Home / Blog / Hardware
Hardware วิเคราะห์จากสเปค + รีวิว

Analysis and Review: Huawei Slows Global AI Chip Expansion After China’s Demand Outpaces Supply Analysis and Review: Huawei Slows Global AI Chip Expansion After China’s Demand Outpaces Supply

Deep dive into a 15,488-chip Huawei Atlas cluster that uses optical networking to take on Nvidia and scale performance up to 120 EFLOPS Deep dive into a 15,488-chip Huawei Atlas cluster that uses optical networking to take on Nvidia and scale performance up to 120 EFLOPS

Huawei Delays Pushing Atlas AI Chips to Global Markets Because Demand in China Still Exceeds Production Capacity

The Atlas cluster uses an optical network to connect a large number of chips, reducing bottlenecks during processing. This gives it potential as an option for large-scale AI workloads that require fast, highly scalable connectivity.

However, challenging Nvidia still requires looking at real-world benchmarks, costs, and software readiness, because powerful hardware alone is not enough for enterprise-level use. Huawei Delays Pushing Atlas AI Chips to Global Markets Because Demand in China Still Exceeds Production Capacity

The Atlas cluster uses an optical network to connect a large number of chips, reducing bottlenecks during processing. This gives it potential as an option for large-scale AI workloads that require fast, highly scalable connectivity.

However, challenging Nvidia still requires looking at real-world benchmarks, costs, and software readiness, because powerful hardware alone is not enough for enterprise-level use.

What Atlas Looks Like at the Cluster Level

Atlas should be viewed as a large cluster that distributes workloads between AI processing units and storage systems, with an optical network serving as the primary link between nodes. This helps data move more efficiently across the system when supporting multiple AI workloads simultaneously.

The key is designing the system so it can be expanded in modules across the chips, memory, and storage. This makes it suitable for data centers that need to increase processing capacity based on actual workloads. The image should clearly use separate arrows to show data paths between compute, the optical network, and storage.

What Atlas Looks Like at the Cluster Level

Atlas should be viewed as a large cluster that distributes workloads between AI processing units and storage systems, with an optical network serving as the primary link between nodes. This helps data move more efficiently across the system when supporting multiple AI workloads simultaneously.

The key is designing the system so it can be expanded in modules across the chips, memory, and storage. This makes it suitable for data centers that need to increase processing capacity based on actual workloads. The image should clearly use separate arrows to show data paths between compute, the optical network, and storage.

When AI Demand Grows Faster Than Chips Can Be Produced

Organizations rushing to increase AI processing capacity may face constraints beyond software, starting with chip availability, delivery, and network systems that depend on infrastructure from Nvidia. When demand grows faster than production capacity, cluster expansion slows immediately.

Huawei has therefore turned to an approach that combines Atlas chips with an optical network to move data more efficiently between processing units. This concept helps reduce queueing in conventional systems and makes it more suitable for organizations that need to support large, continuous workloads.

When AI Demand Grows Faster Than Chips Can Be Produced

Organizations rushing to increase AI processing capacity may face constraints beyond software, starting with chip availability, delivery, and network systems that depend on infrastructure from Nvidia. When demand grows faster than production capacity, cluster expansion slows immediately.

Huawei has therefore turned to an approach that combines Atlas chips with an optical network to move data more efficiently between processing units. This concept helps reduce queueing in conventional systems and makes it more suitable for organizations that need to support large, continuous workloads.

Where Atlas Fits in Huawei’s AI Strategy

Atlas is not merely a chip, but an AI platform that connects chips with servers, management systems, and software for model development. Combined with an optical network, it can support large-scale cluster operations as a unified system.

Huawei positions Atlas at the center of everything from data-center hardware to software and cloud infrastructure. Focusing on the Chinese market therefore aligns with domestic demand and helps the company maintain closer control over both the supply chain and software ecosystem.

Where Atlas Fits in Huawei’s AI Strategy

Atlas is not merely a chip, but an AI platform that connects chips with servers, management systems, and software for model development. Combined with an optical network, it can support large-scale cluster operations as a unified system.

Huawei positions Atlas at the center of everything from data-center hardware to software and cloud infrastructure. Focusing on the Chinese market therefore aligns with domestic demand and helps the company maintain closer control over both the supply chain and software ecosystem.

From the Previous Generation to a 15,488-Chip Cluster

The available information does not specify figures for the previous Atlas generation, so the comparison is mainly based on the design direction. The new generation focuses on combining a large number of chips through optical networking to support large-scale AI workloads in China.

Factor Previous-generation AtlasNew-generation cluster
Chip count Not specified15,488 chips
Processing power Not specifiedUp to 120 EFLOPS
Connectivity Not specifiedOptical networking
System scalability Not specifiedSupports large-scale clusters
Power consumption Not specifiedNot specified
Target use Not specifiedData-center-level AI workloads

The key point is that Huawei is moving from standalone systems toward a cluster-based architecture suited to model training and AI services that require sustained processing capacity.

From the Previous Generation to a 15,488-Chip Cluster

The available information does not specify figures for the previous Atlas generation, so the comparison is mainly based on the design direction. The new generation focuses on combining a large number of chips through optical networking to support large-scale AI workloads in China.

