Tesla's TeraFab and Huawei's Ascend 950PR Launch on Same Day, Escalating AI Compute Sovereignty Race

Global AI Compute Infrastructure Race Accelerates: Musk’s TERAFAB and Huawei’s Ascend 950PR Unveiled on the Same Day—Fueling the Contest for Domestic Alternatives and Supercomputing Sovereignty

On March 22, the evolution of global AI infrastructure reached a highly symbolic “dual-peak moment”: Elon Musk officially announced TERAFAB on X—the world’s first vertically integrated semiconductor fabrication facility designed to achieve an annual production capacity of 1 terawatt (TW) of chip power; almost simultaneously, Huawei held its Ascend Ecosystem Conference in Songshan Lake, Dongguan, unveiling its new AI acceleration chip, the Ascend 950PR. Though rooted in divergent technical approaches and strategic contexts, these two milestones pierced the horizon on the same day—signaling that U.S.–China competition over foundational AI compute sovereignty has fully descended from algorithmic models and software frameworks down into the physical realm: wafer fabs, advanced packaging lines, liquid-cooled server racks, and high-voltage substations. A “dual-track race” concerning national digital survival rights has now entered its most intense phase.

TERAFAB: A Declaration of Compute Sovereignty for the Space Age—Redefining Semiconductor Industry Logic

TERAFAB is no conventional foundry. Musk explicitly defined its core mission as producing dedicated, ultra-reliable, low-latency, radiation-hardened AI inference and control chips for Starlink Gen2, Mars colonization missions, and neural interface devices. Its “1-TW annual capacity” target is profoundly disruptive: assuming current mainstream AI training chips consume 300–700 W per card, full-capacity operation would support the annual supply of over 1.4 million high-end AI accelerator cards—far exceeding TSMC’s total AI-chip foundry output in 2023. More critically, TERAFAB adopts an “IDM 2.0” model—building its own cleanrooms, developing proprietary EUV photoresist formulations, controlling advanced packaging (chiplet-level 3D stacking), and constructing megawatt-scale liquid-cooling clusters and microgrids. This approach bypasses the choke point of ASML lithography machine export controls, instead achieving 3-nm-equivalent energy efficiency at mature nodes (e.g., 7 nm) through heterogeneous integration and system-level optimization. Supply-chain security thus shifts away from reliance on single-node breakthroughs toward full-stack controllability across materials, packaging, thermal management, and power delivery. This directly disrupts global semiconductor equipment vendors’ investment narratives: Applied Materials and Lam Research are rapidly pivoting R&D toward advanced packaging deposition and etching tools; fabless giants like NVIDIA and AMD are showing markedly stronger willingness to diversify away from TSMC; and leading Chinese liquid-cooling firms—including Sugon and Galanz—saw overseas inquiry volumes surge by 300% week-on-week.

Ascend 950PR: A Precision Breakthrough in Low-Precision Compute—HBM Bandwidth as a New Sovereignty Anchor

Mirroring TERAFAB’s “space expedition,” Huawei’s Ascend 950PR represents a meticulously executed “land-based assault.” Its greatest leap lies in delivering 1.25 TB/s of HBM3 bandwidth at FP16 precision—a 210% increase over the prior Ascend 910B—and, for the first time in a domestically developed AI chip, integrating dedicated hardware acceleration units for INT4 sparse computation, achieving an inference energy efficiency of 42 TOPS/W. This means the 950PR can replace portions of A100/H100 compute in latency- and power-constrained applications such as edge-deployed large language models, real-time autonomous driving decisions, and industrial visual inspection—without dependence on NVIDIA’s CUDA ecosystem. Crucially, its HBM3 memory is jointly customized by ChangXin Memory and Huawei’s HiSilicon, with packaging and testing handled by Shenghe Microelectronics—completely severing ties with U.S.-based memory suppliers. This “bandwidth sovereignty” breakthrough empowers domestic AI server OEMs—including Digital China and Inspur—to build truly autonomous technology stacks spanning chip → memory → interconnect → system. According to calculations by the China Academy of Information and Communications Technology (CAICT) under the Ministry of Industry and Information Technology (MIIT), mass deployment of the 950PR will shorten construction timelines for domestic 1,000-GPU-scale intelligent computing centers by 40%, reduce power infrastructure costs by 28%, and directly lift the localization rate of western nodes in China’s “East Data, West Computing” initiative from 35% to 68%.

