Turning AI from Algorithms into an Asset Governance Expert　Johnson Hsieh's talking on the True Logic Behind Industrial AI Deployment

A PhD graduate in statistics from National Tsing Hua University, Hsieh has worked extensively with high–energy-intensity industries―including energy, petrochemicals, and steel―establishing himself as a founder with a distinctly strong technical background.

By Xin-En Wu

At a time when many enterprises still regard AI adoption as little more than a digital transformation slogan or an experimental pilot, Chimes AI has chosen a markedly different path. Rather than racing to showcase the sophistication of its models or the novelty of its algorithms, the company has persistently returned to a far more fundamental question: Why do so many AI deployment projects fail to survive beyond two years on the factory floor?

"We didn't start this company just to build another AI startup," said Johnson Hsieh, Founder and CEO of Chimes AI, as he reflected on the company's origins. Trained as a PhD in statistics from National Tsing Hua University, Hsieh has spent years working closely with energy, petrochemical, steel, and other high–energy- intensity industries. A technologist by background, he has witnessed firsthand countless AI projects that appeared successful on paper, yet ultimately failed to gain long-term traction in real operations.

"Many companies invest enormous resources. Their proof-of-concept results look great," Hsieh observed. "But once the system goes live, no one dares to write in an operational report that 'we are using AI.'" In his view, the problem has never been a lack of willingness. Rather, AI has rarely been designed as a tool that can be trusted, quantified, and used continuously over time.

Starting from Mathematics and the Industrial Front Line　Why Most AI Projects Never Penetrate Deep into the Factory

Although Hsieh's academic roots lie in mathematics and statistics, nearly his entire professional career has been embedded in heavy industrial environments. He is both a statistician whose research software has been cited tens of thousands of times and a long-time advocate of data science and AI democratization¹. Yet instead of remaining in academic abstraction, he chose to work inside petrochemical plants, semiconductor fabs, and energy facilities—places defined by harsh conditions, operational complexity, and zero tolerance for error.

This intersection of theory and practice allowed him to recognize a harsh reality earlier than most AI entrepreneurs: industrial environments do not lack data; they lack manageable judgment.

Hsieh illustrated this with a real-world example from a global top-ten petrochemical complex in Mailiao. The site operates more than 100,000 pieces of equipment, while the entire maintenance workforce numbers fewer than 30 people. "If you do the math," he explained, "each engineer can spend about five seconds per machine per day. Over the course of a year, they barely finish one round of inspection."

In such environments, traditional manual inspection systems have already reached their limits. Even when massive numbers of sensors are installed, the problem does not disappear. Temperature, pressure, and flow data flood in continuously, alarms fire nonstop, yet no one knows which alerts truly matter.

"The old approach is to set threshold values— if something exceeds the limit, you send someone to check," Hsieh said candidly. "That might work when you have a small number of machines. But at the scale of tens of thousands, it overwhelms people and increases the risk of missing the anomalies that actually matter."

From the outset, therefore, Chimes AI was never about "building models for their own sake." Instead, the company treats AI as a decision structure. Using IoT data as a foundation, its system first establishes a behavioral profile for each piece of equipment under healthy operating conditions. AI is then used to quantify deviations from that baseline—going beyond simply flagging an issue to identifying what is likely causing the deviation.

"Our goal is that engineers know exactly what to check before they even arrive on site," Hsieh explained.

AI Is Not Just Early Warning　It Is the Risk Language of Critical Assets

In Chimes AI's definition, AI does far more than predict whether a machine will fail. It translates equipment risk into asset indicators that can be calculated, managed, and communicated. This perspective naturally extends its applications into energy management, virtual power plants (VPPs), and even insurance risk assessment.

Hsieh emphasized that whether one is dealing with energy storage systems, power generation assets, or high-consumption demand-side equipment, these are all critical assets. Once such assets are incorporated into dispatch and control systems, they cease to be purely engineering concerns and become sources of systemic risk—affecting safety, production continuity, financial performance, and insurability.

In semiconductor manufacturing, for example, chilled water systems often account for nearly 20% of total electricity consumption. Any efficiency deviation can simultaneously impact energy costs and production yield. By building behavioral and health models for chillers, Chimes AI enables real-time assessment of whether equipment remains within safe operating envelopes under varying loads, ambient conditions, and dispatch strategies.

"We're not making machines run harder," Hsieh said. "We're making them run in the right state." In practice, this approach has reduced chilled water system energy consumption by approximately 10–15%, while avoiding equipment stress caused by excessive or poorly calibrated adjustments.

On the supply side, cogeneration systems—such as boilers and gas turbines—tend to experience gradual efficiency degradation over time. Under frequent rampup and ramp-down operations, traditional methods struggle to determine whether performance changes represent true anomalies. AI-based modeling, however, provides stable reference benchmarks across operating conditions, enabling adjustments to air-fuel ratios and operating parameters to keep energy performance closer to optimal levels.

