Internet giants' appetite for computing power is growing, but the third-party service providers who nurtured this industry are now facing immense challenges. After years of hard work to mature the AIDC supply chain, telecom operators have entered the market armed with cost and subsidy advantages. Meanwhile, cross-over new energy giants and aggressive listed companies are rewriting the rules of the game with green power and capital leverage. As domestic infrastructure pricing enters a fierce race to the bottom, the only way out is to head overseas to Southeast Asia; with computing power profits squeezed, the only viable option is to move into power assets.
In this issue, Aifenxi interviews Chen Jun, founder of Shishi Capital, to peel back the curtain on the most brutal side of China's AIDC industry and explore how service providers can survive.
Shishi Capital is an industry-finance integration platform deeply engaged in the field of green digital infrastructure. Its business footprint covers domestic core computing hubs, as well as overseas markets such as Southeast Asia and the Middle East. To date, Shishi Capital has achieved a deep closed-loop of computing power leasing, equity investment, asset management, and power trading, making it one of the rare full-chain participants. Its founder, Chen Jun, has 20 years of experience in private equity and real asset investment. He formerly served as Managing Director of Blackstone China, and held key positions in multinational giants and top financial institutions such as Cisco Systems, Bosera Asset Management, and Invesco. With a bachelor's degree in electrical engineering from Tsinghua University and dual master's degrees in finance and management, he has a clearer view of the ultimate industrial landscape where computing power and electricity, capital and assets intersect.
Core Insights The business logic of AIDC has been profoundly rewritten by telecom operators, squeezing the value of third-party service providers. Since 2025, operators have deeply entered the market, winning over half of the orders from internet giants with their cost and subsidy advantages, and directly adopting the supply chain matured by third parties. Traditional third-party service providers must either exit the competition or be downgraded to subcontractors, bearing the risk of secondary settlement. The domestic infrastructure supply chain has been pushed to the extreme, making expansion into Southeast Asia a new window for AIDC. In China, server investment is 10 times that of data center infrastructure, a ratio of 10:1, leaving infrastructure with almost no profit margin. However, construction costs in Southeast Asia are 3 to 4 times higher than in China. Chinese companies, by shipping prefabricated components overseas, can actually earn double-digit profit margins on the infrastructure side. The path to increasing profits in AIDC lies not in "computing" but in "power"; the ultimate game is a contest of green power costs. The leasing revenue of intelligent computing rooms has been capped by internet giants, leaving the only room for profit imagination entirely in the electricity price difference. With the relaxation of the "one-to-many" green power direct connection policy, every fraction of a cent saved per kilowatt-hour will directly translate into doubled net profits for data centers. In the future, only players holding the cheapest and most stable green power will survive. Leveraging new energy genes and the financial leverage of 10,000-GPU pools, cross-over players are delivering a disruptive strike against traditional data centers. New energy giants have secured hundreds of megawatts of large orders from internet giants in the Northwest by utilizing "green power/wind turbine direct connection"; aggressive listed companies, on the other hand, are playing with tens of billions in 10,000-GPU pools through financial leasing to contribute revenue to their financial statements. Video generation and programming are driving a geometric explosion in token consumption. The explosion of AI-generated short dramas is consuming massive computing power with highly realistic videos that require no real human actors, becoming a super engine driving the surge in computing power demand; in the programming track, due to the compliance requirements of domestic large enterprise clients and financial institutions, large model vendors are required to bring chips and servers for private deployment.
The following is the transcript of this interview, slightly modified without changing the original meaning.
01. Telecom Operators Deeply Enter the Game, Rewriting AIDC Business Logic Aifenxi: With massive funds pouring into AIDC, how significant is the impact on the entire industry? Who will be the core beneficiaries? Chen Jun: The impact is enormous. Based on my observation of the industry, the core beneficiaries should be upstream equipment manufacturers, followed by telecom operators. For third-party AIDC service providers, this may not be good news.
Aifenxi: It is easy to understand why upstream equipment manufacturers benefit. What is the specific role of telecom operators in the industry chain, and why do they also benefit from it? Why might third-party AIDC service providers be negatively affected? Chen Jun: Here is the reason. Since the beginning of this year, among the bidding plans initiated by internet giants, half have been won by the three major telecom operators, which means operators have deeply entered the market. This trend actually started in 2025, with China Telecom and China Mobile securing a large number of incremental projects in the Northwest regions such as Inner Mongolia and Ningxia. Now, they completely start construction only after receiving explicit customized orders from internet giants, which is very different from the previous model. Operators possess cost and subsidy advantages, so their quotes are often relatively low, putting third-party service providers in an awkward position. Today, the influx of massive funds into AIDC will accelerate the operators' penetration. Now, third-party service providers mainly have two choices. One is to exit the competition, and the other is to become subcontractors for operators. However, becoming a subcontractor brings another troublesome issue: an additional layer of settlement risk. The overall business logic of AIDC has changed significantly compared to before.
