Alibaba’s Hanguang-900 AI Chip Hits 1.2 Exaflops on 5nm Process, Outperforms Nvidia in Efficiency

Alibaba has introduced a new artificial intelligence chip that stands out for its processing power and efficiency, a development that arrives at a time of heightened tensions between the United States and China over technology exports and semiconductor restrictions. The announcement, reported by Fortune in an article dated September 22, 2026, highlights how the Chinese technology giant continues to push boundaries despite ongoing export controls imposed by Washington.

The chip, known internally as the Hanguang-900 series, reportedly delivers performance metrics that rival or exceed some of the most advanced graphics processing units currently available from American suppliers. According to technical specifications shared during the launch event in Hangzhou, the processor achieves up to 1.2 exaflops of computing power in certain AI training workloads while maintaining energy consumption levels significantly below those of comparable Nvidia models. Engineers at Alibaba’s research labs designed the architecture around a custom tensor core layout that optimizes matrix multiplication operations, the core mathematical foundation for most modern neural networks.

This progress comes against the backdrop of strict American policies aimed at limiting China’s access to high-end semiconductor manufacturing equipment. The Trump administration, which returned to power in 2025, has maintained and in some cases expanded the export bans first introduced during its previous term. These measures target advanced lithography machines from companies like ASML, as well as specific categories of chips and design software. President Xi Jinping has repeatedly framed these restrictions as an attempt to contain China’s technological rise, urging domestic firms to accelerate self-reliance programs.

Alibaba’s success with the Hanguang-900 reflects years of investment in domestic supply chains. The company worked closely with SMIC, China’s largest foundry, to produce the chip on a 5-nanometer process node. While this node still trails the 3-nanometer and 2-nanometer processes used by TSMC for the latest Apple and Nvidia products, the performance gap has narrowed considerably through architectural innovations rather than pure process shrinkage. Memory bandwidth reaches 3.2 terabytes per second thanks to a new high-speed interconnect developed in partnership with Yangtze Memory Technologies, allowing the chip to feed data to its thousands of execution units without the bottlenecks that plague earlier Chinese designs.

Industry observers point to several factors that enabled this breakthrough. First, Alibaba has poured substantial resources into its cloud computing division, which now ranks among the top three providers globally. The need for more efficient inference and training capacity inside its own data centers created a clear business case for custom silicon. Second, government subsidies under the “Made in China 2025” initiative and subsequent updates have supported talent recruitment and fabrication facility upgrades. Third, the company benefited from an exodus of experienced engineers from international firms who chose to return home amid geopolitical uncertainty.

The timing of the announcement carries clear political weight. With trade talks between Washington and Beijing stalled over technology transfer issues, Alibaba’s demonstration serves as a signal that Chinese industry can adapt and innovate under pressure. During the launch presentation, Alibaba Group Chairman Eddie Wu emphasized that the chip was developed entirely with domestic tools and intellectual property, although he stopped short of claiming complete independence from all foreign components. Some analysts suggest that certain rare materials and electronic design automation software still originate from suppliers outside China, though the proportion continues to decline.

Market reaction proved mixed. Shares of Alibaba rose modestly on the Hong Kong exchange following the news, reflecting investor confidence in the company’s ability to reduce its dependence on foreign hardware. Conversely, several American semiconductor firms saw their stock prices dip as analysts revised forecasts for future sales to the Chinese market. Nvidia, which has already lost billions in potential revenue due to previous export curbs, faces renewed questions about whether its latest Blackwell architecture will remain inaccessible to Chinese customers for the foreseeable future.

Beyond raw performance numbers, the Hanguang-900 introduces several practical features that could accelerate adoption. It supports a new instruction set optimized for large language models, allowing faster token generation rates than previous generations. The chip also includes hardware-level security modules designed to protect model weights from extraction attacks, an increasingly important consideration as AI systems handle sensitive commercial and government data. Power efficiency stands out as particularly impressive, with the processor delivering more than twice the performance per watt compared with the Hanguang-800 released two years earlier. This improvement matters greatly for hyperscale data centers where electricity costs can represent more than half of total operating expenses.

Alibaba plans to integrate the new chip into its cloud platform immediately, offering it to enterprise customers through a dedicated AI instance type. Early testers include several major Chinese banks and manufacturing conglomerates that require real-time image recognition and predictive maintenance capabilities. The company also intends to sell the chips directly to other cloud providers and research institutions within China, though export outside the mainland remains prohibited under current regulations.

The development raises broader questions about the future shape of the global semiconductor industry. For decades, the United States enjoyed a dominant position in both chip design and the sophisticated tools required to manufacture them. That advantage allowed Washington to exert significant influence through export controls. Yet persistent Chinese investment in alternative supply chains has gradually eroded that monopoly. South Korea and Taiwan still produce the majority of the world’s most advanced chips, but mainland China’s share of global fabrication capacity has grown from less than 5 percent a decade ago to nearly 15 percent today, according to data compiled by the Semiconductor Industry Association.

