**中国AI基础设施热潮 | 数据中心 | 东数西算 | 字节跳动 | 阿里云 | 腾讯云 | 人工智能 | 中国科技 | 模块化数据中心 | 资本开支 | 华为昇腾 | 昆仑芯 — Transcript & Summary | SozAI** Source: https://sozai.app/transcript/china-ai-infrastructure-data-centers/ An in-depth analysis of China's booming AI-driven data center market, highlighting capacity, growth, key players, and industry dynamics. ## Key Takeaways - China's data center market is much larger and growing faster than commonly estimated, driven by AI demand. - ByteDance is the largest single customer in the wholesale colocation market, influencing industry order cycles. - Legacy telecom infrastructure contributes to high vacancy rates despite strong demand for AI-capable facilities. - State-owned enterprises and telecom carriers still hold significant market share, shaping market structure. - Modular data center construction and regional policy initiatives are key enablers of rapid deployment. ## What the video covers - China's data center capacity is over 24 gigawatts, larger than Europe and the rest of Asia combined, with rapid AI-driven growth. - Major hyperscalers like Alibaba, Tencent, Baidu, and especially ByteDance are driving unprecedented capital expenditure and capacity expansion. - The market features a large number of non-listed players and significant involvement from state-owned enterprises and telecom carriers. - China's data center construction is fast, leveraging modular designs pioneered by companies like Tencent since 2014. - Vacancy rates remain high due to legacy infrastructure from the telecom era, which is incompatible with modern AI computing needs. - The 'East Data, West Computing' initiative promotes regional hubs with favorable policies and cheap green electricity to balance supply. - Chinese hyperscalers are expanding overseas, with leased capacity expected to double by 2029, supplemented by rented GPUs from Western providers. - Price competition is intense, and export controls on chips pose challenges, but AI data center buildout continues at a remarkable pace. - Historical context shows a shift from telecom-dominated retail data centers to hyperscale cloud providers driving market dynamics. - Recent policy and economic shifts have slowed growth rates and capital expenditures but have not stopped the overall expansion. Answers ## Questions about this video How large is China's data center capacity compared to other regions? China's data center capacity exceeds 24 gigawatts, making it larger than all of Europe combined and the rest of Asia excluding China. Who are the major players driving data center growth in China? Major players include Alibaba, Tencent, Baidu, and ByteDance, with ByteDance being the largest wholesale colocation customer. Why are vacancy rates high despite strong AI demand in China? High vacancy rates stem from legacy telecom-era data centers with outdated infrastructure that cannot support modern AI computing needs. ## Full Transcript — Download SRT & Markdown 00:00 Speaker A Hello everyone, this is Best Partners. I’m Da Fei. Just how wide is the gap in external estimates of China's data center capacity? The answer is 15 times. Figures from different institutions vary by more than an order of magnitude, and almost everyone is talking about high vacancy rates. China is huge, and China is empty—these two assumptions seem to hold true without even needing proof. In today's episode, we’re going to talk about an in-depth report released by SMAnalysis. They did one thing: they tracked over 1,000 data centers from more than 60 operators in China from the bottom up, mapping out the real scale, structure, speed, and player distribution. The conclusion is straightforward: China’s data center capacity is larger than everyone imagines, and it is being filled at an extremely high speed by AI. Let’s start with an overall benchmark: in the global data center market, the US has 50 gigawatts of capacity as of the end of 2026, followed by Asia-Pacific excluding China at about 15 gigawatts, and Europe at about 14 gigawatts. Latin America has about 2 gigawatts; meanwhile, China is over 24 gigawatts—larger than all of Europe and larger than the rest of Asia combined. This figure doesn't even include the 20 gigawatts of existing pipeline or the 30 gigawatts already announced. Growth is currently at a turning point. In Q2 2026, the combined capital expenditure of Alibaba, Tencent, and Baidu reached $20 billion, doubling year-on-year. For the first time on record, all three companies saw negative free cash flow—the largest surge in capital expenditure in the industry's history. And this doesn't even count the biggest one of all: ByteDance. ByteDance is a private company and does not disclose financial reports, but according to SMAnalysis’s China data center model, ByteDance alone accounts for about one-fifth of the nation's delivered capacity, and almost all of it is leased. In other words, ByteDance is the single largest customer for all wholesale colocation operators in the country. Their bidding rhythm is essentially the metronome for the entire industry’s orders. Public companies get the most attention, but they are just the tip of the iceberg; GDS and VNET are the only two US-listed Chinese data center operators. In the first half of 2026, they signed 1.3 gigawatts of wholesale orders. That sounds like a lot, doesn't it? But according to the SMAnalysis model, from 2024 to mid-2026, these two companies captured less than one-third of the orders from ByteDance and Alibaba. There are a vast number of non-listed players beneath the surface. Another feature of this market is the share of state-owned entities. Historically, China’s data center market was dominated by telecom carriers, with the three major state-owned operators holding 60-70% of the market in the 2010s, and they still control one-third of the nation’s capacity today. The State Grid Corporation is also investing heavily, with its investment during the 14th Five-Year Plan exceeding original targets by 24%. From 2026 to 2030, this is set to increase by another 40%, reaching a total scale of over $746 billion, or 5 trillion RMB. Construction speed is also frequently underestimated; in China, 100MW data center facilities can be delivered in under 12 months. Modular data centers are standard practice, and Tencent deployed its third-generation modular design as early as 2014. China has been using these methods for years, while the U.S. is only now beginning to adopt them on a large scale. Growth is not limited to the domestic market; the overseas leased capacity of Chinese hyperscalers is projected to double from 2026 to 2029, reaching nearly 4 gigawatts. Furthermore, this figure underestimates their overseas computing power, as it doesn't account for the hundreds of thousands of GPUs they rent from Western cloud providers. The market is not without concerns; vacancy rates are indeed high, price competition among data center developers is intense, and chip supply is constrained by export controls. Yet, none of this has stopped AI data centers from being built and equipped at an astonishing pace. To understand the strange phenomena of today’s Chinese data center market, we must first look at how it came to be. The U.S. data center industry has been hyperscale-oriented from the start, with megawatt-scale campuses existing for many years. These