Why did Zhipu surge? It acquired a domestic AI heterogeneous computing software company and rolled out a 1GW data center

July 21, 2026, Zhipu (02513.HK), the “first global large model stock” in Hong Kong equities, saw an overdue surge. By the lunch close, Zhipu closed at HK$1,116 per share, up 25.32%, with its total market cap recovering to about HK$500 billion. In the prior two trading days, the stock plunged by 28.49% and 19.56% respectively, with its three-day cumulative drop nearing 50%.

The direct catalyst for this strong rebound came from two major industry-level developments: Zhipu officially completed its acquisition of domestic AI heterogeneous compute software company XCore Sigma on the same day; at the same time, Zhipu has rolled out a 1GW-level domestic AI compute data center construction project, with all components using domestic AI chips.

These two moves point to the two key dimensions of “compute supply” and “compute release,” indicating that Zhipu is evolving from a large-model company known for model capabilities into a complete competitive system spanning models, compute, Infra, and ecosystem. The competitive landscape among global Frontier AI players is undergoing profound structural changes.

What is XCore Sigma? Why did Zhipu acquire it?

Founded in 2023, XCore Sigma is an AI heterogeneous compute software infrastructure company focused on compilation technology. Its technical roots trace back to the compilation laboratory of the Institute of Computing Technology, Chinese Academy of Sciences (CAS). Founder Dr. Cui Huimin earned her bachelor’s and master’s degrees from Tsinghua University’s computer department, her PhD from the Institute of Computing Technology, CAS, and previously served as the head of the CAS compilation team. The core team has led or participated in the development of compilers for multiple domestic chip platforms, including Loongson, Sugon, Cambricon, and Huawei Ascend, possessing end-to-end capabilities from virtual instruction set design to operator generation, translation, and optimization.

In terms of product layout, XCore Sigma provides a unified software toolchain for different computing chip types such as CPU, GPU, and NPU. Its product line includes large-model inference optimization tools, automatic operator generation tools, and heterogeneous compute software platforms. The goal is to provide a standardized AI software foundation for China’s domestic AI large-model industry, enabling cross-brand and cross-model standardization. Previously, its SigInfer inference engine claimed it could significantly reduce inference latency and improve throughput efficiency. The core advantage of XCore Sigma lies in virtual instruction set technology—by using an intermediate-layer software stack to unify chips from different brands and models, it effectively turns scattered domestic chips into a unified, ultra-large-scale cluster.

The acquisition deal is worth several hundred million yuan RMB, a rare large M&A move since Zhipu’s listing. On financing, XCore Sigma completed seed, angel, and Pre-A rounds starting in August 2023. Investors include Yuanzhi Capital, Newshine Capital, CAS Venture, Chenshan Capital, BV Baidu Venture, among others.

Why did the stock plunge consecutively before? Was the rebound just technical repair?

To understand the strength of this rebound, we first need to clarify the deeper reasons behind the prior plunge.

On January 8, 2026, Zhipu listed on the Hong Kong Stock Exchange with an issuance price of HK$116.2 per share, becoming the “first global large model stock.” After that, the stock surged steadily, reaching an intraday record high of HK$2,980 on June 22, more than 24 times above the issuance price, with total market cap briefly exceeding HK$1.33 trillion. However, this rally was largely built on a very low free float—before the lock-up period, Zhipu’s market free-floating shares were only 11.74 million shares. Scarcity amplified price volatility, and also planted the risk of severe swings.

The turning point came in July. On July 8, Zhipu saw its first large-scale release of locked-up shares after listing. About 25.68 million shares held by 11 cornerstone investors were released, accounting for about 5.76% of the company’s total share capital. Only a week after the release, on July 13 Zhipu completed the placement of 19.78 million new H shares at HK$1,588 per share, raising about HK$31.4 billion. The placement price was discounted by about 12.99% versus the prior closing price. In the short term, the concentrated release of supply increased pressure on the share price.

Meanwhile, Mian Zhi An Mian (Moonshot) released an open-source model Kimi K3 with 2.8 trillion parameters, and news also surfaced that it would speed up preparations for an IPO in Hong Kong, intensifying market concerns about the intensifying competition in the high-end large-model track. With multiple factors compounding, Zhipu’s share price saw its maximum drawdown from the historical high reach about 70%.

