Z.ai, a major Chinese large-model developer, has completed a 1GW-class data center housing only Chinese-made AI chips and has begun partial operations, Bloomberg reported on July 20, 2026, citing anonymous sources familiar with the matter. What stands out even more than the facility's scale is the shift in how the company secures compute. In its IPO prospectus from late 2025, Z.ai had explicitly stated a policy of not building its own data centers and instead purchasing compute from Chinese cloud providers. If the latest report is accurate, the company reversed its infrastructure strategy in roughly seven months.
1GW-Class, But Not Yet Fully Operational
According to Bloomberg, the facility serves as a base for developing Z.ai's advanced GLM models, and the company has already brought part of it online. Z.ai has reportedly also built or is operating multiple computing clusters composed of more than 10,000 chips each. Z.ai did not respond to Bloomberg's inquiry.
Here, the "1GW" figure represents the facility's eventual power-receiving capacity—it does not mean current power consumption has reached 1GW, nor that the computing equipment is fully operational. It also remains unclear whether the multiple 10,000-plus-chip clusters are located within this facility or at separate sites.
The facility's location, along with the manufacturer, generation, and number of AI chips installed, has not been disclosed. The description "entirely Chinese-made" appears to refer specifically to the AI chips, not to confirmation that all equipment—including CPUs, memory, and networking hardware—is domestically produced. The 1GW power capacity and the actual usable computing performance need to be evaluated as separate matters.
About Seven Months Since IPO Claims of "No Self-Built Facilities"
Z.ai's Hong Kong IPO prospectus, published on December 30, 2025, offers a starkly different picture. The company stated that since being added to the U.S. Commerce Department's Entity List in January 2025, it had not procured AI chips—regardless of whether they were subject to U.S. export control regulations. It further stated it had no plans to develop its own AI data centers, and was instead purchasing compute resources from Chinese cloud providers.
This cloud dependence is also reflected in costs. Compute service fees totaled RMB 1.5528 billion in 2024, accounting for 70.7% of R&D spending. In the first half of 2025, this rose to RMB 1.1451 billion, pushing the ratio up to 71.8%. Full-year R&D spending for 2025 reached RMB 3.1804 billion, while capital expenditure fell 83.8% year-over-year to just RMB 74.7 million. That year, Z.ai secured compute primarily through purchased services, supplementing with equipment leasing.
The prospectus also flagged the risk that if third parties raised prices or restricted supply, it could disrupt development and service delivery. Operating a 1GW-class facility directly would reduce this dependence. The significance of the latest report lies less in how many domestic chips were adopted, and more in the fact that Z.ai appears to be taking direct control over the procurement and operation of its compute resources.
Software Supporting Seven Families of Domestic Chips
Simply lining up large numbers of Chinese-made AI chips doesn't automatically make them function as a unified computing system. As of June 2025, Z.ai stated it had adapted its models to run on more than 40 major chip types, using a proprietary operator library to support training across diverse hardware.
GLM-5, released in February 2026, supports Huawei Ascend, Moore Threads, and Cambricon. It also covers Kunlun Chip and MetaX, along with Enflame and Hygon—supporting a total of seven non-NVIDIA chip families. This reflects a design capable of porting models across multiple domestic accelerator platforms. Even though the chip composition within the facility remains undisclosed, there is clear evidence that Z.ai has built software infrastructure designed to run across multiple hardware platforms.
Compute demands have also grown substantially. GLM-5 activates 40 billion parameters out of a total of 744 billion, with pre-training data reaching 28.5 trillion tokens. This marks a significant increase from GLM-4.5's 355 billion total parameters and 23 trillion tokens. Whether domestic chip support succeeds isn't determined merely by whether a model can be launched. What actually determines real-world performance is how stably the system manages communication and fault handling across numerous chips, and how high an operational uptime it can sustain.
Verification Conditions After Raising Roughly HK$31.4 Billion
One week before the data center report, on July 13, 2026, Z.ai completed a new share issuance, raising net proceeds of approximately HK$31.4 billion (HK$3,137,495,000). The company plans to allocate 55% of this toward general-purpose AI research and development, with uses including the purchase and leasing of compute, related services, repayment of loans used to strategically secure compute, and systems for allocating heterogeneous computing resources. However, there has been no disclosure indicating that these funds were used to build the reported facility.
Even with funding and power capacity secured, it doesn't necessarily follow that the company can continuously develop cutting-edge models using domestic chips. If the company discloses who operates the facility and its chip composition, along with metrics such as the time required to train GLM models there and performance per unit of power, the "1GW" label would transform from a marketing headline into verifiable computing capability. Until such figures are released, "1GW" should be understood as a description of facility scale—not proven computing performance.