PARATERA's Lü Shengliang: Supercomputing is moving from university research to enterprises, connected to 63 computing power centers
Lü Shengliang, general manager of PARATERA's industry business unit, said that five or six years ago university research scenarios accounted for over 70%-80% of its business, while this year demand comes more from enterprises, and the company has connected to about 63 computing power centers (more than forty intelligent computing centers and a dozen supercomputing centers). PARATERA integrates computing power through its "submit at night, view in the morning" on-demand billing model and "PARATERA Computing Network" intelligent scheduling; automotive manufacturing and new energy contribute nearly 40% of its business unit's revenue. IDC expects China's intelligent computing power scale to reach 2576.5 EFLOPS in 2026, up 87.9% year on year.
Body · AI translation
At eleven o'clock at night, in the simulation computing center of a certain new energy vehicle company, the fluid simulation progress bar on the screen is slowly advancing. The city outside the window has already quieted down, but the servers here are still roaring—engineers do analysis and tune parameters during the day, and at night they hand the computing tasks over to the compute power; when they push open the door first thing the next morning, the results are already laid out on the desk.
This "submit at night, review in the morning" working rhythm is becoming the new normal in R&D laboratories across China.
Meanwhile, on the vast lands of Inner Mongolia, Ningxia, and Xinjiang, rows upon rows of server rooms glow with indicator lights, like silent giants swallowing and spitting out data. Electricity is cheap and the climate is suitable in these places, making them ideal for deploying large-scale computing centers. But they are no longer the "ivory towers" of the past that only served university research—more and more automakers, large-model companies, aerospace enterprises, and biopharmaceutical companies are becoming the most active users in these machine rooms.
Supercomputing, a term that once carried a strong academic flavor, is undergoing a profound industrial migration.
Supercomputing "breaks out of the circle": from the ivory tower to the main industrial battlefield
"Five or six years ago, research and university scenarios accounted for more than 70%-80% of our business. But by this year, demand is coming more from enterprises," Lu Shengliang, general manager of the industry business unit at Parallel Technology, described the industry's dramatic change in an interview.
The driving force behind this figure is the sharp growth in demand for compute power across industries. The "2026 China Artificial Intelligence Computing Power Development Assessment Report" released by IDC predicts that in 2026 China's intelligent computing power scale will reach 2576.5 EFLOPS, up 87.9% year on year, and by 2030 it will break through the 10,000 EFLOPS mark.
Faced with this problem, enterprises in the past either built their own computing clusters or relied on university supercomputing centers to "hitch a free ride." But this road is now getting narrower and narrower. Self-built clusters require high investment and heavy operations and maintenance, and are hard-pressed to cope with elastic demand during R&D peak periods; university supercomputing centers themselves have limited resources and often struggle in the face of the volume and timeliness requirements of enterprise-level tasks. There was once a case where an automaker's CAE engineer encountered a backlog of simulation tasks of two weeks, with queuing starting at 12 days.
"In the past, when we served universities, the user group was relatively fixed and business growth was relatively slow. But now it's different," Lu Shengliang said. "Compute demand on the enterprise side is growing sharply, and the timeliness required to satisfy that demand is also rising sharply."
This means that the transformation of the compute usage model is the core of this migration.
In the past, enterprises obtained compute power in a simple way—buying whole machines and renting whole machines. But this approach faces a typical resource mismatch dilemma: buy too much and it's wasted, buy too little and it's not enough. This is because R&D work naturally has a tidal effect—during the day engineers need instant compute power for analysis and debugging, while at night they can set work aside and let the compute power continue running calculations.
"We proposed a business model to customers called 'submit at night, review in the morning,'" Lu Shengliang introduced. "Engineers look at the results calculated the previous day at the company during the day and make anal
ysis and adjustments. After work, they hope the compute power will not rest, and will let the compute power continue running the calculations for them. When they come to work at nine the next morning, they can see the calculation results."
To realize this model, Parallel Technology did a series of complex back-end work: ensuring calculation accuracy, automatically rerunning on errors, intelligently scheduling back-end resources... "For an engineer, there is no longer any need to worry about queuing because of insufficient compute power, or wasting money because of excess compute power. You can use as much as you want, and you are not charged when you don't use it."
This on-demand supply model is likened by Lu Shengliang to "like using electricity"—flip the switch and there is power; turn off the light and there is no charge.
And the deeper change comes from the rise of the token economy. With the explosion of large models and AI applications, the usage model of compute power is beginning to shift from training-dominated to inference-dominated. "With the rapid growth of large-model token call volumes, inference is becoming an important driver of new compute power," Lu Shengliang said. "We have partnered with partners to launch a Token Factory for inference scenarios, providing compute power and platform services for model inference and intelligent applications for customers in different industries, while not intervening in the customers' models and applications themselves."
Inference scenarios have completely different requirements for compute power than training. Some scenarios have extremely high real-time requirements (such as watching short videos or playing online games), requiring compute power to be deployed close to users; while some scenarios (such as AI-assisted programming) can tolerate a certain delay and are suitable for deployment in western data centers with lower compute costs.
