A cloud company's servers need a building with steady power and cooling. GDS finds land, wins the power approvals, builds the hall and charges a monthly fee for space and power.
The company · 2 · What exactly does it sell?
Nine-tenths of sales is renting out space
Each $1 of sales, 2025
91%9%0%
Renting space and powerRunning servers for customersSelling equipment
Rent arrives every month once a customer has moved its servers in.
The company · 3 · How does it make money?
Rent is very profitable once a hall is running
About 45 centskept per yuan before interest
RMB20Mto build each megawatt
About 9 yearsfor a booking to pay back its building
Most of the rent is left after running costs. But the halls are paid for up front, largely with borrowed money.
Is it a good business? · 4 · What gives it an edge?
Customers rarely leave, but rents fall
Under 1%
of rented space ended each quarter
VNET
a rival, calls moving servers hard
About 3%
yearly fall in rent per square metre
Moving thousands of live servers is slow and risky. But staying is not paying more: in 2025 GDS wrote down RMB1.6B of sites because prices came in lower than planned.
Is it a good business? · 5 · What is the catch?
The biggest rival is the customer
RMB380BAlibaba's own cloud and AI spending
37.9% + 11.8%of rented space with the two largest customers
China Telecomwith two other state firms, runs over half the market
China's three state telecom companies also run data centres, and VNET competes for the same halls.
Is it a good business? · 6 · What does its record show?
Sales grow slowly while building costs debt
Sales, B
7.8
9.3
9.8
10.3
11.4
Dec 2021Dec 2022Dec 2023Dec 2024Dec 2025
Sales7.8B → 11.4Byuan of sales, 2021 to 2025
BuildingAbout RMB10Bof building this year, mostly borrowed
Profit per yuanAbout 45 centskept per yuan once halls run
Office costs shrink as a share of sales, but power, rent and wear rise with each building.
Is it a good business? · 7 · How much room is there to grow?
China's data-centre market keeps growing
RMB339.0Bthe market by 2030, Frost & Sullivan
13.2%a year
1,590 → 9,800EFLOPS of AI computing by 2030, the government's goal
Independent landlords like GDS are expected to grow faster than the telecom companies.
Is it a good business? · 8 · Who is in charge?
A founder with most of the votes
Since 2002William Huang as chief executive
About 3%of the shares owned by insiders
58.1%of the votes held by insiders, through special shares
Some Chinese customers now require that control. Pay is mostly shares that vest with time.
The story right now · 9 · Why does China need more data centres?
China's cloud giants are racing to build AI computing
470MWbooked in a half-year, a record
1GWGDS's doubled 2026 target
Alibaba and its peers need halls with huge power for AI. GDS builds them fast, but customers move in only once their chips arrive, and US export rules can delay them.
The story right now · 10 · Why isn't GDS as cheap as it looks?
Debts and preferred shares come before common shares
The halls earn rentAbout 45 cents per yuan.
Lenders are paid firstThe building was borrowed.
Preferred shares nextThey rank ahead too.
Common shares get the restPlus the DayOne stake.
The cheap look comes from counting its stake in DayOne, an overseas data-centre company, on top. Once the debts are counted, little is left over.
The story right now · 11 · What changed lately?
A record 470MW booked, and the target doubled
470MWbooked in the half, a record
1GWthe doubled 2026 target
+6.5%sales, as old contracts repriced
2027when most customers move in
Strong quarter
The shares lost about a third while sales per share rose. DayOne's planned listing would put a public price on GDS's stake.
The story right now · 12 · What is everyone arguing about?
Is the DayOne stake hidden value?
Many investors think
Cheap, with a DayOne stake the price ignores
We think
Worth about build cost, plus DayOne at its last price
We are wrong if 2027 move-ins more than double as promised, or DayOne lists well above its reported target.
The price and our view · 13 · What is a share worth?
No single value here: it is priced as a bet
1.7×if it works
84%kept if it fails
Multiples of the company's value when the bet was last struck, 16 Nov 2026.
On the filed figures its debt outweighs what the business earns, so a single value would need its own count of the buildings. Instead we ask what the shares pay if the plan works, and what is left if not.
The price and our view · 14 · How could we be wrong?
Two customers, borrowed growth and chips
Two customersBiggest
About half the rented space.
Borrowed growthShowing now
Building outruns the cash it earns.
Chips may not arriveWatch
US export rules delay move-ins.
What DayOne earnsUnknown
Only its balance sheet is filed.
GDS runs through contracts Chinese regulators could challenge, and US-China tension touches its chips and investors.
The price and our view · 15 · How will we know?
November shows whether orders reach 1GW
700MW or more booked in nine monthsGood sign
Move-ins of 470MW or more in 2027We turn positive
A quarter's bookings below 120MWWe turn negative
NOV 202625
Third-quarter 2026 results, on 25 November. Companies report every three months; each report is a test of the story.