Factor Previous-generation AtlasNew-generation cluster
Chip count Not specified15,488 chips
Processing power Not specifiedUp to 120 EFLOPS
Connectivity Not specifiedOptical networking
System scalability Not specifiedSupports large-scale clusters
Power consumption Not specifiedNot specified
Target use Not specifiedData-center-level AI workloads

The key point is that Huawei is moving from standalone systems toward a cluster-based architecture suited to model training and AI services that require sustained processing capacity.

How Optical Networking Changes the Real-World Experience

When many chips must work together to train large models, optical networking helps transfer data between nodes more efficiently. Tasks therefore do not have to wait too long for connectivity from any single node.

In data-center workloads, this advantage helps reduce processing bottlenecks, making the cluster better suited to continuous workloads than a standalone machine.

The system can be expanded in groups according to demand, making it suitable for AI services that need more processing capacity when many users access them simultaneously.

For inference workloads, end users may not see the network directly, but they may experience more consistent responsiveness, especially during periods with a high volume of requests.

How Optical Networking Changes the Real-World Experience

When many chips must work together to train large models, optical networking helps transfer data between nodes more efficiently. Tasks therefore do not have to wait too long for connectivity from any single node.

In data-center workloads, this advantage helps reduce processing bottlenecks, making the cluster better suited to continuous workloads than a standalone machine.

The system can be expanded in groups according to demand, making it suitable for AI services that need more processing capacity when many users access them simultaneously.

For inference workloads, end users may not see the network directly, but they may experience more consistent responsiveness, especially during periods with a high volume of requests.

How Far Can Atlas Compete with Nvidia and Other Alternatives?

Factor Huawei AtlasNvidiaAMD InstinctGoogle TPU
Real-world performance Suitable for large-scale AI clustersStrong in both training and inferencePowerful for data-center workloadsSuitable for workloads on Google Cloud
Software Depends on Huawei toolsCUDA is widely adoptedSupports major frameworksTied to Google services
Ecosystem Strong in ChinaBroadest in the marketGrowingStrong in cloud environments
Availability Addresses demand in ChinaDepends on production capacity and export restrictionsMultiple manufacturer optionsPrimarily accessed through the cloud
Use outside China Still limited by software and service availabilityMost readyMore flexibleBest suited to Google Cloud users

Atlas is attractive for organizations in China that want to reduce their dependence on Nvidia. However, if systems need to be moved across countries, Nvidia still has an advantage because of its more mature software and tools.

How Far Can Atlas Compete with Nvidia and Other Alternatives?

Factor Huawei AtlasNvidiaAMD InstinctGoogle TPU
Real-world performance Suitable for large-scale AI clustersStrong in both training and inferencePowerful for data-center workloadsSuitable for workloads on Google Cloud
Software Depends on Huawei toolsCUDA is widely adoptedSupports major frameworksTied to Google services
Ecosystem Strong in ChinaBroadest in the marketGrowingStrong in cloud environments
Availability Addresses demand in ChinaDepends on production capacity and export restrictionsMultiple manufacturer optionsPrimarily accessed through the cloud
Use outside China Still limited by software and service availabilityMost readyMore flexibleBest suited to Google Cloud users

Atlas is attractive for organizations in China that want to reduce their dependence on Nvidia. However, if systems need to be moved across countries, Nvidia still has an advantage because of its more mature software and tools.

Clear Strengths and Constraints to Accept

Atlas stands out for combining a large-scale cluster with optical networking, helping reduce data-transfer bottlenecks and support demanding AI workloads. Its production capacity in China serves the domestic market, but global expansion may face constraints involving supply, software, and cross-border system support.

Pros

  • +The cluster supports large-scale AI workloads
  • +Optical networking helps reduce data bottlenecks
  • +Suitable for organizations seeking to reduce dependence on Nvidia

Cons

  • −Software and tools are not yet as broadly compatible as those of competitors
  • −Maintenance outside China may be more difficult
  • −Global expansion carries supply and after-sales service risks

Clear Strengths and Constraints to Accept

Atlas stands out for combining a large-scale cluster with optical networking, helping reduce data-transfer bottlenecks and support demanding AI workloads. Its production capacity in China serves the domestic market, but global expansion may face constraints involving supply, software, and cross-border system support.

Pros

  • +The cluster supports large-scale AI workloads
  • +Optical networking helps reduce data bottlenecks
  • +Suitable for organizations seeking to reduce dependence on Nvidia

Cons

  • −Software and tools are not yet as broadly compatible as those of competitors
  • −Maintenance outside China may be more difficult
  • −Global expansion carries supply and after-sales service risks

The Cost of a Cluster Goes Beyond the Chips

The true cost also includes servers, power and cooling systems, and optical networking capable of handling large volumes of data transmission. Data centers may therefore need to adjust their floor space, power systems, and security measures to accommodate this workload.

There are also installation costs, software, developer tools, and specialized personnel to maintain the system. If an organization already has an existing system, it may also incur costs to modify or migrate workloads to the new platform.

Therefore, the chip price is only the starting point. Organizations should consider the total cost over the system’s lifetime, including maintenance and future expansion.