Geopolitical Undercurrents: The Dual Contest over the Strait of Hormuz and the “Compute Corridor”

Notably, both technological announcements coincided with escalating Middle Eastern geopolitical risks: Iran responded to Trump’s threat of strikes against its power infrastructure by declaring it would activate an “intelligent management” mechanism if the Strait of Hormuz were blocked—and issued six ceasefire conditions, including establishing a new legal framework for the Strait. Meanwhile, the U.S. quietly activated Jared Kushner-led diplomatic contingency plans for negotiations with Iran, centered squarely on freedom of navigation and nuclear material handling. On the surface, this appears to be an energy corridor dispute—but it reflects a deeper logic: The Strait of Hormuz is the physical-world chokepoint for oil flows; AI compute infrastructure is the digital-world “new strait” for value flows. As the U.S. seeks to erect a “Red Sea” of compute export restrictions via the CHIPS and Science Act, China accelerates its “compute voucher” subsidies and nationwide integrated big-data center strategy—and emerging nations like Iran are launching national AI compute hubs (e.g., Tehran’s “Persian Gulf AI Cloud”). A three-tiered global compute sovereignty architecture is now taking shape:

• The “Rule-Setting Tier”, led by the U.S. and Europe;
• The “Manufacturing Execution Tier”, anchored by China, Japan, and South Korea;
• The “Access & Usage Tier”, represented by the Middle East, Latin America, and Africa.
  Whoever controls end-to-end standards—from electricity input and chip manufacturing to model training—will hold agenda-setting power over the next generation of digital order.

Investment Logic Reassessed: A Paradigm Shift from “Chip Stocks” to “Compute Infrastructure ETFs”

Market reactions confirm the depth of this foundational shift. On March 22, A-share compute leasing stocks (e.g., Hongbo Co., Sugon) surged over 12% in a single day; optical module/CPO概念股 (e.g., Innolight, New Optics) saw volume break above their yearly moving averages; meanwhile, valuations of traditional GPU design firms diverged sharply—IP-license-dependent names faced pressure, while Cambricon and Biren Technology—both possessing advanced packaging capabilities—received countercyclical inflows from Northbound funds. A more telling signal came from financial innovation: the ChinaAMC “Artificial Intelligence Compute Infrastructure ETF” was approved the following day. For the first time, its constituent stocks include liquid-cooling equipment manufacturers, ultra-high-voltage supporting enterprises, and domestic HBM supply-chain players—with weightings deliberately decoupled from traditional semiconductor index frameworks. This signals that capital has clearly recognized: future AI investing is no longer about betting on a single chipmaker’s product, but about systematically pricing the integrated capability across five dimensions—power, land, cooling, networking, and silicon. For domestic investors, three key themes warrant close attention:

1. Green power infrastructure supporting western nodes of the “East Data, West Computing” initiative (e.g., China Three Gorges Corporation, Longyuan Power);
2. Progress on domestic HBM commercialization to meet the Ascend 950PR’s high-bandwidth demands (e.g., ChangXin Memory, Jiangfeng Electronics target materials);
3. Megawatt-scale liquid-cooling solutions compatible with TERAFAB-class hyperscale compute clusters (e.g., Galanz, Envicool).

Compute sovereignty is not an abstract concept. It is the temperature-control curve engineers at Songshan Lake Lab fine-tune on the Ascend 950PR; it is the millisecond-level load response of TERAFAB’s microgrid in the Texas desert; it is the newly plotted “digital shipping lane” coordinates overlaid onto Strait of Hormuz nautical charts. When physical-world energy corridors and digital-world compute corridors are assigned strategic priority on the same day, we finally grasp the truth: genuine technological autonomy has never resided solely in isolated lab breakthroughs—it lives instead in the silent yet tenacious coupling points between wafer fabs and substations, between packaging lines and cooling towers, between code and cables.