Last updated: September 28, 2026
At its Apsara conference on September 22, 2026, Alibaba announced the Zhenwu V900 AI chip (216 GB of memory, mass production in Q1 2027), supernode clusters of up to 500,000 chips, Qwen 4.5 and Qwen 5 models scaling to 5–10 trillion parameters, and a goal of more than 20 GW of data-center capacity by 2032. The strategy is real; most of the performance numbers are Alibaba’s own and have not been independently benchmarked. The central question is whether owning every layer lowers the cost of AI, or replaces dependence on NVIDIA with dependence on one integrated operator.
Most coverage of Apsara 2026 split it into three stories: a chip, a model roadmap and a capacity target. They are one story. Alibaba is building a vertically integrated AI stack in which each layer is designed to make the next one cheaper, the same logic that Google has pursued with TPUs, Gemini and Google Cloud. What makes Alibaba’s version distinctive is the political context around it, and an unusual mix of a closed bottom layer with an open top layer.
What did Alibaba announce at Apsara 2026?
CEO Eddie Wu used the September 22 keynote in Hangzhou to lay out a roadmap from silicon to agents. According to Alibaba’s press release and Reuters, the main commitments are:
| Announcement | Detail | Status |
|---|---|---|
| Zhenwu V900 accelerator (T-Head) | For training and inference; 216 GB memory; 1,200 GB/s inter-chip bandwidth; native FP8 and FP4 | Mass production and commercial release planned for Q1 2027 |
| Performance | “Three times the performance” of the Zhenwu M890 | Vendor claim; no FLOPS figure published |
| Supernode clusters | More than 1,000 V900s working as one system; clusters of “up to 500,000 cards” | Vendor claim; supernodes entering commercial operation at scale this quarter, per Wu |
| Qwen roadmap | Qwen 4 in training; Qwen 4.5 and Qwen 5 “projected to scale up to 5 to 10 trillion parameters” | No release dates, prices or weight availability announced |
| Capacity target | Global data-center capacity operated by Alibaba Cloud to “surpass 20GW” by 2032 | Target; current capacity not disclosed |
Two corrections to early coverage. The V900’s direct predecessor is the Zhenwu M890, launched in May 2026, not the older 810E. And Alibaba did not say which of Qwen 4.5 and Qwen 5 will be 5 trillion and which 10 trillion parameters; it gave a single range for both. For scale, Reuters reports that the current flagship, Qwen3.8-Max, has 2.4 trillion parameters.
How good is the Zhenwu V900 really?
Nobody outside Alibaba can say yet. The company published memory capacity, interconnect bandwidth and supported precisions, but not the numbers that let engineers compare accelerators: peak FLOPS at each precision, memory type and bandwidth, power per chip, process node or foundry. There are no MLPerf results and no third-party tests.
What can be reconstructed is the generational trend, entirely from Alibaba’s own claims:
| Zhenwu 810E | Zhenwu M890 | Zhenwu V900 | |
|---|---|---|---|
| Announced | 2025 | May 20, 2026 | Sept 22, 2026 |
| Memory | 96 GB (HBM2e) | 144 GB | 216 GB (type not disclosed) |
| Inter-chip bandwidth | 700 GB/s | 800 GB/s | 1,200 GB/s |
| Relative performance (vendor) | Baseline | 3× 810E | 3× M890 |
| Availability | Shipping | Shipping | Q1 2027 |
Compounded, Alibaba is claiming roughly a ninefold gain over two generations in about eighteen months. That is plausible if the baseline was modest (TrendForce has compared the 810E to NVIDIA’s export-restricted H20), but a multiplier over an undisclosed baseline is not a benchmark.
Watch for five numbers when the V900 ships: dense and sparse FP8/FP4 FLOPS, HBM generation and bandwidth, power per chip, the foundry and node, and independent results on a common workload. Until then, treat “the most powerful AI chip in China”, Wu’s description, as a marketing claim with a plausible direction.
Why build the whole stack?
Because each layer is someone else’s margin and someone else’s veto. Alibaba’s stack now looks like this:
- Accelerator Zhenwu M890 today, V900 from Q1 2027 (T-Head)
- Interconnect ICN Switch chips linking more than 1,000 accelerators
- Server Panjiu supernode servers, with in-house NICs and SSD controllers
- CPU Yitian server CPUs; Yitian 730, its first fully in-house core, planned for 2027
- Data center Alibaba Cloud regions, targeting 20 GW by 2032
- Model Qwen, largely open-weight, from small models to Qwen3.8-Max
- Agents and apps Alibaba Cloud Model Studio and consumer apps
The economic argument for integration is simple. A merchant GPU carries its vendor’s margin, and a model trained on hardware you do not control has to tolerate that hardware’s constraints. If Alibaba designs the chip, the interconnect and the model together, it can tune the model architecture to the chip’s strengths, drop features it does not need, and keep the margin. Google has shown that this can work at scale: TPUs underpin Gemini, and even Anthropic has committed to gigawatts of Google TPU capacity.
