US-China AI Compute Sovereignty Escalates: TERAFAB and Ascend 950PR Signal a New Era of Infrastructure-First Competition
Global AI Compute Infrastructure Enters a New Phase of National Competition: Hardware Sovereignty Emerges as the Strategic High Ground
On March 21, Elon Musk officially announced the TERAFAB project in Austin—a super-fab targeting an annual production capacity of 1 terawatt (1 TW) of AI compute. Jointly launched by Tesla, SpaceX, and xAI, the facility plans to deeply integrate logic chip manufacturing, high-bandwidth memory (HBM) assembly, and advanced 2.5D/3D packaging processes within a single physical site—enabling a closed-loop workflow spanning “design → tape-out → packaging → system-level validation.” Crucially, 80% of its output will be reserved for upgrading the Starlink satellite constellation, powering AI-based navigation systems for Mars missions, and training xAI’s large language models. Almost simultaneously, on March 22, Huawei unveiled its Ascend 950PR AI acceleration chip in Dongguan: built on Huawei’s self-developed Da Vinci Architecture 3.0, it delivers a peak FP16 performance of 1.2 petaFLOPS, achieves an energy efficiency ratio of 42 TOPS/W, supports lossless interconnectivity across thousands of accelerator cards, and—uniquely—integrates HBM3e memory with liquid-cooled microchannel heat sinks in a single package. These two milestones are not isolated technical breakthroughs; rather, they signal a fundamental paradigm shift in global AI competition: the race for computing power is rapidly evolving—from iterative algorithmic improvements in research labs into a state-led contest for sovereignty over foundational infrastructure.
United States: Vertical Integration + Space Anchoring — Reconfiguring the Geopolitics of Compute
At its core, TERAFAB represents a disruptive deconstruction of the traditional semiconductor industry’s division of labor. Over the past decade, foundries like TSMC and Samsung have sustained global demand for AI chips through “pure-play” manufacturing. Musk, however, bypasses this mature foundry ecosystem entirely—opting instead to build end-to-end capabilities covering Chiplet-based heterogeneous integration, silicon photonics interconnects, and 3D stacking packaging. This model reflects a clear strategic anchor: space missions impose extreme compute requirements—ultra-low latency, radiation hardening, and autonomous decision-making—that commercial foundries cannot guarantee. Starlink Gen2 satellites must process terabytes of remote-sensing data onboard in real time; Mars landers must autonomously navigate obstacles during communication blackouts—scenarios demanding chips engineered from inception for specific physical environments. TERAFAB is thus more than a factory: it is a pivotal U.S. initiative to “militarize and space-qualify” AI compute infrastructure. Industry estimates suggest that 1 TW of annual capacity equates to deploying roughly 200,000 H100-class GPUs, entailing capital expenditures exceeding $30 billion—and directly stimulating domestic demand for EUV lithography tools, advanced packaging equipment, and HBM supply-chain restructuring. More profoundly, TERAFAB accelerates implementation of the U.S. CHIPS and Science Act, pushing Intel Foundry Services (IFS) to absorb part of the order volume—thereby forging a closed-loop ecosystem: “design (xAI) → manufacturing (TERAFAB/IFS) → application (SpaceX).”
China: Ascend 950PR’s Breakout Strategy and Ecosystem Offensive
Huawei’s Ascend 950PR launch embodies China’s alternative path under intense sanctions: leveraging architectural innovation to offset process-node limitations, and pursuing system-level optimization to mitigate supply-chain risks. Constrained by export bans on sub-7nm advanced fabrication equipment, Ascend 950PR avoids chasing raw transistor density. Instead, it achieves quantum-leap performance gains through three key innovations: First, the Da Vinci 3.0 architecture introduces a dynamic sparsity computation engine, boosting real-world large-model inference throughput by 40%. Second, its pioneering “compute-in-memory” HBM3e interface delivers 1.2 TB/s bandwidth—reducing data-movement power consumption by 35%. Third, integrated micro-pumps and heat-pipe arrays cap per-card TDP below 650W, enabling mega-scale clusters of 10,000+ cards to achieve a PUE as low as 1.08. This marks a decisive pivot in China’s AI chip development—from “spec-by-spec benchmarking” toward “use-case-defined engineering.” Its core value lies in enabling large domestic models (e.g., Pangu, HunYuan) to scale across mission-critical sectors: government services, energy, and transportation. Notably, Ascend 950PR launched alongside CANN 8.0—the latest software stack supporting one-click migration of PyTorch/TensorFlow models and offering an open library of 1,200 industry-specific operators. Hardware breakthroughs are now accelerating deep ecosystem penetration.