Aifenxi: From the perspective of internet giants, computing power infrastructure might be locked in by a single supplier. How do they cope with this risk? Chen Jun: From the business strategy of internet giants, they certainly hope to maintain a balance among suppliers. They cannot allow a single supplier to monopolize the market, otherwise, they will eventually lose the initiative. Therefore, there must be considerations for checks and balances in their commercial arrangements. But whether this strategy will work remains a question mark, because operators are developing very fast, leaving a very short time window for internet giants. Over the past five or six years, after third-party service providers brought up the entire industry, operators had almost no presence in this field. Because operators are state-owned enterprises, their internal processes are time-consuming, making it difficult to keep up with the explosive computing power demand of internet giants on a monthly basis. But this time is completely different; the operators' actions are much cleaner and neater, showing great improvement from before. For example, on a project in Ningxia, it took them only 56 days from civil engineering to delivery. This level of service and delivery speed is the result of third parties maturing the entire design, procurement, and construction supply chain system over the years, which operators directly took over and pushed to the extreme.
02. The 1:10 Infrastructure Code: When Domestic Competition is Fierce, Head to Southeast Asia Aifenxi: What are the essential differences between AIDC and traditional data centers? Chen Jun: If we break down AIDC, it is essentially AI plus DC. The data center part, whether in terms of power density or physical space, is largely similar to traditional internet data centers in terms of overall logic and supply chain, except that the current power density is much higher than before. The reason everyone calls it AIDC now is that most people playing in the AIDC field have to deeply participate in the buying, holding, and leasing of servers. When service providers did traditional data centers in the past, they didn't need to worry about servers at all; internet giants would procure chips and servers themselves, and service providers were only responsible for providing the data center infrastructure. But now, due to the technological friction between China and the US, the acquisition of various Nvidia chips faces restrictions, forcing many data center companies to step in and hold servers on clients’ behalf, bundling servers with infrastructure for sale, leading to a significant increase in computing power capital expenditures.
Aifenxi: What is the approximate ratio between pure infrastructure investment and server hardware investment in AIDC? Chen Jun: In China, the investment in servers is basically 10 times the investment in data center infrastructure. Let's do the math. Currently, a standard intelligent computing cabinet is about 10KW. The infrastructure cost to build one cabinet in China is about 150,000 RMB. But putting in high-performance servers with Nvidia's power rating, the hardware cost alone is over 2 million RMB. The bulk of the investment is entirely in hardware and chips.
Aifenxi: Why can the proportion of infrastructure investment in overseas computing center construction reach as high as 30% to 40%, while the gap in China is so large? Chen Jun: The core reason is that China has pushed the infrastructure supply chain to the extreme, and domestic manufacturing costs are too cheap. For data center infrastructure, what costs 1 RMB to get done in China basically requires 1 USD in the US, which is a huge difference. Therefore, in the US, data center infrastructure investment excluding servers is very substantial and comparable to server costs. Domestic infrastructure is too cheap, so now China's data center supply chain is going overseas frantically. They assemble the entire computer room using prefabricated components directly inside containers and then ship them all over the world.
Aifenxi: If AIDC service providers choose to go overseas, are they also following this logic of low-cost Chinese manufacturing? Chen Jun: Yes, everyone in China has been competing so fiercely that they are barely making money, so they can only take the same funds to build in Southeast Asia. Construction costs in Southeast Asia are basically 3 to 4 times higher than in China. Even though shipping increases costs, there is still a decent profit margin. In China, infrastructure costs 1 unit and servers cost 10 units; now, that 1 unit for infrastructure has been competed down to 0.9 units. But in Southeast Asia, infrastructure can be sold for 3 to 4 units, while servers remain at 10 units. Going overseas actually makes more money on the data center infrastructure side.
Aifenxi: Holding servers on clients’ behalf requires a solid financial foundation. Besides operators, are only the established top service providers like GDS, Chindata, and ChinaRun capable of playing this game? Chen Jun: Besides these two types of players, many A-share listed companies are also frantically rushing in. Currently, those playing well in the AIDC field in China are either state-owned enterprises with deep backgrounds or extremely aggressive listed companies. Some listed companies that originally did not come from the data center background want to build 10,000-GPU pools as soon as they enter. A 10,000-GPU pool requires an investment of tens of billions. They use the leverage of listed companies, only needing to put up 10% to 20% of their own funds, with the rest entirely relying on financial leasing. Since the supply of high-quality computing power is currently limited, and the clients renting computing power are internet giants, as long as this money is invested and the servers go online, it can immediately bring extremely substantial cash flow and revenue, which looks very good on the financial statements of listed companies. In contrast, traditional pure computer room construction takes at least over a year from land acquisition and energy assessment to completing all processes, which is too long a cycle.
Aifenxi: Who are the top service providers in the domestic AIDC field? What are their competitive barriers? Chen Jun: From the perspective of large-scale base-type, next-generation AIDC, ChinaRun, China Golddata, China Unidata, and Chindata are several with strong capabilities. VNET has also been catching up very fast recently; its controlling shareholder holds a very substantial portfolio of new energy assets, which is also a huge advantage. Another category that absolutely cannot be ignored is the new quality productivity players who crossed over with new energy genes from day one. For example, a certain new energy giant secured a bidding project of hundreds of megawatts from a top internet giant on its very first attempt, very cleanly and decisively. If you have the opportunity to visit the Gobi Desert in Ulanqab or Zhongwei, Ningxia, the vast expanse of intelligent computing bases and new energy power stations are no less visually and industrially shocking than the real estate era back in the day. Therefore, whether for old-school service providers or cross-over players, the core competitive barrier is energy, specifically green power. The ultimate outcome will definitely be that whoever holds the cheapest and most stable green power will be the one to survive and see the dawn in this round of massive computing power reshuffle.
