Chinese tech companies are piling into the Middle East and Southeast Asia to build data centers, and they are writing checks worth hundreds of millions, even billions, of dollars.

From the outside, it looks like China copy-pasting its reputation as an infrastructure machine onto foreign soil: whoever has the deepest pockets and builds the fastest walks away with the next round of AI money.

On the ground, it works nothing like that. Power, water and land all have to be fought for. Host governments are tightening policy, and U.S. export controls on advanced chips keep shifting. In March 2026, data centers were pulled directly into the physical battlefield of the Middle East conflict.

So what exactly is this wave of Chinese computing power seeking overseas, where is it getting stuck, and can it actually make money?

An Entire Industry Is Relocating

This is not one or two companies taking a gamble. It is the whole Chinese computing power supply chain moving abroad together.

IDC estimates that the market for cloud services used by Chinese companies operating overseas reached $5.23 billion in 2025, up 31% year on year. In a survey of 120 Chinese companies with overseas operations, 84.2% were using Chinese-backed cloud providers.

The biggest buildout is in Southeast Asia. In May 2026, Thailand's Board of Investment approved six large projects worth roughly $29 billion in total investment. Three of them were data center and data hosting projects, accounting for about $27 billion of the approved amount. The single largest came from TikTok System (Thailand), approved at around $25 billion, mainly to expand servers and data storage and processing capacity.

The following month, Alibaba Cloud opened two more data centers in Malaysia's Johor state, bringing its total there to five. Back in November 2024, GDS and Google had each secured projects in Thailand worth roughly $1 billion.

Colocation operators are moving out too. Chindata's Bridge Data Centres already has operations in Malaysia, India and Thailand, while GDS's international arm DayOne has opened campuses in Malaysia and on Indonesia's Batam Island.

Tencent went to the Middle East. In 2025, Tencent Cloud announced its first Middle East cloud region in Saudi Arabia, pledging more than $150 million over the coming years for infrastructure, resources and business operations.

From the big cloud providers and colocation operators to vendors selling servers and liquid cooling equipment, China's computing power industry is entering the Middle East and Southeast Asia as a full formation. Why does everyone feel they have to go?

Capacity That Cannot Fit, and Capacity That Cannot Fill

A set of figures published by Science and Technology Daily in 2025 laid bare the awkward state of domestic computing power: at one 1,000-card AI data center in western China, fewer than half the racks were occupied and actual server utilization was under 30%. Even so, annual operating costs ran as high as 30 million yuan. Estimates from Inspur's AI research institute put average computing power utilization at Chinese AI data centers at around 30%.

So at home, idle and effectively stranded computing power is burning cash, while the major players spend heavily abroad. The instinctive reaction is that domestic facilities are already gathering dust, so there is no need to pour more money overseas.

But those facilities cannot take today's AI orders. Most were designed for standard servers, and their power and cooling cannot support high-density AI racks. Conventional racks mostly draw under 20 kilowatts; AI racks can reach tens of kilowatts, up to roughly 120 kilowatts. The halls sit empty, yet new machines cannot be moved in. Retrofitting means recalculating power capacity, cooling loops, floor loading and the cost of downtime. Put plainly, a batch of facilities was built too early and built wrong, and sitting on them will not conjure up orders.

Overseas is the opposite: demand is still climbing. A 2026 report from China's CAICT found that total computing power infrastructure capacity in key Southeast Asian countries already exceeded 4.4 gigawatts in 2025, and is expected to pass 10 gigawatts by 2030.

Demand comes from two kinds of customers. The first is Chinese companies going abroad. BYD is building overseas plants, Meituan is delivering food outside China, and wherever the business goes, the data and the compute have to follow. If a cloud provider does not follow its customers out, it loses their domestic business too. The second is local heavyweights overseas. Indonesian internet group GoTo spent eight months preparing to move more than 1,000 microservices onto Tencent Cloud; the actual cutover took 4 hours and 54 minutes. These large local accounts are what Chinese cloud providers most want to win abroad.