were tailor-made for a few giant clients. China is different; its industry grew out of telecom carriers and thousands of small retail deployments. This history explains why high vacancy rates and AI computing shortages can coexist today, and why carriers still hold one-third of the market. Early on, it was the era of carrier dominance. Before roughly 2015, unlike in other parts of the world, China’s data center market was historically dominated by telecom firms. China Mobile, China Telecom, and China Unicom, the three state-owned carriers, collectively held 60% to 70% of the market share. At the time, the largest third-party players, GDS and VNET, each had less than 5% market share. Carriers had structural advantages in licensing and networks, along with natural vendor lock-in; the government, state-owned enterprises, and regulated industries naturally preferred renting from state-owned operators. Data centers at that time were essentially telecommunications real estate, with low-density retail racks where 90% of equipment had a power density under 2,000 watts. Clients were mainly websites and CDNs, usually bundled with bandwidth, and services were provided by small urban exchange nodes around Beijing, Shanghai, and the Yangtze River Delta. Third-party operators were largely just resellers of carrier space and bandwidth. Hyperscalers hadn't really taken off back then; Alibaba Cloud was only founded in 2009, and Tencent Cloud didn't open to the public until 2013. Why is this history important today? Because the capacity built back then is precisely what is sitting empty now. Old retail rack specifications can no longer meet the requirements for H20 or high-end servers, and the cost of retrofitting partially occupied space is too high, making upgrades unrealistic. The roughly 50% utilization rate often seen in the market is largely a legacy left over from that era. Next came the land-grab phase of the cloud era. From 2015 to 2021, industry leaders Alibaba and Tencent both believed this was a winner-takes-all market, sparking a capital expenditure arms race. Alibaba Cloud's revenue between 2015 and 2019 saw a compound annual growth rate of about 111%, at the cost of brutal price wars and persistent EBITDA losses. Core product prices dropped by over 50%, with 17 price cuts occurring over 12 months. Tencent also followed suit, pushing the combined capital expenditure of hyperscalers to a peak in 2020. Third-party colocation platforms expanded alongside them—GDS and VNET rose, while retail developers kept building small-scale urban data centers, betting that cloud demand would be infinite. By 2022, the music stopped, and the frantic expansion came to an abrupt halt. With policy headwinds, a macroeconomic slowdown, and the rising market share of state-owned clouds, these factors combined to drop Alibaba Cloud's growth rate from 50% in 2021 to 23% in 2022, and later to almost zero growth. Tencent's management defined it as a shift from pursuing revenue growth to healthy growth, with its own growth rate also slowing significantly. In mid-2022, they slashed their capital expenditure budget by nearly half, and made similar adjustments in 2023. At almost the same time, in February 2022, the NDRC issued guidelines for the "East Data, West Computing" project, establishing eight national computing hubs and ten clusters. These regions, especially in the west, enjoy fast-track approval channels and cheap green electricity. The result is that even though the delivery pace has slowed down, supply remains in excess. The biggest impact was on pricing, where market rent 00:11 Speaker A , and almost everyone is talking about high vacancy rates. China is huge, and China is empty—these two assumptions seem to hold true without even needing proof. In today's episode, we’re going to talk about an in-depth report released by SMAnalysis. They did one 00:23 Speaker A thing: they tracked over 1,000 data centers from more than 60 operators in China from the bottom up, mapping out the real scale, structure, speed, and player distribution. The conclusion is straightforward: China’s data center capacity is larger than everyone 00:37 Speaker A imagines, and it is being filled at an extremely high speed by AI. Let’s start with an overall benchmark: in the global data center market, the US has 50 gigawatts of capacity as of the end of 2026, followed by Asia-Pacific 00:48 Speaker A excluding China at about 15 gigawatts, and Europe at about 14 gigawatts. Latin America has about 2 gigawatts; meanwhile, China is over 24 gigawatts —larger than all of Europe and larger than the rest of Asia combined. This figure doesn't even include the 20 01:02 Speaker A gigawatts of existing pipeline or the 30 gigawatts already announced. Growth is currently at a turning point. In Q2 2026, the combined capital expenditure of Alibaba, Tencent, and Baidu reached $ 20 billion, doubling year-on-year. 01:14 Speaker A For the first time on record, all three companies saw negative free cash flow —the largest surge in capital expenditure in the industry's history. 01:22 Speaker A And this doesn't even count the biggest one of all: ByteDance. ByteDance is a private company and does not disclose financial reports, but according to SMAnalysis’s China data center model, ByteDance alone accounts for about one-fifth of the nation's delivered 01:34 Speaker A capacity, and almost all of it is leased. In other words, ByteDance is the single largest customer for all wholesale colocation operators in the country. Their bidding rhythm is essentially the metronome for the entire industry’s orders. Public companies get the most attention, but 01:48 Speaker A they are just the tip of the iceberg; GDS and VNET are the only two US-listed Chinese data center operators. In the first half of 2026, they signed 1.3 gigawatts of wholesale orders. That sounds like a lot, doesn't it? But 01:59 Speaker A according to the SMAnalysis model, from 2024 to mid-2026, these two companies captured less than one-third of the orders from ByteDance and Alibaba. 02:06 Speaker A There are a vast number of non-listed players beneath the surface. Another feature of this market is the share of state-owned entities. Historically, China’s data center market was dominated by telecom carriers, with the three major state-owned operators holding 60-70%of the market in the 02:18 Speaker A 2010s, and they still control one-third of the nation’s capacity today. The State Grid Corporation is also investing heavily, with its investment during the 14th Five-Year Plan exceeding original targets by 24%. From 2026 to 2030, this is set to increase 02:30 Speaker A by another 40%, reaching a total scale of over $ 746 billion, or 5 trillion RMB. Construction speed is also frequently underestimated; in China, 100MW data center facilities can be delivered in under 12 months. Modular data centers are standard practice, and 02:43 Speaker A Tencent deployed its third-generation modular design as early as 2014. China has been using these methods for years, while the U.S. is only now beginning to adopt them on a large scale. Growth is not limited to the domestic market; the 02:54 Speaker A overseas leased capacity of Chinese hyperscalers is projected to double from 2026 to 2029, reaching nearly 4 gigawatts. Furthermore, this figure underestimates