On July 21, 2026, Hong Kong-listed Zhipu (02513.HK), known as the "world’s first large model stock," saw a long-awaited surge in its share price. By midday, Zhipu closed at HKD 1,116 per share, up 25.32%, with its total market capitalization rebounding to approximately HKD 500 billion. In the previous two trading days, the stock had plunged 28.49% and 19.56%, with a cumulative three-day drop nearing 50%.
The immediate catalyst for this sharp rebound was two major industry developments: Zhipu officially completed its acquisition of XCore Sigma, a domestic AI heterogeneous computing software company, and simultaneously put into operation a 1GW-class domestic AI computing data center, fully equipped with Chinese-made AI chips.
These two moves target the dual dimensions of "computing supply" and "computing enablement," signaling Zhipu’s evolution from a model-centric large model company to a comprehensive competitive system covering models, computing power, infrastructure, and ecosystem. The global competitive landscape for Frontier AI is undergoing profound structural change.
What is XCore Sigma? Why did Zhipu acquire it?
Founded in 2023, XCore Sigma is an AI heterogeneous computing software infrastructure company focused on compiler technology, originating from the Compiler Laboratory of the Institute of Computing Technology, Chinese Academy of Sciences. Founder Dr. Huimin Cui earned her bachelor’s and master’s degrees from Tsinghua University’s Department of Computer Science and her Ph.D. from the Institute of Computing Technology, where she also led the compiler team. The core team has led or participated in compiler development for several domestic chips, including Loongson, Sunway, Cambricon, and Huawei Ascend, possessing end-to-end capabilities from virtual instruction set design to operator generation, translation, and optimization.
In terms of product offerings, XCore Sigma provides a unified software toolchain for various compute chips such as CPUs, GPUs, and NPUs. Its products include large model inference optimization tools, operator auto-generation tools, and a heterogeneous computing software platform. The goal is to deliver standardized AI software foundations across brands and models for China’s AI large model industry. Its SigInfer inference engine claims to significantly reduce inference latency and boost throughput. XCore Sigma’s core advantage lies in its virtual instruction set technology—using middleware to unify different chip ecosystems, assembling fragmented domestic chips into a single, large-scale cluster.
The acquisition was valued at several hundred million RMB, making it one of Zhipu’s largest deals since going public. On the financing side, XCore Sigma completed seed, angel, and pre-A rounds since August 2023, with investors including Oriza Seed, NewShang Capital, CAS Venture Capital, Morning Hill Capital, and BV Baidu Ventures.
Why did the stock crash before? Is this rebound just a technical correction?
To understand the strength of this rebound, we must first clarify the underlying causes of the previous sharp decline.
On January 8, 2026, Zhipu debuted on the Hong Kong Stock Exchange at HKD 116.2 per share, becoming the world’s first large model stock. The share price soared, hitting an all-time intraday high of HKD 2,980 on June 22—a more than 24-fold increase from the IPO price, with market capitalization briefly surpassing HKD 1.33 trillion. However, this rally was largely built on an extremely limited free float—before lock-up expiration, only 11.74 million shares were freely tradable. This scarcity amplified price elasticity but also sowed the seeds for wild volatility.
The turning point came in July. On July 8, Zhipu faced its first major post-IPO lock-up expiration, with about 25.68 million shares held by 11 cornerstone investors becoming tradable, representing roughly 5.76% of total shares. Just a week later, on July 13, Zhipu placed 19.78 million new H-shares at HKD 1,588 each, raising about HKD 31.4 billion. The placement price was about 12.99% below the previous close. This sudden increase in tradable shares created significant short-term pressure on the stock.
Meanwhile, Moonshot AI released the 2.8 trillion-parameter open-source model Kimi K3 and was rumored to be accelerating its Hong Kong IPO, intensifying concerns about fierce competition in the high-end large model space. Under these combined pressures, Zhipu’s share price fell nearly 70% from its peak.
Against this backdrop, the July 21 rebound reflects both a technical correction after an oversold drop and a market reassessment of Zhipu’s strategic positioning. The simultaneous strengthening of computing supply (data centers) and computing enablement (software optimization) provides the foundation for the market to reprice Zhipu’s long-term competitiveness.