The Cost of a Cluster Goes Beyond the Chips

The true cost also includes servers, power and cooling systems, and optical networking capable of handling large volumes of data transmission. Data centers may therefore need to adjust their floor space, power systems, and security measures to accommodate this workload.

There are also installation costs, software, developer tools, and specialized personnel to maintain the system. If an organization already has an existing system, it may also incur costs to modify or migrate workloads to the new platform.

Therefore, the chip price is only the starting point. Organizations should consider the total cost over the system’s lifetime, including maintenance and future expansion.

Conclusion: Can 120 EFLOPS Really Shift the Competitive Balance?

The 120 EFLOPS figure makes Atlas appear capable of challenging Nvidia, but success is not measured by performance alone. Production capacity, software, delivery, and the developer ecosystem are equally important.

The fact that demand in China still exceeds production capacity may reflect both supply constraints and the need to establish a foothold in the home market. Delaying global marketing could therefore be a temporary defensive move or preparation for a major capacity expansion. The key question is how effectively Huawei can turn the potential of a 15,488-chip cluster into a system that works in practice.

Conclusion: Can 120 EFLOPS Really Shift the Competitive Balance?

The 120 EFLOPS figure makes Atlas appear capable of challenging Nvidia, but success is not measured by performance alone. Production capacity, software, delivery, and the developer ecosystem are equally important.

The fact that demand in China still exceeds production capacity may reflect both supply constraints and the need to establish a foothold in the home market. Delaying global marketing could therefore be a temporary defensive move or preparation for a major capacity expansion. The key question is how effectively Huawei can turn the potential of a 15,488-chip cluster into a system that works in practice.

What Atlas Looks Like at the Cluster Level

The image should show Atlas as a group of multiple processing systems connected by an optical network, enabling continuous data transfer between chips. The key point is to clearly separate the data path from the storage system into the processing section.

This type of network reduces dependence on conventional internal connectivity and makes cluster expansion appear more systematic. The image should emphasize the connection architecture rather than simply showing a large pile of chips.

What Atlas Looks Like at the Cluster Level

The image should show Atlas as a group of multiple processing systems connected by an optical network, enabling continuous data transfer between chips. The key point is to clearly separate the data path from the storage system into the processing section.

This type of network reduces dependence on conventional internal connectivity and makes cluster expansion appear more systematic. The image should emphasize the connection architecture rather than simply showing a large pile of chips.

When AI Demand Grows Faster Than Chips Can Be Produced

Organizations rushing to increase AI processing capacity may be constrained not only by budgets, but also by delayed chip deliveries and excessive dependence on Nvidia infrastructure. When demand grows rapidly, system expansion plans can stall from the outset.

Huawei is therefore attempting to establish its own approach by combining Atlas chips with optical networking, allowing clusters to expand in a more systematic way. The key is reducing dependence on conventional internal connectivity and responding to the Chinese market’s continuous AI demand, which has outpaced production capacity.

When AI Demand Grows Faster Than Chips Can Be Produced

Organizations rushing to increase AI processing capacity may be constrained not only by budgets, but also by delayed chip deliveries and excessive dependence on Nvidia infrastructure. When demand grows rapidly, system expansion plans can stall from the outset.

Huawei is therefore attempting to establish its own approach by combining Atlas chips with optical networking, allowing clusters to expand in a more systematic way. The key is reducing dependence on conventional internal connectivity and responding to the Chinese market’s continuous AI demand, which has outpaced production capacity.

Where Atlas Fits in Huawei’s AI Strategy

Atlas is not merely a chip, but a platform that combines chips, servers, software systems, and cloud infrastructure, enabling customers to build AI systems from the data center through to real-world applications.

Huawei is positioning Atlas as a complete option for the Chinese market by connecting processing with optical networking and large-scale clusters. This allows the company to control both hardware and software while addressing domestic demand, which is growing faster than production capacity, rather than immediately entering global markets.

Where Atlas Fits in Huawei’s AI Strategy

Atlas is not merely a chip, but a platform that combines chips, servers, software systems, and cloud infrastructure, enabling customers to build AI systems from the data center through to real-world applications.

Huawei is positioning Atlas as a complete option for the Chinese market by connecting processing with optical networking and large-scale clusters. This allows the company to control both hardware and software while addressing domestic demand, which is growing faster than production capacity, rather than immediately entering global markets.

From the Previous Generation to a 15,488-Chip Cluster

The new cluster moves from the previous Atlas system toward combining a large number of chips, using optical networking as its core. Its strengths include support for data-center-level AI workloads and a target processing capacity of up to 120 EFLOPS, although the available information does not specify power consumption details.

Factor Previous-generation AtlasNew-generation cluster
Chip count Not specified15,488 chips
Processing power Not specifiedUp to 120 EFLOPS
Connectivity Not specifiedOptical networking
System scalability Existing systemLarge-scale cluster
Power consumption Not specifiedNot specified
Target use AI workloadsData-center-level AI workloads

From the Previous Generation to a 15,488-Chip Cluster

The new cluster moves from the previous Atlas system toward combining a large number of chips, using optical networking as its core. Its strengths include support for data-center-level AI workloads and a target processing capacity of up to 120 EFLOPS, although the available information does not specify power consumption details.