their overseas computing power, as it doesn't account for the hundreds of thousands of GPUs they rent from Western cloud providers. The 03:06 Speaker A market is not without concerns; vacancy rates are indeed high, price competition among data center developers is intense, and chip supply is constrained by export controls. Yet, none of this has stopped AI data centers from being built and equipped 03:17 Speaker A at an astonishing pace. To understand the strange phenomena of today’s Chinese data center market, we must first look at how it came to be. The U.S. data center industry has been hyperscale-oriented from the start, with megawatt-scale campuses existing 03:28 Speaker A for many years. These were tailor-made for a few giant clients. China is different; its industry grew out of telecom carriers and thousands of small retail deployments. This history explains why high vacancy rates and AI computing shortages can coexist today, 03:43 Speaker A and why carriers still hold one-third of the market. Early on, it was the era of carrier dominance. Before roughly 2015, unlike in other parts of the world, China’s data center market was historically dominated by telecom firms . China Mobile, China Telecom, and 03:56 Speaker A China Unicom, the three state-owned carriers, collectively held 60%to 70%of the market share. At the time, the largest third-party players, GDS and VNET, each had less than 5%market share . Carriers had structural advantages in licensing and networks, along with 04:09 Speaker A natural vendor lock-in; the government, state-owned enterprises, and regulated industries naturally preferred renting from state-owned operators. Data centers at that time were essentially telecommunications real estate, with low-density retail racks where 90%of equipment had a power density under 2,000 watts. Clients were mainly 04:23 Speaker A websites and CDNs, usually bundled with bandwidth, and services were provided by small urban exchange nodes around Beijing, Shanghai, and the Yangtze River Delta. Third-party operators were largely just resellers of carrier space and bandwidth. Hyperscalers hadn't really taken off back then; Alibaba 04:38 Speaker A Cloud was only founded in 2009, and Tencent Cloud didn't open to the public until 2013. Why is this history important today? Because the capacity built back then is precisely what is sitting empty now. Old retail rack specifications can no longer meet the 04:50 Speaker A requirements for H20 or high-end servers, and the cost of retrofitting partially occupied space is too high, making upgrades unrealistic. The roughly 50%utilization rate often seen in the market is largely a legacy left over from that era. Next came the 05:02 Speaker A land-grab phase of the cloud era. From 2015 to 2021, industry leaders Alibaba and Tencent both believed this was a winner-takes-all market, sparking a capital expenditure arms race. Alibaba Cloud's revenue between 2015 and 2019 saw a compound annual growth rate of 05:16 Speaker A about 111%, at the cost of brutal price wars and persistent EBITDA losses. Core product prices dropped by over 50%, with 17 price cuts occurring over 12 months. Tencent also followed suit, pushing the combined capital expenditure of hyperscalers to a peak 05:30 Speaker A in 2020. Third-party colocation platforms expanded alongside them—GDS and VNET rose, while retail developers kept building small-scale urban data centers, betting that cloud demand would be infinite. By 2022, the music stopped, and the frantic expansion came to an abrupt halt. With policy 05:44 Speaker A headwinds, a macroeconomic slowdown, and the rising market share of state-owned clouds, these factors combined to drop Alibaba Cloud's growth rate from 50%in 2021 to 23%in 2022, and later to almost zero growth. Tencent's management defined it as a shift from 05:57 Speaker A pursuing revenue growth to healthy growth, with its own growth rate also slowing significantly. In mid-2022, they slashed their capital expenditure budget by nearly half, and made similar adjustments in 2023. At almost the same time, in February 2022, the NDRC issued 06:10 Speaker A guidelines for the "East Data, West Computing" project, establishing eight national computing hubs and ten clusters. These regions, especially in the west, enjoy fast-track approval channels and cheap green electricity. 06:21 Speaker A The result is that even though the delivery pace has slowed down, supply remains in excess. The biggest impact was on pricing, where market rents for dedicated power were cut in half from about $ 80 per kW per month. GDS 06:31 Speaker A management stated on their August 2026 earnings call that it would take about 18 more months to fully digest the price reset. While this had actually been going on for years, everything changed starting in 2024; the rise of DPUs/GPUs alongside hyperscalers 'own 06:46 Speaker A generative AI ambitions reawakened the entire market between 2024 and 2025. The combined capital expenditure of ByteDance, Alibaba, Tencent, and Baidu rose from about $ 35 billion in 2024 to over $ 50 billion in 2025. The trend line for 2026 is heading toward $ 100 07:02 Speaker A billion, wholesale orders have returned at an unprecedented scale, and western hubs have finally found their tenants. 07:07 Speaker A GDS and VNET signed a combined total of about 1.3 GW of wholesale capacity in the first half of 2026, while retail orders were less than 10 MW. The AI era is an era of wholesale; the old retail inventory simply hasn't been part of it 07:18 Speaker A . Looking back at these four periods of history basically explains the Chinese data center market today. There is a batch of retail-oriented legacy racks around first-tier cities with utilization rates between 30%and 50%, which are physically unsuitable for high-density AI demands. At the same 07:32 Speaker A time, wholesale-driven AI construction is taking place entirely elsewhere. While legacy retail is in a structural decline, hyperscaler self-builds and third-party wholesale have taken over the growth baton. Next, let's talk about "East Data, West Computing." This policy planned eight major national 07:47 Speaker A computing hubs, landing into ten clusters; it covers both demand centers , such as Beijing-Tianjin-Hebei, the Yangtze River Delta, and the Greater Bay Area, as well as resource-rich western regions like Inner Mongolia, Ningxia, Gansu, and Guizhou. This is the top-level design blueprint for the 08:00 Speaker A entire industry, and unlike many guidelines, this policy has truly changed the way the industry develops. 