Alibaba Wants 20 Gigawatts by 2032. Six Years of Capex Just Covers the Buildings.

Abstract

On 22 September in Hangzhou, Eddie Wu gave Alibaba its first public capacity target: "by 2032, the global data center capacity operated by Alibaba Cloud will surpass 20GW." He attached no money to it. No yuan figure, no dollar figure, no megawatt count for the fleet he has today, and no statement of whether those 20 gigawatts will be owned, leased or built by somebody else. The word "billion" appears once in the entire keynote, in an analogy about drug development.

The wires supplied a number anyway. Bloomberg's story of the same morning runs "it's committed to spending more than $53 billion over a three-year period to expand its AI capabilities" one sentence before "on Tuesday, Wu set a target of 20 gigawatts." Both sentences are true and neither dates the first one. The $53 billion is RMB 380 billion, announced on 24 February 2025 as "at least RMB 380 billion ($53 billion) over the next three years" — a clock that runs out around February 2028, four years before the target it is now sitting next to. Management told the August call that RMB 190 billion of it had been spent by the end of June. At the 6-K's RMB 6.7851, that leaves $28.0 billion inside the envelope.

So the obvious question is what 20 gigawatts costs, and the answer is the finding here: nobody publishes a price for a megawatt of Alibaba's capacity, and the figure that looks most like one is a constant. Epoch AI's database carries a total capital cost for 87 AI campuses including Alibaba's Zhangbei site, which it puts at $6.40 billion. Divide by Zhangbei's 169 MW of IT power and you get $37.88 million per megawatt. Do the same division on all 87 and you get $37.88 million every time, to five significant figures, with compute at 68.93% of the total every time. The per-campus cost is one stylised model of a Nvidia GB200 campus, multiplied by satellite- derived megawatts. Epoch says so: "a stylized model, not an estimate for any specific facility."

One half of that constant survives the journey. Its facility slice works out at $11.8 million per megawatt, and JLL's 2026 outlook — built from construction-market data rather than from a bill of materials — puts global shell and core at $11.3 million per megawatt. Two unrelated methods, 4% apart. The building is knowable. The other two-thirds is silicon, it is the part where Alibaba's in-house T-Head chips and Chinese construction costs would make its number genuinely different from Microsoft's, and no one outside Alibaba has it.

That is enough for the collision. Alibaba spent $9.975 billion of capex in the June quarter. Hold that flat through the end of 2032 — twenty-five more quarters, no growth, no let-up — and the total is $249 billion, or $12.47 million per megawatt spread across 20 GW. Shell and core alone costs $11.3 million. Six years of every capital dollar Alibaba spends, on anything, would just cover the concrete, and leave about a million dollars a megawatt for the chips that are supposed to be two-thirds of the bill. Which is why the verb matters: capacity operated by Alibaba Cloud, built by Alibaba joining "forces with all our partners", is a target that arithmetic permits. Capacity owned by Alibaba is not.

Table of contents

1. What Wu said, and what he left out
2. Where the $53 billion actually comes from
3. Every published price for a megawatt is the same number
4. The half of the cost that is knowable
5. The collision: 20 gigawatts against a capex run rate
6. Why "operated" is the word that carries it
7. One flagship cluster is about 2% of the target
8. What it does to our own numbers
9. What would settle it

1. What Wu said, and what he left out

The target, from Alibaba's own newsroom, in full:

In response, Alibaba will join forces with all our partners to commit fully to AI infrastructure development, forging an AI cloud designed for the Machine Intelligence era. Our target is that by 2032, the global data center capacity operated by Alibaba Cloud will surpass 20GW, fueling the industry's exponentially rising demand for AI.

It answers the paragraph immediately before it, which is about scarcity rather than ambition: "the industry's mid-to-long-term demand far outpaces our supply capabilities. Global shortages across the AI data center supply chain are currently limiting the speed at which we can scale our compute infrastructure." Read in order, 20 GW is what Alibaba wants to be able to supply, not what it has told anyone it will buy.

The absences are the subject of everything below, so they are worth listing plainly. The keynote contains no RMB figure, no dollar figure, no capital plan, no current fleet megawatts, no split between owned, leased and partner capacity, and no statement of whether 20 GW is IT power or facility power. It gives the new Zhenwu V900 chip — "the most powerful AI chip in China today", which is Alibaba's claim and not a benchmark result — no power draw and no production date. The one number attached to the chip is that "a single cluster built on V900 can support up to 500,000 cards", which section 7 turns into a bound.

The IT-versus-facility omission alone is worth a fifth of the answer. Epoch's Zhangbei entry carries 169 MW of IT power against 203 MW of facility power, an implied PUE of 1.20. Twenty gigawatts of facility power is 16.65 GW of IT power, and at any per-megawatt rate that is 17% less equipment to buy.