How does the acquisition address Zhipu’s engineering bottlenecks?
This acquisition directly targets Zhipu’s previous systemic engineering challenges.
Due to insufficient computing capacity, in January Zhipu had to cut daily new user slots for the GLM Coding Plan to just 20% of normal levels. After the March launch of GLM-5, the Coding Agent’s daily call volume reached hundreds of millions, and some users reported abnormal outputs on complex tasks. The company’s review found engineering challenges in inference architecture and KV Cache management under high concurrency. Recently, after GLM-5.2 launched on aggregation platforms, daily token call volume surged 27-fold in the first week, further exposing bottlenecks in inference infrastructure for high-concurrency, long-context scenarios.
After being added to the U.S. Entity List, Zhipu accelerated domestic substitution and has now completed inference adaptation for eight major Chinese computing platforms, including Huawei Ascend, Pingtouge, and Moore Threads. However, the fragmented domestic chip ecosystem—with significant differences in instruction sets, operator libraries, and programming frameworks—has made model deployment and inference optimization much more difficult.
This is where XCore Sigma’s value lies. With unified compilers, runtime, and inference engines, XCore Sigma enables Zhipu to efficiently migrate and optimize models across different Chinese chips, greatly reducing repetitive adaptation costs. Analysts believe the acquisition aims to shore up the critical "computing enablement" link, using foundational software to significantly improve heterogeneous chip utilization and reduce inference costs while boosting deployment efficiency.
Reportedly, Zhipu completed XCore Sigma’s integration in the first half of this year, further strengthening its capabilities in compilers, runtime, inference engines, and heterogeneous computing software. For GLM-5.2, Zhipu has built a CUDA-equivalent middleware ecosystem based on domestic computing power, now covering the full inference software stack from models, compilers, inference engines, runtime, to heterogeneous computing scheduling.
What does a 1GW data center mean?
Alongside the acquisition, Zhipu has launched a 1GW-class domestic AI computing data center, fully equipped with Chinese AI chips. Given ongoing structural bottlenecks in domestic computing supply, this move brings computing autonomy into Zhipu’s competitive system, rather than relying entirely on external procurement.
A 1GW computing center ranks among the top tier of domestic AI data centers. Reports indicate Zhipu has built or is operating several compute clusters, each with over 10,000 chips. This data center will be one of the largest server hubs ever constructed by a Chinese AI lab.
The data center provides the computational resources needed for large-scale model training, while the AI Infra team uses compilers, runtime, and inference engines to maximize heterogeneous chip utilization. These two elements address "available computing power" and "efficient computing utilization," respectively. This dual "hardware + software" approach is becoming the standard for leading AI companies to build competitive moats.
From Model Competition to System Competition: The Landscape Is Being Redefined
Industry insiders point out that global Frontier AI competition is shifting from single-model performance to system-level competition across models, computing power, infrastructure, and ecosystem. Zhipu’s recent efforts to shore up foundational capabilities and expand its model, agent, MaaS, and ecosystem strategies indicate it is building a complete competitive system as a foundational model company—not just competing on individual model releases.
This view is well supported by industry logic. In the era of large models, advances in model capabilities depend heavily on the scale of computing power and infrastructure efficiency. Without sufficient computing resources, model training and iteration cannot continue; without efficient infrastructure, computing power cannot be transformed into usable model capabilities. Both are indispensable.
From a commercialization perspective, Zhipu’s growth logic is also being validated. As of July 2026, Zhipu’s ARR (annual recurring revenue) has reached $1 billion, all from API and Coding Plan revenue, with no contribution from consumer-facing products. From January to July 2026, Zhipu’s ARR grew 15-fold year-on-year. While it took Anthropic 15 months to grow ARR from $100 million to $1 billion, Zhipu achieved this in just five months.
With large-scale computing, a mature infrastructure system, and long-term post-training capabilities, Zhipu’s next-generation foundation model is expected to advance toward larger parameter counts and higher intelligence, while maintaining inference efficiency and engineering deployability.