Factor Previous-generation AtlasNew-generation cluster
Chip count Not specified15,488 chips
Processing power Not specifiedUp to 120 EFLOPS
Connectivity Not specifiedOptical networking
System scalability Existing systemLarge-scale cluster
Power consumption Not specifiedNot specified
Target use AI workloadsData-center-level AI workloads

How Optical Networking Changes the Real-World Experience

When many chips must work together to train large models, optical networking helps transfer data between nodes more quickly. Tasks therefore do not have to wait too long for communication between chips.

For large-scale clusters, the main advantage is reducing bottlenecks between nodes, making it easier to add machines and expand the system without allowing connectivity to become the limiting factor.

Inference workloads and AI services with many users also benefit because the system can distribute tasks across different nodes more efficiently, supporting simultaneous requests in a way suited to data-center workloads.

Overall, this is not merely about powerful chips, but about designing the entire system so the chips can work together at full capacity, with optical networking serving as a key enabler.

How Optical Networking Changes the Real-World Experience

When many chips must work together to train large models, optical networking helps transfer data between nodes more quickly. Tasks therefore do not have to wait too long for communication between chips.

For large-scale clusters, the main advantage is reducing bottlenecks between nodes, making it easier to add machines and expand the system without allowing connectivity to become the limiting factor.

Inference workloads and AI services with many users also benefit because the system can distribute tasks across different nodes more efficiently, supporting simultaneous requests in a way suited to data-center workloads.

Overall, this is not merely about powerful chips, but about designing the entire system so the chips can work together at full capacity, with optical networking serving as a key enabler.

How Far Can Atlas Compete with Nvidia and Other Alternatives?

The available research does not provide details about Atlas, Nvidia, AMD, or Google TPU, so it is not possible to make definitive conclusions about real-world performance, cost, or use outside China. This table should be viewed as highlighting information gaps before making an investment decision.

Factor Huawei AtlasNvidiaAMD InstinctGoogle TPU
Real-world performance No verified dataNo verified dataNo verified dataNo verified data
Software No verified dataNo verified dataNo verified dataNo verified data
Ecosystem No verified dataNo verified dataNo verified dataNo verified data
Availability No verified dataNo verified dataNo verified dataNo verified data
Infrastructure cost No verified dataNo verified dataNo verified dataNo verified data
Use outside China No verified dataNo verified dataNo verified dataNo verified data

How Far Can Atlas Compete with Nvidia and Other Alternatives?

The available research does not provide details about Atlas, Nvidia, AMD, or Google TPU, so it is not possible to make definitive conclusions about real-world performance, cost, or use outside China. This table should be viewed as highlighting information gaps before making an investment decision.

Factor Huawei AtlasNvidiaAMD InstinctGoogle TPU
Real-world performance No verified dataNo verified dataNo verified dataNo verified data
Software No verified dataNo verified dataNo verified dataNo verified data
Ecosystem No verified dataNo verified dataNo verified dataNo verified data
Availability No verified dataNo verified dataNo verified dataNo verified data
Infrastructure cost No verified dataNo verified dataNo verified dataNo verified data
Use outside China No verified dataNo verified dataNo verified dataNo verified data

Clear Strengths and Constraints to Accept

Pros

  • +Optical networking has the potential to reduce bottlenecks between machines in the cluster
  • +Dependence on demand in China may provide a clearly defined initial market

Cons

  • −There is still no verified data on performance, production capacity, or maintenance
  • −Software compatibility outside China remains unclear, and global expansion carries risks

Clear Strengths and Constraints to Accept

Pros

  • +Optical networking has the potential to reduce bottlenecks between machines in the cluster
  • +Dependence on demand in China may provide a clearly defined initial market

Cons

  • −There is still no verified data on performance, production capacity, or maintenance
  • −Software compatibility outside China remains unclear, and global expansion carries risks

The Cost of a Cluster Goes Beyond the Chips

The chip price is only the starting point, because organizations must also pay for servers, power systems, cooling systems, and data centers capable of supporting continuous operation.

Optical networking also involves equipment and installation costs, along with software, personnel, and system maintenance. If existing systems need to be adapted to the new cluster, costs and downtime may increase further. This means evaluating value requires looking at the entire system rather than only the chip price.

The Cost of a Cluster Goes Beyond the Chips

The chip price is only the starting point, because organizations must also pay for servers, power systems, cooling systems, and data centers capable of supporting continuous operation.

Optical networking also involves equipment and installation costs, along with software, personnel, and system maintenance. If existing systems need to be adapted to the new cluster, costs and downtime may increase further. This means evaluating value requires looking at the entire system rather than only the chip price.

Conclusion: Can 120 EFLOPS Really Shift the Competitive Balance?

The 120 EFLOPS figure sounds impressive, but it is not enough to determine the competitive outcome. Atlas must prove its production capacity, practical software, reliable delivery, and ecosystem that makes adoption easy for customers.

Delaying global marketing could therefore be either a defensive move to support demand in China or preparation for a major capacity expansion. The outcome will become clearer when Huawei demonstrates how effectively it can turn paper performance into a system that works in practice.

Conclusion: Can 120 EFLOPS Really Shift the Competitive Balance?

The 120 EFLOPS figure sounds impressive, but it is not enough to determine the competitive outcome. Atlas must prove its production capacity, practical software, reliable delivery, and ecosystem that makes adoption easy for customers.