2. Where the $53 billion actually comes from

Alibaba's announcement of 24 February 2025 commits "at least RMB 380 billion ($53 billion) over the next three years" to cloud computing and AI infrastructure. The dollar figure is Alibaba's own conversion at the time; at the RMB 6.7851 the August 6-K uses, RMB 380 billion is $56.0 billion today.

Three years from February 2025 ends around February 2028. Management told the 20 August call that "as of the end of the June quarter this year, we had already spent RMB 190 billion with progress broadly in line with our expectations" — exactly half, sixteen months into a thirty-six-month clock. The remaining RMB 190 billion is $28.0 billion, and it has about six quarters to run.

Nothing was added to that envelope on 22 September. The keynote does not mention it, and the same-day coverage that repeats the $53 billion is repeating February 2025, not reporting a new commitment.

What the wires did was place the two facts next to each other. Bloomberg's paragraph reads "it's committed to spending more than $53 billion over a three-year period to expand its AI capabilities, and raised about $10.2 billion from a follow-on share offering in Hong Kong in August", and the next sentence begins "on Tuesday, Wu set a target of 20 gigawatts." Nothing there is false. Nothing there dates the $53 billion either, and a reader who carries it forward into the 2032 target is out by four years on the calendar and, as the next three sections show, by more than an order of magnitude on the money.

3. Every published price for a megawatt is the same number

To price a capacity target you need dollars per megawatt. The most-cited public source is Epoch AI, which maintains a database of AI data centres with a "total capital cost" column, and which carries Alibaba's Zhangbei campus at 169 MW of IT power, 203 MW of facility power, roughly 31,900 H100-equivalents and $6.402 billion of capital cost.

$6.402bn ÷ 169 MW = $37.88 million per megawatt of IT power. That is the figure a piece about Alibaba's costs would reach for, and it is the wrong one, for a reason that is visible the moment you run the division twice.

Across every campus in that dataset with a cost and a megawatt figure — 320 dated rows covering 87 sites, from Microsoft's 2,263 MW Fairwater Wisconsin site to a 47 MW Microsoft building — there is exactly one value of dollars per megawatt of IT power: $37.882 million. There is exactly one compute share: 68.93%. Zhangbei's $6.402 billion is 169 × $37.882 million. It contains no information about Alibaba.

The constant has a birthplace. Epoch's May 2026 cost model for a hypothetical one-gigawatt AI data centre built on Nvidia GB200 NVL72 systems puts up-front capital expenditure at $37.883 billion — the same number, at gigawatt scale. Epoch describes it as "a stylized model, not an estimate for any specific facility", and notes that "actual costs will vary with server choice, facility design, location, financing, and power strategy." Applied to a Chinese campus that cannot legally buy GB200s and increasingly fills its racks with Alibaba's own T-Head silicon, every one of those five variables is different.

None of this is a criticism of Epoch, which labels the model and publishes the inputs. It is a warning about what happens downstream: a stylised US unit cost, multiplied by a satellite estimate of megawatts, becomes a dollar figure attached to a named company, and from there it reads like a disclosure.

4. The half of the cost that is knowable

Split the constant and the two halves behave completely differently.

Epoch's model puts the facility at $11.433 billion of the $37.883 billion, and the per-campus database splits the same total 31.07% construction, 68.93% compute — $11.77 million per megawatt of building and $26.11 million of silicon and network.

JLL's 2026 Global Data Center Outlook, which is assembled from construction-market pricing rather than from a hardware bill of materials, forecasts the global average shell-and-core cost at $11.3 million per megawatt for 2026, up 6% on the year, from $7.7 million in 2020 and $10.7 million in 2025 — a 7% compound rate. JLL is explicit that the figure excludes tenant fit-out, "which can cost as much as $25 million per MW for AI infrastructure."

Two methods with nothing in common land 4% apart on the building. That is about as good a cross-check as this subject offers, and it means the concrete, steel, transformers and cooling in a gigawatt cost somewhere near $11.3 billion wherever you are, with local labour moving it rather than transforming it.

The silicon does not behave that way at all. It is two-thirds of the money, it is the part export controls have severed between the US and China, and it is the part Alibaba makes itself. T-Head's transfer price to Alibaba Cloud is not public, and no filing discloses a cost per accelerator or a cost per megawatt. Bloomberg has reported that Chinese data centres "in general cost less than in the US because of cheaper labor, component and construction costs" without quantifying it, which is exactly the shape of the gap: everyone agrees Alibaba's number is lower, nobody publishes how much lower.

