Alibaba dates mass production of its most powerful AI chip to the first quarter of 2027
T-Head has published a basic spec sheet for the Zhenwu V900 and Eddie Wu claimed clusters of up to 500,000 chips, yet the foundry behind the part is still unnamed, and Wu himself named supply shortages as the limit on Alibaba's compute.
The Product Desk
What happened
- Alibaba's chief executive Eddie Wu unveiled the Zhenwu V900 at the company's Apsara cloud conference in Hangzhou, a processor from its T-Head design unit built for both high precision training and low precision inference.
- T-Head published a basic spec sheet listing 216GB of memory and 1,200 GB/s of inter-chip interconnect bandwidth, with native support for precisions from FP32 through FP4.
- Wu said a single cluster can hold up to 500,000 of the chips, and that the resulting machine will run frontier model training and inference.
- Mass production and commercial release are set for the first quarter of 2027, with Bloomberg reporting that Alibaba will refresh the lineup every year.
- Bloomberg put Zhenwu family deliveries at 560,000 units to more than 400 external customers, while Alibaba's own post claims more than 650 customers for the same family.
Why it matters
- What it limits constraint Anyone betting on the 500,000-chip cluster is also betting on a fabrication route no one has named, with Nvidia's top parts and TSMC both closed off and the domestic alternative short of capacity.
- Who has to decide decision Teams planning next year's compute cannot order this part at all, so the live choice is how much Alibaba Cloud capacity to commit to on the strength of a chip due in 2027.
- What doesn't add up contradiction Wu's growth promises and his supply warning point in different directions: a yearly chip refresh and larger models, alongside his own statement that shortages cap how fast Alibaba can add compute.
- What it costs cost Shareholders carry the build-out while the silicon is still 2027 inventory, with quarterly profit down 75% after almost $10bn of spending.
Next story · Invest · Sep 22
Alibaba pulls its V900 chip forward two quarters to the first quarter of 2027Confirmed13 publishers
A team deciding how much Alibaba Cloud capacity to book for 2027 can work with part of the spec sheet. Per-chip memory and interconnect bandwidth show up the first time a training job runs [7]. The cluster ceiling is a different kind of number. Multiply the 500,000 chips Wu cited by the memory on T-Head's sheet and one machine holds roughly 108 petabytes of attached memory [17]. "The most powerful AI chip in China today, delivering three times the performance of its predecessor, the Zhenwu M890," Wu said of the processor at Apsara [1].
The delivery figures describe something smaller. Bloomberg's count works out at fewer than 1,400 chips for the average external customer, and fewer than 900 using Alibaba's own higher customer number [18]. The cluster ceiling is about 89% of every Zhenwu unit Bloomberg says has been delivered so far [21]. Filling one would take nearly everything the family has shipped.
Nobody has said who makes the wafers. United States export controls keep Nvidia's most advanced accelerators out of China and largely block Chinese firms from using TSMC, and Bloomberg reported that it is unclear whether Alibaba works with SMIC, a capacity constrained rival [12]. Wu said global shortages across the AI data centre supply chain are limiting how fast Alibaba can scale its compute, and that mid to long term industry demand far outpaces what the company can supply [6].
What arrives sooner is rented capacity. Alibaba Cloud starts bringing its AI supernodes online at commercial scale this quarter [11]. Qwen 4 is already training, and Wu put Qwen 4.5 and Qwen 5 at five to 10 trillion parameters against the 2.4 trillion in Qwen 3.8 Max [13]. At the top of that range, that is about four times the current flagship [20]. The capacity goal for 2032 is around 53% of the 37.7GW that Cushman and Wakefield counts under construction in the United States alone [5][10][19]. Citigroup estimates that much capacity could drive more than $160bn of external revenue for the cloud division, according to Bloomberg [16]. Set that against the more than $53bn Alibaba has committed over three years and about $10.2bn raised in an August share sale [14].
For a buyer, three dates carry more weight than the spec sheet. First, when the part can be rented or bought. Second, when a fabrication route gets named or second-sourced. Third, when a customer outside Alibaba trains a frontier model on more than a pilot allotment. Only the first has a published date [4]. Until the second one is answered, the 500,000-chip cluster is a capacity promise from a cloud vendor. Price it the way you would any other 2027 supplier commitment, with a fallback and an exit if the part slips. Wu did not say where the new data centres will be built [11].
