Alibaba Cloud has a 2032 capacity goal and a next accelerator. The figures needed to size either, from power basis to peak FLOPS, are not yet public.
At its Apsara Conference in Hangzhou on September 22, 2026, Alibaba set two markers for the next stage of China's AI infrastructure. Group CEO Eddie Wu said Alibaba Cloud aims to operate more than 20 GW of data center capacity worldwide by 2032. Alibaba's chip unit, T-Head, introduced the Zhenwu V900, an accelerator the company says delivers three times the performance of the Zhenwu M890 it launched in May, carries 216 GB of memory and 1,200 GB/s of inter-chip bandwidth, and enters mass production in the first quarter of 2027. According to Alibaba Cloud's release, a supernode server built around the V900 can support a cluster of up to 500,000 cards.
The announcement came five days after Huawei launched its Atlas 960E SuperPoD at Huawei Connect in Shanghai. Huawei says the pod can be linked into clusters of up to 512,000 NPUs over a two-tier Clos network, or up to one million with a multi-rail topology. Huawei first put a cluster of more than 500,000 NPUs on its roadmap at Huawei Connect in September 2025. Alibaba's 500,000-card figure lands at the level Huawei published a year ago, while NVIDIA says it has shipped only a fraction of the H200s it is licensed to sell into China.
Alibaba's two announcements leave out the figures a reader would need to size them: whether 20 GW counts IT load or total facility power, what Alibaba Cloud operates today, and the V900's peak throughput, memory type, process node and foundry. What follows separates what Alibaba has published from what it has not, and compares the published numbers with figures other operators and Huawei have disclosed.
The release includes one sentence from Wu: "our target is that by 2032, the global data center capacity operated by Alibaba Cloud will surpass 20GW." TechNode Global describes it as a future target rather than deployed capacity, and the wording matters in two places. "Operated" could include facilities Alibaba leases or runs with partners, not only sites it owns. And "capacity" is not qualified. Operators usually quote either IT load, the power that reaches servers, storage and networking, or facility power, which adds cooling and electrical losses on top. At 20 GW, the gap between the two readings is itself measured in gigawatts, and neither the English release nor the coverage SCN reviewed says which one Alibaba means. Wu also spoke about supply on stage. According to Reuters, he said the industry's mid- to long-term demand far exceeds Alibaba Cloud's supply capabilities, and that the company would begin bringing its AI supernodes online at commercial scale this quarter.
Alibaba also gave no baseline. There is an earlier marker. At Apsara 2025, Wu said Alibaba Cloud's global data center energy consumption in 2032 would be ten times its 2022 level, as 36Kr reported at the time, in a translation carried by KrASIA. SCN did not find that sentence in Alibaba's own English releases from the 2025 event. Energy consumed over a year and installed capacity are different quantities, so the two statements cannot be combined to work out what Alibaba Cloud ran in 2022 or runs now.
The capex record is easier to trace. On February 24, 2025, Alibaba said it would invest at least RMB 380 billion (US$53 billion) over three years in cloud and AI infrastructure. At Apsara 2025, it said it planned to go beyond that figure without naming a new total. The September 22 release attaches no capex number to the 20 GW goal.
Quarterly filings show the pace. In the June quarter 2026 results, group capital expenditure was RMB 67.7 billion (US$9.98 billion), up 75% from RMB 38.7 billion a year earlier. Alibaba gave three reasons: shifts in procurement cycles, more CPU capacity for expected growth in AI agent workloads, and "higher pricing of a broad range of chip components." That last clause lines up with what SCN has reported about memory pricing as data center AI absorbs supply, and the CPU line echoes AMD's case for general-purpose cores in agentic data centers. The same filing folds T-Head into a reporting segment with the cloud business, now called AI Cloud and Compute Services, which posted RMB 48.4 billion in quarterly revenue, up 45%.
Analysis: The RMB 380 billion pledge averages about RMB 127 billion a year. The June quarter alone, annualized, runs above twice that pace. One quarter is a thin basis for annualizing, and the capex figure covers the whole group, but the direction is consistent with Alibaba's statement that it would exceed the original pledge.
No peer publishes a directly comparable figure, so any comparison here is about order of magnitude. Amazon's 2025 shareholder letter says AWS added 3.9 GW of new power capacity in 2025 and expects to double its total power capacity by the end of 2027. Those are power-capacity figures, as Amazon states them: one an annual addition and the other a doubling of a total Amazon does not state, so they show scale without supporting a ranking. Set side by side, a 20 GW total for 2032 and 3.9 GW added in a single year put Alibaba's goal in the multi-gigawatt range that the largest US operators now work in. How much of any such target becomes running capacity depends on grid-connected megawatts, which SCN has described as the build-out's scarcest asset wherever power has to be contracted years ahead.
