At the 2026 Apsara Conference, Alibaba said the global data center capacity it operates will exceed 20GW by 2032. At the same event, its chip unit T-Head unveiled a new AI chip, the Zhenwu V900, claiming a single cluster built on it can scale to as many as 500,000 cards. Alibaba Group CEO Eddie Wu said bluntly in his keynote that a tight global AI data center supply chain is limiting how fast Alibaba Cloud can add computing power.

The 20GW target, the 500,000-card scaling ceiling and the compute supply crunch arriving together point to a shift in where AI competition is being fought: away from showcasing model capability and toward a systematic rebuild of the underlying infrastructure.

Zhenwu V900: Three Times the Computing Power of the M890, Mass Production Set for Q1 2027

Alibaba says the Zhenwu V900 delivers three times the computing power of the previous-generation M890, and, more significantly, that a single cluster built on the chip can scale to 500,000 cards. The company frames that capability as more than stacking hardware: it rests on full-stack optimization it calls "chip-model co-design," where effective computing power depends less on how many cards you own than on wringing maximum efficiency out of every card and every data transfer.

The 500,000-card figure, though, is the disclosed scaling ceiling of the cluster architecture. Mass production of the Zhenwu V900 is planned for the first quarter of 2027, and how that schedule is actually tracking remains unknown.

500,000 Cards Is Five Times the Current Ceiling, a Scale No Domestic Chip Has Publicly Delivered

Super nodes are where AI infrastructure vendors are concentrating their spending. NVIDIA, Huawei, AMD, Alibaba, Intel, Meta, Microsoft and Inspur are all betting on them, with the same aim: use high-speed interconnects to enlarge a single compute domain and link as many chips as possible. Super nodes are what make 100,000-card clusters possible in the first place.

A 100,000-card cluster is already the entry ticket for the world's top players, and China counts only a handful of fully domestic builds at that scale. Alibaba's 500,000-card target is five times the current ceiling. Scaling a cluster has never been a matter of adding cards: interconnect, scheduling and fault tolerance are systems engineering, and domestic chips have never publicly delivered at this magnitude.

By setting the bar so far out, Alibaba is staking a claim on future demand in advance. Other cloud providers will likely move earlier to lock up power, land and critical equipment, which means part of the competition in the compute market will play out before customers place a single order.

From Card Counts and Peak Numbers to Steady-State Effective Compute Output

One change is clear: the industry is moving past the early phase of counting cards and comparing peak computing power. The core measure of a large cluster now is steady-state effective compute output.

The industry used to define cluster strength by GPU count and paper PFLOPS. In real large-model training and mixed-workload inference, though, the losses stack up — waiting on network communication, storage IO bottlenecks, fragmented compute scheduling, hardware failure interruptions — and clusters with identical hardware can end up several times apart in actual business output.

The series of AI compute service standards and benchmarking frameworks CAICT released in 2026 argues against judging chips on theoretical peak computing power alone, pushing instead from "paper compute" toward the token output a system can reliably supply to customers. That means the 20GW target requires full-stack breakthroughs on metrics such as per-rack compute density, rack-level compute utilization, high-speed interconnect networking efficiency and speed of recovery from failures.

Capex Heading to 200-300 Billion Yuan a Year, Squeezing Smaller Players Out of the Core Race

Behind the 20GW target sits capital spending in the trillions of yuan. Goldman Sachs estimates that taking Alibaba from roughly 5-6GW today to 20GW requires adding 1-2GW a year over the coming years, translating into annual capital expenditure of 200 billion to 300 billion yuan.

For comparison, before Alibaba began spending heavily on AI, its normal annual capex ran about 30 billion to 50 billion yuan. In fiscal 2026 that rose to 126.1 billion yuan, and capex in the first quarter of the current fiscal 2027 hit a record 67.678 billion yuan, with the earnings report stating the money went primarily into AI infrastructure.

Spending at that intensity is itself a very high barrier to entry. Smaller players will be pushed out of the core race, leaving the compute contest to a handful of giants.

A Split Market: Racks Hard to Come By at the High End, While 2025 Card-Hour Utilization Ran Just 36.8%

China's AI compute industry sits in a structurally split phase: high-end advanced computing power is in short supply while large volumes of mid- and low-end installed capacity sit idle.

On one side are new-generation domestic high-performance compute clusters built for large-model training and high-concurrency agent inference, locked up by major customers the moment they come online; in some sought-after regions the balance is tight enough that a single rack is hard to come by, as advanced supply expands more slowly than the token consumption driven by the boom in AI agents. On the other side are smaller clusters rushed into service years ago on older chips or niche-ecosystem compute products, unable to handle the workloads of today's mainstream large models and idling at low load.

CAICT monitoring data shows card-hour utilization of China's AI computing power ran only about 36.8% in 2025, with more than 60% of built capacity idle, even as the supply-demand gap for mainstream advanced computing power keeps widening. This split will not fully clear in the near term.

Worth noting: the industrial opportunity created by the 20GW plan will not be shared out among suppliers in proportion to the investment. Power equipment, high-density cooling, chips, interconnect and operations services each have their own openings, and their own challenges.