There is also a hard technical limit: if the customer is overseas, you cannot serve them remotely from computing power sitting in China. For real-time gaming or cross-border payments, data shuttling across an ocean means latency and link stability that directly degrade the experience. On top of that, markets such as Indonesia impose localization and cross-border transfer requirements on public sector and financial data, which have to be assessed case by case. Put a facility too far from the customer and some business simply cannot be taken on.

Layer on U.S. chip controls and constrained domestic supply of high-end AI chips, and building facilities abroad became, for a time, a supplementary route to advanced computing power for some companies. Misallocated domestic capacity, rising overseas demand, the need to sit close to customers, and chip and compliance pressure have all pushed Chinese computing power companies out of the country at once.

Once abroad, the business splits into four distinct models: cloud providers building cloud regions and selling cloud services, GPU operators selling training and inference capacity, data center operators investing in buildings and electromechanical systems and leasing by the kilowatt, and equipment vendors selling power distribution, liquid cooling and engineering delivery.

Overseas, the Scramble Is for Power First, Then Water, Then Land

Once abroad, Chinese companies quickly discover that their strongest cards, technology and construction speed, are not what decides the game.

On building and fitting out, Chinese firms really are fast. GDS delivered its NTP1 project in Johor, Malaysia in 14 months. Using liquid cooling and indirect evaporative cooling, Chindata brought the average annual PUE of its China data centers down to 1.21 in 2025, an industry-leading level.

The problem is that no matter how fast the building goes up, nothing switches on without power. In Johor, the standard queue for a data center grid connection runs three to four years. With a government fast-track channel, the best case is still about a year.

The queue is that long because too many players are chasing the same electricity. Malaysia's national utility has already locked in 61 data center projects with combined maximum demand of 8.35 gigawatts, of which Johor alone accounts for roughly 5.6 gigawatts, close to two-thirds. But 8.35 gigawatts is a future number. In June 2026, actual data center load in Malaysia was just 1,256 megawatts, with many projects still under construction or barely started. What everyone is competing for right now is a power allocation for the next few years.

With allocations tight, the Malaysian government has started pushing costs onto project developers: investors pay for grid upgrades themselves, and once supply capacity is reserved for you, you pay for it whether you use it or not. The facility takes on a fixed electricity bill before it earns a cent.

Settle the power question and water is next. High-density AI servers throw off enormous heat; liquid cooling carries it away from the chips and racks, but the heat ultimately has to be rejected outdoors through cooling towers, and those towers need constant makeup water. Research institute AMRO estimates that a 100-megawatt data center using a water-intensive cooling design can consume about 4.2 million liters of water a day.

The applications Johor has received go far beyond that. According to RHB, data center water requests filed for 2024 through 2035 total roughly 440 million liters per day, while the water utility supported only about 316 million liters per day, leaving 28% unapproved. With water short, Johor raised the bar: a separate water tariff for data centers, a suspension of approvals for some water-intensive projects, and delays for certain water-cooled expansions.

Accept the water and power prices and you still need somewhere to put the machines. AI-era data centers routinely start at hundreds of megawatts, and beyond the server halls they need substations, water treatment plants and energy storage, all of which eat land. Singapore paused data center approvals years ago for lack of land and power, and the demand spilled straight over to neighboring Johor, where prices for prime industrial land have risen fast.

And the companies competing with Chinese firms for that power, water and land are the Silicon Valley giants. Microsoft is investing a cumulative $15.2 billion in the UAE from 2023 to 2029; it has taken a stake in local AI firm G42 and, in September 2025, secured a U.S. government export license allowing it to ship in advanced NVIDIA chips such as the GB300. Amazon has committed roughly $6.2 billion to its Malaysian cloud region from 2024 to 2038. Google is buying land in Thailand and Malaysia. Established Western and local operators, meanwhile, long ago claimed most of the best sites.

By comparison, Chinese companies are late arrivals: prime locations, power allocations and anchor customers have already been divided up once by those who got there first. Money is only the entry ticket. Who wins the power, water and land is what decides who survives.

Policy, Chips and Artillery

Even after fighting through the water, power and land, three more obstacles remain, each harder than the last.

