, migration, and tuning are extremely scarce.

Liu Guang, Head of the System Research Group at BAAI (Beijing Academy of Artificial Intelligence), stated frankly at the "Frontier of Intelligent Computing · AI Empowering Systems" sub-forum of the 2026 Zhiyuan Conference: "Kernel" development is the hardest, most foundational work in AI infrastructure. Talents are extremely scarce, and the talent bottleneck has reached a critical point.

It can be said that the supply of kernel talents who understand software-hardware collaboration failing to keep pace with industrial expansion is the real hurdle for domestic compute.

04Cultivating Talent Through Competitions, Establishing Unified Standards,and Expanding the Pool of Domestic Kernel Talents

Talent gaps do not close automatically. As large model scales continue to grow and domestic compute hardware expands, the talent gap at the kernel level will only widen further in the future.

There are many people in the industry who can write prompts and develop model applications, but those capable of underlying kernel development, migration, and tuning are truly scarce resources.

Under the wave of large models, most people rushed to the upper layers where achievements come more easily, while few were willing to sit on the cold bench for the hardest, toughest, and most critical underlying layers. For the next stage of domestic compute, the primary contradiction is no longer whether suitable compute cards exist, but whether there are people to truly release hardware performance.

For this reason, this competition deserves greater attention from the industry. The China Telecom Star Cup · University AI Kernel Development Challenge is not just a contest; by turning kernel development capabilities into a measurable standard, it is building a "reservoir" for industrial "kernel" talents.

For example, post-competition university kernel talent seminars and talent promotion and recruitment by the Star General Artificial Intelligence Laboratory are doing the same thing: gathering scattered "kernel" talents from campuses one by one, and through competitions, highlighting the importance of "kernel" talents to the industry, thereby accelerating industry investment in and cultivation of "kernel" talents.

When operators and enterprises build intelligent computing centers, they cannot bypass the underlying proposition of "kernels." Compute power determines how fast China's AI can theoretically run; kernel adaptation determines how far it can realistically go.

And whether kernels can be continuously polished to release compute value ultimately depends on software-hardware collaborative talents willing to sit on the cold bench. The second half of the domestic compute competition comes down to a competition for underlying talents.