What to watch
- Any disclosure of which foundry makes the V900, or of a second source, before the 2027 release.
- Whether the first quarter of 2027 holds, given Alibaba's stated plan to refresh the Zhenwu line every year.
- An external customer naming a Zhenwu order in the tens of thousands of units rather than the low thousands.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence45
- Adoption30
- Hype gap+35
- Incentives75
- Confidence55
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Wu described the Zhenwu V900 as "The most powerful AI chip in China today, delivering three times the performance of its predecessor, the Zhenwu M890."
- [2]
Alibaba's chief executive Eddie Wu unveiled the Zhenwu V900 processor on Tuesday at Apsara, the company's annual cloud conference in Hangzhou; Alibaba said in a post on X that the chip comes from T-Head, its chip design unit, and that it handles both high precision model training and ultra low precision inference.
- [3]
Wu said a single cluster can hold up to 500,000 of the chips, and that the resulting machine will power frontier model training and inference.
- [4]
Mass production and commercial release of the Zhenwu V900 are set for the first quarter of 2027.
- [5]
Wu set a target for Alibaba Cloud to pass 20GW of global data centre capacity by 2032.
- [6]
Wu said global shortages across the AI data centre supply chain are limiting how fast Alibaba can scale its compute, and that mid to long term industry demand far outpaces what the company can supply.
- [7]
T-Head has published a basic spec sheet for the chip listing 216GB of memory and 1,200 GB/s of inter-chip interconnect bandwidth; Simon Sharwood reported the numbers for The Register.
- [8]
The V900 supports FP32 down to FP4 natively, and T-Head reworked the Tensor Core arithmetic unit for better instruction precision in FP8 and FP4, with more scaling factor formats and block size configurations under MXFP8 and MXFP4.
- [9]
Alibaba will refresh the chip lineup every year, Bloomberg reported.
- [10]
The commercial property firm Cushman and Wakefield counts 37.7GW of data centres under construction in the United States alone.
- [11]
Alibaba Cloud will start bringing its AI supernodes online at commercial scale this quarter, and Wu did not say where the new data centres will be built.
- [12]
United States export controls keep Nvidia's most advanced accelerators out of China and largely block Chinese firms from using TSMC; Bloomberg reported that it is unclear whether Alibaba works with SMIC, a capacity constrained rival.
- [13]
Alibaba is already training Qwen 4; Wu said Qwen 4.5 and Qwen 5 should scale to between five and 10 trillion parameters, against 2.4 trillion in the current flagship Qwen 3.8 Max.
- [14]
Alibaba has committed more than $53bn over three years to its AI expansion and raised about $10.2bn in a follow-on share sale in August.
- [15]
Alibaba's quarterly profit fell by 75% after a spree of almost $10bn.
- [16]
Citigroup estimates that 20GW of capacity could drive more than $160bn of external revenue for Alibaba's cloud division, according to Bloomberg.
- [17]
At the spec sheet's 216GB per chip, a 500,000-chip cluster would hold about 108 petabytes of attached memory.
- [18]
Bloomberg's 560,000 delivered Zhenwu units spread across more than 400 external customers averages fewer than 1,400 chips per customer, and fewer than 900 against Alibaba's count of more than 650 customers.
- [19]
Alibaba's 20GW global target for 2032 is about 53% of the 37.7GW of data centre capacity now under construction in the United States.
- [20]
A 10 trillion parameter Qwen 5 would be about 4.2 times the size of Qwen 3.8 Max at 2.4 trillion parameters.
- [21]
The 500,000-chip cluster ceiling equals about 89% of the 560,000 Zhenwu units Bloomberg says have been delivered.
- [22]
Bloomberg put deliveries of the Zhenwu family at 560,000 units to more than 400 external customers, while Alibaba's own post puts the figure higher, at more than 650 customers.
Sources
2 independent publishers whose own reporting we read for this story.
- datacenterdynamics.comAlibaba unveils Zhenwu V900, says cloud capacity will hit 20GW by 2032
2 articles · September 22, 2026
- thenextweb.comAlibaba says its new Zhenwu V900 is China’s most powerful AI chip
1 article · September 22, 2026
Topics and entities
Topics
- China Chip Design vs Export Controls
- AI Accelerators
- Data Centre Capacity