First, host country policy can change overnight. A couple of years ago, countries were competing for investment and welcomed almost anything. Now, with power and water stretched, half the door has swung shut. Johor had already rejected a batch of data center applications, and from February 2026 it simply restricted new applications for non-AI data centers while raising the energy efficiency bar for tax incentives. Thailand has gone further into the details, setting up a working group to scrutinize data center water and power consumption, and its new investment promotion guidelines tie tax breaks directly to metrics such as PUE and water management. Indonesia and Saudi Arabia are tightening data localization and cross-border transfer rules. Not all data is barred from leaving, but companies have to work through compliance item by item, by business line and data type. The result is that a project approvable yesterday needs a redesign today, and the payback plan already written goes out the window.

Policy, at worst, costs a few months and some extra money. The second obstacle is more serious: even with the building finished, you may not be able to put chips inside it.

Many Chinese companies used to assume that registering a subsidiary in Singapore or Malaysia was enough to buy NVIDIA's high-end chips. At the end of May 2026, the U.S. Commerce Department spelled it out: chip purchases are judged not only on place of registration and shipping address, but on who the headquarters is, who the ultimate parent is, and who the end user is. If a company falls within the scope of the controls, its overseas subsidiaries need licenses all the same.

Needing a license does not mean every application is refused. In January 2026, the U.S. shifted review of certain direct exports of H200 and MI325X chips to China and Macau to a case-by-case basis, subject to strict conditions. Other chip tiers and other transaction routes may still face a presumption of denial. What has ended is the era when an offshore shell was enough to buy chips freely.

Renting overseas computing power instead of buying chips is also getting harder. In January 2026, the U.S. House of Representatives passed the Remote Access Security Act, which would bring remote access via the internet and cloud services under the same controls. As of October 2026 the bill had not completed the rest of the legislative process, so it is not in force, but the direction is clear.

The last obstacle is the harshest: the facility can simply be destroyed. After conflict broke out in the Middle East in late February 2026, two Amazon data center facilities in the UAE were hit directly by drones in March, and a site in Bahrain was affected by a nearby strike. There was structural damage and loss of power, and some fire suppression measures caused additional water damage. Iran then named several U.S. tech companies as targets. By September, Amazon acknowledged that in the Bahrain region and the UAE's mec1-az2 availability zone, some customer resources and data that existed only locally were permanently unrecoverable; most customers rebuilt from backups or other regions.

High walls, guards and backup generators do not stop drones. The attacks also exposed a subtler flaw: multiple availability zones inside one region is not the same as cross-region disaster recovery. If data and backups sit in the same failure domain, no tier rating will save them. Facilities used to be underwritten on eight to ten years of peaceful operation. Now geopolitical site risk, insurance liability and exclusions, and the cost of cross-border, cross-region backup all have to be rebuilt into the model.

Southeast Asia Pays Back Slowly; the Middle East Pays Well but Demands Shared Control

With that many pitfalls, it can look like staying home is the better call. But for cloud providers and data center operators, not going means handing the new overseas customers and contracts to rivals.

Going abroad does not mean every company has to buy land, put up buildings and stockpile GPUs itself. Building a cloud region, leasing third-party facilities, forming a joint venture, or exporting equipment and engineering capability each involve very different capital and very different risk. The real questions are three: how to go, which slice of the value chain to monetize, and how the numbers work.

Southeast Asia and the Middle East, the two hottest destinations, operate on completely different economics.

Southeast Asia's problem is slow payback. A ResearchAndMarkets report puts the average construction cost of 1 megawatt of data center capacity in Malaysia in 2025 at roughly $8.5 million to $10 million, so 50 megawatts runs $400 million to $500 million. That covers facility construction only, not total cluster investment once GPUs are installed; whether land, external grid works, financing and IT equipment are included swings the total substantially. And the building is only the start. Electricity, water, land and reserved grid capacity charges keep coming, while customers and revenue may not arrive on schedule. Malaysia's tariff revision is a case in point: industry figures and government officials have estimated power costs for large users could rise 10% to 14%, before certain surcharges. When costs climb, payback stretches out further.

