MaKe | MaHui Member "Deep Intelligent Control": HVAC Energy-Saving SaaS Technology Powers Dual-Carbon Goals

#MaKe Key Updates from Source Code Capital and MaHui Members

Few people realize that optimizing air conditioning energy use is a one-in-a-hundred-thousand proposition.

The central air conditioning systems we commonly see consist of a heat source/cooling system and an air handling system, comprising components like cooling towers, chillers, cooling water pumps, and chilled water circulation pumps. Every temperature adjustment is the result of these components working in concert. The way these components are configured and the parameters they use mean there are hundreds of thousands of possible operational pathways for controlling an HVAC system. And since air conditioning typically accounts for 20%–50% of industrial and building energy consumption, the most energy-efficient pathway is just one among countless others.

In public buildings like office towers, shopping malls, and subway stations, HVAC systems are relatively straightforward and humidity/temperature control requirements are less stringent, so energy optimization isn't particularly difficult. But precision factories and data centers impose strict requirements on temperature, humidity, and pressure—maintaining temperatures within ±0.5°C, ensuring stable room pressure to prevent airborne contamination, and so on. Meeting these demands requires complex central air conditioning control systems, which in turn drives high energy consumption.

With carbon neutrality targets looming over every enterprise, how can companies achieve precise, efficient low-carbon operation of their air conditioning systems? 36Kr discovered that a three-year-old startup is tackling this exact problem. DeepCtrls uses AI data-driven optimization combined with physics-based modeling to help factories and data centers belonging to CATL, BOE, and the National Supercomputing Center reduce energy consumption by an additional 10%–40% on top of what mainstream energy management systems already achieve. At one large factory, DeepCtrls' system control software optimization saved the client over 8 million kWh in a single year.

The SaaS business of air conditioning energy optimization has brought this three-year-old company—with not a single salesperson on staff—to the brink of nearly RMB 100 million in annual software contract revenue, a milestone that many enterprise software companies six or seven years old fail to reach. DeepCtrls has received investment from a major internet strategic investor, HSG, and Source Code Capital.

01 Don't Underestimate Air Conditioning Control

Human control of air conditioning systems has evolved through several iterations.

The earliest and most widespread approach is a fixed set of "IF-THEN" rules—for instance, IF the target temperature is X, THEN activate five chillers. These rules are typically based on the experience of veteran technicians, vary by manufacturer and product line, and require professionals to program case by case. The standardization is low, data interoperability is poor, and most critically, this approach cannot deliver real-time energy-optimal solutions. Over 90% of air conditioning control systems domestically and internationally still use this expert experience-based method.

After AI emerged, Google and major domestic internet companies began rolling out AI energy optimization algorithms. These systems feed all historical adjustment data into AI training algorithms to calculate an optimal solution. The flaw in this approach is that the algorithm's output is derived from summarizing and training on past "veteran technician" adjustment behaviors—it's a passive algorithm constrained by the quality of historical operational data and conditions. Its ceiling is the collective wisdom of all veteran technicians. If the system never operated at an optimal state historically, the AI cannot conjure one from nothing.

"Could we combine physics-based modeling with AI data-driven methods to find the true optimal solution?" In 2018, Dr. Hui Li, then a researcher at Lawrence Berkeley National Laboratory in the United States, found the answer.

A "physics-based model," using chillers as an example, captures the inherent performance characteristics of the equipment. What Li needed to develop was a high-precision modeling method that preserved these chiller characteristics while remaining trainable by AI—then, based on this high-precision model, exhaustively search and compute the real-time optimal solution for system operation.

After validating the algorithm across several projects, Li founded DeepCtrls with several graduate students and developed the product "DeepCtrls." DeepCtrls first performs high-precision simulation modeling and prediction of air conditioning system equipment. Based on real-time cooling loads, it calculates the energy consumption corresponding to each of the hundreds of thousands of possible operational combinations in a simulated environment, identifies the control parameters with the lowest energy consumption, and dispatches them to the equipment controllers to achieve real-time optimal system operation.

DeepCtrls consists of three components:

First, the hardware DeepBox edge server, which handles all data collection and output command execution locally. DeepBox supports numerous communication protocols, enabling data collection and integration with third-party equipment and systems.

Second, the DeepLogic algorithm module development platform—the modeling and AI computation platform responsible for core data processing, model building and prediction, scheduling, diagnostics, and alarms. Notably, DeepLogic is a no-code platform where users can build applicable algorithm modules through drag-and-drop.