The strategic argument is stronger still. US export controls have made NVIDIA’s best chips hard or impossible to buy in China, and in 2025 Chinese authorities reportedly required new state-funded data centers to use only domestic AI chips. NVIDIA said in August that Beijing had restricted sales of the H200 units it was licensed to ship. For a Chinese hyperscaler, a domestic accelerator is not an option to evaluate; it is a condition of growth. Alibaba says Zhenwu chips are already used by more than 650 external customers.
Does full integration make AI cheaper?
It can, but only under conditions that Alibaba has not yet shown it meets.
Integration lowers cost when utilization is high. Custom silicon has large fixed costs: design teams, masks, software stacks, compilers and kernels. Those costs are recovered only if the chips run busy for years. Alibaba’s scale helps here. In the June 2026 quarter its AI cloud segment grew revenue 45% to RMB 48.4 billion, and AI-related product revenue grew at triple digits for a twelfth consecutive quarter, according to its SEC filing.
It raises cost when the software lags. NVIDIA’s advantage is as much CUDA and its ecosystem as silicon. Every model Alibaba runs on Zhenwu needs kernels, compilers and debugging tools that must be built and maintained in-house. This hidden cost does not show up in a chip announcement.
And it is expensive before it is cheap. Alibaba’s official plan, announced in February 2025, is to invest RMB 380 billion in AI and cloud infrastructure over three years. Its June-quarter capital expenditure alone was RMB 67.7 billion, up 75% year on year, and free cash flow was an outflow of RMB 44.7 billion. A reported plan to raise the budget to RMB 480 billion comes from a single insider account in LatePost and has not been confirmed. The bet may pay off; it has not yet.
So the honest answer to “does the full stack lower the price of inference?” is: it lowers Alibaba’s cost of supply if utilization stays high, and it protects Alibaba from supply shocks. Whether customers see lower prices depends on competition, not on the stack.
Who gets locked in: closed at the bottom, open at the top
Here is the feature of Alibaba’s strategy that most coverage misses. Google’s stack is closed at both ends: TPUs are available only through Google Cloud, and Gemini’s weights are not public. Alibaba’s is closed at the bottom and largely open at the top.
Qwen is the most-used open-weight model family in the world by several measures. Hugging Face’s State of Open Models report (August 2026) counts 151,448 Qwen-based derivative models on the Hub, 2.6 times Meta’s footprint, and about 2.0 billion Qwen downloads between January and August 2026. Alibaba’s own figures, which include other platforms, are higher. Those open models run on NVIDIA, AMD, Apple Silicon and Huawei hardware, not just on Zhenwu.
That has two consequences for customers:
- At the model layer, lock-in is low. A company that builds on open Qwen weights can move them to any cloud or run them locally. Our guide to running open models locally applies to Qwen as much as to any other family.
- At the infrastructure layer, lock-in is high. The cheapest way to run the largest Qwen models at scale may end up being Alibaba Cloud on Zhenwu hardware, tuned by the same company that trained the model. If Qwen 5 at 5–10 trillion parameters is not released with open weights, or runs efficiently only on Zhenwu supernodes, the open top layer becomes a funnel into the closed bottom one.
In other words, Alibaba does not ask customers to trade NVIDIA dependence for Alibaba dependence all at once. It lets them adopt Qwen freely and competes to host it. For companies outside China, data residency, sanctions compliance and procurement rules will usually decide the question before price does.
Is parameter count still a useful measure?
Less and less. A 5–10 trillion parameter headline sounds like a leap from Qwen3.8-Max’s 2.4 trillion, but frontier models are now mostly mixture-of-experts systems, in which only a fraction of parameters work on any given token. Kimi K2.6 has one trillion total parameters and 32 billion active; DeepSeek V4-Pro has about 1.6 trillion total and 49 billion active. Inference cost tracks active parameters and memory footprint; capability depends on data, training compute, reinforcement learning and evaluation discipline as much as on size.
Total parameters still matter for one practical reason: memory. A 10-trillion-parameter model at four bits needs around five terabytes just for weights, before any cache. That is why the model roadmap and the 216 GB chip with 1,000-chip supernodes belong in the same announcement. The model is designed for the machine, and the machine for the model.
What does 20 GW actually mean?
A gigawatt target describes power capacity, not consumption, so it helps to convert it. Twenty gigawatts running at full load for a year is about 175 terawatt-hours. Real data centers do not run at full load, so treat this as an upper bound. Even so, the International Energy Agency estimates that all data centers worldwide consumed about 415 TWh in 2024 and projects about 945 TWh by 2030. A single company’s 2032 target at full load would equal roughly 42% of global data-center electricity use in 2024, and roughly the IEA’s entire projected growth in Chinese data-center demand between 2024 and 2030 (about 175 TWh).
