Alibaba Increases Cambricon GPU Order to 150,000 Units
Computing Power Crisis? Not an Issue! Alibaba's 150,000 Cambricon Order Fills H20 Supply Gap
The global AI computing power race intensifies as NVIDIA's H20 supply restrictions deal a heavy blow to the domestic industry. At this critical moment, Alibaba steps forward with decisive action to break through the computing power bottleneck, sending shockwaves through the industry with every move!
Emergency Order of 150,000 Units! Cambricon Becomes Alibaba's "Computing Power Fire Brigade"
As NVIDIA's H20 supply disruption hits suddenly, creating an instant computing power shortage domestically, Alibaba Cloud's Tongyi Qianwen large language model faces a "supply crisis." In this critical moment, Alibaba decisively places a massive order - increasing Cambricon Siyuan 370 units to 150,000 pieces, staging a "computing power rescue operation"!
This Siyuan 370 is no ordinary chip: featuring 7nm process + Chiplet technology, integrating 39 billion transistors, with tested 300TOPS (INT8) computing power that directly competes with H20, while achieving 40% superior energy efficiency! As of Q2 2025, it already handles 60% of Alibaba Cloud's inference demands. Through PCIe 5.0 multi-card interconnection, it successfully withstands the user growth pressure of Tongyi Qianwen, rapidly "stopping the bleeding" of the computing power gap!
Alibaba's Trump Card! Self-Developed AI Chips and Domestic Wafer Foundries!
Not satisfied with external support alone, Alibaba directly launches a "self-research counterattack"! The new generation AI chip targets the market void left by H20's exit, backed by domestic 14nm and more advanced processes, with strong support from local forces like Yangtze Memory Technologies.
The heterogeneous architecture integrates high-density computing units with ultra-large memory, LPDDR5X bandwidth exceeding 1TB/s, single-card computing power of 300-400TOPS (INT8) matching H20. Even more impressive is its compatibility: not only supporting FP8/FP16 mixed precision, but also directly connecting to the CUDA ecosystem through dynamic instruction translation layers, allowing engineers zero-pressure code reuse and cutting migration costs by over 70%! Previously, Alibaba Cloud used Siyuan 370 + MagicMind toolchain to triple model conversion efficiency, firmly refuting the "domestic chip ecosystem is inferior" narrative.
Dual-Foundry Backup + Three-Phase Strategy! Alibaba Establishes Domestic Computing Power "Master Plan"
To ensure chip mass production, Alibaba employs "dual insurance": SMIC's 14nm production line maintains 95%+ yield rate with stable monthly output of 50,000 units; Hua Hong Semiconductor's 7nm process targets 2026 mass production, aiming for computing power exceeding 500TOPS with 30% improved energy efficiency.
Furthermore, a "three-phase strategy" determines the overall direction: short-term (2025-2026) leverages 7nm/14nm inference chips to capture market share; medium-term (2027-2028) introduces 4nm training chips with thousand-card cluster computing power targeting 1EFLOPS to rival H100; long-term (post-2030) bets on photonic computing, with already released commercial photonic AI chips achieving speeds 1000 times faster than GPUs while reducing power consumption by 90%, directly targeting the future!
From emergency procurement to self-research breakthroughs, then to long-term planning, Alibaba is breaking NVIDIA's monopoly through the "compatibility - substitution - transcendence" path. The 2026 mass production of 4nm chips will be a crucial battle - if delivered on schedule, domestic computing power will surely rise to dominance on the global stage!