Our view moves on these business events, not on a price line, since we put no single value on a share here.
The price and our view · 16 · So, what do we think?
China's data-centre landlord, at the cost of building it
NeutralMedium conviction
GDS rents data-centre halls to China's cloud giants. Customers stay and AI orders hit a record, but old contracts reprice down and borrowing pays for the building. Neutral.
↑ ↓ to move · C for contents
What it does
More about the company
Who it is
A cloud company's servers need a building with steady power and cooling. GDS finds land, wins the power approvals, builds the hall to the customer's design within about a year, and then charges a monthly fee for the space and the power used.
Its customers are China's biggest cloud and internet companies. They could build for themselves, and some do, but GDS sells speed and sites they cannot easily get: a land bank of nearly 4GW in the cities and western hubs they want.
How it makes money
Rent arrives every month once a customer has moved its servers in, and most of it is left over after running costs: about 45 cents of each yuan of revenue before interest and depreciation. A new hall costs about RMB20M per megawatt to build and management expects about RMB2.2M a year of that profit from each megawatt, so each new booking pays for itself in about nine years.
Growth and profit
What growth costs
Growing does not make GDS much more profitable. Office and sales costs shrink as a share of revenue, but they were small to begin with; power, rent and the wear on each new building rise with it. What growth really costs is the building itself, paid for up front with borrowing and new shares.
Its record
The numbers
Is it growing?Steady
Sales grew 11% in the year to Dec 2025, faster than the 5.5% in the year before.
Is it profitable?Losing
Operating margin is −0.5%, down from 7.7% in the year to Dec 2022.
Does growth pay off?Flat
Running costs take 23¢ of each sales dollar, up from 15¢ in the year to Dec 2021.
Does it earn good returns?Low
Shareholders' money earned 3.9% a year, up from −4.7% in the year to Dec 2021.
How long can it keep going?Long runway
It used CNY 1.33B of cash in the latest year, so the CNY 14.3B it holds lasts about 10.8 years at that pace.
Is the balance sheet safe?Heavy debt
It owes CNY 12.0B more than the cash it holds, about 3.6 years of the cash its operations bring in.
Are shares being diluted?Diluting
The share count rose 13% in 4.0 years, and pay in shares costs 2.5% of sales.
Why it wins
The evidence
Moving thousands of live servers to another building is slow and risky, so customers rarely leave: under 1% of the rented space ended each quarter in 2025. Even a rival, VNET, says relocating servers is operationally difficult.
But staying is not the same as paying more. As old contracts renew, the rent per square metre is falling about 3% a year, and in 2025 GDS wrote down RMB1.6B of sites because prices came in lower than planned.
Who could beat it
The rivals in full
The biggest rival is the customer itself. Alibaba has committed at least RMB380B over three years to its own cloud and AI buildings, and two customers account for about half of the space GDS has rented out. The three state telecom companies also run more than half of the market's data centres, and VNET competes for the same buyers.
Each rival
China Telecom, China Unicom and China Mobile data centres0728.HK0762.HK0941.HK
the whole lease, sold with the carrier's own network; 55.7% of the 2025 market by Frost & Sullivan's count
VNET's wholesale data centresVNET
the same wholesale lease to the same hyperscale buyers, from 889MW in service
Sinnet and Range Intelligent Computing300383.SZ300442.SZ
wholesale capacity in Beijing, Hebei and the northern hubs, listed in Shenzhen
Chindataprivate
hyperscale campuses built for a single anchor tenant in the western hubs
Private data-centre platforms building in Singapore, Malaysia and Indonesiaprivate
DayOne's hyperscale tenants and the power and land approvals its growth needs
Alibaba's own data centresBABA
the lease itself, when the cloud provider builds for its own servers
How big it could get
The market and who has signed
The market
Frost & Sullivan expects China's data-centre services market to reach RMB339.0B by 2030, growing 13.2% a year, with independent landlords like GDS growing faster than the telecom companies. The government's own target is bolder: it wants AI computing power to grow about sixfold by 2030. Nobody publishes how many megawatts of buildings that needs, so GDS's share cannot be measured.
Who has signed
In the first half of 2026 customers signed for 470MW on take-or-pay terms, meaning they pay whether or not they fill the space, and reserved another 600MW beside it. GDS does not name them, beyond saying each of the three largest cloud companies signed.
Who runs it
Leaders, pay and the cash
Who runs it
William Huang founded GDS and has run it since 2002; Daniel Newman has run its finances since 2011. Insiders own only about 3% of the company, but Huang's special shares give him well over half the votes, which some Chinese customers now require. Pay is mostly shares that vest with time and with share-price or other goals; the filing does not say which goals they must hit.
