The success of modern artificial intelligence is now dictated by the physics of the datacenter. Power, cooling, and interconnect decide what models you can run, how fast they respond, and what the final operational cost will be. China’s Ascend platform from Huawei is leaning into that reality with a scale-first approach: ship liquid-cooled Atlas 900 A3 SuperPoDs that knit up to 384 Ascend 910C chips into one logical machine today, then step up to multi-thousand-chip supernodes over the next product cycles.
For readers, the takeaway is simple. This is not a chip-versus-chip duel. It is a systems story about clustered compute, efficient cooling, and an opening software stack that aims to loosen dependence on any one vendor.
Along the way, we will keep two promises. First, we separate what is confirmed and deployed from what is announced and upcoming. Second, we connect the engineering to everyday impact, from lower energy waste to more accessible AI services. Where outside reporting or primary disclosures inform a claim, we link in context so you can check the source yourself.
Huawei Ascend Numbers that Matter Right Now
- 384-Chip Pods Exist Today: Huawei launched the Atlas 900 A3 SuperPoD in 2025, describing performance up to roughly 300 PFLOPS (FP16) and operation as “a single computer” built from interconnected racks.
- Deployed at Scale: Huawei reports 300+ SuperPoDs shipped in 2025 across more than 20 customers in internet, finance, carriers, electricity, and manufacturing. Treat this as vendor-reported until independently audited.
- Supernodes Are Next: Public statements set the Atlas 950 around 8,192 Ascend NPUs in Q4 2025, with a larger Atlas 960 around 15,488 NPUs targeted later. These are roadmap disclosures, not completed deployments.
- Cooling Drives Efficiency: Huawei’s full liquid-cooled cabinets claim a PUE drop from ~2.2 to ~1.1 and about 500,000 kWh annual savings per 50-kW cabinet compared with conventional air cooling. These figures come from Huawei’s sustainability explainer.
- Ecosystem Momentum: The vLLM-Ascend project continues shipping releases, while DeepSeek introduced a model variant with first-class support for Ascend + CANN, signaling broader software portability.
The Atlas 900 A3 SuperPoD: Deployed and In Use
What “SuperPoD” Means in Practice
A SuperPoD is Huawei’s term for a tightly networked cluster that behaves like one computer. The Atlas 900 A3 SuperPoD binds up to 384 Ascend 910C chips through a high-bandwidth fabric. This allows memory and compute to be orchestrated as a single logical system. For workloads like large-language-model inference or fine-tuning, this architecture prioritizes predictable scaling over raw single-die supremacy. Huawei cites up to ~300 PFLOPS (FP16) per SuperPoD and promotes CloudMatrix384 as a cloud instance built on top.
Where it is Deployed
Huawei disclosed that more than 300 Atlas 900 A3 SuperPoDs shipped from 2025 to 2026 to 20+ customers across multiple sectors. While this is what the vendor says, the industries mentioned match where liquid-cooled, high-performance computing is already beneficial: busy internet platforms, telecom networks, financial risk analysis, and electric utilities that need reliable prediction services.
Why this Matters Beyond China
Even if you never touch an Ascend chip, the pattern is global. Datacenters everywhere are pushing toward denser racks and liquid cooling because they reduce wasted power and tame hot spots that limit performance. High-voltage power backbones are shifting toward GaN-based 800-volt DC racks that streamline conversion stages, improve electrical safety, and free more headroom for dense AI compute. One radical answer to those same physics constraints is Google’s Project Suncatcher, which put its first four TPUs into orbit in October 2026 to test whether solar-powered satellites can host AI workloads without a data hall at all.
The Efficiency Gains of Liquid Cooling and PUE
PUE, Explained without the Jargon
Power Usage Effectiveness (PUE) is a data center’s total power divided by the power used by the IT equipment. A perfect PUE is 1.0, meaning every watt feeds the computers, not the chillers or fans. Most air-cooled halls run hotter and waste energy on air movement. Liquid, by contrast, moves heat out through cold plates and sealed loops that are far more efficient at high densities.
The Claimed Gains
Huawei’s full liquid-cooled cabinet design claims a PUE drop from about 2.2 to 1.1 and around 500,000 kWh saved annually for each 50-kW cabinet, with a corresponding cut in CO₂ emissions. Treat the exact values as vendor-provided, but the directional physics align with industry experience as racks climb past 50 to 100 kW.
Why it Aligns with Global Trends
As inference volumes rise, operators adopt several key technologies to curb thermal throttling and reclaim capacity, including:
- Liquid loops
- Rear-door heat exchangers
- Immersion systems
One caveat belongs alongside that list, since the fastest-growing immersion designs depend on fluorinated fluids from the PFAS family. Our coverage of PFAS data center cooling and its environmental risk sets out the trade-off those systems make.