Delaying global marketing could therefore be either a defensive move to support demand in China or preparation for a major capacity expansion. The outcome will become clearer when Huawei demonstrates how effectively it can turn paper performance into a system that works in practice. Huawei Delays Pushing Atlas AI Chips to Global Markets Because Demand in China Still Exceeds Production Capacity

The Atlas cluster uses an optical network to connect a large number of chips, reducing bottlenecks during processing. This gives it potential as an option for large-scale AI workloads that require fast, highly scalable connectivity.

However, challenging Nvidia still requires looking at real-world benchmarks, costs, and software readiness, because powerful hardware alone is not enough for enterprise-level use. Huawei Delays Pushing Atlas AI Chips to Global Markets Because Demand in China Still Exceeds Production Capacity

The Atlas cluster uses an optical network to connect a large number of chips, reducing bottlenecks during processing. This gives it potential as an option for large-scale AI workloads that require fast, highly scalable connectivity.

However, challenging Nvidia still requires looking at real-world benchmarks, costs, and software readiness, because powerful hardware alone is not enough for enterprise-level use.

What Atlas Looks Like at the Cluster Level

Atlas should be viewed as a large cluster that distributes workloads between AI processing units and storage systems, with an optical network serving as the primary link between nodes. This helps data move more efficiently across the system when supporting multiple AI workloads simultaneously.

The key is designing the system so it can be expanded in modules across the chips, memory, and storage. This makes it suitable for data centers that need to increase processing capacity based on actual workloads. The image should clearly use separate arrows to show data paths between compute, the optical network, and storage.

What Atlas Looks Like at the Cluster Level

Atlas should be viewed as a large cluster that distributes workloads between AI processing units and storage systems, with an optical network serving as the primary link between nodes. This helps data move more efficiently across the system when supporting multiple AI workloads simultaneously.

The key is designing the system so it can be expanded in modules across the chips, memory, and storage. This makes it suitable for data centers that need to increase processing capacity based on actual workloads. The image should clearly use separate arrows to show data paths between compute, the optical network, and storage.

When AI Demand Grows Faster Than Chips Can Be Produced

Organizations rushing to increase AI processing capacity may face constraints beyond software, starting with chip availability, delivery, and network systems that depend on infrastructure from Nvidia. When demand grows faster than production capacity, cluster expansion slows immediately.

Huawei has therefore turned to an approach that combines Atlas chips with an optical network to move data more efficiently between processing units. This concept helps reduce queueing in conventional systems and makes it more suitable for organizations that need to support large, continuous workloads.

When AI Demand Grows Faster Than Chips Can Be Produced

Organizations rushing to increase AI processing capacity may face constraints beyond software, starting with chip availability, delivery, and network systems that depend on infrastructure from Nvidia. When demand grows faster than production capacity, cluster expansion slows immediately.

Huawei has therefore turned to an approach that combines Atlas chips with an optical network to move data more efficiently between processing units. This concept helps reduce queueing in conventional systems and makes it more suitable for organizations that need to support large, continuous workloads.

Where Atlas Fits in Huawei’s AI Strategy

Atlas is not merely a chip, but an AI platform that connects chips with servers, management systems, and software for model development. Combined with an optical network, it can support large-scale cluster operations as a unified system.

Huawei positions Atlas at the center of everything from data-center hardware to software and cloud infrastructure. Focusing on the Chinese market therefore aligns with domestic demand and helps the company maintain closer control over both the supply chain and software ecosystem.

Where Atlas Fits in Huawei’s AI Strategy

Atlas is not merely a chip, but an AI platform that connects chips with servers, management systems, and software for model development. Combined with an optical network, it can support large-scale cluster operations as a unified system.

Huawei positions Atlas at the center of everything from data-center hardware to software and cloud infrastructure. Focusing on the Chinese market therefore aligns with domestic demand and helps the company maintain closer control over both the supply chain and software ecosystem.

From the Previous Generation to a 15,488-Chip Cluster

The available information does not specify figures for the previous Atlas generation, so the comparison is mainly based on the design direction. The new generation focuses on combining a large number of chips through optical networking to support large-scale AI workloads in China.

Factor Previous-generation AtlasNew-generation cluster
Chip count Not specified15,488 chips
Processing power Not specifiedUp to 120 EFLOPS
Connectivity Not specifiedOptical networking
System scalability Not specifiedSupports large-scale clusters
Power consumption Not specifiedNot specified
Target use Not specifiedData-center-level AI workloads

The key point is that Huawei is moving from standalone systems toward a cluster-based architecture suited to model training and AI services that require sustained processing capacity.

From the Previous Generation to a 15,488-Chip Cluster

The available information does not specify figures for the previous Atlas generation, so the comparison is mainly based on the design direction. The new generation focuses on combining a large number of chips through optical networking to support large-scale AI workloads in China.

Factor Previous-generation AtlasNew-generation cluster
Chip count Not specified15,488 chips
Processing power Not specifiedUp to 120 EFLOPS
Connectivity Not specifiedOptical networking
System scalability Not specifiedSupports large-scale clusters
Power consumption Not specifiedNot specified
Target use Not specifiedData-center-level AI workloads

The key point is that Huawei is moving from standalone systems toward a cluster-based architecture suited to model training and AI services that require sustained processing capacity.

