Alibaba Group CEO Eddie Wu (吴泳铭) announced on September 22 at the company's annual Apsara Conference in Hangzhou a goal to expand Alibaba Cloud's global data center capacity to more than 20GW by 2032. On the same day, chip subsidiary T-Head unveiled a new AI accelerator called the Zhenwu V900, claiming triple the performance of its predecessor.
As the industry-wide metric for AI infrastructure scale shifts from accelerator counts to power consumption and capacity, the 20GW figure places Alibaba alongside other companies' large-scale AI infrastructure plans. However, attempting to evaluate this number from the outside reveals several critical pieces of information that are missing—information needed as a basis for comparison.
20GW by 2032, and a V900 with 216GB of Memory
What Wu presented was a target for cloud infrastructure expressed in power capacity: expanding Alibaba Cloud's global data center footprint to over 20GW by 2032. Wu explained that the company's long-term strategy rests on three pillars: AI models, AI chips, and AI cloud.
The choice to foreground power capacity rather than investment figures reflects a broader shift in what constrains AI training and inference—not just the number of accelerators a company can procure, but the power and cooling capacity needed to run them.
The Zhenwu V900, unveiled on the same stage, comes equipped with 216GB of memory and inter-chip interconnect bandwidth reaching 1,200GB/s. It also natively supports FP8 and FP4 computation. T-Head stated the chip delivers three times the performance of its predecessor, the Zhenwu M890, and Wu described it as "currently China's most powerful AI chip."
However, these are all internal comparisons provided by Alibaba itself; no absolute computational performance figures or third-party benchmark results have been published.
A roadmap for related technologies was also presented. For Arm-based server CPUs, the company plans to launch the Yitian 720 and 730 in Q3 2027, with the 730 reportedly delivering 1.4 times the single-core performance of the Yitian 710, as measured by SPECint2017/GHz.
On the model side, Qwen 4 is currently in training, with the subsequent Qwen 4.5 and Qwen 5 series reportedly planned at the scale of 5 to 10 trillion parameters. The fact that AI chips, CPUs, models, and cloud infrastructure were all presented together at a single event itself signals that Alibaba views these as a unified AI infrastructure strategy.
At the same time, Wu acknowledged that supply is currently failing to keep pace with demand. "The industry's medium- to long-term demand far exceeds our supply capacity," he said, describing customer demand as extremely robust. The 20GW target also represents a plan to resolve this supply shortfall over the next six years.
Following the announcement, Alibaba's Hong Kong-listed shares reportedly rose 4%.
What's the Difference Between 1,000 Chips and 500,000 Chips?
T-Head's explanation includes two distinct figures—"over 1,000" and "up to 500,000"—which carry different meanings and need to be considered separately.
A configuration in which more than 1,000 Zhenwu V900 chips are tightly interconnected to operate as a single system is called a "super node." Separately, using the company's proprietary ICN Switch, a single compute cluster can reportedly be scaled up to 500,000 chips.
The former refers to a setup where numerous accelerators are connected at high speed to function as if they were one giant computer. The latter refers to networking many such machines together to scale up into an even larger cluster.
Therefore, if one extracts only the larger figure and interprets it as "running 500,000 chips as a single system," the meaning of the announcement changes.
Models at the scale of 5 to 10 trillion parameters cannot possibly fit within the memory of a single accelerator. When a model is split across multiple chips, large volumes of data must then be exchanged between those chips. If this communication time exceeds computation time, adding more chips yields diminishing performance gains. This is why configurations like super nodes become necessary.
The 216GB memory capacity allows a single accelerator to hold a larger portion of the model. The 1,200GB/s interconnect exists to rapidly transfer that data between multiple accelerators.
Though they may appear to be separate performance metrics at first glance, both address the same underlying challenge: how efficiently a large number of accelerators can be treated as a single, unified computing resource.
Native support for FP8 and FP4 is an extension of this same theme. These are low-precision formats that represent numbers using 8 or 4 bits; in exchange for reduced computational precision, they allow more parameters and operations to be handled within the same memory capacity and bandwidth.
"Native support" means the hardware can process these formats directly, rather than converting them via software. The plan for models with up to 10 trillion parameters, the 216GB memory, support for low-precision computation, and the super-node configuration are all connected by a consistent design philosophy.
The Zhenwu V900 is scheduled to actually reach the market in Q1 2027—that is the timeframe given for the start of mass production and commercial availability. The Yitian 720 and 730 will follow two quarters later. More than a year will pass between the announcement and the first product actually reaching the market.
The Numbers Disclosed, and Three Missing Reference Points
Reviewing articles published on the day of the announcement by The Register, Tech Wire Asia, China Daily, and IT Home, three things remain unclear.
First, the power capacity of the data centers Alibaba currently operates. Second, the foundry and manufacturing process that will be used to produce the Zhenwu V900. Third, the chip's absolute computational performance, or any third-party benchmark results.
In other words, Alibaba has not yet disclosed three benchmarks critical to externally verifying the 20GW plan and the Zhenwu V900: current operating capacity, the manufacturing foundry and process node, and third-party performance verification.