Third, the DeepSight functional visualization platform, where clients can observe real-time energy savings and modularly build and extend various functions and visualization modules.

In simple terms, the workflow is: DeepBox collects data, DeepLogic calculates the optimal control commands and returns them to DeepBox for dispatch to controllers for execution, and results are viewed and validated on DeepSight.

Li told 36Kr that based on high-precision simulation modeling, DeepCtrls must perform hundreds of thousands of comparative calculations for each operating condition to find the energy-optimal control commands—something expert experience-based rules simply cannot achieve.

02 What Does a Profitable Enterprise Software Company Look Like?

Compared to many companies that overemphasize core technology without commercial traction, DeepCtrls' distinguishing feature is translating core technology into product form with a sound business model. Li told 36Kr that DeepCtrls was profitable within its first year, and this year's software contract revenue is poised to break the RMB 100 million threshold.

In summary, DeepCtrls' commercial strength stems from several factors: making products standardized and lightweight, enabling cross-industry scalable deployment, and letting clients see results.

Li told 36Kr that both DeepLogic and DeepSight are low-code platforms, with standardization as their chief benefit. The standard deployment and commissioning timeline for DeepCtrls is two person-months, yet delivers energy savings worth millions to clients. DeepCtrls is positioned as a PaaS+SaaS platform—no R&D personnel are needed for project customization, only project configuration.

Meanwhile, the DeepCtrls product is applicable across industries: factories, data centers, hospitals, commercial buildings, campuses, subway systems, and more. The application scenarios appear diverse, but all are implemented on the same underlying system—master one system, then scale and repeat.

DeepCtrls aims to deliver more than energy optimization across multiple scenarios. Conventional AI products have long been applied in safety management and operations maintenance, but mostly for advisory recommendations. By combining physics-based modeling, DeepCtrls can perform more precise prediction for safety management, preventive diagnostic maintenance, and execute commands accordingly.

Energy savings, safety management, and operations maintenance results are all visible in real time on DeepSight. Li told 36Kr that DeepLogic's equipment performance predictions typically achieve errors within 3% versus actual values—meaning the system can approach the theoretical energy optimization limit within a 3% margin. Clients can view energy data on DeepSight and switch to pre-retrofit baselines for comparison; the switch takes just 3–5 minutes, allowing clients to self-verify energy savings. Li noted that real-time verifiable energy savings are a key driver of client renewals in actual project sales.

For commercial monetization, DeepCtrls primarily promotes two models. One is energy performance contracting, where DeepCtrls shares a percentage of actual energy savings generated. The second is an annual subscription fee, where the parties agree on energy-saving targets and the client pays a fixed annual software license fee.

In its three years since founding, DeepCtrls' products and solutions have been deployed across over 110 projects nationwide. Notable cases include data centers such as the National Supercomputing Center and Tencent data centers; precision factories including BOE, CATL, ATL, and Guobo Electronics; as well as rail transit systems, energy stations, Class III hospitals, and landmark commercial buildings across multiple cities. Among these industries, factory and data center systems are the most complex, have the most stringent requirements, and face the most urgent decarbonization needs—DeepCtrls is concentrating its efforts on these two sectors.

Close examination reveals that this company's business and products are not particularly complex, but perhaps simplicity is precisely its advantage—identifying a genuine need, delivering solutions that demonstrably create value, and enabling lightweight, scalable replication and promotion. Enterprise software has long been a difficult business to profit from, yet in this highly vertical赛道, we see another possible path for enterprise services.

• MaKe | Source Code Capital Named to "2022 Most Founder-Friendly Investment Institutions / Investor Directory"

• MaKe | Source Code Capital's Yungang Huang Named to "Top 10 Young Partners"

• MaKe | Nearly 40 MaHui Members Named to "X·36Under36" S-Tier Young Founder Directory

• MaKe | Source Code Capital's Xingchen Zhang Named to Cyzone "40 Under 40 Investors"

• MaGuan | Source Code Capital's Xingchen Zhang: Technology-Driven Innovation from 0 to 1

• MaGuan | Source Code Capital's Yungang Huang: The Most Important Quality for Founders Is Resilient, Bold "Pioneering Spirit"

• MaGuan | The "Tech + Biomedical" Crossover: A New Consensus on Convergent Innovation Investment

• MaKe | Source Code Capital Completes Fifth RMB Fund Raise: Driving Technological Innovation, Empowering SMEs

• 2021: Original Aspiration Is the Future

For more MaHui members, click "Read Original" to view