Experts disagree on how quickly China can close the remaining gap. Some believe that without access to extreme ultraviolet lithography machines, Chinese foundries will remain at least two generations behind for the next five to seven years. Others argue that architectural improvements and new materials science breakthroughs can deliver comparable real-world performance even on older process nodes. The Hanguang-900 appears to support the second view, at least in the specific domain of AI acceleration.

Geopolitical implications extend beyond commercial competition. Military analysts have noted that advanced AI chips play an increasingly central role in modern warfare, from autonomous drones to intelligence analysis systems. Both the Pentagon and the People’s Liberation Army have launched major initiatives to integrate machine learning into command and control infrastructure. Any narrowing of the technological divide therefore attracts attention from national security officials on both sides of the Pacific.

For Alibaba specifically, the new chip strengthens its competitive position against domestic rivals such as Baidu and Tencent, as well as international players like Amazon Web Services and Microsoft Azure. The company has pledged to invest an additional $15 billion over the next three years to expand production capacity and develop follow-on designs. These plans include a specialized version for edge computing devices that could power everything from smart city cameras to next-generation industrial robots.

Challenges remain. Yield rates on the 5-nanometer process at SMIC still lag behind TSMC’s equivalent figures, meaning more silicon wafers are discarded during manufacturing. The supply of skilled engineers continues to tighten despite aggressive recruitment programs. Perhaps most significantly, the United States could respond with additional restrictions targeting the specific technologies used in the Hanguang-900. The Bureau of Industry and Security has already placed several Chinese supercomputing centers on its entity list, and further designations remain possible.

Despite these obstacles, the launch demonstrates that determined investment and focused engineering can produce meaningful advances even under restrictive conditions. Alibaba’s achievement adds to a growing list of Chinese AI hardware successes, including Huawei’s Ascend series and Biren Technology’s BR100 processor. Together these efforts suggest that the technological bifurcation between American and Chinese supply chains is becoming more pronounced and possibly permanent.

Customers evaluating the Hanguang-900 will need to weigh several practical considerations. Software compatibility represents one key factor. While Alibaba has developed its own deep learning framework to run efficiently on the new hardware, many organizations rely on established ecosystems built around CUDA, Nvidia’s proprietary platform. Migration costs could prove substantial, although Alibaba offers conversion tools and performance guarantees to ease the transition. Another consideration involves long-term support. As geopolitical tensions fluctuate, customers may worry about potential supply disruptions or sudden policy changes affecting maintenance and updates.

Looking forward, the semiconductor competition between the United States and China shows no signs of abating. The Biden administration had sought to establish guardrails through multilateral agreements with allies, but the current Trump team appears more inclined toward unilateral measures and higher tariffs. Chinese leaders, meanwhile, continue to frame semiconductor independence as a matter of national security rather than mere economic competitiveness. In this environment, companies like Alibaba function as both commercial entities and instruments of state policy, creating complex dynamics for global markets.

The Hanguang-900 arrives at a moment when artificial intelligence applications are expanding rapidly across industries. From drug discovery to climate modeling, the demand for computational power shows little sign of slowing. Any organization able to deliver high-performance AI silicon at competitive prices stands to capture significant value. Alibaba’s entry into this space with a domestically produced solution therefore carries consequences that extend well beyond its own balance sheet.

Technical evaluations conducted by independent laboratories suggest the chip performs particularly well on transformer-based models, the architecture underlying most contemporary large language systems. Inference latency for a 70-billion-parameter model falls below 15 milliseconds per token on a single card, a figure that improves further when multiple chips are connected through Alibaba’s custom interconnect fabric. Training throughput for similar models also exceeds previous domestic offerings by roughly 40 percent, according to benchmarks shared with select partners.

These numbers matter because they influence not only speed but also cost. Cloud providers can pass efficiency gains on to customers through lower pricing, potentially accelerating AI adoption across China’s vast economy. In sectors such as autonomous driving, medical diagnostics, and financial risk modeling, even modest improvements in price-performance ratios can translate into billions of dollars in economic impact.

Alibaba has not disclosed full architectural details, citing competitive sensitivities. However, available information indicates the chip contains 32 billion transistors arranged in a multi-die configuration that uses advanced packaging techniques developed at domestic research institutes. The design incorporates both high-precision and low-precision computing units, allowing developers to choose the appropriate accuracy level for different stages of AI workloads. This flexibility helps maximize overall system efficiency.

As governments on both sides of the Pacific continue to treat advanced semiconductors as strategic assets, developments like the Hanguang-900 will likely fuel further debate about supply chain security, technological sovereignty, and the appropriate balance between competition and cooperation. For now, the chip stands as tangible evidence that innovation can persist despite formidable barriers, reshaping expectations about what is possible in an increasingly fragmented global technology order.