Alibaba’s full-stack AI bet: Can chips, Qwen and 20GW of cloud power the next re-rating?

• Alibaba is building a vertically integrated AI ecosystem spanning chips, models and cloud infrastructure, with the aim of lowering compute costs, reducing supply-chain dependence and capturing more value across the AI stack.
  
  
• The new Zhenwu V900 marks another step forward in proprietary AI hardware, with performance potentially approaching Nvidia’s H200 range based on our estimates.
  
  
• Qwen remains central to Alibaba’s AI flywheel. The upcoming Qwen 4 series is expected to scale towards 5–10 trillion parameters, supporting greater AI adoption, higher cloud utilisation and stronger monetisation.
  
  
• Alibaba’s 20GW data-centre target materially increases its long-term cloud revenue potential, but also requires substantial investment that could keep capex elevated and free cash flow under pressure for several years.
  
  
• We remain constructive on Alibaba’s long-term growth outlook, although higher near-term capex is likely to dilute earnings. We therefore lower our target price for Alibaba (SGX: HBBD) to SGD5.52, implying 61.0% upside based on our FY2029 valuation. This corresponds to HKD172 for HKEX: 9988 and USD176 for NYSE: BABA.

From proprietary chips and increasingly powerful Qwen models to a major expansion in data-centre capacity, Alibaba laid out a clearer full-stack AI roadmap at the Apsara conference held from 22 to 24 September 2026. Investors responded positively, with Alibaba’s Hong Kong-listed shares rising 3.7% to HKD116.80 on 22 September.

The strategy is becoming increasingly clear: build its own chips to lower compute costs and reduce supply-chain dependence, use more capable Qwen models to drive AI adoption, and expand cloud infrastructure to monetise that demand. If executed well, this could allow Alibaba to capture more value across the AI stack rather than relying on any single layer.

In this article, we unpack the three major announcements and assess what they could mean for Alibaba’s earnings, cash flow and valuation.

Table 1: The key announcements of Alibaba’s AI products at the conference

AI products	Announcements	Timing
Chip	Zhenwu V900: Alibaba's latest AI training and inference processor, with 216GB memory, 1,200GB/s chip-to-chip bandwidth and native FP8/FP4 support. Alibaba claims 3x the performance of the M890. Yitian 720/730: next-generation proprietary server CPUs, with the Yitian 730 featuring Alibaba's own microarchitecture.	Zhenwu V900 Mass production 1Q27 (pulled forward from 3Q27) Yitian series release in 3Q27
Models	Qwen 4 is currently in training. Alibaba also outlined a roadmap for Qwen 4.5 and Qwen 5 to scale to 5–10 trillion parameters.	Qwen 4 imminent
Data centers	Alibaba Cloud targets more than 20GW of global data-centre capacity operated by 2032. New cloud regions are planned in Türkiye, Finland and the Netherlands, alongside capacity expansion in Malaysia, Germany, UAE, France and Hong Kong.	>20 GWs by 2032
Source: Alibaba Apsara Conference 2026. iFAST compilations. Data as of 22 September 2026.

The Zhenwu V900: Strategic value matters more than benchmark leadership

The most tangible hardware announcement was the Zhenwu V900, developed by Alibaba's T-Head semiconductor unit. Alibaba claims that the new accelerator delivers three times the performance of the previous-generation Zhenwu M890, while increasing memory to 216GB and chip-to-chip bandwidth to 1,200GB/s. It is designed to scale into clusters of up to 500,000 accelerators.

These specifications point to a significant improvement in Alibaba’s ability to train and serve larger AI models. While Alibaba claims the V900 is China’s most powerful AI chip, we remain cautious about this claim as Alibaba has not disclosed its official FLOPS, power consumption or process node, making direct comparisons difficult.

As a rough guide, we can build an indicative estimate. Alibaba has described the previous-generation M890 as an “Nvidia’s A100-class” chip. Nvidia’s A100 can perform around 0.31 quadrillion AI calculations per second (measured by PFLOPS) using a common 16-bit computing format. Applying Alibaba’s claimed 3x performance uplift implies an indicative ~0.9 PFLOPS for the V900, assuming the two performance measures are broadly comparable. This would put the V900 around H200-level on theoretical compute, and above Huawei’s Ascend 950DT at around 0.5 PFLOPS FP16.

That said, the V900’s 216GB memory, native FP8/FP4 support and 1,200GB/s chip-to-chip bandwidth suggest that its overall capabilities could extend beyond its headline compute figure. We believe its overall performance could broadly sit between Nvidia’s H200 and earlier Blackwell-generation chips, subject to independent benchmarks.