How Optical Networking Changes the Real-World Experience

When many chips must work together to train large models, optical networking helps transfer data between nodes more efficiently. Tasks therefore do not have to wait too long for connectivity from any single node.

In data-center workloads, this advantage helps reduce processing bottlenecks, making the cluster better suited to continuous workloads than a standalone machine.

The system can be expanded in groups according to demand, making it suitable for AI services that need more processing capacity when many users access them simultaneously.

For inference workloads, end users may not see the network directly, but they may experience more consistent responsiveness, especially during periods with a high volume of requests.

How Optical Networking Changes the Real-World Experience

When many chips must work together to train large models, optical networking helps transfer data between nodes more efficiently. Tasks therefore do not have to wait too long for connectivity from any single node.

In data-center workloads, this advantage helps reduce processing bottlenecks, making the cluster better suited to continuous workloads than a standalone machine.

The system can be expanded in groups according to demand, making it suitable for AI services that need more processing capacity when many users access them simultaneously.

For inference workloads, end users may not see the network directly, but they may experience more consistent responsiveness, especially during periods with a high volume of requests.

How Far Can Atlas Compete with Nvidia and Other Alternatives?

Factor Huawei AtlasNvidiaAMD InstinctGoogle TPU
Real-world performance Suitable for large-scale AI clustersStrong in both training and inferencePowerful for data-center workloadsSuitable for workloads on Google Cloud
Software Depends on Huawei toolsCUDA is widely adoptedSupports major frameworksTied to Google services
Ecosystem Strong in ChinaBroadest in the marketGrowingStrong in cloud environments
Availability Addresses demand in ChinaDepends on production capacity and export restrictionsMultiple manufacturer optionsPrimarily accessed through the cloud
Use outside China Still limited by software and service availabilityMost readyMore flexibleBest suited to Google Cloud users

Atlas is attractive for organizations in China that want to reduce their dependence on Nvidia. However, if systems need to be moved across countries, Nvidia still has an advantage because of its more mature software and tools.

How Far Can Atlas Compete with Nvidia and Other Alternatives?

Factor Huawei AtlasNvidiaAMD InstinctGoogle TPU
Real-world performance Suitable for large-scale AI clustersStrong in both training and inferencePowerful for data-center workloadsSuitable for workloads on Google Cloud
Software Depends on Huawei toolsCUDA is widely adoptedSupports major frameworksTied to Google services
Ecosystem Strong in ChinaBroadest in the marketGrowingStrong in cloud environments
Availability Addresses demand in ChinaDepends on production capacity and export restrictionsMultiple manufacturer optionsPrimarily accessed through the cloud
Use outside China Still limited by software and service availabilityMost readyMore flexibleBest suited to Google Cloud users

Atlas is attractive for organizations in China that want to reduce their dependence on Nvidia. However, if systems need to be moved across countries, Nvidia still has an advantage because of its more mature software and tools.

Clear Strengths and Constraints to Accept

Atlas stands out for combining a large-scale cluster with optical networking, helping reduce data-transfer bottlenecks and support demanding AI workloads. Its production capacity in China serves the domestic market, but global expansion may face constraints involving supply, software, and cross-border system support.

Pros

  • +The cluster supports large-scale AI workloads
  • +Optical networking helps reduce data bottlenecks
  • +Suitable for organizations seeking to reduce dependence on Nvidia

Cons

  • −Software and tools are not yet as broadly compatible as those of competitors
  • −Maintenance outside China may be more difficult
  • −Global expansion carries supply and after-sales service risks

Clear Strengths and Constraints to Accept

Atlas stands out for combining a large-scale cluster with optical networking, helping reduce data-transfer bottlenecks and support demanding AI workloads. Its production capacity in China serves the domestic market, but global expansion may face constraints involving supply, software, and cross-border system support.

Pros

  • +The cluster supports large-scale AI workloads
  • +Optical networking helps reduce data bottlenecks
  • +Suitable for organizations seeking to reduce dependence on Nvidia

Cons

  • −Software and tools are not yet as broadly compatible as those of competitors
  • −Maintenance outside China may be more difficult
  • −Global expansion carries supply and after-sales service risks

The Cost of a Cluster Goes Beyond the Chips

The true cost also includes servers, power and cooling systems, and optical networking capable of handling large volumes of data transmission. Data centers may therefore need to adjust their floor space, power systems, and security measures to accommodate this workload.

There are also installation costs, software, developer tools, and specialized personnel to maintain the system. If an organization already has an existing system, it may also incur costs to modify or migrate workloads to the new platform.

Therefore, the chip price is only the starting point. Organizations should consider the total cost over the system’s lifetime, including maintenance and future expansion.

The Cost of a Cluster Goes Beyond the Chips

The true cost also includes servers, power and cooling systems, and optical networking capable of handling large volumes of data transmission. Data centers may therefore need to adjust their floor space, power systems, and security measures to accommodate this workload.

There are also installation costs, software, developer tools, and specialized personnel to maintain the system. If an organization already has an existing system, it may also incur costs to modify or migrate workloads to the new platform.