03. The Path to Increasing AIDC Profits Lies in Power Aifenxi: For AIDC service providers, what is the profit situation? Chen Jun: For pure data center operations in China, a return on net assets of 6% is already very good, while going overseas can achieve double digits.
Aifenxi: Are there still newly approved AIDC quotas now, or is everyone digesting the past inventory? Chen Jun: Regarding quotas, they are still newly approved once a quarter. What everyone is currently building and digesting is mostly the window-period quotas reported previously.
Aifenxi: Besides traditional computer room and server leasing, there are now many downstream manufacturers specializing in computing-power-electricity coordination software. How much value will they have in the industry chain in the future? Chen Jun: This opportunity entirely depends on the transformation of the power architecture of data centers, namely green power direct connection, which means directly connecting computing centers with new energy power stations. Previously, policies stipulated that direct connection must be under the same entity, meaning the data center and the new energy power station are under the same company name; now the policy has been relaxed, allowing two different entities to conduct one-to-many direct connection and settlement. This brings up a new question: should those playing in AIDC step in to invest in new energy power stations themselves? Since the leasing yield of the computer room itself has been tightly capped by internet giants, the only room for profit imagination lies entirely in the electricity price difference. If you can save 0.1 RMB per kilowatt-hour through green power direct connection, even if you give 0.05 RMB to internet giants, the 0.05 RMB you keep is pure profit. In data centers, costs are fixed; the few cents saved per kilowatt-hour, accumulating over time, can directly translate into doubled net profits for the data center. Therefore, everyone is willing to step in and get involved in the power business.
Aifenxi: Currently, ByteDance alone consumes nearly half of the domestic computing power demand, and Alibaba takes half of the remaining half. The rest fight for the meager remaining share. Will this monopoly pattern by internet giants persist in the future? Chen Jun: Looking at the global scope, this is basically the trend. In the US, it is also monopolized by a few clouds like Amazon, Microsoft, and Google. However, if domestic large model companies really run their applications successfully, the explosion in their computing power demand cannot be underestimated. More and more enterprises in China can now build 10,000-GPU pools. A 10,000-GPU pool is roughly a thousand high-performance servers. This is not only a standard configuration for large model companies but also a rigid demand for cutting-edge industries like autonomous driving and robot brains. A 10,000-GPU pool sounds intimidating, but if converted to actual power, it only requires about 15 to 20 megawatts of power capacity. In contrast, internet giants' bidding often involves large orders of hundreds of megawatts, so the volume of a single 10,000-GPU pool is actually not that large.
Aifenxi: Following the demand of internet giants, ByteDance internally has very high expectations for token consumption and revenue brought by AI applications. How feasible is this? Chen Jun: Aggressive expectations are not groundless, because the token consumption and business growth brought by large models are not changing on a yearly basis, but undergoing earth-shaking changes on a daily and monthly basis. Currently, the domestic AI applications that can truly generate massive cash flow and high-frequency user consumption focus on two core directions: one is programming, and the other is video generation. Especially in video generation, an extremely shocking industry transformation is taking place. Since the beginning of this year, domestic AI-generated short dramas have begun to explode completely. Many crews that originally filmed traditional short dramas, and even a large number of mid-tier actors, have faced collective unemployment in a short time. Now, a large number of hit short dramas on the market no longer require real humans to act at all; they are entirely highly realistic videos generated by computers. The token consumption brought by video generation, compared to traditional text and dialogue, is a geometric multiple increase. The commercialization path in the programming track is very clear. Because domestic large enterprise clients and financial institutions do not accept the billing model of public cloud APIs, every bidding requires large model vendors to bring domestic chips and servers to bid, deploying the models offline and privately within the enterprises.
Aifenxi: Since downstream profits are tightly capped by internet giants and upstream chips are monopolized, are there still new opportunities in the AIDC track in the future? Chen Jun: Opportunities still exist. The national annual infrastructure investment in data center infrastructure is hundreds of billions of RMB. If servers are included, it is an even larger procurement market of trillions of RMB. In this huge cake, those doing liquid cooling, prefabricated components, and new types of power supplies can all make big money by adapting to technological changes in the next two or three years, giving rise to many listed companies with tens of billions in market value. As an investment institution, our strategy in this industry is becoming increasingly focused, mainly carrying out standardized operations across the entire life cycle around asset portfolios: from early development permits, construction, leasing loads to final asset securitization. At the same time, we extend to upstream and downstream. Upstream looks at energy; besides the current wind and solar green power, we have already seen top players getting involved in micro nuclear reactors. One reactor is 200 megawatts, which can provide extremely cheap and stable green power; downstream directly looks at chips.