The Middle East is the inverse: the money is there, and so is the risk. The region is not short of cash, and the buyers are national AI platforms and sovereign capital. In 2025, NVIDIA announced it would supply 18,000 GB300 Blackwell chips to Saudi state AI company HUMAIN for the first phase of a 500-megawatt data center project. Abu Dhabi sovereign fund Mubadala invested $12.9 billion in AI and digitalization in a single year.

That money comes with conditions. One is that the region can be drawn into war and geopolitical conflict at any time. The other is that local partners want more than your services; they want local assets, operating capability, data governance and talent development. Alibaba Cloud entered Saudi Arabia through a joint venture, with stc Group, Alibaba Cloud and other parties forming Saudi Cloud Computing Company, which manages and operates two local data centers. The price of admission is sharing control, sharing profit, and genuinely building up a local team.

Both markets can be profitable, but not for the same players. The first to see revenue are the suppliers winning engineering, power distribution, cooling and modular orders; they do not wait ten years of campus operation to recoup, but they carry acceptance testing, warranties and receivables. The steadiest cash flows belong to operators that sign high-credit customers and medium- to long-term pre-leases before breaking ground; with minimum commitments and power cost pass-through clauses in the contract, revenue can be modeled in advance. The most volatile are the GPU compute operators, where occupancy, hourly price per card, renewals, chip generations and depreciation assumptions can each rewrite the income statement. Cloud providers depend on long-term customers and higher-layer services; shipping servers abroad only buys a seat at the table, and without databases, models, platform software and a local ecosystem, even heavy capex struggles to earn a decent return.

Who builds is not who earns the most, and who books revenue first is not who ends up with the highest return.

Power Is the Variable That Decides It

And the variable that decides all of it is electricity.

Land without a confirmed power allocation and grid connection date is just a site option waiting to be redeemed. Only land that gets energized on schedule and can support high-density IT load counts as real capacity.

Take an example: a project planned for 100 megawatts of IT load, at full load and an assumed PUE of 1.25, implies roughly 125 megawatts of facility-side demand, with capacity headroom and redundancy design to be assessed on top. Treating a "100-megawatt campus" as 100 megawatts sellable to servers overstates revenue and understates power requirements. Land can be expanded, buildings can be added, customers can be negotiated over time, but if the grid connection slips two years, construction, equipment procurement and customer onboarding all have to be resequenced.

It is precisely because power is the bottleneck that opportunities open up for the Chinese supply chain. EPC and modular vendors can shift power skids, cooling skids and more of the electromechanical scope into factory prefabrication and pre-commissioning, reducing exposure to the uncertainties of overseas site work. For power distribution vendors, the business goes beyond UPS and transformers to substations, switchgear, busways, backup generation, energy storage and power management. For liquid cooling vendors, selling cold plates and CDUs is not enough; they have to master outdoor heat rejection, water quality management, leak detection, joint commissioning and local operations and maintenance.

Orders, of course, are not profit, and profit is not cash. Local certification, exchange rates, customer defaults, cross-border logistics and warranty liability can all eat the margin that looked good on paper.

So the hard part of this computing power push abroad is not putting up the building. It is getting power, water, customers and policy to land at the same time. If any one of them falls through, the money already spent sits locked in concrete.

Some companies broke ground on impulse, then failed to close the power deal or sign customers, turning hundreds of millions of dollars into assets that generate no cash and simply depreciate. The smarter approach is to lock in solid long-term contracts before building in phases, secure water and power approvals before breaking ground, use joint ventures to buy local policy protection, and have an off-site data recovery plan in place from day one.

The press conferences will keep coming. Next time a Chinese company announces a billion-dollar overseas computing power project, six things matter more than the size of the building: how much saleable IT load there is and when it gets energized; whether the paying customer has signed, and whether the contract includes minimum commitments and power cost pass-through; what the per-megawatt construction cost actually includes, and whether the project still earns a return if tariffs, financing costs or exchange rates move; at what load and with what cooling design the PUE and WUE figures were measured; whether the chips, data and local shareholding are compliant; and whether customer data survives an actual conflict or region-level failure.

The competition in AI computing power has shifted from who can get GPUs to who can assemble power, engineering, capital and customers into a durable business.