Therefore, the chip price is only the starting point. Organizations should consider the total cost over the system’s lifetime, including maintenance and future expansion.

Conclusion: Can 120 EFLOPS Really Shift the Competitive Balance?

The 120 EFLOPS figure makes Atlas appear capable of challenging Nvidia, but success is not measured by performance alone. Production capacity, software, delivery, and the developer ecosystem are equally important.

The fact that demand in China still exceeds production capacity may reflect both supply constraints and the need to establish a foothold in the home market. Delaying global marketing could therefore be a temporary defensive move or preparation for a major capacity expansion. The key question is how effectively Huawei can turn the potential of a 15,488-chip cluster into a system that works in practice.

Conclusion: Can 120 EFLOPS Really Shift the Competitive Balance?

The 120 EFLOPS figure makes Atlas appear capable of challenging Nvidia, but success is not measured by performance alone. Production capacity, software, delivery, and the developer ecosystem are equally important.

The fact that demand in China still exceeds production capacity may reflect both supply constraints and the need to establish a foothold in the home market. Delaying global marketing could therefore be a temporary defensive move or preparation for a major capacity expansion. The key question is how effectively Huawei can turn the potential of a 15,488-chip cluster into a system that works in practice.

What Atlas Looks Like at the Cluster Level

The image should show Atlas as a group of multiple processing systems connected by an optical network, enabling continuous data transfer between chips. The key point is to clearly separate the data path from the storage system into the processing section.

This type of network reduces dependence on conventional internal connectivity and makes cluster expansion appear more systematic. The image should emphasize the connection architecture rather than simply showing a large pile of chips.

What Atlas Looks Like at the Cluster Level

The image should show Atlas as a group of multiple processing systems connected by an optical network, enabling continuous data transfer between chips. The key point is to clearly separate the data path from the storage system into the processing section.

This type of network reduces dependence on conventional internal connectivity and makes cluster expansion appear more systematic. The image should emphasize the connection architecture rather than simply showing a large pile of chips.

When AI Demand Grows Faster Than Chips Can Be Produced

Organizations rushing to increase AI processing capacity may be constrained not only by budgets, but also by delayed chip deliveries and excessive dependence on Nvidia infrastructure. When demand grows rapidly, system expansion plans can stall from the outset.

Huawei is therefore attempting to establish its own approach by combining Atlas chips with optical networking, allowing clusters to expand in a more systematic way. The key is reducing dependence on conventional internal connectivity and responding to the Chinese market’s continuous AI demand, which has outpaced production capacity.

When AI Demand Grows Faster Than Chips Can Be Produced

Organizations rushing to increase AI processing capacity may be constrained not only by budgets, but also by delayed chip deliveries and excessive dependence on Nvidia infrastructure. When demand grows rapidly, system expansion plans can stall from the outset.

Huawei is therefore attempting to establish its own approach by combining Atlas chips with optical networking, allowing clusters to expand in a more systematic way. The key is reducing dependence on conventional internal connectivity and responding to the Chinese market’s continuous AI demand, which has outpaced production capacity.

Where Atlas Fits in Huawei’s AI Strategy

Atlas is not merely a chip, but a platform that combines chips, servers, software systems, and cloud infrastructure, enabling customers to build AI systems from the data center through to real-world applications.

Huawei is positioning Atlas as a complete option for the Chinese market by connecting processing with optical networking and large-scale clusters. This allows the company to control both hardware and software while addressing domestic demand, which is growing faster than production capacity, rather than immediately entering global markets.

Where Atlas Fits in Huawei’s AI Strategy

Atlas is not merely a chip, but a platform that combines chips, servers, software systems, and cloud infrastructure, enabling customers to build AI systems from the data center through to real-world applications.

Huawei is positioning Atlas as a complete option for the Chinese market by connecting processing with optical networking and large-scale clusters. This allows the company to control both hardware and software while addressing domestic demand, which is growing faster than production capacity, rather than immediately entering global markets.

From the Previous Generation to a 15,488-Chip Cluster

The new cluster moves from the previous Atlas system toward combining a large number of chips, using optical networking as its core. Its strengths include support for data-center-level AI workloads and a target processing capacity of up to 120 EFLOPS, although the available information does not specify power consumption details.

Factor Previous-generation AtlasNew-generation cluster
Chip count Not specified15,488 chips
Processing power Not specifiedUp to 120 EFLOPS
Connectivity Not specifiedOptical networking
System scalability Existing systemLarge-scale cluster
Power consumption Not specifiedNot specified
Target use AI workloadsData-center-level AI workloads

From the Previous Generation to a 15,488-Chip Cluster

The new cluster moves from the previous Atlas system toward combining a large number of chips, using optical networking as its core. Its strengths include support for data-center-level AI workloads and a target processing capacity of up to 120 EFLOPS, although the available information does not specify power consumption details.

Factor Previous-generation AtlasNew-generation cluster
Chip count Not specified15,488 chips
Processing power Not specifiedUp to 120 EFLOPS
Connectivity Not specifiedOptical networking
System scalability Existing systemLarge-scale cluster
Power consumption Not specifiedNot specified
Target use AI workloadsData-center-level AI workloads

How Optical Networking Changes the Real-World Experience

When many chips must work together to train large models, optical networking helps transfer data between nodes more quickly. Tasks therefore do not have to wait too long for communication between chips.