08:05 Speaker A The policy itself has been reported on many times, but what few have explained clearly is its transmission mechanism: after a data center developer in China acquires land, the first permit to apply for is an energy consumption review submitted to the local 08:15 Speaker A Development and Reform Commission. This energy consumption quota is the limiting factor for how large a facility can be built. Any project with an annual comprehensive energy consumption exceeding 10,000 tons of standard coal—which converts to roughly 10 MW of power or more. Without 08:27 Speaker A passing this review, construction cannot legally begin. This review system has been in place since 2010, and it later gradually evolved into a quota tool. After the "East Data, West Computing" project was introduced in 2022, the coordination system became 08:37 Speaker A the most effective implementation mechanism, as the departments issuing the quotas were the same ones enforcing the policy. The result was that non-hub regions basically stopped receiving any new construction quotas. Guangdong is the clearest example; the province suspended data center energy quotas 08:51 Speaker A from 2021 to 2022. When they reopened, it was only for the Shaoguan hub, with similar restrictions applied to other major eastern cities. In contrast, western hub regions offered generous energy quotas, land approvals, and tax incentives. While the mechanism was in 09:05 Speaker A place, migration remained slow for the first two years. Why? Because mainstream demand—like cloud computing, gaming, and e-commerce— required proximity to users and low latency, so hyperscalers continued to scramble for quotas around Tier 1 cities. There is one overlooked event 09:20 Speaker A that perfectly illustrates how this coordination system operates in practice. In mid-2023, the Guangdong Energy Bureau inspected two of Tencent’s self-built parks in Qingyuan and found that each was operating at about 13 to 14 times its approved scale. This was labeled as " 09:34 Speaker A small approval, large construction" and ordered to be rectified within six months. The local power grid was also instructed to supply electricity based only on approved quotas, meaning even for Tencent, new capacity had to be built in province-designated hub nodes. 09:44 Speaker A Those parks are still operating today, but tracking models show that Tencent’s subsequent new capacity has indeed moved to Shaoguan. The reason isn't just regulatory pressure. Two things fundamentally changed the landscape: first, telecom-led fiber investment slashed latency. Shaoguan is 09:58 Speaker A a prime example; since 2022, the local government has invested about 600 million USD (4 billion RMB) in network infrastructure, building 13 ultra-high-speed 400G all-optical lines covering the Greater Bay Area. Now, one-way latency from Shaoguan to Guangzhou is 1.3 milliseconds, to 10:12 Speaker A Shenzhen 1.66 milliseconds, and to Hong Kong less than 3 milliseconds. The plan also includes a Shenzhen-Guangzhou-Shaoguan hollow-core fiber corridor, which could further reduce latency by about 30%after 2027. 10:22 Speaker A At this level, except for the most latency-sensitive workloads, a server rack in Shaoguan is essentially equivalent to one in Shenzhen. Then, AI demand finished the job; training workloads don't care about latency at all, but they care deeply about 10:33 Speaker A electricity prices. Inner Mongolia is the best example. This region boasts China's largest installed capacity of renewable energy, and crucially, western Inner Mongolia uses the Mengxi Power Grid—the only provincial-level independent grid outside the State Grid and China Southern Power Grid, 10:46 Speaker A featuring its own multilateral trading mechanism and independent pricing system. The result is that Inner Mongolia has become China's Johor, the default destination for hyperscale AI capacity. There are abundant energy consumption quotas here, land measured by the square kilometer, and most 10:57 Speaker A importantly, electricity prices that are about half those of first-tier cities. The comparison to Johor goes beyond just land and power. In terms of economic returns, Chinese rents are much lower, but construction costs are also lower; in Johor, rents are higher, 11:08 Speaker A but construction costs are higher too. Data shows that the cost-yield ranges of the two regions actually overlap. 11:14 Speaker A The best parks in Inner Mongolia have returns comparable to international developers in Johor, but the trends differ: Inner Mongolia's construction costs are still falling, driven by prefabrication and domestic substitution, while developers in Johor face capital expenditure inflation and 11:26 Speaker A cost-overrun risks from an overheated supply chain. Chinese operators building in Johor can also leverage the Chinese supply chain to enjoy geographic arbitrage and achieve higher yields. Zooming out from Inner Mongolia , the stories of the ten clusters are 11:38 Speaker A actually quite different. If we weight them by actual deployed and committed capacity, rather than policy-planned scope, the differences become even more apparent. There is a common narrative in the media that the "East Data, West Computing" project is just a vanity 11:51 Speaker A project, with a bunch of empty shells built in the west; the simplest rebuttal is to look at where hyperscale vendors are actually investing their money. According to tracking data, ByteDance currently operates three self-built clusters: Datong in Shanxi was the earliest; Wuhu in Anhui belongs 12:03 Speaker A to the "East Data, West Computing" cluster, serving East China inference; and there are the dual-site projects in Inner Mongolia, located in Horinger and Ulanqab. In July 2026, ByteDance also registered two Ningxia entities in Zhongwei; since Zhongwei is also a node 12:14 Speaker A of "East Data, West Computing," this points to a fourth self-built cluster. Beyond Inner Mongolia, the development of each cluster is indeed highly divergent, but much of this divergence was by design. Gui'an, for instance, was positioned as a storage hub from 12:25 Speaker A the very beginning; cold data doesn't care if the mountains are high and the roads are far. The most famous anchor tenant is Apple's iCloud China data center, operated by Cloud Guizhou. Less reported is the fact that both Huawei 12:34 Speaker A and Tencent have self-built clusters in this region. Chengdu and Chongqing present a different situation. They were positioned as national computing hubs, but in the competition for AI demand, they lost to Inner Mongolia due to electricity prices and power grid 12:45 Speaker A flexibility. The "East Data, West Computing" project successfully concentrated construction within the planned map, and the market then selected winners within that territory. 12:52 Speaker A Shanxi is the opposite; it is the largest data center region outside the "East Data, West Computing" hub system, yet it has attracted a committed capacity of nearly two gigawatts from ByteDance, Baidu, and JD.com combined. 13:01 Speaker A These include both self-built facilities and colocation leases. The reason is straightforward: it has abundant, low-cost electricity. Shanxi is a major coal-producing province in China, close to Beijing, and holds a first-mover advantage. Hyperscalers went there even before "East Data, West 13:12 Speaker A Computing" was launched and have been expanding ever since. This shows that the quota system does guide construction, but ultimately, power and geography are the deciding factors. Now , let's talk about speed and costs. 13:23 Speaker A Data center construction in the U.S. is bottlenecked by the power side, including grid interconnection queues, long transformer delivery lead times, and various behind-the-meter solutions. 13:29 Speaker A China’s bottleneck is on the chip side. Power generation is sufficient, and utility companies view data center loads as growth drivers. Approval is a policy pathway rather than a bottleneck , and construction labor is abundant. 