Table 2: Zhenwu V900’s performance is likely land between Nvidia’s H200 and early Balckwell series

Spec	Zhenwu V900	Huawei Ascend 950DT	Nvidia H200 (Hopper)	Nvidia B200 (Blackwell)
Announced	22 Sep 2026	26 August 2026	13 Nov 2023	18 Mar 2024
Shipping	Mass production 1Q27	Q4 2026	2024	From Q4 2024
Memory	216GB	144GB	141GB HBM3e	192GB HBM3e
Chip-to-chip bandwidth	1,200GB/s	2,000GB/s	900GB/s (NVLink 4)	1,800GB/s (NVLink 5)
Precision support	Native FP8 and FP4	FP8, MXFP8, HiF8, MXFP4	Down to FP8	Down to FP4
Manufacturing	Not disclosed	SMIC N+3	TSMC 4N	TSMC 4NP
FP16 dense	Estimated 0.9	0.5	0.84-0.99	2.25
FP8 dense	Estimated 1.8	1.0	2.0	4.5
Nvidia equivalent	Not stated by Alibaba; between H200 and B200 (our estimate)	—	—	—
Source: Alibaba Apsara Conference 2026. iFAST estimations and compilations. Data as of 22 September 2026.

For investors, however, absolute benchmark leadership may not be the most important issue. Alibaba does not need to build the world's fastest accelerator to create economic value. Its strategic objective is to control more of the computing stack used by Alibaba Cloud and Qwen, allowing it to optimise hardware, networking, storage and software together. This could lower the cost of AI compute, reduce exposure to restrictions on imported accelerators and give Alibaba greater control over the availability of computing capacity.

This is already beginning to show up commercially. In the June 2026 quarter, Alibaba said its Zhenwu series AI chips had been deployed through Alibaba Cloud to more than 650 external customers across over 20 industries, demonstrating that its proprietary silicon is moving beyond internal use.

The Yitian CPU roadmap reinforces the same strategy. The Yitian 720 and 730 are expected in 2027, with the Yitian 730 being Alibaba's first server CPU based on a proprietary microarchitecture. Alibaba says the Yitian 730 can deliver up to a 40% improvement in SPECint2017/GHz over the Yitian 710.

The significance is therefore broader than simply replacing Nvidia GPUs. Alibaba is gradually building its own computing stack across accelerators, CPUs, networking and storage. If successful, this could improve cost efficiency and supply security while allowing Alibaba Cloud to offer a more integrated AI infrastructure platform.

Qwen 4: Larger models strengthen the AI flywheel, but parameters are not everything

At the model layer, Alibaba revealed that Qwen 4 is currently in training, while future Qwen 4.5 and Qwen 5 models could scale to 5–10 trillion parameters. This compares with the 2.4-trillion-parameter Qwen3.8-Max, meaning the upper end of the roadmap would represent roughly four times the parameter count.

Larger models can potentially improve reasoning, generalisation and agentic capabilities, but they also require substantially more compute during training and inference. This creates an important link between Alibaba's model and infrastructure strategies: more capable Qwen models can increase demand for Alibaba Cloud's compute, while Alibaba's proprietary chips and infrastructure can potentially lower the cost of running those models.

This creates a potential AI flywheel, with early evidence already visible in the June quarter. AI Cloud and Compute Services revenue increased 45% YoY to RMB48.4 billion, while AI-related product revenue reached RMB12.4 billion and continued to grow at triple-digit rates for the 12th consecutive quarter. Adjusted EBITA for the AI Cloud and Compute Services segment rose 133% YoY to RMB5.63 billion, lifting its EBITA margin to 12%. For investors, this emerging monetisation is arguably more important than the headline increase in model parameter counts.

At the same time, parameter count should not be treated as a direct measure of model quality. AI performance increasingly depends on model architecture, training data, post-training, reasoning techniques and inference efficiency. Therefore, the investment significance of Qwen 4 will ultimately depend less on whether it has five or ten trillion parameters and more on whether higher model capability translates into greater enterprise adoption, token consumption, agent workloads and monetisation on Alibaba Cloud.

How realistic is the 20GW data-centre target? 

The most significant announcement was Alibaba Cloud's target to operate more than 20GW of global data-centre capacity by 2032. This represents a major expansion of Alibaba's infrastructure footprint and reflects management's view that demand for AI compute will continue to exceed available supply.

The target is ambitious. Alibaba has not disclosed a detailed current global capacity figure in its official announcement, although several investment banks estimate it at 4GW. Reaching 20GW would therefore imply roughly a fivefold expansion by 2032. This would place Alibaba among the world's most aggressive data-centre builders. Microsoft added just over 2GW in FY2025, while Nvidia has announced plans with Australian partners to support up to 2GW of AI infrastructure by 2027. Meta's planned Alberta data centre, meanwhile, is a 1GW project expected to take up to three years to complete. The comparison suggests that Alibaba's target is achievable from an industry-capacity perspective but would require a sustained build-out at a top-tier global pace.