Without knowing the current capacity, it's impossible to assess the practical meaning of the 20GW target. How much expansion is needed to reach the goal, what pace of annual growth is required, and how much investment that would demand—none of this can be calculated without a starting figure.
It would be easier to understand the scale of expansion if expressed as "a multiple of current capacity," but since Alibaba has not disclosed the denominator—current capacity—no such multiple can be calculated. While we know six years remain until 2032, how much expansion is actually needed to get there remains unclear.
The lack of disclosed manufacturing conditions poses a more direct problem when it comes to assessing the feasibility of the mass-production schedule.
Amid ongoing U.S. export restrictions, where Chinese companies can secure advanced logic semiconductor manufacturing capacity significantly affects whether production plans can succeed. Which foundry will handle production, at what process node, and at what volume? Without answers to these three questions, only a Q1 2027 mass-production start date has been given, making it difficult for outside observers to judge how realistic that schedule is.
The same applies to performance. The "3x" figure is merely a relative value compared to the M890. Since the M890's absolute performance was never disclosed either, it's impossible to judge how it compares to AI accelerators from companies like NVIDIA or AMD.
That said, this only reflects what could not be confirmed in coverage published on the day of the announcement; it doesn't rule out the possibility that additional information may be disclosed later through technical documentation or investor materials. Also, since full articles from outlets requiring paid subscriptions could not be reviewed, this is not a definitive claim that "this information exists nowhere."
Still, the fact that all three points remain unclear across the major global tech outlets that reported on the same day offers a useful clue as to what Alibaba chose to prioritize in this announcement.
20GW, 10GW, 5GW: Same Unit, Different Scope
The practice of publishing power capacity as a target for AI infrastructure has spread among hyperscalers over the past few years. However, even when companies use the same GW unit, what exactly gets counted varies significantly by company.
Organizing the major publicly announced plans by target value, target date, and scope reveals these differences.
| Plan | Target | Target Date | Scope |
|---|---|---|---|
| Alibaba Cloud | 20GW+ | 2032 | Entire globally operated data center fleet |
| OpenAI Stargate | 10GW | 2029 | Combined total across multiple U.S. sites |
| Meta Hyperion | 5GW | Phased expansion | Single campus in Louisiana |
Alibaba's 20GW is a capacity target for its cloud business as a whole. By contrast, Meta's Hyperion (5GW) refers to a single campus, while Stargate (10GW) is a combined figure across multiple U.S. sites—the scopes differ. As a result, the numbers alone don't allow for a simple size comparison.
Additionally, all three figures are future targets set by the companies themselves, not current operating capacity. The target dates also don't align—2029, 2032, and phased expansion—so these aren't comparisons anchored to the same point in time either.
What this table reveals is not a ranking, but a difference in scope of measurement.
Alibaba's 20GW figure covers the entire business, including existing regions, making it closer to a company-wide power and data center strategy than an individual project—unlike the other companies' figures. Therefore, it's more accurate to understand 20GW not simply as "the largest plan," but as "the figure with the broadest scope of aggregation."
Simply concluding that this is "four times the scale" of Hyperion's 5GW would mean comparing the wrong things. Furthermore, since Alibaba hasn't disclosed its current operating capacity, it's also unclear how much of the 20GW represents genuinely new capacity versus existing infrastructure.
The First Opportunity for Verification: Q1 2027
This announcement did not include many items with specific dates attached. Arranged chronologically, they look like this:
| Timing | Details |
|---|---|
| September 22, 2026 | 20GW target and Zhenwu V900 announced at Apsara Conference |
| Q1 2027 | Zhenwu V900 mass production and commercial availability begin (planned) |
| Q3 2027 | Yitian 720 and 730 launch (planned) |
| 2032 | Alibaba Cloud's global data center capacity reaches 20GW+ |
None of the plans Alibaba presented can be verified as actual results within 2026. The first opportunity for verification will come in Q1 2027, when Zhenwu V900 mass production and commercial availability are scheduled to begin, with the ultimate target being 20GW by 2032.
Regarding Qwen 4, the company only stated that it is "currently in training"; since no release date was given, it has been excluded from this table.
Additionally, no interim milestones have been set for the six years leading up to 2032. As a result, there is currently no yearly indicator available from the outside to confirm whether data center expansion is proceeding as planned.
For cloud users in Japan as well, this announcement is unlikely to affect actual services or procurement conditions until at least Q1 2027.
Moreover, Wu himself has acknowledged that industry demand far exceeds Alibaba's supply capacity. Even once mass production of the Zhenwu V900 begins, that won't necessarily translate immediately into improved availability in overseas regions.
Still, it's clear what information would make this easier to evaluate going forward.
That would include: mass-produced Zhenwu V900 units reaching external users and producing benchmark results comparable to competing products; disclosure of the foundry and process node responsible for manufacturing; and continued public disclosure of Alibaba Cloud's actual data center capacity figures and expansion progress.
Once these three pieces of information are available, the 20GW figure will transform from a mere long-term goal into a plan whose progress investors and customers can verify from the outside.