Risks and Outlook: Caution Still Needed After the Rebound
Despite the strong rebound on July 21, Zhipu’s valuation still faces multiple uncertainties.
First, the impact of lock-up expirations and share placements on the shareholding structure has yet to be fully absorbed. The first batch of unlocked shares totaled about 25.68 million, plus 19.78 million new H-shares placed, creating significant short-term supply pressure.
Second, competition in the Frontier AI space is intensifying. After Moonshot AI released Kimi K3, the market reassessed the competitive landscape for large models. Multiple Chinese large model companies are actively seeking funding and preparing for IPOs, drawing capital market attention in different directions.
Third, domestic AI chip capacity and performance are still ramping up. Whether a 1GW data center running solely on domestic chips can match mainstream international solutions in training efficiency and inference cost remains to be seen.
From a broader perspective, Zhipu’s dual moves of "acquisition + infrastructure" mark a shift among China’s leading AI companies from model-level competition to a systemic race encompassing computing infrastructure, foundational software stacks, and model capabilities. The outcome of this race will depend not just on standout model launches, but on who can build a complete, autonomous, and sustainable technology ecosystem.
Summary
On July 21, 2026, Zhipu officially completed its acquisition of XCore Sigma, a domestic AI heterogeneous computing software company, in a deal worth several hundred million RMB. At the same time, it launched a 1GW-class AI data center fully equipped with Chinese-made chips. Boosted by this news, Zhipu’s share price soared over 25% that day, with market capitalization rebounding to around HKD 500 billion. The acquisition of XCore Sigma shores up Zhipu’s weaknesses in compilers, runtime, inference engines, and other infrastructure layers, while the 1GW data center addresses structural bottlenecks in computing supply. These two initiatives target "computing supply" and "computing enablement," signaling that Frontier AI competition is shifting from single-model performance to system-level competition across models, computing, infrastructure, and ecosystem. With ARR already at $1 billion and up 15-fold in half a year, Zhipu is building one of the few complete competitive systems among China’s leading AI companies. However, changes in shareholding structure, intensifying industry competition, and the maturity of the domestic computing ecosystem remain variables to watch.
FAQ
Q1: What was the transaction value of Zhipu’s acquisition of XCore Sigma?
The acquisition was valued at several hundred million RMB, making it one of Zhipu’s largest deals since going public. Founded in 2023, XCore Sigma had previously completed seed, angel, and pre-A rounds, with investors including Oriza Seed, CAS Venture Capital, BV Baidu Ventures, and others.
Q2: What are XCore Sigma’s technical advantages?
XCore Sigma’s technology originates from the Compiler Laboratory of the Chinese Academy of Sciences. Its core team has contributed to compiler development for domestic chips such as Loongson, Sunway, Cambricon, and Huawei Ascend. The company provides a unified software toolchain for CPUs, GPUs, NPUs, and other compute chips. Through virtual instruction set technology, it integrates chips of different brands and models into collaborative clusters, significantly improving computing utilization.
Q3: What does a 1GW data center mean for Zhipu?
A 1GW data center, fully equipped with domestic AI chips, signals that Zhipu is bringing computing autonomy into its competitive system amid ongoing structural bottlenecks in domestic supply. This data center will be among the largest server hubs built by Chinese AI labs, providing massive computational resources for Zhipu’s large model training and iteration.
Q4: What is Zhipu’s current ARR?
As of July 2026, Zhipu’s ARR (annual recurring revenue) reached $1 billion, all from API and Coding Plan revenue. From January to July 2026, ARR grew 15-fold year-on-year. While it took Anthropic 15 months to grow ARR from $100 million to $1 billion, Zhipu did it in just five months.
Q5: How is the competitive landscape for Frontier AI changing?
Global competition among Frontier AI companies is evolving from single-model performance to system-level competition across models, computing, infrastructure, and ecosystem. Leading AI companies must excel in model R&D, have robust computing infrastructure, maintain efficient foundational software stacks, and possess sustainable commercialization capabilities to gain an edge in the next phase of competition.