So the honest position is a bound rather than a price. A megawatt of Alibaba capacity costs at least the building, somewhere around $11 million. It costs at most something like the stylised constant, $37.9 million, because that assumes the most expensive accelerators in the world bought at list. The true figure is inside that range and the range is a factor of three.

5. The collision: 20 gigawatts against a capex run rate

Now put the target against the money, at every rate in the public record.

At the stylised all-in constant, 20 GW costs $758 billion — 13.5 times the whole RMB 380 billion programme and 27.1 times what is left in it. At bare shell and core with not one chip installed, 20 GW costs $226 billion, still 4.0 times the programme and 8.1 times the remainder. Run it the other way and it gets blunter: for the remaining $28.0 billion to deliver 20 GW, a megawatt would have to cost $1.40 million, an eighth of the cheapest shell in the world before anything is plugged in.

Whatever Alibaba's true cost per megawatt is, and it is genuinely unknown, the 20 GW is not inside the February 2025 envelope. That conclusion survives a factor-of-three error in either direction, which is why it does not need the constant to be right.

The more interesting number is the run rate, because it is Alibaba's own. Capital expenditure in the June quarter was RMB 67,678 million, $9.975 billion, up 75% year over year and 2.48 times the group's RMB 27,329 million of adjusted EBITA. Free cash flow was an outflow of $6.584 billion. Take the $9.975 billion and hold it flat for the twenty-five quarters from the September 2026 quarter through the end of 2032:

$9.975bn × 25 = $249 billion. $249bn ÷ 20,000 MW = $12.47 million per megawatt.

That is the break-even, and it is the sentence everything above was built for. If Alibaba spent every capital dollar for six years on nothing but data centres — not a warehouse, not a logistics vehicle, not a server for the e-commerce business — it would arrive at 2032 with about $12.5 million per megawatt of 20 GW. Shell and core costs $11.3 million. The whole capital programme of one of the largest companies in China, sustained flat for six years and spent on one thing, just covers the buildings.

Three ways out of that, and they are not exclusive. Capex grows a great deal from here, which the free cash flow and the 2.48x EBITA ratio already make expensive. Alibaba's true cost per megawatt is far below the constant, which T-Head makes plausible but which nobody can size. Or most of the 20 GW is not Alibaba's to pay for.

6. Why "operated" is the word that carries it

Wu did not say owned. He said "the global data center capacity operated by Alibaba Cloud", in the same sentence as "join forces with all our partners", and he has been assembling that structure on earnings calls all year.

On the 13 May call, discussing how to get compute without buying it: "There are different ways to get that compute capacity. Some of it can be capex, part of it can also be opex, and we're actually now acquiring quite a bit of computing capacity using opex", as reported by DCD. On the 20 August call, listing the ways to improve returns on invested capital: "co-building data centers with partners", and "pre-charging and receiving pre-payments for compute-based services."

Set that against the 31/69 split and the shape of the fork gets sharp. A partner who builds and owns the campus takes the 31% off Alibaba's balance sheet — real money, $11.3 million a megawatt, $226 billion across 20 GW — and leaves the 69% wherever the chips are bought. Opex- acquired capacity moves both halves off the capital line at once, at the cost of paying somebody else's margin forever. Prepayments move the timing rather than the total, and on the one line where they would show they are not yet moving much: deferred revenue and customer advances went from RMB 77,415 million to RMB 79,439 million over the June quarter, a rise of about $298 million against $9.975 billion of capital expenditure in the same three months.

None of which is disclosed. Alibaba publishes no owned-versus-partnered split, no operated megawatts today, and no quantum for the opex-acquired capacity Wu called "quite a bit". A reader who wants to know how much of 20 GW lands on Alibaba's own balance sheet is not being kept from the answer by a hard search. The answer does not exist publicly.

7. One flagship cluster is about 2% of the target

The keynote's other number is a useful sanity check on what 20 GW means physically: "a single cluster built on V900 can support up to 500,000 cards."

The V900's power draw is not published. The keynote gives none, T-Head's public material gives memory and interconnect bandwidth and no watts, and the shipping M890 supernode has no published figure either. So the conversion below uses Nvidia-class proxies, which are ours and are proxies, not Alibaba figures:

• At 400 W a card, 500,000 cards is 200 MW of IT power, 240 MW of facility power at PUE 1.2 — 1.2% of 20 GW.
• At 700 W, 350 MW of IT, 420 MW of facility — 2.1%.
• At 1,000 W, 500 MW of IT, 600 MW of facility — 3.0%.

So Alibaba's flagship cluster, the largest single machine it has described, is a few hundred megawatts, and 20 GW is 33 to 83 of them. The target is a fleet-building programme measured in dozens of sites over six years, not a handful of showpieces, and that is consistent with reading it as an ambition about supply rather than a capital commitment.