"This is not just about saving energy," Hsieh stressed. "It's about preventing efficiency degradation from being ignored."

In the petrochemical sector, reciprocating compressor cases point even more directly to public safety. By analyzing vibration, current, and flow data, AI systems can issue warnings weeks before equipment approaches failure, enabling proactive maintenance, preventing unplanned shutdowns, and potentially averting catastrophic accidents.

"Sometimes, preventing a single incident is worth far more than all the electricity you save," Hsieh noted.

Making Distributed Resources Reliable Power　Behind-the-Meter VPPs Need Better Management, Not Just More Equipment

When discussing how Chimes AI differentiates itself from peers, Hsieh openly identified speed as a key factor. Traditional equipment risk models often require six months to over a year of data preparation and expert interviews. Through its proprietary AutoML platform, Chimes AI has shortened model-building cycles to just a few weeks.

More importantly, AI is not treated as a black box accessible only to data scientists. Through structured training, equipment engineers themselves participate in model development. This "horizontal scaling" capability allows large enterprises to deploy thousands of equipment models in a short period, enabling true scalability.

The introduction of deep AI assistants further addresses knowledge gaps on the factory floor. When alerts occur, the system does not merely display abnormal indicators. It integrates historical maintenance records, operation manuals, and past cases to suggest likely causes and recommended actions—reducing cognitive burden on less experienced engineers.

"Our intention is for AI to support people, not replace them," Hsieh said.

Commercially, Chimes AI adopts a subscription model. At a cost far below hiring a dedicated AI engineer, enterprises receive continuous updates and longterm support. This transforms AI from a one-off project into an ongoing service.

Turning to the development of behind-the-meter virtual power plants, Hsieh argued that the core challenge lies not in market size or business models, but in asset management capability. As distributed generation, storage, and demand-side resources are increasingly integrated into grid dispatch systems, the real question is no longer whether assets can be interconnected— but whether they can be managed safely, predictably, and over the long term.

Without intelligent asset management, scaling VPPs becomes untenable as asset counts grow. Reliance on manual oversight and traditional monitoring simply cannot sustain system performance at scale.

Hsieh emphasized that whether one is dealing with energy storage systems, power generation assets, or high-consumption demand-side equipment, these are all critical assets. Once such assets are incorporated into dispatch and control systems, they cease to be purely engineering concerns and become sources of systemic risk―affecting safety, production continuity, financial performance, and insurability.

Expanding on the definition of "critical assets," Hsieh stressed that the term does not refer solely to high-energy-consumption equipment. It encompasses supply-side, storage-side, and demand-side assets— any facility that carries both value and risk.

Energy storage systems, for example, face performance degradation and safety concerns. If batteries are to participate in dispatch, their health and aging status must be clearly understood. Power generation assets such as gas turbines and steam turbines similarly experience efficiency declines over time and require continuous monitoring. On the demand side, chilled water systems are indispensable for maintaining stable operations and yield in semiconductor fabs and other critical industries.

Without integrating these assets into a unified, quantifiable management framework, large-scale VPP deployment remains unrealistic.

From Hsieh's practical observations, most Taiwanese enterprises remain in the IoT stage—collecting massive volumes of data but still relying heavily on human monitoring and experience-based judgment. As asset numbers grow, management costs quickly spiral out of control, constraining market expansion.

By translating equipment risk into calculable, communicable indicators, AI enables behind-the-meter VPPs to evolve into systems that can be trusted over the long term. Such systems can even integrate with industrial insurance, reducing the likelihood of unplanned outages and major incidents.

Compared with the tens of billions of dollars in losses a single accident can cause—such as a large-scale energy storage fire—modest monthly investments in intelligent risk management can generate far greater value for industry and the financial system alike.

"If you don't have reliable asset management and risk quantification, the market simply cannot scale," Hsieh concluded. "As the number of critical assets continues to grow, only intelligent management can sustain long-term system operation."

"The value of AI is not how smart it is, but whether enterprises dare to use it—and how long they dare to keep using it."

For Chimes AI, this is not a slogan. It is a path of practice that begins with mathematics, runs through the factory floor, and ultimately reaches the core of energy transition itself.

¹ The iNEXT software he developed in the field of biodiversity research has been cited more than 13,000 times by the global academic community to date. In 2020, it was selected for inclusion in the Arctic Code Vault, where it was archived alongside open-source systems of enduring significance to human civilization, such as Linux, and preserved within the Arctic permafrost. Over the years, Hsieh has also collaborated with leading institutions including the Taiwan AI Academy and the Artificial Intelligence Technology Foundation, helping to train thousands of AI talents. Through these efforts, he has remained committed to advancing AI from the realm of knowledge-based education into real-world, practical application.