Supply-Chain Catalysts: A Comprehensive Reassessment Across Equipment, Materials, and Cooling
These two flagship projects are driving structural transformation across the global semiconductor supply chain.
First, advanced packaging equipment has reached an inflection point: TERAFAB explicitly adopts CoWoS-L and SoIC technologies, directly boosting orders for hybrid bonding tools from Applied Materials and Tokyo Electron (TEL); meanwhile, Ascend 950PR’s 2.5D packaging needs benefit domestic players—including Advanced Micro-Fabrication Equipment (AMEC) and Shenghe Precision—in RDL (redistribution layer) and TSV (through-silicon via) substitution efforts.
Second, HBM3e has become the new strategic battleground: SK Hynix has announced mass production of HBM3e by 2026—but both TERAFAB and Ascend demand customized bandwidth and reliability specifications, accelerating R&D efforts by ChangXin Memory Technologies and GigaDevice in HBM-compatible DRAM and interface ICs.
Third, liquid cooling infrastructure has shifted from optional to mission-critical: With 10,000-card clusters generating over 100 MW of thermal load, air cooling has hit its physical limits. Companies including Invecas and Galanz Power disclose that liquid-cooled server penetration is projected to reach 35% by 2026—driving demand for cold plates, CDUs (cooling distribution units), and fluorinated coolants to double.
Fourth, optical module speed upgrades are now urgent: Ascend 950PR supports 800G CPO (co-packaged optics), while TERAFAB’s roadmap references silicon photonics integration—accelerating industrialization of 1.6T optical modules and LPO (linear-drive pluggable optics) technology by companies such as Innolight and Accelink.
Geopolitical Risks: Dual Pressures of Regulatory Escalation and Technological Decoupling
Yet this national sprint is intensifying technological fragmentation. The U.S. Department of Commerce has initiated a special review of TERAFAB concerning potential “dual-use technology leakage,” while preparing to add more advanced packaging tools—including laser annealing and atomic layer deposition (ALD) systems—to its export control list. More seriously, Washington is upgrading its “small yard, high fence” strategy: following restrictions on A100/H100 exports to China, new regulations expected in 2026 may cover all AI chips with FP16 performance exceeding 500 TFLOPS—the very metric Ascend 950PR targets. If implemented, China’s large-model training could confront a hard “compute ceiling.” Meanwhile, Iran’s recent statements regarding the Strait of Hormuz—though rooted in regional geopolitics—underscore an underlying anxiety about energy transport security, reflecting AI infrastructure’s acute dependence on electricity and cooling resources: global data centers already consume ~3% of worldwide electricity, and TERAFAB alone is projected to draw over 2 billion kWh annually—making grid stability an invisible cornerstone of compute sovereignty.
Conclusion: Infrastructure Sovereignty Beyond the Silicon Die
As Musk sketches his vision of a “space-based AI compute hub” in Austin, and Huawei unveils the “domestic large-model heart” in Dongguan, humanity stands at a new watershed: AI’s future will no longer hinge solely on algorithmic superiority—but on who controls wafer fabs, who sets packaging standards, and who masters coolant formulations. Premier Li Qiang’s remarks at the China Development Forum—“upholding openness and driving innovation”—carry heightened significance precisely now: genuine technological sovereignty has never meant isolationist self-reliance, but rather building irreplaceable value nodes within shared rule-based frameworks. The competition between TERAFAB and Ascend 950PR is, at root, a clash of two infrastructure philosophies: one anchoring technological evolution in outer space as its ultimate use case; the other cultivating ecosystem resilience across thousands of industries on Earth. The decisive factor may not lie in the raw compute number on any single chip—but in which nation first transforms computing power into a strategically managed, trusted, and universally accessible national resource—as dependable and ubiquitous as water or electricity.