The bigger question is the cost. We estimate China-based AI data centres could cost less than USD20 billion per GW, including buildings and hardware, supported by lower construction and labour costs and Alibaba's increasing use of proprietary T-Head chips. Turner & Townsend's 2025 index puts Shanghai construction costs at USD6.12/W, roughly half the level of many developed markets. Given Alibaba's planned expansion across Europe, the Middle East and Asia, we assume USD40 billion/GW overseas. With an even split, this implies a blended cost of around USD30 billion/GW.

As Alibaba intends to work with partners to develop the required infrastructure, assuming partners finance 50% of incremental capacity, we estimate Alibaba would need to fund around USD40 billion annually to reach 20GW by 2032. This is substantial: June-quarter capex was already USD10 billion (CNY67.6 billion), or roughly USD40 billion annualised, and includes spending across e-commerce and existing infrastructure. Free cash flow was consequently a CNY44.7 billion outflow, suggesting that such an expansion could keep cash flow under pressure for several years.

The revenue opportunity, however, is also significant. We estimate Alibaba Cloud currently generates around USD5 billion of annualised revenue per GW, about 25% higher than FY2025. Assuming stronger AI adoption and improving utilisation lift this to an average USD8 billion/GW during FY2027–FY2029, 20GW of fully utilised capacity could theoretically support up to USD160 billion of annual cloud revenue.

Under our assumptions, each GW requires around USD30 billion of total investment, while Alibaba's direct share falls to roughly USD15 billion if partners fund half. Against potential annual revenue of around USD8 billion/GW, the economics are unlikely to be attractive during the initial build-out phase, when utilisation remains low and depreciation, power and financing costs are high. This is the central trade-off behind Alibaba's 20GW strategy. Higher capacity could materially increase Cloud's long-term revenue and EBITA potential, but at the cost of elevated capex and weaker near-term free cash flow. The ultimate return will therefore depend less on capacity expansion itself and more on whether Alibaba can improve utilisation, revenue per GW and Cloud margins fast enough to justify the investment.

Higher AI capex tempers near-term valuation

The positive share-price reaction on 22 September reflects the market's recognition that Alibaba's AI strategy is becoming more coherent. Alibaba's chips, Qwen models and cloud infrastructure are increasingly designed to work together.

We believe this is strategically important because it could give Alibaba three potential advantages:

• Lower compute costs through proprietary silicon and system-level optimisation;
• Greater supply-chain resilience by reducing dependence on imported AI accelerators; and
• Higher monetisation potential by connecting Qwen models and AI agents directly to Alibaba Cloud infrastructure and its broader consumer and enterprise ecosystem.

However, the Apsara Conference does not eliminate the key investment risks. The V900 has yet to enter mass production, Qwen 4 has yet to demonstrate meaningful commercial impact, and the 20GW infrastructure build-out is highly capital-intensive and could weigh on free cash flow for several years. In other words, Apsara increases the long-term earnings ceiling, but it also increases the amount of capital Alibaba needs to deploy before reaching that ceiling.

For the Cloud Intelligence Group, we maintain our 3.0x fair P/S multiple despite stronger AI revenue growth, as the significantly higher infrastructure requirements increase capital intensity and delay free-cash-flow conversion. Multiple expansion would require evidence of higher revenue per GW, improving utilisation and stronger Cloud margins. At the same time, while the additional capex does not immediately reduce earnings, it is expected to result in higher depreciation, financing and infrastructure-related costs over the next few years.

As such, we lower the earnings growth in the next 2 years and our target price for Alibaba (SGX: HBBD) from SGD6.04 to SGD5.59, representing 61.0% upside based on our FY2029 valuation. The corresponding target prices are also lowered to HKD172 for Alibaba’s HKEX-listed shares (HKEX: 9988) and USD176 for its NYSE-listed ADS (NYSE: BABA).

We remain constructive on Alibaba’s long-term growth outlook. The reduction in our target price does not reflect a weaker view of Alibaba's AI strategy. Rather, it reflects the fact that building a full-stack AI ecosystem creates both greater earnings potential and greater near-term capital requirements.

The next phase of Alibaba’s AI story will therefore depend less on headline chip or model launches and more on three measurable outcomes: sustained AI revenue growth, higher infrastructure utilisation and improving returns on invested capital. If these continue to strengthen, Alibaba’s AI investments could evolve from a capital-intensive build-out into a meaningful long-term earnings engine.

Table 3: Earnings table

Metric	FY2026A	FY2027E	FY2028E	FY2029E
P/E (x)	30.7x	19.1x	16.1x	13.1x
EPS (SGD per SDR)	0.11	0.18	0.22	0.26
EPS YoY growth (%)	-61.6%	60.8%	19.0%	22.6%
Dividend yield (%)	1.0%	1.1%	1.2%	1.3%
Target Price (SGD)				5.59
Upside potential				61.0%
Source: iFAST Estimates. Data as of 30 September 2026.

Declaration:

For specific disclosure, at the time of publication of this report, IFPL (via its connected and associated entities) hold a NIL position in the abovementioned securities. The analyst who produced this report hold positions in Alibaba (HKEX:9988).