13:40 Speaker A There is no physical friction in facility construction; the cycle is driven entirely by internet platform demand. In terms of speed, the standard delivery time for a 100-megawatt facility has shrunk from about 18 months in the cloud era to about 12 13:50 Speaker A months in the AI era, with top-tier projects being even faster. For reference, 300Analysis previously conducted in-depth research on U.S. 13:57 Speaker A modular data centers. Even the best modular construction in the U.S. still takes 12 to 18 months, plus an additional 12 to 13 months for approvals, which cannot overlap. In China, pre-construction approvals typically take three to six months, and 14:09 Speaker A even less in "East Data, West Computing " regions. Several structural factors contribute to this faster speed: construction begins before the contract is finalized. Since hyperscaler tender requirements now demand a "T + 6" delivery—meaning delivery within six months of winning the bid—developers 14:24 Speaker A take on the risk and start civil works first. As a result, tenant competition has effectively turned into a race between semi-finished shells. GDS signed a massive 152-megawatt AI deal in early 2025; the client required delivery within six months, and they 14:36 Speaker A actually pulled it off. Greenfield construction in western hubs has also changed the way things are built. 14:41 Speaker A Moving away from the land constraints of tier-one cities, developers have shifted from traditional multi-story concrete structures to low-rise, long-span steel factories. The steel is produced in factories while foundation excavation happens on-site, after which it is transported to the site and 14:52 Speaker A hoisted into place. Concrete is only used for the underground portion, so there is no need to wait for one floor to dry before building the next. A project video from Vnet in Ulanqab shows that an entire factory building 15:01 Speaker A was completed from bare land in just one winter season. Tencent’s T-block takes this logic to the extreme: the entire data center is a factory-prefabricated module. 15:10 Speaker A Architecture is compressed into a leveled site and a simple steel shell; sometimes, no building is even needed. 15:16 Speaker A With the synergistic effect of prefabrication, most mechanical and electrical work is moved from the site to the factory, so M&E installation does not need to wait for the building to be completed. The most typical example is Alibaba’s QW 5.0, known as 15:25 Speaker A the "100-day data center": 30 days for factory production and site preparation , 50 days for on-site installation, and 20 days for commissioning. This timeline is calculated from the completion of the shell to server deployment; the M&E and commissioning 15:35 Speaker A phases are about half the length of the industry standard of six months. If calculated from ground-breaking to the completion of commissioning, it takes approximately seven to nine months. All these factors point in the same direction: more factory-based production and less on-site work. 15:47 Speaker A Following this logic to the end, you might think that Tencent’s fully modular container parks should have become the national standard long ago, but in reality, they haven't. There are two reasons for this, both with distinct Chinese characteristics. First 15:57 Speaker A , the labor math is different; in the West, the greatest benefit of prefabrication is avoiding scarce and expensive skilled labor. In China, on-site labor is abundant and cheap; the traditional approach of steel structures plus prefabrication can already meet very aggressive 16:08 Speaker A construction schedule requirements. The marginal speed improvement brought by containerization is minimal. In fact, the logic of financing is also at play. 16:15 Speaker A Colocation developers in China mortgage data centers to banks for loans. A permanent building designed for a 20- year lifespan has a high appraised value, but a container park has a much lower appraised value and lower residual value. When your business 16:27 Speaker A model relies on using assets to borrow money to lease to yourself, you will choose to build a real building. 16:31 Speaker A Therefore, today, fully modular architecture remains mainly limited to self-built projects by hyperscalers and projects in Southeast Asia, but the overall construction speed is already fast enough. Regarding costs, while speed is increasing, capital expenditure is steadily declining. 16:43 Speaker A Continued supply chain compression and rising localization rates have brought the cost of shells and M&E down to a fraction of wholesale costs in the U.S. 16:49 Speaker A Alibaba says their modular approach, compared to earlier versions, has not only shortened the construction period but also cut costs by over 10%. We’ve discussed scale, history, East-to-West computing, speed, and cost. Now, let’s look at the five major demand-side players. Each has a 17:01 Speaker A different approach, driven by completely different AI strategies and business models. Let’s start with ByteDance. In China, whenever you talk about wholesale data centers, you eventually circle back to one question: What does ByteDance want? ByteDance is China’s largest single AI buyer, and 17:15 Speaker A because most of their infrastructure is leased, they are the demand engine for the entire third-party colocation industry. This demand is real, paid for by the world's most profitable consumer app cash cows: Douyin and TikTok. 17:25 Speaker A ByteDance’s flagship large model product is Doubao, the Chinese version of ChatGPT. It may not be the strongest model in benchmark tests, but as a friendly, low-cost consumer product, it serves countless users. As of March 2026, Doubao’s monthly active users 17:39 Speaker A reached 345 million, making it China’s most-used AI product—more than Tongyi Qianwen’s 166 million and DeepSeek’s 127 million combined. In the video model space, ByteDance is also leading the pack. In March 2026, SCD 2.0 topped both the text-to-video and image-to-video tracks. Unlike 17:54 Speaker A almost every other Chinese lab, ByteDance does not open-source its video models. When you are in the lead and have sufficient compute, you don't open-source. In B2B, ByteDance created AI programming tools and the agent platform Coze, but they haven't been 18:06 Speaker A very successful. This August, these two teams were merged into the Doubao division, and Doubao was officially designated as the company’s primary AI business. Volcengine is ByteDance’s external cloud business; it’s growing fast and competing directly with Alibaba Cloud. Volcengine 18:17 Speaker A is also heavily pushing private deployment, targeting high-ARPU production workloads in finance, government, and industry—territory previously dominated by Alibaba, Huawei , Tencent, and Baidu. ByteDance hasn't released a frontier-capable large open-source model yet, but they are currently in the early stages of 18:30 Speaker A training a model with a total of 10 trillion parameters. What does that even mean? For comparison, Tongyi 3.8 Max is 2.4 trillion parameters and Kimi k3 is 2.8 trillion; ByteDance's target scale is much larger than both. If you 18:42 Speaker A stack all these business lines on top of Douyin's recommendation and ad inference, along with TikTok's global workloads, the total training and inference demand easily exceeds anything they have seen before. The trajectory of capital expenditure has also been consistently revised upward. 