Huawei Accelerates Ascend 960DT AI Chip Launch to Q1 2027, Intensifying Competition
AI chip development accelerates with new launches and performance benchmarks, while the industry grapples with power constraints and evolving data center infrastructure demands.
The story
In a significant move to challenge global leaders, Huawei has accelerated its AI chip roadmap, announcing that its flagship Ascend 960DT chip will now launch in the first quarter of 2027, several months ahead of its original late 2027 schedule. A variant, the Ascend 960PR, is slated for Q3 2027.
The company also detailed its Atlas 960 SuperPoD computing cluster. This acceleration underscores China's push for technological self-reliance amidst global restrictions and highlights Huawei's ambition to compete with Nvidia on the global stage. The move implies a faster timeline for deploying advanced AI compute capacity, particularly within China, as Huawei aims to boost the nation's fundamental computing power and offer an alternative to US-origin semiconductors.
Silicon
Ascend 960DT
Maker: Huawei
What: Flagship AI chip, accelerated launch to Q1 2027, targeting AI training and inference.
For Whom: Chinese AI developers and data centers, aiming to replace Nvidia in China.
Ascend 960PR
Maker: Huawei
What: AI NPU variant, accelerated launch to Q3 2027, offering 8 FP4 PFLOPS for inference.
For Whom: AI prefill and recommendation workloads, primarily within China.
MTIA 450 (Arke)
Maker: Meta
What: Third-generation in-house AI chip, currently in testing, aims for better performance per watt/dollar for inference.
For Whom: Meta's own data centers to power Facebook and Instagram AI applications, deploying in H1 2027.
Vera Rubin NVL72
Maker: Nvidia
What: AI supercomputer system, performance results highlighted for up to 40% more GPU capacity per megawatt and 2.5x throughput over GB300 NVL72.
For Whom: Large-scale AI deployments and operators seeking to expand output within existing power budgets.
The build-out
Integrated Energy-to-Compute Infrastructure
Who: Delta Electronics
Scale: 800 VDC power delivery (98% efficiency), liquid-cooling to 3 MW
Where: For NVIDIA DSX AI Factories globally, announced from Fremont, California.
Data Center Services Restructuring
Who: EverOn Data Center Services (spun off from T5 Services), JK Technology Services (acquired 360 Rigging), Limbach Holdings (acquired CYMCOR)
Scale: Creating integrated platforms for data center lifecycle, specialized workforce, logistics, and heavy-lift capabilities.
Where: North American data center market, driven by record absorption of 25 GW in H1 2026.
Supply & policy signals
US House passes Ratepayer Protection Act (417-3 vote)
Implication: Establishes a federal standard to prevent households and small businesses from subsidizing data center energy costs.
Oregon PUC approves Schedule 96 for large-load data center customers of Portland General Electric
Implication: Ensures data centers pay for their specific energy costs, protecting other consumers from explosive growth.
South Dakota legislation requires data centers (over 10 MW) to reimburse utilities for specific costs
Implication: Addresses direct power costs and the risk of stranded infrastructure investments for large energy users.
Average data center rack loads more than doubled from 12 kW in 2024 to 26 kW by 2026
Implication: Higher rack densities demand upgrades in power delivery, cooling, and structural design, making electricity a key constraint alongside chip supply.
What we'll be watching
- Meta's deployment of its MTIA 450 (Arke) custom AI chips in its data centers during the first half of 2027.
- Huawei's commercial launch of the Ascend 960DT AI chip in Q1 2027.
- Huawei's release of the Ascend 960PR AI NPU in Q3 2027.
- Meta's MTIA 500 (Astrid) chip completing design work within a month and entering data centers by the end of 2027.
Reporting + analyst voices: grounded via Google Search at publish time.