Friday, September 11, 2026
AI 인프라 · 뉴스 & 분석
데이터센터리포트
데이터센터 · 리포트

중국 AI 개발사 Z.ai가 1기가와트 데이터 센터 완공(NVIDIA 실리콘 제외 국산 칩 독점 운영); 일부 구간 현재 운영 중이며 10,000칩 클러스터 서빙

미국 수출 통제 하의 중국 공급망 독립성 입증; 지정학적 AI 인프라 비대칭성 재편성; 대규모 추론용 대체 실리콘 실행 가능성 검증
업계 전문지Slicast · July 22, 2026 · 글로벌 · 출처: Tom's Hardware
중요도 92

Chinese AI developer Z.ai (formerly Zhipu) has completed construction of a 1-gigawatt data center stocked exclusively with domestically manufactured chips and has begun operating portions of it, Bloomberg reported Monday, citing sources familiar with the matter. The facility will train the company's GLM model family, and Z.ai now operates several computing clusters containing more than 10,000 chips each. A gigawatt of electrical capacity—enough to power approximately 750,000 homes—places the site among the largest AI infrastructure projects built by any Chinese developer.

While Z.ai has not publicly named its chip supplier, recent developments point to Huawei. The company released GLM-5.2 in June, an open-weight model trained entirely on Huawei Ascend accelerators with no Nvidia involvement, which reached the top of open-weight model leaderboards within a week. Z.ai's constraint is structural: the company has been on the U.S. Commerce Department's entity list since January 2025, which cuts off legal access to Nvidia processors and leaves Chinese suppliers as its only viable option.

However, raw power consumption paints an incomplete picture. Chinese accelerators, including Huawei's Ascend line, deliver lower performance per watt than Nvidia's current Blackwell processors. A gigawatt of domestic silicon therefore provides substantially less usable training compute than a gigawatt of Nvidia systems.

China is drafting a nationwide AI data center grid initiative targeting approximately 2 trillion yuan ($295 billion) in spending over five years, with a requirement that at least 80 percent of underlying technology come from Chinese suppliers. Domestic capacity constraints pose a major challenge. SMIC's most advanced stable node—the 7-nanometer-class N+2 process—is running above 93 percent utilization. Meanwhile, scarce domestic high-bandwidth memory (HBM) limits Ascend accelerator production, and Huawei shipped approximately 812,000 AI chips in the past year. China can construct a 1-gigawatt facility shell faster than domestic fabs can supply the chips needed to fill it.

Z.ai is on track for $1 billion in annual recurring revenue after meeting its 2026 sales target in July, according to people familiar with the matter. The company recently raised billions in capital through a Hong Kong IPO and follow-on share sale. Competitor Moonshot suspended new subscriptions Sunday, prioritizing compute allocation for existing members following the launch of its Kimi K3 model. Z.ai has not disclosed the new data center's location, construction cost, or build timeline.

원문 보기