Against this backdrop, the rebound on July 21 included not only a technical rebound after overselling, but also reflected the market’s reassessment of Zhipu’s industry deployment logic. The simultaneous strengthening of two capabilities—compute supply (data centers) and compute release (software optimization)—forms the logical foundation for the market to reprioritize Zhipu’s long-term competitiveness.

How does the acquisition address Zhipu’s engineering weak spots?

This acquisition directly targets the systemic engineering bottlenecks Zhipu faced previously.

Because compute expansion could not keep up with demand, in January this year Zhipu reduced the daily new user purchase quota for the GLM Coding Plan to 20% of normal levels. After the GLM-5 release in March, Coding Agent reached hundreds of millions of daily calls, and some users reported abnormal outputs in complex tasks. In its post-incident review, the company found that under high-concurrency scenarios, there were engineering challenges in the inference architecture and KV Cache management. Recently, in the first week after the GLM-5.2 large model launched on the aggregation platform, daily token calls surged 27 times on average. The bottlenecks in the inference infrastructure under high concurrency and long-context scenarios became even more evident.

After being added to the U.S. entity list, Zhipu proactively promoted domestic substitution and has already completed inference adaptation for eight major domestic compute platforms, including Huawei Ascend, Pingtouge, and Moore Threads. However, the fragmentation of the domestic chip ecosystem—significant differences among brands and models in instruction sets, operator libraries, and programming frameworks—has increased the difficulty of model deployment and inference optimization substantially.

This is where the value of XCore Sigma lies. By unifying the compiler, Runtime, and inference engine, XCore Sigma enables Zhipu to achieve efficient model migration and inference optimization across different domestic chips, greatly reducing redundant adaptation costs. Analysts believe the acquisition aims to fill the key “compute release” link by substantially improving heterogeneous chip utilization through core software capabilities, effectively lowering inference costs and improving model deployment efficiency.

It is understood that in the first half of this year, Zhipu completed the integration of XCore Sigma and further filled capabilities in compilers, runtimes, inference engines, and the heterogeneous compute software foundation. On GLM-5.2, Zhipu has already built a CUDA-comparable intermediate-layer ecosystem based on domestic compute, starting to cover the full inference software stack from model—compiler—inference engine—runtime—heterogeneous compute scheduling.

What does a 1GW data center mean?

Rolling out alongside the acquisition is a 1GW-level domestic AI compute data center, using only domestic AI chips. Against the backdrop that structural bottlenecks still exist in domestic compute supply, this means Zhipu is bringing compute autonomy into its own competitive framework rather than relying entirely on external procurement.

A 1GW compute scale places it in the top tier among current domestic AI data centers. Reports say Zhipu has built or is operating multiple compute clusters, with each cluster equipped with more than 10k chips. This data center is set to become one of the largest server hubs built by China’s AI laboratories.

Data center construction provides the compute resources required for large-scale model training. Meanwhile, the AI Infra team improves heterogeneous chip utilization through foundational software capabilities such as compilers, Runtime, and inference engines—together addressing the two layers of the problem: “compute is available” and “compute can be used well.” This “hardware + software” dual-wheel drive model is becoming the standard path for leading AI companies to build competitive moats.

From model competition to system competition: the landscape is being reshaped

Industry insiders point out that competition among global Frontier AI companies is gradually evolving from a single-model capability race into system competition across models, compute, Infra, and ecosystem capabilities. Zhipu has recently been continuously filling gaps in underlying capabilities and keeps advancing its layout for models, agents, MaaS, and the industrial ecosystem. This suggests it is building a complete competitive system of a foundational model company, not just competing around a one-off model release.

This assessment is supported by strong industry logic. In the era of large models, improvements in model capabilities depend heavily on compute scale and infrastructure efficiency. Without sufficient compute supply, model training and iteration cannot be sustained; without efficient Infra, compute cannot be converted 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) reached $1 billion, all from API and Coding Plan revenue, excluding any C-end product revenue. Only from January to July 2026, Zhipu’s ARR grew year over year by as much as 15 times. From $100 million to $1 billion, Anthropic took 15 months; Zhipu did it in just 5 months.

Supported jointly by large-scale compute, a mature Infra system, and long-term post-training capabilities, Zhipu’s next-generation foundational model is expected to continue evolving toward larger parameter scale and higher intelligence, while balancing inference efficiency and engineering deployability.