"If compute power is deployed in Inner Mongolia, Ningxia, and Xinjiang, electricity is cheaper and the overall compute cost is lowered," Lu Shengliang analyzed. "But if it is a scenario with high real-time requirements, we need to deploy compute power close to customers, such as Shanghai and Guangzhou."
From "computing faster" to "computing better and more economically," this is not just an adjustment of the technology route, but a profound reconstruction of the business model of the entire compute industry.
The "super scheduling" advantage of the compute black soil
In this transformation, the "parallel compute network" model proposed and practiced by Parallel Technology provides a unique way of solving the problem.
Up to now, Parallel Technology has connected about 63 computing centers, including more than 40 intelligent computing centers and more than a dozen supercomputing centers. These compute resources distributed across the country are integrated into a unified "compute pool" through an intelligent scheduling system.
"When a customer submits a task, how can I determine which back-end compute resource is most suitable for him?" Facing this core question, Parallel Technology's answer is a self-developed "application runtime characteristic analysis" software.
Parallel Technology has put forward a concept: it hopes that for every new customer who comes, within four hours it can determine what kind of application he has, whether it is compute-intensive or memory-access-intensive, and then, based on actual data and our experience, find the most cost-effective computing resource."
But the challenges this system faces go far beyond the technical level. Of the 62 computing power centers, some are self-built data centers, some are co-built with local governments, and some are third-party computing power resources. "How do we achieve a win-win model? We need to let everyone make money," said Lü Shengliang. "This involves pricing and settlement, and requires a model architecture that combines technology and business."
Parallel Technology's strategy is to do a good job on the "black soil"—only doing the infrastructure and platform layers of IaaS and PaaS, and not touching the customer's application layer. "Our customers are engineers doing applications, developing large models, and doing automotive R&D; they understand their own business better. We first do the infrastructure well and let them grow their own chimney on top of it."
This neutral third-party positioning appears especially precious in the AI era. "Many vendors doing large models, if they build their computing power entirely on top of some big vendor, and that big vendor has its own large model, then if for competitive reasons they cut off your computing power, your business will be greatly affected," Lü Shengliang pointed out. "So Parallel Technology is more willing to choose a neutral supplier that does not do models and only does computing power."
More than 1,000 cores working simultaneously, the AI computing power future reflected by smart cars
To understand how supercomputing penetrates industry, the new energy vehicle industry is an excellent observation window.
"Automotive manufacturing is the largest part of our business unit's revenue, with nearly 40% of revenue coming from automotive manufacturing and new energy," Lü Shengliang introduced. The R&D of a new energy vehicle company involves thousands of engineers, from vehicle design to autonomous driving algorithm training, and computing power demand runs through the entire R&D chain.
In the traditional CAE simulation field, the design of a new car needs to go through multiple stages such as fluid simulation (aerodynamics), solid simulation (structural strength), battery simulation (electrochemical performance), and full-vehicle simulation. These computations require large-scale parallel computing capability, sometimes requiring 512 or even more than a thousand cores to work simultaneously.
"It is impossible to build a new car every time and then do various experiments, crash tests, and fluid tests; the cost is too high," Lü Shengliang said. "So using computer-aided engineering and high-performance computing to assist engineering design, this method has already been developed for fifty years."
But the arrival of AI has brought entirely new computing power demands to the automotive industry. Training autonomous driving large models requires GPU computing power, and AI-assisted programming is changing the way software engineers work. "Now if engineers still cannot use AI to assist in writing code, it is basically hard to have room to survive," Lü Shengliang observed. "Previously you typed code line by line; now you may use natural language to tell AI what function to implement and let AI help write the code."
A more cutting-edge change comes from embodied intelligence. "People doing intelligent driving go into embodied intelligence because they have experience that can be transferred," Lü Shengliang analyzed. "But embodied intelligence needs not only manufacturing simulation capability, but also a 'brain' and a 'cerebellum'—world models and AI scenarios. So we have recently proposed integrating supercomputing with intelligent computing again, which we call super-intelligent integration."
From fluid simulation to autonomous driving training, from AI-assisted programming to embodied intelligence, the new energy vehicle industry's demand for computing power is showing a trend of diversification and exponential growth. The value of Parallel Computing Network lies in being able to find the most suitable computing power configuration for these completely different scenarios.
And when we look back from the node of industrial transformation, the trend of supercomputing moving from the laboratory to the industrial frontline is already irreversible.
"AI can bring improvements in people's actual production and living efficiency; this is beyond doubt," Lü Shengliang said. "The existing scenarios, these so-called AI applications, are still at a very preliminary stage. In the next three years, five years, or even ten years, the efficiency improvements brought by AI may be beyond what we can imagine now."
When computing power changes from a scarce scientific research resource into industrial infrastructure that can be accessed on demand like water and electricity, a deep revolution in production efficiency is taking place. From "computing faster" to "computing better and more economically," from a single cluster to a distributed computing network, from serving universities to empowering thousands of industries—the industrial migration of supercomputing has only just begun.
Leiphone Leiphone