The last report
The latest report
GDS books a record 470MW in the half and doubles its 2026 target
What it did to the thesis Strong quarter
It proved the demand turn is real and binding, and left open the part the view hangs on: whether the new space fills on time in 2027 while older contracts keep repricing down.
The report
The second quarter of 2026 brought steady numbers and a big change in orders. Revenue rose 6.5% and profit before interest and depreciation 2.5%, as older contracts renewed at lower prices.
Orders were the story: 260MW booked in the quarter and 470MW in the half, a record, so management doubled its target for the year to 1GW and raised its revenue and profit guidance. The catch is timing. Customers move in over about a year after delivery, so most of this lands in the second half of 2027.
What's happening now
The background
The shares have kept falling since 2 October, when they were already near their low for the year, with no company news we could find to explain it.
Two tests are close. DayOne made its listing papers public on 5 October, but no price is set and it was reported weighing a later or smaller listing. GDS reports its third quarter on 25 November, the first look at whether orders are on track for 1GW.
Where we disagree
Both sides in full
The market
Most analysts see a cheap AI play with a valuable DayOne stake the price ignores.
We believe
The China business is worth about what it costs to build and DayOne about its last price, which together make today's price.
The argument
The cheap look comes from counting DayOne on top of a business whose debts and preferred shares come before the common shares. Once those are counted, little is left over.
We would be wrong if 2027 move-ins more than double as promised and each megawatt earns what management says, or if DayOne lists well above its reported target.
What it's worth
How we value it
No fair value yet: on the filed figures its debt outweighs what the business earns, so a value needs its own build of the data-centre assets before it can move the view.
The bet
This bet is on one thing: the DayOne stake. The China business is counted separately, at what it would cost to build again.
If DayOne lists at the value reported in the press, GDS's stake is worth a little more than the value GDS already puts on it, so today's price rises only slightly. If the listing fails, the China business still covers about half of today's price. That is an expensive bet: the good outcome is already mostly paid for.
The bet
Seen as a bet
This bet is on one thing: the DayOne stake. The China business is counted separately, at what it would cost to build again.
If DayOne lists at the value reported in the press, GDS's stake is worth a little more than the value GDS already puts on it, so today's price rises only slightly. If the listing fails, the China business still covers about half of today's price. That is an expensive bet: the good outcome is already mostly paid for.
What could go wrong
Every risk
The risks
Two customers
The two largest customers hold about half the space rented out; either building more for itself would leave halls empty.
Borrowed growth
This year's building costs about RMB10B against under RMB2B of cash earned in the first half, so debt keeps rising.
Chips may not arrive
US export rules limit China's access to AI chips; customers without chips move in late, and rent starts late.
Prices keep falling
Old contracts renew lower, and the 2025 write-down shows some sites earning less than planned.
DayOne prices low
A listing below the last private price, or a delay, would cut the stake GDS values at about two-fifths of the market value.
Structure and politics
GDS runs through contracts Chinese regulators could challenge, and US-China tension touches both its chips and its investors.
What we still do not know
What DayOne earns and owes: GDS's filings carry only its balance sheet and capacity, and its own listing papers are not public yet.
Who the two largest customers are, and how much of this year's orders they signed.
What a megawatt rents for on new contracts compared with old ones.
What changes our view
Every route to a new view
Every route to a new view
→ Attractive
Customers move into at least 470MW of new space in 2027, double this year. DayOne lists at or above its last private valuation of about $10.9B.
→ Unattractive
Orders in any quarter fall below 120MW, under half of the second quarter's. Debt, net of cash, rises above 6 times a year's profit before interest and depreciation, from 4.7 times.
The next test
It supports the case if
Nine-month bookings of 700MW or more and Q3 adjusted EBITDA of RMB1,400M or more.
It pressures the case if
Nine-month bookings below 700MW or Q3 adjusted EBITDA below RMB1,400M.
We would also be wrong if
DayOne's listing slips past 2027.
Rent per square metre stops falling by the end of 2027, which would help.
The verdict
Scores and how we weighed it
The scores
Growth
3 of 4
Business quality
3 of 4
Financial position
2 of 4
Execution record
3 of 4
Per share capture
2 of 4
Potential
2 of 4
Risk profile
3 of 4
Valuation
3 of 4
Evidence quality
3 of 4
How we weighed it
ValueMixed
No fair value is drawn: debt outweighs what the business earns on the filings, so the price is read against replacement cost and the overseas stake.
BusinessMixed
Customers stay, but older contracts reprice down and the building is funded with debt.
GrowthPositive
AI orders hit a record and the demand turn is binding.
RisksNegative
Two customers hold about half the space, building costs far exceed cash earned and US chip rules may slow customers.
RivalsMixed
No filed rival compares directly, so the price is judged on its own.
The groupMixed
It sits in none of our ranked groups; Chinese cloud spending moves it.
Latest newsPositive
The last report proved the demand turn real; the timing of new space filling stays open.