Skip to content

In Markets

Alibaba sets a 20-gigawatt cloud target and unveils the Zhenwu V900 chip

Alibaba CEO Eddie Wu on Tuesday, September 22, said Alibaba Cloud will surpass 20 gigawatts of global data-center capacity by 2032 and introduced T-Head's Zhenwu V900 at the Apsara Conference in Hangzhou.

A row of modern server racks illuminated with blue light in a large data center.
Photo by panumas nikhomkhai on Pexels

· 4 min read · ETF.net Research

KWEBAIQ

Eddie Wu put three long-dated calendars on one stage in Hangzhou: a Qwen model of 5 trillion to 10 trillion parameters, a T-Head accelerator due in the first quarter of 2027, and more than 20 gigawatts of Alibaba Cloud capacity by 2032. He also said Alibaba Cloud would begin bringing its AI supernodes online at commercial scale this quarter.

Alibaba Cloud's 20-gigawatt target

"The industry's mid-to-long-term demand far outpaces our supply capabilities," Wu said. Shortages across the data-center supply chain, he added, are already limiting how fast Alibaba Cloud can expand. Wu said the company's global data-center footprint will surpass 20 gigawatts by 2032. He did not attach a new spending figure to that target.

Alibaba has not disclosed how many gigawatts it operates today. Goldman Sachs estimated in October 2025 that Alibaba had 3 to 4 gigawatts of live data-center capacity.

In February 2025 Alibaba said it would invest at least RMB380 billion ($53 billion) over three years in cloud and AI infrastructure. In the quarter ended June 30, capital expenditures were RMB67.7 billion ($9.98 billion), up 75% from a year earlier. In August it placed HK$80 billion of new Hong Kong shares, with 100% of net proceeds earmarked for full-stack AI, including infrastructure; those shares were offered to non-U.S. persons and were not registered under the U.S. Securities Act.

The spend is showing up in the cloud line. Alibaba said AI Cloud and Compute Services revenue in that June quarter was RMB48.4 billion ($7.14 billion), with external-customer growth of 45%. Alibaba Cloud already lists sites in the United States, Europe, the Middle East and Southeast Asia. Tuesday's 20-gigawatt figure did not map those regions onto a build schedule, a grid plan, or a permit list.

Zhenwu V900, on Alibaba's numbers

T-Head, Alibaba's chip-design unit, called Zhenwu V900 a training-and-inference processor, not an inference-only part. Alibaba said it delivers three times the performance of Zhenwu M890, the chip T-Head released in May, and that a cluster can scale to 500,000 cards. Mass production and commercial release are scheduled for the first quarter of 2027.

The company specified 216 GB of GPU memory and 1,200 GB/s of inter-chip bandwidth, against 144 GB and 800 GB/s on the M890, with native support for FP8 and FP4, the low-precision formats used to cut the cost of inference. Wu called V900 the most powerful AI chip in China. That is Alibaba's claim, not an independent benchmark. The company did not name a foundry, a process node, or a wafer supplier, and it offered no apples-to-apples comparison with Nvidia's China-legal parts or Huawei's Ascend line.

Existing Zhenwu chips, Alibaba said, are already used by more than 650 customers. Wu said the company's M890 AI supernode already handles inference for models above 2 trillion parameters, the generation behind this quarter's commercial-scale ramp. V900 mass production remains a 2027 event. He also said Alibaba expects significant growth in annual AI chip shipments. U.S. export rules still require licenses for advanced accelerators destined for China-headquartered groups. What V900 does, or does not, substitute in that constrained market is a question Alibaba left unanswered.

Qwen 4 is in training. The 5 trillion to 10 trillion figure is Alibaba's projection for the Qwen 4.5 and Qwen 5 series, not a model it released on Tuesday. Qwen3.8-Max, the current flagship, has 2.4 trillion total parameters.

What a holder actually owns

Hong Kong-listed Alibaba shares traded at HK$115.10, up 2.2%, as of 1:30 a.m. in New York, with the session still running and a high of HK$118.30. The New York ADRs last closed Monday at $115.76, up 2.2% in a session that ended before Wu spoke.

Panel not shown

Chart: Alibaba ADRs remain lower this year

As of Monday, a fund of offshore-listed Chinese internet companies, KWEB, holds Alibaba's Hong Kong shares in a $4.74 billion portfolio.

KWEB holdings as of Monday

Alibaba is KWEB's second-largest holding

  • Tencent10.0%
  • Alibaba8.7%
  • PDD8.2%
  • Meituan7.1%
  • Lenovo6.6%
  • NetEase6.3%
  • KE Holdings4.2%
  • JD.com3.9%
  • Trip.com3.6%
  • Baidu3.2%

The top six each hold more than 6%.

etf.net grades it C in its China category. A global artificial-intelligence equity fund, AIQ, holds 2.15% in the New York ADRs, a $219 million position; etf.net grades it B.

Seventy-nine funds in etf.net's graded universe hold the ADRs. The concentrated China-internet sleeve is still KWEB.

New York's regular session later Tuesday is the first U.S. print on the speech. Until the V900 ships in 2027 and the 20-gigawatt fleet is more than a target, a holder of BABA or of KWEB owns a larger planning number, not a larger data center.

Frequently asked

How much data-center capacity does Alibaba have now?

Alibaba has not disclosed it; Goldman Sachs estimated 3 to 4 gigawatts of live capacity.

Can you buy the V900 chip yet?

No: mass production and commercial release are scheduled for the first quarter of 2027.

Is the claim that V900 is China's most powerful AI chip verified?

No, it is Alibaba's own claim, with no named foundry, process node, or comparison with Nvidia's China-legal parts or Huawei's Ascend line.

Which funds hold Alibaba?

Seventy-nine funds in etf.net's graded universe hold the ADRs, and Alibaba is the second-largest holding in the China-internet fund KWEB, graded C.