Z.ai's GLM-5.2 Beats GPT-5.5 on Key Benchmarks as China Pushes for AI Chip Self-Sufficiency

On other benchmarks, GLM-5.2 scored 89.5% on GPQA Diamond (graduate-level science Q&A), 81.0 on Terminal-Bench 2.1, and 40.1 on Humanity's Last Exam. NVIDIA's API documentation describes GLM-5.2 as delivering state-of-the-art performance across reasoning, coding, and agentic benchmarks.

Market performance has been dramatic but uneven.

• Z.ai posted a 132% rise in 2025 revenue, reaching RMB 724 million (~$100 million), making it China's largest independent large model provider by revenue.
• The company raised $552–558 million in its January 2026 Hong Kong IPO at a $6.6 billion valuation, becoming the world's first publicly traded LLM company.
• Shares surged as much as 35% on its first post-IPO earnings report and briefly pushed the company's market capitalization above HK$1 trillion (~$128 billion).
• However, those gains rest on thin fundamentals. Z.ai reported a net loss of RMB 4.718 billion (~$650 million) in 2025, with R&D expenses of RMB 3.18 billion. Reuters Breakingviews notes that Chinese LLM specialists like Z.ai face "significant financial losses, steep R&D expenses, intense price competition, and still-developing business models."

Custom Chip Development and the Domestic Silicon Pivot

Z.ai's push toward compute self-sufficiency falls into two categories: confirmed adaptation to existing domestic chips and preliminary exploration of a custom design.

Confirmed: GLM models already run on domestic chip infrastructure. Since February 2026, Z.ai's GLM-5 series has been adapted to run on domestic semiconductors after the U.S. tightened access to advanced Nvidia chips. GLM-5.2 was released with inference adaptation for a wide variety of domestic chip infrastructure, including Huawei Ascend clusters. In August 2025, Z.ai announced that its GLM models are compatible with Huawei's Ascend processors and Kirin chips. The company also trained GLM-Image, a 9-billion-parameter image generation model, entirely on Huawei's Ascend chips — the first major open-source model developed without any U.S. semiconductors.

Exploratory: Z.ai is in early discussions about a custom ASIC. According to The Information, as reported by Yahoo Finance, Z.ai recently made preliminary inquiries with several Chinese chip design houses about building a bespoke AI processor optimized specifically for its GLM model family. The company has not yet selected a partner, and no tape-out or finalized program has been confirmed. The reported trigger: a 27-times growth in daily token usage for GLM-5.2 colliding with tightening U.S. export restrictions.

Broader China AI Chip Sourcing and Data Center Trends

Z.ai's domestic-chip pivot is unfolding within a wider policy environment that is reshaping China's entire AI infrastructure.

U.S. export controls and the Entity List. Z.ai was added to the U.S. Department of Commerce Entity List on January 16, 2025, making the pivot to domestic chips a strategic necessity, not merely a cost-saving measure.

China's domestic chip mandate for state-funded data centers. In November 2025, Reuters reported that Chinese regulators issued directives requiring all state-funded data center projects to use only domestically produced AI chips. Facilities less than 30% complete were instructed to eliminate foreign chips already installed or forgo plans to acquire them.

Massive national investment. Multiple sources report that China is preparing to spend roughly 2 trillion yuan (~$295 billion) over the next five years building AI data centers, with a stated requirement that at least 80% of the chips come from domestic suppliers, primarily Huawei. China's data center and AI computing market reached approximately ¥500 billion in 2025, with AI computing capacity growing nearly 40% year-on-year.

Impact on Z.ai specifically. Z.ai's compute strategy already aligns with these policy directions. The company has demonstrated it can train models entirely on domestic chips (GLM-Image on Huawei Ascend), adapt inference for multiple domestic chip platforms (GLM-5.2 on Huawei Ascend), and is now exploring whether a custom ASIC makes sense at scale. CEO Zhang Peng stated: "We are trying our best to improve our infrastructure and to make the model more efficient on different kinds of chips."

The Bottom Line

Z.ai's confirmed compute strategy is best described as a domestic-chip pivot with a possible custom-ASIC exploration, not a verified custom-chip program. GLM-5.2's supported benchmark results show strong open-weight coding performance that beats GPT-5.5 on SWE-bench Pro (62.1 vs. 58.6), while the company's commercial story is defined by rapid revenue growth, public-market fundraising, and ongoing profitability pressure. The broader policy environment — U.S. export controls, Entity List designation, and China's state-funded data center mandates — makes Z.ai's domestic-chip alignment a strategic necessity as well as a technical achievement.