The target is less of a new ambition than it looks. At Apsara 2025, Wu pledged that Alibaba Cloud’s data-center energy consumption would grow tenfold between 2022 and 2032. The 20 GW figure puts a number on that path. Alibaba has not disclosed its current capacity; a Goldman Sachs estimate cited by analysts puts it at about 5–6 GW, which would mean roughly tripling or quadrupling in six years.
The constraint is not only chips. It is grid connections, power purchase agreements and cooling, which take years to secure. Energy is the bottleneck every hyperscaler is now competing for, and gigawatt announcements should be read as options on future capacity rather than as capacity under construction.
The Q1 2027 race with Huawei
One date deserves attention. Alibaba’s V900 enters mass production in Q1 2027. At Huawei Connect on September 17, Huawei said its Ascend 960DT is expected to be ready in Q1 2027 too, and its Atlas 950 SuperCluster is designed to connect up to 256,000 accelerator cards. In May, TrendForce had listed the V900 for the third quarter of 2027, so Alibaba’s schedule appears to have moved forward, although the company has not described it that way.
Two domestic accelerators landing in the same quarter, as NVIDIA’s China sales shrink under controls from both Washington and Beijing, sets up the first real competition in Chinese AI silicon. The winner will be decided less by peak FLOPS than by software maturity, supply from domestic foundries and how many external customers each ecosystem can win.
What to watch next
- Independent V900 benchmarks once chips reach customers in 2027, ideally on MLPerf or a common open model such as a Qwen release.
- Qwen 4 release terms: whether it ships with open weights, and at what sizes.
- Capex disclosures in Alibaba’s next quarterly results, including whether the RMB 380 billion plan is formally increased.
- Power agreements and new region announcements that turn the 20 GW target into sites.
- Export-control changes in the US and China that could reopen or further close access to NVIDIA hardware.
For a broader view of how inference costs are moving on the buyer’s side, see our analysis of local hardware versus cloud APIs and the LLM inference cost comparison.
FAQ
What is the Alibaba Zhenwu V900?
It is an AI accelerator designed by Alibaba’s chip unit T-Head for both training and inference. Alibaba says it has 216 GB of memory, 1,200 GB/s of chip-to-chip bandwidth and native FP8 and FP4 support, with mass production planned for Q1 2027.
Is the V900 faster than NVIDIA’s chips?
There is no way to know yet. Alibaba claims three times the performance of its previous Zhenwu M890 but has not published FLOPS figures or independent benchmarks against NVIDIA, AMD or Huawei accelerators.
How many parameters will Qwen 5 have?
Alibaba says Qwen 4.5 and Qwen 5 are projected to scale to 5 to 10 trillion parameters. It gave one range for both models and no release dates. The current flagship, Qwen3.8-Max, has 2.4 trillion parameters, according to Reuters.
What does Alibaba’s 20 GW target mean?
Alibaba aims for its cloud to operate more than 20 GW of data-center capacity by 2032. At full load that is about 175 TWh a year, equal to roughly 42% of global data-center electricity use in 2024 as estimated by the IEA.
Is Qwen open source?
Many Qwen models are released with open weights, and Hugging Face counts more than 151,000 Qwen-based derivative models. Whether the planned Qwen 4.5 and Qwen 5 models will be released openly has not been announced.
Why is Alibaba making its own AI chips?
To control costs and supply. US export controls restrict NVIDIA’s best chips in China, and Chinese policy favors domestic chips in state-funded data centers. Owning the chip also lets Alibaba design models and hardware together.
Sources and further reading
- Alibaba Cloud, Alibaba unveils roadmap on full-stack AI strategy.
- Reuters (via Investing.com), Alibaba plans AI model with 5 to 10 trillion parameters, unveils new chip.
- Alibaba Group, Zhenwu M890 and Panjiu AL128 announcement (May 2026).
- Alibaba Group, June quarter 2026 results (SEC Form 6-K).
- Alibaba Cloud, RMB 380 billion AI and cloud investment plan.
- The Register, Alibaba Cloud plans six-year stroll to 20 GW.
- TrendForce, Alibaba unveils Zhenwu V900 for 1Q27 mass production.
- TechNode, T-Head unveils Zhenwu V900.
- Hugging Face, State of Open Models (Summer 2026).
- International Energy Agency, Energy and AI.
- TechCrunch, Huawei plans Q1 2027 launch of new AI chip.
- US Bureau of Industry and Security, revised licence review policy for semiconductors exported to China.
- Reuters (via Yahoo Finance), China bans foreign AI chips from state-funded data centers.
- KrASIA, Apsara Conference 2025 energy pledge.
Chip performance, cluster size and capacity figures are Alibaba’s own claims as of September 28, 2026. The 20 GW energy conversion is our calculation at full load.