Risks and outlook: still need caution after the rebound

Although the rebound on July 21 was impressive, pricing of Zhipu by the market still faces multiple uncertainties.

First, the adjustment in the equity supply structure caused by the release of lock-up shares and placements has not been fully absorbed. The initial release size is about 25.68 million shares, and combined with the placement of 19.78 million new H shares, the market’s ability to absorb the increased supply in the short term should not be underestimated.

Second, competition in the Frontier AI track is intensifying. After Mian Zhi An Mian released Kimi K3, the market re-priced the competitive landscape in the large-model track. Multiple domestic large-model companies are actively pushing forward financing and listing processes, and attention in the capital market is being dispersed.

Third, the capacity and performance of domestic AI chips are still ramping up. Whether the full domestic chip solution for a 1GW data center can form effective competition—on training efficiency and inference costs—with mainstream international solutions still needs time to be verified.

From a more macro perspective, Zhipu’s “acquisition + infrastructure” dual move signals that leading Chinese AI companies are upgrading their competition from the model layer to a system-level contest encompassing compute infrastructure, the underlying software stack, and model capabilities. The outcome of this contest will depend not only on how astonishing a single model release is, but also on who can build a complete, independent, and sustainable technical ecosystem.

Summary

Zhipu officially completed its acquisition of the domestic AI heterogeneous compute software company XCore Sigma on July 21, 2026, with the deal value in the range of several hundred million yuan RMB. At the same time, it rolled out a 1GW-level AI compute data center that uses all domestic chips. Boosted by this news, Zhipu’s stock price surged more than 25% on the day, and total market cap recovered to about HK$500 billion. The acquisition of XCore Sigma fills Zhipu’s Infra-layer shortcomings in compilers, Runtime, and inference engines, while the 1GW data center addresses the structural bottlenecks in compute supply. Together, these two moves target “compute supply” and “compute release,” marking Frontier AI competition evolving from a single-model capability showdown to system competition spanning models, compute, Infra, and ecosystem. In the commercialization context where ARR has reached $1 billion and growth of 15x in half a year, Zhipu is building one of the few complete competitive systems among China’s leading AI firms. However, variables such as changes in the supply structure of shares, intensifying industry competition, and the maturity of the domestic compute ecosystem still need continuous monitoring.

FAQ

Q1: What is the transaction amount for Zhipu’s acquisition of XCore Sigma?

The deal value for this acquisition is several hundred million yuan RMB, which is a rare large M&A move since Zhipu listed. XCore Sigma was founded in 2023. It previously completed seed, angel, and Pre-A rounds, with investors including Yuanzhi Capital, CAS Venture, and BV Baidu Venture, among others.

Q2: What are XCore Sigma’s technical advantages?

XCore Sigma’s technical roots come from the compilation laboratory of the Institute of Computing Technology, CAS. The core team has participated in compiler R&D for multiple domestic chip platforms, including Loongson, Sugon, Cambricon, and Huawei Ascend. The company provides a unified software toolchain for different computing chips such as CPU, GPU, and NPU. Using virtual instruction set technology, it unifies domestic chips from different brands and models into a cluster that can work together, greatly improving compute utilization.

Q3: What does a 1GW-level data center mean for Zhipu?

A 1GW-level data center uses all domestic AI chips. Against the backdrop that structural bottlenecks still exist in domestic compute supply, it indicates that Zhipu is bringing compute autonomy into its own competitive system. The data center is expected to become one of the largest server hubs built by China’s AI laboratories, providing large-scale compute resources for Zhipu’s large-model training and iteration.

Q4: What level has Zhipu’s ARR reached currently?

As of July 2026, Zhipu’s ARR (annual recurring revenue) has reached $1 billion, all coming from API and Coding Plan revenue. From January to July 2026, Zhipu’s ARR grew year over year by as much as 15 times. From $100 million to $1 billion, Anthropic took 15 months, while Zhipu took only 5 months.

Q5: What changes are happening in the competitive landscape of Frontier AI?

Competition among global Frontier AI companies is gradually evolving from a race focused on single-model capabilities into system competition across models, compute, Infra, and ecosystem capabilities. Leading AI companies need to have both strong model research and development capabilities and ample compute infrastructure, along with an efficient underlying software stack and sustainable commercialization capabilities, to gain an advantage in the next stage of competition.