The effect compounds at scale: lower PUE reduces operating costs, makes siting easier in power-constrained regions, and opens room for cleaner energy portfolios. Efficient thermal design now blends airflow engineering with liquid loops and controls described in modern data center cooling solutions, aligning hardware layouts with predictable PUE targets.
Interpreting Vendor Claims Accurately
- Treat vendor PUE claims as best-case in designed conditions.
- Look for utility-grade monitoring and third-party site audits to validate sustained PUE under mixed workloads.
- Compare energy per token or queries per kWh over time, not only peak TFLOPS.
From SuperPoD to Supernode: 8,192 to 15,488 Chips and the Interconnect Story
What Changes When You Scale by an Order of Magnitude?
The next step is not a faster single chip. It is a bigger logical computer. Huawei has outlined the Atlas 950 at about 8,192 Ascend NPUs, with a larger Atlas 960 at about 15,488 NPUs slated later, describing gains in compute, memory capacity, and fabric bandwidth. These disclosures set expectations, but they remain roadmap announcements until fielded.
Why Interconnect and Memory Pooling Drive Real-World Speed
Once you jump from hundreds to thousands of accelerators, model throughput is dominated by several factors:
- Latency
- Network topology
- Collective operations
A supernode’s value is the ability to schedule mixture-of-experts, cache long-context states, and synchronize gradients without drowning in network overhead. That is why Huawei emphasizes that a SuperPoD “works like a single computer” and positions supernodes as the same idea at a much larger scale. The point is not just raw FLOPs. It is a usable performance at the cluster scope.
How this Compares to Other Scale-Out Systems
Public comparisons in vendor materials sometimes position supernodes against other scale-out offerings. Sensible due diligence is to track independent disclosures, early customer case studies, and open-source performance notes as they appear. The vLLM-Ascend release feed on GitHub provides software primitives for scheduling, paged attention, and tensor parallelism that reach production maturity on Ascend.
Software Portability as the Enabler
Two recent developments matter for day-to-day builders. First, vLLM-Ascend continues to add features, which lowers friction for serving large models on Ascend hardware. Second, DeepSeek’s latest model line launched with first-class support for Ascend + CANN, a signal that major third parties are actively validating the stack. These changes do not eliminate growing pains, especially for frontier-scale training, but they do expand practical deployment choices in inference and fine-tuning.
A Reality Check on Training Versus Inference
Independent reporting indicates that some teams have struggled to train next-gen models entirely on Ascend due to software maturity and interconnect behavior, then switched to Nvidia for training while keeping Ascend for inference. This nuance matters. It suggests an ecosystem in transition, with rapid progress on serving stacks and a more extended development timeline required for end-to-end training.
Building an Open Software Ecosystem for AI
Why a Portable Software Stack Matters
Training and serving modern AI models depends on software layers that sit between the model code and the accelerator hardware. For years, most AI teams optimized around Nvidia’s CUDA, which made switching vendors difficult. China’s Ascend ecosystem is pushing toward portability by opening CANN, advancing the MindSpore framework, and shipping a vLLM-Ascend plugin so popular inference servers can run on Ascend NPUs with minimal code changes.
CANN’s Open-Source Turn
Huawei executives publicly committed to open access and open source for CANN by December 31, 2025, with compiler interfaces and a virtual instruction set exposed so developers can write lower-level kernels and tooling. Follow-up briefings reinforced that CANN operators would land on GitCode first, with additional core components staged through year-end. This is a pivotal shift because it reduces vendor lock-in and invites outside contributors to review, optimize, and maintain performance-critical pieces of the stack.
MindSpore as the Native Framework
MindSpore remains the native model framework for Ascend. In state-owned and industrial deployments, MindSpore’s graph execution, distributed strategies, and domain libraries have been paired with Ascend hardware to stand up large models with sector-specific goals. That foundation gives Chinese utilities, manufacturing, and telecoms a route to deploy models on domestic chips without a full re-platform.
vLLM-Ascend for Fast Inference
On the serving side, vLLM-Ascend is now the community-maintained hardware plugin that makes the widely used vLLM server run on Ascend NPUs. Recent releases highlight stability work, kernel coverage, and performance features that close the gap with mature CUDA paths. For practitioners, this means existing vLLM workflows can often be moved to Ascend by swapping backends and checking operator support, rather than rewriting the entire stack.
Third-Party Validation Signals
Signals from outside Huawei matter. DeepSeek’s newest model line debuted with first-day support for Ascend + CANN, alongside other Chinese accelerators. That kind of independent integration indicates that model developers see enough stability and demand to prioritize Ascend in their launch matrix. It also gives downstream teams confidence that popular models can be served natively on Ascend hardware. Precision formats like FP8 reduce energy per token, a tradeoff illustrated by DeepSeek’s FP8 efficiency work that links numerical accuracy to lower power and faster inference.