At the 2026 Apsara Conference, Alibaba said the global data center capacity it operates will exceed 20GW by 2032. At the same event, its chip unit T-Head unveiled a new AI chip, the Zhenwu V900, claiming a single cluster built on it can scale to as many as 500,000 cards. Alibaba Group CEO Eddie Wu said bluntly in his keynote that a tight global AI data center supply chain is limiting how fast Alibaba Cloud can add computing power.

The 20GW target, the 500,000-card scaling ceiling and the compute supply crunch arriving together point to a shift in where AI competition is being fought: away from showcasing model capability and toward a systematic rebuild of the underlying infrastructure.

Zhenwu V900: Three Times the Computing Power of the M890, Mass Production Set for Q1 2027

Alibaba says the Zhenwu V900 delivers three times the computing power of the previous-generation M890, and, more significantly, that a single cluster built on the chip can scale to 500,000 cards. The company frames that capability as more than stacking hardware: it rests on full-stack optimization it calls "chip-model co-design," where effective computing power depends less on how many cards you own than on wringing maximum efficiency out of every card and every data transfer.

The 500,000-card figure, though, is the disclosed scaling ceiling of the cluster architecture. Mass production of the Zhenwu V900 is planned for the first quarter of 2027, and how that schedule is actually tracking remains unknown.

500,000 Cards Is Five Times the Current Ceiling, a Scale No Domestic Chip Has Publicly Delivered

Super nodes are where AI infrastructure vendors are concentrating their spending. NVIDIA, Huawei, AMD, Alibaba, Intel, Meta, Microsoft and Inspur are all betting on them, with the same aim: use high-speed interconnects to enlarge a single compute domain and link as many chips as possible. Super nodes are what make 100,000-card clusters possible in the first place.

A 100,000-card cluster is already the entry ticket for the world's top players, and China counts only a handful of fully domestic builds at that scale. Alibaba's 500,000-card target is five times the current ceiling. Scaling a cluster has never been a matter of adding cards: interconnect, scheduling and fault tolerance are systems engineering, and domestic chips have never publicly delivered at this magnitude.

By setting the bar so far out, Alibaba is staking a claim on future demand in advance. Other cloud providers will likely move earlier to lock up power, land and critical equipment, which means part of the competition in the compute market will play out before customers place a single order.

From Card Counts and Peak Numbers to Steady-State Effective Compute Output

One change is clear: the industry is moving past the early phase of counting cards and comparing peak computing power. The core measure of a large cluster now is steady-state effective compute output.

The industry used to define cluster strength by GPU count and paper PFLOPS. In real large-model training and mixed-workload inference, though, the losses stack up — waiting on network communication, storage IO bottlenecks, fragmented compute scheduling, hardware failure interruptions — and clusters with identical hardware can end up several times apart in actual business output.

The series of AI compute service standards and benchmarking frameworks CAICT released in 2026 argues against judging chips on theoretical peak computing power alone, pushing instead from "paper compute" toward the token output a system can reliably supply to customers. That means the 20GW target requires full-stack breakthroughs on metrics such as per-rack compute density, rack-level compute utilization, high-speed interconnect networking efficiency and speed of recovery from failures.

Capex Heading to 200-300 Billion Yuan a Year, Squeezing Smaller Players Out of the Core Race

Behind the 20GW target sits capital spending in the trillions of yuan. Goldman Sachs estimates that taking Alibaba from roughly 5-6GW today to 20GW requires adding 1-2GW a year over the coming years, translating into annual capital expenditure of 200 billion to 300 billion yuan.

For comparison, before Alibaba began spending heavily on AI, its normal annual capex ran about 30 billion to 50 billion yuan. In fiscal 2026 that rose to 126.1 billion yuan, and capex in the first quarter of the current fiscal 2027 hit a record 67.678 billion yuan, with the earnings report stating the money went primarily into AI infrastructure.

Spending at that intensity is itself a very high barrier to entry. Smaller players will be pushed out of the core race, leaving the compute contest to a handful of giants.

A Split Market: Racks Hard to Come By at the High End, While 2025 Card-Hour Utilization Ran Just 36.8%

China's AI compute industry sits in a structurally split phase: high-end advanced computing power is in short supply while large volumes of mid- and low-end installed capacity sit idle.

On one side are new-generation domestic high-performance compute clusters built for large-model training and high-concurrency agent inference, locked up by major customers the moment they come online; in some sought-after regions the balance is tight enough that a single rack is hard to come by, as advanced supply expands more slowly than the token consumption driven by the boom in AI agents. On the other side are smaller clusters rushed into service years ago on older chips or niche-ecosystem compute products, unable to handle the workloads of today's mainstream large models and idling at low load.

CAICT monitoring data shows card-hour utilization of China's AI computing power ran only about 36.8% in 2025, with more than 60% of built capacity idle, even as the supply-demand gap for mainstream advanced computing power keeps widening. This split will not fully clear in the near term.

Worth noting: the industrial opportunity created by the 20GW plan will not be shared out among suppliers in proportion to the investment. Power equipment, high-density cooling, chips, interconnect and operations services each have their own openings, and their own challenges.