China's national figures use a different unit. The Ministry of Industry and Information Technology put the country's intelligent computing capacity at 2,185 EFLOPS at the end of June 2026, with nationwide facility utilization at 71.4%. MIIT's 15th Five-Year Plan for the information and communications industry, which MIIT described in a September 7 explainer, targets 9,800 EFLOPS by 2030 and RMB 3.8 trillion of information infrastructure investment from 2026 to 2030, including clusters of 100,000 or more accelerator cards, according to TechNode, which also reports the 2,185 EFLOPS figure as a 177% rise year on year. Neither the ministry figure as carried by Xinhua nor TechNode's report of the plan states the numerical precision behind the EFLOPS count. Because the state targets are in FLOPS and Alibaba's is in watts, no honest conversion exists between them. Alibaba has published no power figure for the V900 either, so there is no way to state what share of 20 GW a maximum-size V900 cluster would draw.
Every V900 figure so far comes from Alibaba. No datasheet, independent benchmark, or MLPerf result exists yet. The release lists four things: three times the performance of the M890, 216 GB of what it calls GPU memory, 1,200 GB/s of inter-chip bandwidth, and native FP8 and FP4 support, with mass production and commercial release in Q1 2027. The chip ships in an upgraded supernode alongside three companion parts, the ICN Switch, the Panmai SmartNIC, and the Zhenyue SSD controller.
Measured against the M890 as described in Alibaba's May 20 release, which lists 144 GB of memory and 800 GB/s of inter-chip bandwidth, the V900 adds 50% to both. The "3x" figure comes without a metric, precision, or workload, so it can't be converted into FLOPS.
The list of what is missing is longer. Alibaba has published no peak FLOPS at any precision, no thermal design power, no process node or foundry, and no scale-up domain size for the V900. TechWire Asia noted that Alibaba did not say who would manufacture the chip or on which node. The release does not say whether 1,200 GB/s is measured in one direction or both, or per link or in aggregate, and each choice changes the comparison by a factor of two or more. Much of Hot Chips 2026 turned on the widening gap between compute and the memory and interconnect bandwidth that feeds it, so pinning down the definition matters.
Memory type is the most consequential gap. The release does not say HBM. Chinese coverage of the May launch and an on-site post by analyst Poe Zhao describe the M890's memory as HBM without naming a generation; Wccftech reported it as HBM3. Alibaba has not confirmed either, and nothing has been published about the V900's memory generation or supplier. The question matters because the US Commerce Department's Bureau of Industry and Security added controls on high-bandwidth memory in its December 2, 2024 rule, and the three leading HBM makers are Samsung, SK hynix, and Micron, according to TrendForce. SCN has traced how that supply concentrates in Korea for NVIDIA's Vera Rubin, where the HBM4 tier is Samsung and SK hynix only. With HBM allocation already shaping which accelerators ship in volume, 216 GB per chip across hundreds of thousands of chips is a large memory order, whatever its source.
The claim that the V900 is China's most powerful AI chip has two versions. Reuters attributes it to Wu's keynote. Global Times attributes a narrower phrasing to T-Head: "the most powerful Chinese AI chip in terms of computing performance." With no compute figure published, neither version can be checked against a rival part, and Alibaba's English release keeps to the relative 3x figure without making the claim at all.
In May, Alibaba gave the V900 a different date. Reuters reported on May 20 that the company said the V900 would follow the M890 in the third quarter of 2027, with a Zhenwu J900 in the third quarter of 2028. Chinese trade coverage of the Alibaba Cloud summit carries the roadmap slide T-Head supplied (Jiemian) and reports T-Head vice president Gao Hui presenting it with the same headline specs Alibaba published on September 22 (chinaaet). TrendForce carried the same dates, citing Mydrivers. The September release says mass production and commercial release will come in Q1 2027, about two quarters sooner.
The May date appears in Alibaba's on-stage roadmap and in Reuters' account of the company's statement, but not in Alibaba's English release from May. The two dates could reflect a pull-in, or they could describe different milestones, with the May date referring to general availability. Alibaba has not said which reading is correct.
Accelerator clusters are built in two tiers. A scale-up domain is a group of chips joined by a high-bandwidth fabric so they can share memory and, for many workloads, behave like one large device. A scale-out network then links many of those domains across racks and buildings. The two numbers answer different questions, and vendors increasingly publish both.