Distinguishing Deployed Systems from Future Roadmaps
What is Deployed Today?
Atlas 900 A3 SuperPoD systems with up to 384 Ascend 910C chips are shipping now. Huawei states that more than 300 SuperPoDs have been delivered in 2025 across 20+ customers in internet, finance, carrier, electricity, and manufacturing. Treat these as vendor-reported figures until independent audits are published, but the consistency of the number across multiple outlets raises confidence that broad deployment is underway.
What is Announced for Tomorrow
- CANN: public pledge to open access and open source by December 31, 2025, including compiler interfaces and core software components. This is a time-bound commitment, not a retroactive publication.
- Scale-Up to Supernodes: Huawei has outlined multi-thousand-NPU supernode targets for the next cycle. As with any roadmap, treat chip counts, bandwidth claims, and timing as forward-looking statements until customers disclose production usage and independent benchmarks appear.
Interpreting Vendor Claims Accurately
- Check the scope: cluster-level performance depends on interconnect topology and software maturity, not just peak FLOPs.
- Look for repeatability: vendor keynotes often present best-case runs. Independent site reports and utility-grade telemetry confirm whether results hold at scale.
- Track software parity: follow vLLM-Ascend and other backends for feature completeness, not just initial availability.
Where AI Touches Daily Life: Power Grids and Industrial Models
A Utility-Scale Case: China Southern Power Grid
One of the clearest examples comes from China Southern Power Grid, which built a sector model nicknamed “MegaWatt” on the Ascend computing platform with the MindSpore framework. The model runs on a Mixture-of-Experts setup and combines computer vision with language understanding to scan images from line inspections, flag likely defects, and triage work orders faster than manual review.
Public case notes from this deployment cite a fivefold improvement in defect and risk identification efficiency and accuracy above 90% in image recognition tasks.
Why Industrial AI Prefers Dense, Liquid-Cooled Clusters
Power grids, rail networks, and large factories generate continuous streams of telemetry and imagery. Inference happens around the clock, often with strict latency and reliability requirements. Liquid-cooled, high-density racks let operators pack more inference capacity into fewer cabinets while keeping Power Usage Effectiveness (PUE) low, which cuts operating costs and frees room in sites where power is capped. Those gains translate into shorter outage windows, quicker fault detection, and more stable services for end users.
The Practical Checklist for Industrial Buyers
- Thermal plan: requires liquid loops or rear-door heat exchangers for >50 kW per rack operations.
- Serving stack: confirm vLLM-Ascend or MindSpore coverage for your specific model family.
- Telemetry: insist on live PUE tracking and component-level alarms to verify real savings after go-live.
The Future of AI: Systems, Software, and Sustainability
China’s Ascend platform demonstrates that the future of artificial intelligence is a systems-level challenge, not just a chip-versus-chip race. The deployment of liquid-cooled Atlas 900 A3 SuperPoDs today, with a clear roadmap toward multi-thousand-chip supernodes, underscores a global shift. Performance is now defined by the efficiency of the interconnect, the intelligence of the software stack, and the sustainability of the physical infrastructure.
For developers and industry observers, the most critical development may be the push toward an open software ecosystem with CANN and vLLM-Ascend. These initiatives represent a deliberate move away from vendor lock-in, creating pathways for greater hardware choice and innovation. As these platforms mature, the true measure of success will be seen in real-world metrics: lower Power Usage Effectiveness (PUE) in data centers, broader access to powerful AI tools, and the development of a truly sustainable, planetary-scale AI infrastructure.
Your Questions on Liquid-Cooled Supernodes Answered
What is The Main Difference Between a SuperPoD and a Standard Server Rack?
A standard server rack contains multiple independent computers. In contrast, an Atlas 900 A3 SuperPoD is a tightly networked cluster of up to 384 chips that functions as a single, logical computer. This architecture is crucial for large-scale AI, as it allows for massive datasets and complex models to be processed in a unified and efficient manner without the bottlenecks of traditional networking.
Why Is Liquid Cooling So Important for Modern AI Hardware?
Modern AI chips generate an immense amount of heat, especially when packed densely into racks. Traditional air cooling becomes inefficient and costly at this scale, leading to a high Power Usage Effectiveness (PUE) where much of the energy is wasted on cooling. Liquid cooling is far more efficient at transferring heat, allowing for denser racks, lower energy consumption, and a PUE close to the ideal 1.0, which is critical for sustainable and cost-effective AI operations.
What Does an Open-Source Software Stack Like CANN Mean for AI Developers?
For years, the AI industry has been heavily reliant on Nvidia’s proprietary CUDA software, creating vendor lock-in. An open-source stack like Huawei’s CANN provides a “CUDA escape hatch.” It allows developers to write and optimize AI models that can run on different hardware platforms with minimal changes, fostering greater competition, innovation, and flexibility in the AI ecosystem.