18:56 Speaker A The 2026 AI infrastructure budget started at approximately $ 24 billion, or 160 billion RMB, with about half dedicated to chips. By May, this figure was adjusted upward by at least 25%, exceeding $ 30 billion, or 200 billion RMB. The latest figure circulating in 19:10 Speaker A the media is $ 70 billion. But that is an upper limit under discussion, not a formal budget. Behind these capital expenditure figures, ByteDance's domestic data center footprint is built almost entirely on wholesale colocation , with over 4 gigawatts of leased 19:23 Speaker A capacity currently delivered across Chindata, DataPipe, VNET, RZ, Yunchangsheng, and a host of other players. ByteDance is the anchor tenant for the entire wholesale market in China. It maintains long-term and deep relationships with suppliers while diversifying across operators to manage 19:37 Speaker A supply risks, pricing, and regional coverage. Because the orders are so massive, suppliers develop a reliance on ByteDance; many derive over 80%of their revenue from them, with Chindata being the most prominent example as ByteDance's most visible landlord. As of early 2026, ByteDance still 19:51 Speaker A accounted for approximately 90%of Chindata's revenue. Chindata was once listed in the U.S., was taken private by Bain Capital in 2023, and sold its China operations to a local consortium in 2026. The real hidden winner is DataPipe, or ZDA; this company has 20:04 Speaker A never been public, is funded by U.S. dollar funds including ByteDance’s early investors, and has very little external data. However, according to model tracking, its operational and under-construction scale has now surpassed Chindata, making it one of ByteDance's largest suppliers. VNET is 20:17 Speaker A a latecomer, pivoting from retail to wholesale on the back of ByteDance's demand, with sites in Taicang, Huailai, and Ulanqab. In the first half of 2026, VNET signed 862 megawatts of new wholesale orders, with its wholesale capacity in use breaking through 1 20:30 Speaker A gigawatt for the first time. The vast majority of this is from ByteDance, including a record-breaking single order of approximately 500 megawatts to be delivered in phases through 2028. 20:37 Speaker A After reaching a certain scale, ByteDance began shifting toward self-owned facilities through Volcengine. This is done to lock in long-term capacity for AI training and inference, while lowering unit costs and reducing reliance on third parties. 20:48 Speaker A ByteDance currently operates three major data center clusters: Shanxi Datong is planned for approximately 500 megawatts, Anhui Wuhu is also planned for about 500 megawatts to serve inference workloads in East China, and the Inner Mongolia dual-node in Ulanqab and Helingeer is each planned for 21:00 Speaker A around one gigawatt. With two to three gigawatts already online, a fourth cluster is now taking shape in Ningxia. 21:05 Speaker A According to corporate registration tracking, ByteDance established two subsidiaries in Ningxia in July 2026, following the same pattern as their Inner Mongolia project: register first, announce later. The Inner Mongolia cluster is the most recently built and the fastest-growing one. In August 2026 21:18 Speaker A , official government portals disclosed the approval of five Volcano Engine data center projects: three in Helingeer and two in Ulanqab. One is an expansion phase of the current campus, while the other four are newly approved with a construction cycle spanning from 21:32 Speaker A 2025 to 2029. After ByteDance, let’s take a look at Alibaba. Alibaba is China's largest cloud provider and the player in this country most similar to AWS. Its strategy mirrors AWS, leveraging a strong IaaS foundation to upsell AI capabilities, and management 21:44 Speaker A has clearly positioned Model-as-a-Service as a value-add atop the cloud. However, the geographic footprint differs, as Alibaba Cloud’s overseas revenue accounts for only about 20%, whereas analysts estimate AWS's international market contributes 30%to 40%of its revenue. Simply put, Alibaba is using a hyperscale playbook 21:59 Speaker A to serve a single, addressable market. The transition to MaaS is already reflected in the income statement; AI-related product revenue currently accounts for approximately 35%of Alibaba Cloud’s external revenue, with an annualized run rate nearing $ 10 billion. In March of this year, the 22:12 Speaker A group specifically established the Alibaba Tokenhub division, headed by the Group CEO, to integrate and align models, MaaS, and applications. The other half of the strategy is its investment network, with Alibaba investing in five of China’s major model unicorns: Moonshot AI, MiniMax, 22:24 Speaker A 01.AI, Baichuan Intelligent, and Zhipu AI. Furthermore, the structure of these investments is primarily designed to keep these models within the Alibaba Cloud ecosystem rather than allowing them to become competitors. The Moonshot AI deal is a template: Alibaba invested about $ 800 million for 22:36 Speaker A roughly 36%of preferred shares, with nearly half paid in the form of Alibaba Cloud credits. After MiniMax listed in Hong Kong, Alibaba remains its largest external shareholder with a 12%stake. 22:46 Speaker A In a sense, every time these unicorns run their training workloads, it translates into demand for Alibaba Cloud. In the open-source model space, Alibaba’s Qwen series has become a leader, with cumulative downloads exceeding 3 billion and over 300,000 models deployed. As an e-commerce giant 22:58 Speaker A , Alibaba has integrated Tongyi Qianwen into its existing business lines to create intelligent shopping agents and search bots, while also selling AI inference services. It is the only player in China truly following a large-scale model path. Including foundation models, cloud computing, and 23:10 Speaker A e-commerce distribution, everything is drawing from a single computing power pool. In terms of hardware, Alibaba's T-Head has moved past the experimental phase typical of many chip projects and is now serving Alibaba Cloud workloads. 23:20 Speaker A Reports indicate that Alibaba is partnering with telecom carriers to build multiple data centers fully deployed with T-Head chips. In April 2026, China Telecom and Alibaba Cloud powered on a 50,000-card T-Head cluster in the Shaoguan region, a hub node for 23:32 Speaker A the "East Data, West Computing" project , with plans to expand to 100,000 cards . The division of labor is clear: Telecom provides the building, power, and government/SOE sales channels, while Alibaba provides full-stack capabilities from chips to cloud platforms and models, bundled as a 23:45 Speaker A complete solution. This joint effort is just one example of Alibaba's flexible data center strategy. Data shows that even in its self-built clusters, Alibaba mixes third-party customized leasing with its own proprietary infrastructure. This helps to alleviate the burden of capital expenditures. The 23:59 Speaker A Zhangbei flagship complex is the best example of this. Within its five campus zones, both models coexist side-by-side . Some are built by Alibaba, while others are leased from third-party operators like AtHub, GDS, and VNET; third parties handle the shell and 24:10 Speaker A power access, while Alibaba fills them with servers. Next is Tencent, who holds in its hands. The most powerful distribution assets of the era, yet it has been the slowest to move. WeChat is China's largest social platform, with over 1.4 billion monthly active users. 