Huawei just dropped a chip that draws 600 watts.
The Ascend 950PR, packaged inside Huawei's new Atlas 350 accelerator card, delivers 1.56 petaflops of FP4 compute and 112GB of in-house HBM running at 1.4 TB/s bandwidth. Those numbers put it at roughly 2.8 times the performance of Nvidia's China-legal H20. The power draw? About 1.5 times the H20 as well. That 600W figure is not a ceiling. It is a floor for what comes next.
And two of China's largest cloud buyers want in. Reuters reported on March 27 that ByteDance and Alibaba are planning to place orders for the 950PR after receiving samples in January. Huawei aims to ship around 750,000 units this year, with mass production expected in H2 2026.
The cooling industry should be paying very close attention to this.
The Western AI infrastructure conversation revolves around Nvidia's B200 and B300 platforms, chips expected to push past 1,000 watts each. That power density is forcing every hyperscaler from Microsoft to Google into liquid cooling. The conversation in China is following an identical thermodynamic trajectory, just with different silicon.
At 600W per card, the Ascend 950PR cannot be air-cooled at scale. Full stop. Huawei knows this. The company has already launched the Atlas 950 SuperPoD, a fully liquid-cooled server architecture built around the 950PR. When you pack these cards into racks of eight or more, you are looking at cabinet power densities well north of 50kW, with PUE targets as low as 1.12. Air cannot do that.
Now multiply that by 750,000 chips. Even at conservative rack densities, that volume of 600W accelerators translates to an enormous cooling infrastructure buildout across Chinese data centers over the next 18 months.
The spending numbers tell part of the story. Alibaba Cloud is weighing an increase in its three-year AI infrastructure capex to $69 billion, up from a $52.4 billion commitment made in early 2025. ByteDance planned to spend roughly 100 billion yuan ($14 billion) on AI chips alone in 2026. Some meaningful fraction of those budgets will flow directly into thermal management.
The question that matters for the cooling supply chain: who gets that money?
In China, the answer is increasingly domestic. Sugon Data Innovation (a subsidiary of Dawning Information Industry, or Zhongke Sugon) held a 55.7% share of China's liquid cooling and temperature control equipment shipments in 2024, ranking first domestically for four consecutive years. Sugon specializes in immersion phase-change liquid cooling, a technology well-suited to the thermal profiles of high-wattage AI accelerators.
Then there is Envicool. Shenzhen-based, publicly traded, and pulling in roughly $663 million in trailing twelve-month revenue. Envicool manufactures modular coolant distribution units (CDUs) entirely out of its Guangdong facilities, with delivery cycles under four weeks for domestic customers. That speed advantage matters when you are racing to deploy 750,000 chips. Google has reportedly tapped Envicool as a core CDU supplier, granting it about 25% of Google's 2026 cooperation orders. If Envicool is good enough for Mountain View, it is certainly good enough for Hangzhou and Beijing.
Goldman Sachs and UBS have both raised price targets on Envicool stock, citing demand tied to AI workloads and energy efficiency. The investor class sees what the cooling industry should already know.
Here is what this Huawei chip order means at a systems level. The global data center cooling market is not consolidating. It is forking.
On one side, you have the Nvidia-driven stack. GB200, GB300, whatever Jensen Huang announces next. That stack pulls Vertiv, Schneider Electric, CoolIT, and a growing roster of Western liquid cooling vendors into hyperscaler supply chains. On the other side, Huawei's Ascend ecosystem is pulling Sugon, Envicool, and Hongsheng into an entirely parallel cooling supply chain, one built for Chinese hyperscalers running Chinese silicon.
The thermal engineering problems are identical. The vendor ecosystems are diverging.
China's data center cooling market is projected to reach $830 million by 2030, growing at a 17.45% CAGR from its 2025 base of $371 million. China's broader liquid cooling industry is expected to hit the 100 billion yuan mark in 2026 alone. And AI data centers are forecast to account for more than 30% of liquid cooling demand by year's end.
Those are not small numbers. And they exist almost entirely outside the Western cooling vendor ecosystem.
Huawei has published a three-year Ascend chip roadmap targeting 4 ZettaFLOPS of FP4 performance by 2028, backed by next-generation in-house HBM with up to 1.6 TB/s bandwidth. The memory comes from Huawei's own HiBL 1.0 interconnect standard. Every generation on that roadmap will draw more power. Every generation will demand more from the cooling stack.
The 950PR runs at 600W today on SMIC's 7nm process. Huawei is chasing compute density gains on a process node that Nvidia abandoned years ago. That means brute-force power scaling to close the gap. For cooling vendors, that translates to guaranteed demand growth. Physics does not care about geopolitics.
ByteDance and Alibaba placing orders for the 950PR is a procurement decision. But it is also a vote of confidence in a fully domestic AI compute stack, from silicon to server to cooling. Western vendors like Vertiv still sell into China, but the policy direction is clear: localize everything. When Huawei's Ascend ecosystem scales, the thermal infrastructure scales with it. And that thermal infrastructure is being built by Chinese companies, for Chinese data centers, on Chinese timelines.
The cooling TAM just got bigger. But the addressable part depends entirely on which side of the chip stack you serve.