The 2026 World Intelligent Industry Expo has officially opened its doors, boasting the participation of 741 companies and institutions—the largest turnout in its history. The unveiling on stage of the Ascend 384 super node and Sugon’s scaleX 10,000-card supercluster, representing two generations of flagship solutions, provided a definitive response to the industry’s lingering query: ‘What is the trajectory for domestic computing power?’

The Pivotal ‘Track Switch’ in the Computing Power Race

In a nutshell: Huawei’s 384 is dedicated to the ‘in-depth battle’ of AI training and inference, while scaleX’s 10,000-card supercluster aims for ‘comprehensive coverage’ in the era of scientific intelligence.

The turning point emerges from a trend known as ‘hyper-intelligence convergence.’ For years, scientific computing and AI have pursued entirely distinct paths. Scientific computing has strived for ultimate dual-precision performance, centering on mathematical model-driven numerical simulations that demand ‘accuracy.’ Conversely, AI large models have prioritized massive parallel throughput at low precision, focusing on data model-driven pattern recognition that requires ‘speed.’ One seeks precision, the other speed—fundamentally different precision requirements at their core.

However, AI for Science is dismantling this binary framework. In cutting-edge research domains such as biopharmaceuticals, new material simulations, weather forecasting, and quantum chemistry, computational tasks now necessitate both high-precision numerical simulations and low-precision inference acceleration simultaneously. For instance, employing AI to expedite structural searches for protein folding, followed by verifying structural stability through high-precision validation. Single-precision computing solutions are no longer adequate.

This is not a far-fetched theoretical assertion. At this expo, Sugon presented several compelling data points: a 30,000-card protein folding simulation achieved a 1,000-fold acceleration over traditional algorithms; a 45,000-card liquid water molecular dynamics simulation enabled trillion-atom-scale computation, breaking world records while enhancing efficiency by three orders of magnitude. DFT-precision simulations of 41.47 billion atoms, with billion-grid efficiency improvements from ‘weekly’ to ‘hourly’—these figures signal not merely ‘faster’ but the emergence of a fundamentally different computing paradigm.

The global AI4S market is projected to surge from approximately $4.5 billion in 2025 to $26.2 billion by 2032, with a Compound Annual Growth Rate (CAGR) nearing 29%. The six major downstream industries could collectively encompass a market size of nearly $11 trillion. The industry’s future main battleground is undergoing a complete transformation.

From ‘In-Depth Battle’ to ‘Comprehensive Coverage’

Huawei’s 384 super node has distinct design boundaries: optimized for AI training and inference scenarios, its core metrics concentrate on BF16 low-precision computing power—a nominal single-cluster value of 300 PFlops, already utilized to train the 718-billion-parameter Pangu Ultra MoE large model. However, its capacity to support scientific computing scenarios requiring dual-precision or high-precision coverage remains uncertain.

This is not to say it cannot achieve this, but rather that it was never intended for that battlefield. A high-performance vehicle designed for highways struggles in off-road terrain—regardless of the car’s excellence, the scenario simply doesn’t align.

Meanwhile, Sugon’s 60,000-card AI4S computing cluster officially launched in April at Zhengzhou’s National Supercomputing Internet core node. Engineered to support full-precision computing from FP8 to FP64, it harmonizes traditional supercomputing’s high-precision scientific computing needs with large-scale AI training’s low-precision, high-throughput requirements.

The entire computing power industry is undergoing profound transformations. Demand in the era of scientific intelligence is rapidly shifting from single low-precision to full-precision coverage. Any solution ‘specialized’ in precision dimensions, regardless of its past triumphs, may encounter systemic challenges in the next era.

Reflecting on the past two years, the domestic computing power industry’s most significant accomplishment has been addressing part of the ‘sufficiency’ question. When single-card performance lagged, system engineering capabilities pieced together a viable domestic computing power foundation—a value that cannot be overstated.

As computing power transitions from laboratories to industries, from ‘buildable’ to ‘usable,’ architectural inclusivity, precision coverage breadth, and systemic engineering collaboration depth become the critical variables determining ceilings. This enables the same foundation to drive low-power AI inference while also powering high-load scientific simulations.

Looking forward, the truly compelling question is: Who can genuinely construct a computing power system capable of simultaneously handling high-precision scientific computing and low-precision intelligent acceleration?

The answer is becoming increasingly evident.