24:21 Speaker A With such a massive surface area, even a minor change would impact over a billion users. Management admitted at the January 2026 annual meeting that the company was nine months to a year behind in AI. The catch-up began with 24:32 Speaker A organizational restructuring. By July 2026, all model work was consolidated into a unified foundation model department, led directly by the head of the mobile division. The model itself has also caught up; Hunyuan 3 was officially launched on July 6, 2026, 24:44 Speaker A featuring a 95-billion parameter MoE architecture with 21 billion active parameters and a 256K context window. 24:49 Speaker A It now powers Tencent's AI chatbot " Yuanbao," the coding agent "CodeBuddy," and various social platforms. However, Tencent's strategy is clear: own the interface, but do not necessarily own every model. Its AI products run not only its own models but also DeepSeek, 25:02 Speaker A Kimi, and GRM simultaneously. Tencent's data center development perfectly mirrors its business strategy. During the downturn from 2022 to 2023, Tencent focused on consolidation, closing small leased urban facilities and centralizing capacity into its own mega-campuses in Qingyuan, Huailai, and 25:16 Speaker A Yizheng. Then, starting in 2024, both capital expenditures and committed capacity began to recover. Capital expenditure in the first half of 2026 has already exceeded the full year of 2025, and free cash flow has turned, yet our model tracking shows steady 25:27 Speaker A capacity growth. It is not a sudden, sharp spike. This gap best illustrates that the incremental spending is mainly going toward chip and storage prepayments, and current capacity additions now rely more on wholesale leasing than building proprietary data centers. This reverses the previous 25:40 Speaker A trend of favoring self-built facilities , against a backdrop of tighter cash investment. Another interesting point is that while everyone else is moving west, Tencent has remained in the east; compared to ByteDance and Alibaba, its western footprint is the smallest. This 25:52 Speaker A aligns with its latency-sensitive workload structure, such as WeChat, gaming, and video. The only exception is the collaboration with China Unicom in Ningxia Zhongwei, where the first data center went live in April 2025. It should be said that we are still in the 26:03 Speaker A early stages of Tencent's "all-in" push ; unlike press releases and benchmark scores, data center capacity is a hard, physical, and real indicator of computing power, so we must continue to track Tencent's data center moves, as it is the clearest litmus test for how 26:15 Speaker A serious this catch-up really is. Next is Baidu, which pioneered China’s AI narrative but is now the one spending the least. It was the first company in China to bet on AI, spanning everything from large models to autonomous driving 26:26 Speaker A . Its model, Ernie Bot, was released as early as 2019, and in March 2023, it was the first Chinese tech giant to launch a ChatGPT-style product; it did lead for a few months, but its own Doubao soon overtook it and took the 26:37 Speaker A crown in the consumer segment. Baidu used to be closed-source, but it open-sourced Ernie 4.5 in June 2025, then launched Ernie 5.0 in January 2026 with 4 trillion parameters, and updated it to Ernie 5.1 in May 2026. Like Alibaba, Baidu has integrated Ernie Bot 26:50 Speaker A into search, maps, and coding assistants, but this cannot save its core business. In the second quarter of 2026, online advertising revenue fell by about 20%year-on-year, with the only bright spot being private deployments, as Ernie is selling well among 27:00 Speaker A government and state-owned enterprises. The capital expenditure curve tells the same story: Baidu spends the least among the four major internet hyperscalers, with its capital expenditure for the first half of 2026 being roughly equal to what Alibaba spends in a single month. However, 27:12 Speaker A Baidu is running the deepest vertical integration, with its Kunlun Xin chips having trained Wenxin 5.1. In contrast, Alibaba T-Head's Zhenwu chip is still primarily used for inference. Kunlun Xin is also sold externally, particularly for government and financial sector procurement. Kunlun 27:24 Speaker A Xin is currently planning an A + H dual listing, with rumored valuations already exceeding that of its parent company. The data center map also corroborates this assessment. HL's capital expenditure in Q2 2026 grew by 201%year-on-year, with core capex 27:36 Speaker A accounting for 45%of revenue—the highest intensity among all Chinese platforms—yet models show its physical infrastructure footprint is the smallest of the five, with third-party tender scales at only one-tenth of ByteDance's, and no gigawatt-level parks in the western hubs. Baidu's first Inner Mongolia 27:50 Speaker A computing power order didn't arrive until 2025, and it has no overseas computing power; the gap between its capital expenditure figures and physical construction is mainly due to chips and storage. Yet, Baidu isn't hoarding inventory to chase frontier training runs, but is increasingly 28:01 Speaker A acting like a chip company, which is also evident in its income statement. In the second quarter of 2026, GPU cloud revenue grew by 283%year-on-year, accelerating from 184%in the first quarter. Management focused on return on investment rather than capacity 28:14 Speaker A during the earnings call, stating that GPU cloud margins outperform traditional CPU clouds, attributing this to the cost advantages of their self-developed chips. The final one is Huawei. Huawei appears in this report for reasons different from the others: it is both the second-smallest 28:27 Speaker A hyperscaler and the supplier that determines the ceiling for everyone else. As a data center operator, Huawei builds almost everything itself, with a self-build ratio exceeding 90%, the highest among all players. Its layout is anchored in three cloud hubs, each 28:39 Speaker A designed to start with million-level server capacity; the Guizhou Gui'an hub , launched in 2021, is currently Huawei Cloud's largest operational park, with a design capacity of 1 million servers. 28:47 Speaker A The second hub is in Inner Mongolia with two sites: Ulanqab is Huawei's earliest self-built project, dating back to 2016, while the neighboring Linggeer park, started in 2023, is designed to be Huawei's largest footprint, including 1,000 acres of land, 50 planned data hall buildings, 29:01 Speaker A six dedicated 220kV substations, and a full capacity of over 1 million servers ; the Anhui Wuhu park launched in 2024, completing the three-hub layout. 