Alibaba's May material gives the scale-up figures for the current generation: ICN Switch 1.0 provides 25.6 Tbps of aggregate bandwidth and full-bandwidth interconnection across 64 accelerators, and the Panjiu AL128 supernode puts 128 accelerators in a single rack. The September release calls the 500,000-card figure a supernode cluster and gives no V900 scale-up domain size. It does give a scale-out figure: the HPN 8.0 Pro network, announced alongside the chip, supports more than 130,000 800G ports in a single cluster, by Alibaba's description.
Analysis: read against the May numbers, 500,000 most likely describes a scale-out ceiling across a network fabric, and Alibaba has not said whether it is a demonstrated or a designed limit. Building a single supercomputer from half a million accelerators spread over multiple halls raises the scale-across networking problems SCN has covered for gigawatt-class US systems. Alibaba's decision to ship its own switch, NIC, and storage controller alongside the accelerator follows the same logic as d-Matrix's move into rack-scale systems: the vendor sells the fabric along with the chip.
Huawei's September 17 disclosure is more complete on these points. It says a single Atlas 960E SuperPoD scales to 4,096 NPUs in one UnifiedBus domain, rated at 8 EFLOPS at FP8 and 16 EFLOPS at FP4, and that multiple pods form a SuperCluster of up to 512,000 NPUs over a two-tier, four-plane Clos network, or up to one million with a multi-rail topology. Alibaba's ceiling sits just below the lower of Huawei's two figures. Huawei has also published the scale-up domain size and pod throughput behind its number; Alibaba has published neither for the V900.
The nearest comparison is Huawei's Ascend 950DT, which Huawei's rotating chairman Eric Xu presented at Huawei Connect 2025 for Q4 2026. By Huawei's figures, it has 144 GB of Huawei-developed HBM, 4 TB/s of memory bandwidth, 2 TB/s of interconnect bandwidth, and 1 PFLOPS at FP8. The Ascend 960, dated Q4 2027 on the same roadmap, is to double the 950's compute, memory capacity, memory bandwidth, and interconnect port count, by Huawei's description. The 950PR, an earlier variant using Huawei's own lower-cost HBM, is already running inference for DeepSeek V4-Pro, and Huawei's parts are now being weighed outside China for sovereign AI projects.
Comparison using vendor figures only: on memory capacity, the V900's 216 GB is larger than the 950DT's 144 GB, and Huawei names its memory type where Alibaba does not. On compute, there is nothing to compare, because Alibaba has published no FLOPS. On interconnect, the V900's 1.2 TB/s is below the 950DT's 2 TB/s, though the two companies may define the figure differently.
NVIDIA's quarterly report for the period ended July 26, 2026 describes how the licensed channel has worked in practice. Starting in February 2026, the US government licensed small amounts of H200 products to specific China-based customers, "but such sales were restricted by the PRC government." NVIDIA took a $0.4 billion H200 charge for excess inventory and purchase obligations in the first half of fiscal 2027, and says the shipments it has made account for less than 1% of Data Center revenue in the latest quarter. The licensed parts must pass US inspection before delivery, making them subject to a 25% tariff on import into the United States; NVIDIA says it has absorbed that tariff and expects to keep doing so.
Analysis: with the licensed channel this narrow, new Chinese AI capacity at the scale Alibaba describes has to come largely from domestic designs. SCN has tracked that shift through the audited financials of China's listed GPU companies and through LineShine, the CPU-only system that returned China to the top of the TOP500. Alibaba reported more than 560,000 Zhenwu units delivered to date in its May announcement; an IDC estimate for calendar 2025 alone, which SCN reported in July, put T-Head's shipments at about 265,000. Alibaba counts more than 650 external customers for the chips as of its June-quarter results.
Several disclosures would let the 20 GW target and the V900 be sized properly: whether the target counts IT or facility power, and what Alibaba Cloud operates now; the V900's FP8 and FP4 peak throughput, memory type and supplier, process node and scale-up domain size; and which V900 milestone the May and September dates each refer to. The first independent check on the chip will come when V900 systems reach customers, which Alibaba has scheduled for the first quarter of 2027.
The SCN Staff is a small AI editorial squad working under human direction. Each agent owns one job.
Scout does the research. It runs down primary sources and checks what's already been published, on SCN and everywhere else, before a story gets written. If a claim can't be traced back to a real document, Scout flags it.
Forge writes. It takes what Scout found and turns it into a draft, argument and sentences and all. Every SCN piece starts here, then gets sharpened.
Cipher handles search: the titles, descriptions, and keyphrase work that decides whether a good article ever gets found. Least glamorous job on the squad. Also one that matters more than it looks.
Pixel makes the visuals. Images, charts, the occasional diagram, all built to SCN's brand instead of pulled from a stock library. When something's easier to see than to read, it goes to Pixel.
Editorial judgment and the final call stay with the humans. So does the fact-checking.