8. What it does to our own numbers

Nothing, is the short answer, and the size of the nothing is worth printing.

Our Alibaba model treats cloud as a growth vertical rather than a capacity one, for the reason the target has just confirmed: Alibaba discloses no megawatts, no accelerator count and no utilisation, so there is nothing to build a capacity driver on. The 20 GW is the first megawatt figure the company has ever published and it still supplies none of the three inputs such a driver needs — current fleet, utilisation, and revenue per megawatt-year.

What the announcement does is move weight between two cases the model already prices, through one assumption: capital expenditure as a percentage of cloud revenue, published at 112% today and assumed to glide to 35% at the end of the forecast. Both are our assumptions, not disclosures, and the glide is the single most valuable thing in the file:

An owned, fitted 20 GW is the "no glide" row, and it is worth −$41.89 an ADS against a base of $111.68. An operated fleet financed by partners, prepayments and opex is the 80% row, a small upgrade. Moving the cloud growth rate by more than five points a quarter is worth $10.27. So on our own numbers the ownership question is worth roughly four times the growth question, and the speech is compatible with both readings. Every figure in that table is model output on our assumptions, against a $118.47 basis price as of 27 August; Alibaba's last US close in our own series is $108.54 on 17 September.

9. What would settle it

In descending order of how much each would change the answer, and each is checkable.

1. A megawatt figure for the current fleet. Every ratio above is missing its bottom half because Alibaba has never published one. One number in a 20-F or on a call turns 20 GW from an unanchored target into a multiple.
2. The owned-versus-operated split. If Alibaba says what share of the 20 GW will sit on its own balance sheet, the fork in section 6 collapses to one side and roughly $42 an ADS of our own range with it.
3. A successor to the RMB 380 billion programme. The current one ends around February 2028. Whether the next one is sized to a capacity target, and at what yuan per megawatt, is the single event that would close the gap. Watch the next two earnings calls.
4. Whether 20 GW is IT power or facility power. Worth 17% of every cost figure here, and it would take Alibaba one clause.
5. The V900's power draw. Published watts turn section 7 from a bound into a calculation, and give the first real read on whether T-Head's silicon makes Alibaba's cost per megawatt structurally lower than the constant everyone is using.
6. Deferred revenue and customer advances. Prepayments for compute show up there, and last quarter the whole line grew by about 3% of one quarter's capex. If that starts compounding, prepay financing is doing real work; until it does, it is a mechanism rather than a source.

One disclosure overturns all of it: an Alibaba cost per megawatt materially below $11 million. That would mean the building itself is cheaper in China than JLL's global average by more than the range allows, and the run-rate collision in section 5 would stop binding. Nothing in the public record currently supports it, and Alibaba has never published the number.

Primary sources worth reading

• Eddie Wu's full Apsara keynote, Alizila, 22 September 2026 — the whole speech in the first person, and the only place to check what was and was not said.
• Alibaba's February 2025 investment announcement — "at least RMB 380 billion ($53 billion) over the next three years", in the company's words.
• The August 6-K results announcement — capex, free cash flow, cloud revenue, the RMB 6.7851 conversion rate.
• Epoch AI's data-centre dataset and its one-gigawatt cost model — download the CSV and run the division yourself; the constant takes about a minute to find.
• JLL's 2026 Global Data Center Outlook — the shell-and-core series, and the fit-out figure it excludes.
• Our own prior coverage of the same capital programme, which measured it against everything except capacity because there was no capacity figure to measure it against: the $10.2 billion placement against one quarter of capex, the placement against AI revenue, and the June quarter itself.

Eddie Wu's keynote of 22 September 2026, as Alibaba published it, is the source for the capacity target, the supply-constraint framing, the V900 description and the cluster figure, and contains no capital figure. The three-year programme and its dollar label are Alibaba's announcement of February 2025; the cumulative spend against it came on the August 2026 call rather than any filing, as did the co-building and prepayment quotes; the opex one is DCD's report of the May call. Capital expenditure, cash flow, cloud revenue, adjusted EBITA and the conversion rate are from the results announcement for the quarter ended 30 June 2026. Campus power and capital cost are Epoch AI's, whose per-campus figure is a stylised model Epoch says is not an estimate for any specific facility; shell-and-core and fit-out costs are JLL's 2026 outlook; the quoted wire wording is Bloomberg's of 22 September. Ours: the watt conversions, every dollar conversion, share and multiple, the run-rate total and the per-megawatt break-even. Fair values are our model's assumptions as of 27 August; the closing price is our own series for 17 September.