29:08 Speaker A domestic layout, which is the most ambitious of the three, designed to house ten thousand servers on 1,000 acres of land. SNASIS has identified a clear pattern in these campuses: designed capacity is far ahead of actual deployment, with an 29:20 Speaker A operational-to-planned ratio of roughly 1 to 6. So, the correct way to understand Huawei Cloud is perhaps not to view it as a cloud business at all, but as a distribution channel for raw computing power. For customers who cannot buy or operate CloudMatrix 29:32 Speaker A hardware, it offers rentals by the hour or by the token, turning Huawei's chip advantage into recurring revenue. In 2025, cloud revenue accounts for less than 5%of the group's total, with shrinking market share. Its clients are mainly government, finance, and 29:45 Speaker A state-owned enterprises, not internet-native workloads. Policies are also pushing in this direction; starting in November 2025, new state-owned data centers are required to use domestic AI chips. This leads to the conclusion that among domestic hyperscalers, Huawei has the second-smallest footprint, but its 29:58 Speaker A exposure to everyone else is the largest. From now on, every national intelligent computing project is essentially a bid for Huawei's supply, pitting the five major domestic players against each other. Let's pull the camera back and see that the 30:07 Speaker A construction boom is not limited to within domestic borders. Beyond the domestic building spree, Chinese tech giants are also investing heavily globally. According to our model, overseas leased capacity will reach approximately 4 gigawatts by 2029. 30:18 Speaker A ByteDance is, of course, the leader. To achieve local processing of user data, TikTok signed a massive number of capacity contracts overseas between 2020 and 2024. Although ByteDance does not build its own facilities overseas, Sanalysis has tracked its major leasing 30:30 Speaker A and cloud leasing activities. In the U.S., TikTok expanded aggressively from 2020 to 2023, mainly concentrated in Northern Virginia and Portland, but has added very little capacity since then. 30:40 Speaker A In Europe, TikTok opened data centers in Ireland and Norway in 2023. A third site in Finland breaks ground in 2025. 30:45 Speaker A The latest move is in Brazil, starting with about 200 megawatts. Unlike the slow ramp-up in the U.S. and Europe, ASEAN is the current bright spot. 30:53 Speaker A Acceleration in ASEAN has been evident since 2023. Because export controls limit which chips can run domestically, ByteDance's Nvidia fleets are stationed overseas, using developers who bring the "China construction speed" to build the facilities. Key owners include GDS, VDC, and a wave of developers backed by 31:07 Speaker A private equity, with capacity concentrated in the Singapore, Johor, and Batam hubs. These operators build at an incredible pace; projects over 100 megawatts are delivered in under a year, while some mature Western wholesale developers take twice as long . Alibaba and Tencent are also actively 31:21 Speaker A expanding overseas, albeit on a smaller scale. Alibaba is the more determined of the two, currently operating 46 availability zones across 18 international regions, compared to 60 availability zones and 13 regions domestically. Over the past two years, the pace of launching new regions and 31:33 Speaker A availability zones has clearly accelerated, with this expansion backed by hard currency. In August 2026, Alibaba completed a $ 10 billion new share placement—its first since listing in Hong Kong in 2019 and the largest follow-on offering in Hong Kong 31:45 Speaker A stock market history. The prospectus clearly states that the funds will go toward an all-in AI strategy, with about 60%used to expand global computing infrastructure and the rest for hyperscale AI data centers and cloud upgrades. The drivers include its 31:58 Speaker A AI business, which has dual growth targets: raising the share of AI-related services in cloud revenue from 35%to over 50%, and increasing overseas revenue from about 20%to between 25%and 30%in three years. 32:09 Speaker A Another driver is securing advanced chips, as renting GPUs to Chinese hyperscalers is currently a legal and significantly large business; Tencent's overseas expansion is smaller, initially serving Chinese game and internet companies going global, but the pivot is similar. Since 2025, it 32:25 Speaker A has added three overseas regions— Saudi Arabia, Japan, and Malaysia—and eight international availability zones, driven by the international promotion of model service gateways and AI agent products, with Southeast Asia being a clear focus. International business currently accounts for about 10%of 32:37 Speaker A Tencent Cloud's revenue, but it is growing by over 40%year-over-year. Their publicly stated goal is to exceed 30%by 2030. Finally, to summarize: high vacancy rates coexist with AI capacity shortages, and a record-breaking capital expenditure cycle is accompanied by falling facility costs. 32:49 Speaker A The market’s largest tenants disclose almost nothing, yet they anchor the entire wholesale industry. Like their Western peers, most capital expenditure ends up spent on servers, but servers require buildings, and data center construction is the earliest signal of when and where they will land. Because 33:03 Speaker A this model tracks projects under construction and the planned pipeline, the confirmed parts of the build-out are actually already in the data. 33:09 Speaker A Alibaba, Tencent, Huawei, Bytedance, and Baidu—the combined technical capacity of these five leading hyperscale operators is expected to more than triple from the start of the AI cycle to 2030, growing from approximately 6 gigawatts at the end of 2023 to about 21 gigawatts by 2030. 33:22 Speaker A Annual new capacity is expected to accelerate from roughly one gigawatt per year during the cloud era to three to four gigawatts per year by 2026 and 2027. Beyond that, there is also a pipeline of approximately 6 gigawatts in speculative campus reserves set to 33:34 Speaker A break ground. Supporting this growth curve is demand driven by a structurally protected domestic market, state-led infrastructure, and an open-source flywheel that keeps Chinese models globally relevant even despite export controls. These three factors combined form the underlying logic behind this construction boom. So, what 33:50 Speaker A do you think is the biggest variable in the race for AI infrastructure in China ? Is it chip supply, model capabilities , or electricity costs? Feel free to leave your thoughts in the comments section. Thanks for watching, and see 34:02 Speaker A you next time. Topics: China data centers AI infrastructure Alibaba Cloud Tencent Cloud ByteDance modular data centers East Data West Computing state-owned enterprises capital expenditure Chinese tech industry Study this video - [Flashcards from this video](https://sozai.app/tools/ai-flashcard-generator/?from=transcript&id=78386) - [Quiz on this video](https://sozai.app/tools/quiz-generator/?from=transcript&id=78386) Free, made by AI from this transcript. No signup. --- This is the markdown twin of https://sozai.app/transcript/china-ai-infrastructure-data-centers/ — the same content, without the markup. Published by SozAI (https://sozai.app). Reuse and quotation are allowed with attribution and a link back. Machine-readable index: https://sozai.app/llms.txt · data API: https://sozai.app/api/