For large-scale clusters, the main advantage is reducing bottlenecks between nodes, making it easier to add machines and expand the system without allowing connectivity to become the limiting factor.

Inference workloads and AI services with many users also benefit because the system can distribute tasks across different nodes more efficiently, supporting simultaneous requests in a way suited to data-center workloads.

Overall, this is not merely about powerful chips, but about designing the entire system so the chips can work together at full capacity, with optical networking serving as a key enabler.

How Optical Networking Changes the Real-World Experience

When many chips must work together to train large models, optical networking helps transfer data between nodes more quickly. Tasks therefore do not have to wait too long for communication between chips.

For large-scale clusters, the main advantage is reducing bottlenecks between nodes, making it easier to add machines and expand the system without allowing connectivity to become the limiting factor.

Inference workloads and AI services with many users also benefit because the system can distribute tasks across different nodes more efficiently, supporting simultaneous requests in a way suited to data-center workloads.

Overall, this is not merely about powerful chips, but about designing the entire system so the chips can work together at full capacity, with optical networking serving as a key enabler.

How Far Can Atlas Compete with Nvidia and Other Alternatives?

The available research does not provide details about Atlas, Nvidia, AMD, or Google TPU, so it is not possible to make definitive conclusions about real-world performance, cost, or use outside China. This table should be viewed as highlighting information gaps before making an investment decision.

Factor Huawei AtlasNvidiaAMD InstinctGoogle TPU
Real-world performance No verified dataNo verified dataNo verified dataNo verified data
Software No verified dataNo verified dataNo verified dataNo verified data
Ecosystem No verified dataNo verified dataNo verified dataNo verified data
Availability No verified dataNo verified dataNo verified dataNo verified data
Infrastructure cost No verified dataNo verified dataNo verified dataNo verified data
Use outside China No verified dataNo verified dataNo verified dataNo verified data

How Far Can Atlas Compete with Nvidia and Other Alternatives?

The available research does not provide details about Atlas, Nvidia, AMD, or Google TPU, so it is not possible to make definitive conclusions about real-world performance, cost, or use outside China. This table should be viewed as highlighting information gaps before making an investment decision.

Factor Huawei AtlasNvidiaAMD InstinctGoogle TPU
Real-world performance No verified dataNo verified dataNo verified dataNo verified data
Software No verified dataNo verified dataNo verified dataNo verified data
Ecosystem No verified dataNo verified dataNo verified dataNo verified data
Availability No verified dataNo verified dataNo verified dataNo verified data
Infrastructure cost No verified dataNo verified dataNo verified dataNo verified data
Use outside China No verified dataNo verified dataNo verified dataNo verified data

Clear Strengths and Constraints to Accept

Pros

  • +Optical networking has the potential to reduce bottlenecks between machines in the cluster
  • +Dependence on demand in China may provide a clearly defined initial market

Cons

  • −There is still no verified data on performance, production capacity, or maintenance
  • −Software compatibility outside China remains unclear, and global expansion carries risks

Clear Strengths and Constraints to Accept

Pros

  • +Optical networking has the potential to reduce bottlenecks between machines in the cluster
  • +Dependence on demand in China may provide a clearly defined initial market

Cons

  • −There is still no verified data on performance, production capacity, or maintenance
  • −Software compatibility outside China remains unclear, and global expansion carries risks

The Cost of a Cluster Goes Beyond the Chips

The chip price is only the starting point, because organizations must also pay for servers, power systems, cooling systems, and data centers capable of supporting continuous operation.

Optical networking also involves equipment and installation costs, along with software, personnel, and system maintenance. If existing systems need to be adapted to the new cluster, costs and downtime may increase further. This means evaluating value requires looking at the entire system rather than only the chip price.

The Cost of a Cluster Goes Beyond the Chips

The chip price is only the starting point, because organizations must also pay for servers, power systems, cooling systems, and data centers capable of supporting continuous operation.

Optical networking also involves equipment and installation costs, along with software, personnel, and system maintenance. If existing systems need to be adapted to the new cluster, costs and downtime may increase further. This means evaluating value requires looking at the entire system rather than only the chip price.

Conclusion: Can 120 EFLOPS Really Shift the Competitive Balance?

The 120 EFLOPS figure sounds impressive, but it is not enough to determine the competitive outcome. Atlas must prove its production capacity, practical software, reliable delivery, and ecosystem that makes adoption easy for customers.

Delaying global marketing could therefore be either a defensive move to support demand in China or preparation for a major capacity expansion. The outcome will become clearer when Huawei demonstrates how effectively it can turn paper performance into a system that works in practice.

Conclusion: Can 120 EFLOPS Really Shift the Competitive Balance?

The 120 EFLOPS figure sounds impressive, but it is not enough to determine the competitive outcome. Atlas must prove its production capacity, practical software, reliable delivery, and ecosystem that makes adoption easy for customers.

Delaying global marketing could therefore be either a defensive move to support demand in China or preparation for a major capacity expansion. The outcome will become clearer when Huawei demonstrates how effectively it can turn paper performance into a system that works in practice.