国产模型供应商算力保供新解:智谱 1GW 国产芯片数据中心+中科加禾收购,算力交易中心扮演什么角色?

As a powerful computing infrastructure, the 1GW peak power supply capacity can meet the daily electricity demand of about 750,000 ordinary households simultaneously, delivering super-large-scale and stable independent computing support.


1GW Domestic Chip Data Center & CAS Infra Acquisition: Zhipu’s Grand Computing Layout, With Trading Platform as the Final Puzzle

On July 21, 2026, Zhipu AI completed two strategic layouts that reshaped the domestic AI computing competition landscape.

First, it officially launched a 1GW-level domestic AI computing data center fully equipped with self-developed domestic AI chips. As a powerful computing infrastructure, the 1GW peak power supply capacity can meet the daily electricity demand of about 750,000 ordinary households simultaneously, delivering super-large-scale and stable independent computing support.

Second, Zhipu AI fully acquired Zhongke Jiahe, a core technical team spun off from the Compiler Laboratory of the Institute of Computing Technology, Chinese Academy of Sciences. Its core members have led and participated in the compiler R&D of almost all mainstream domestic chips, including Loongson, Sunway, Cambricon and Huawei Ascend, holding top-tier domestic underlying Infra technical capabilities.

Driven by the two major positive news, Zhipu’s stock price surged by more than 35% intraday and closed up over 36%, with its total market value rebounding to HK$567.6 billion. The capital market has reached a consistent judgment with real money: Zhipu has completely completed the closed-loop layout of computing supply capacity and computing efficiency release capacity.

Self-Built Scale vs. Efficient Utilization: Fixing Core Engineering Shortcomings

After the launch of GLM-5, Zhipu’s long-standing engineering bottlenecks in computing deployment were fully exposed.

In January 2026, due to the mismatch between computing expansion speed and explosive business growth, Zhipu compressed the daily user purchase quota of the GLM Coding Plan to only 20% of the normal level. After the official release of GLM-5 in March, the daily calling volume of Coding Agent surged to the hundreds of millions level, and many users reported abnormal model output in complex task scenarios. After in-depth review, the core problems were locked: the inference architecture and KV Cache management could not adapt to ultra-high concurrency scenarios, resulting in insufficient computing efficiency release.

All these problems point to one core crux: computing resources are sufficient, but far from being fully utilized.

Zhipu’s 1GW full-domestic-chip data center solves the scale problem. Against the backdrop of structural shortages in domestic high-quality computing supply, this layout enables Zhipu to take the initiative in computing supply, getting rid of excessive reliance on external procurement and building independent core resource barriers.

The acquisition of Zhongke Jiahe solves the efficiency problem. Its underlying technical capabilities including self-developed compilers, Runtime frameworks and inference engines can significantly improve the utilization rate of heterogeneous domestic chips, reduce model inference costs, and accelerate large-scale model deployment efficiency. To put it simply, self-built data centers realize "having sufficient computing resources", while CAS Infra technology realizes "maximizing the value of computing resources". The two complement each other and are indispensable.

Zhongke Jiahe boasts irreplaceable scarcity in the domestic AI Infra track. Founded in 2023, it inherits the technical heritage of the Compiler Laboratory of the Chinese Academy of Sciences. Its product matrix covers large model inference optimization tools, automatic operator generation tools, and heterogeneous computing software platforms. Different from general AI infrastructure vendors, it focuses exclusively on compilation adaptation and efficiency optimization for heterogeneous domestic chips.

Through this acquisition, Zhipu has acquired the core capability to unify the underlying scheduling logic of diverse domestic chips, enabling different types of domestic hardware to "speak the same language" and achieve collaborative efficient operation.

Computing Trading Center: The Last Mile From Self-Built Closed Loop to Market Circulation

The 1GW super-large computing base and CAS Infra technical capabilities have completely solved Zhipu’s own computing supply and efficiency pain points. However, the prosperity of the entire domestic AI industry cannot rely solely on the self-built closed loop of a single enterprise. What the industry urgently needs is the large-scale circulation and universal sharing of domestic computing resources.

As the core carrier of industry resource allocation, computing trading centers are fundamentally transforming static idle computing resources into tradable, profitable dynamic assets.

According to public data, domestic national computing platforms have realized linkage with 15 provincial and municipal sub-platforms, accessing more than 130 computing service providers and over 200 basic and vertical large models. The JingSuan Token Factory has built seven core capabilities including PD-separated intelligent scheduling architecture, full-stack inference acceleration framework and full ecological adaptation of domestic large models. The Southeast Token Operation Center serves as a core token service hub covering southern Zhejiang, northern Fujian and eastern Jiangxi.

The core value of these regional platforms is to integrate fragmented and scattered domestic computing resources into standardized, low-threshold AI public service capabilities. For small and medium-sized model developers and suppliers with insufficient capital strength, there is no need to invest heavily in building self-owned GW-level data centers. They can purchase high-quality domestic computing resources on demand through trading platforms to support business iteration.

On this basis, Unexhub has further upgraded the industrial model, moving the computing token trade from the traditional fixed-price "supermarket sales model" to a fully market-oriented "open trading market model", bringing inclusive benefits to all industry participants through three core mechanisms:

Real-Time Full Settlement. It completely breaks the industry’s long-standing pain points of long account periods and slow capital recovery. The professional financial clearing system enables all computing and model suppliers to recover funds in real time based on actual resource supply, completely zeroing out corporate cash flow pressure and converting idle hardware resources into sustainable stable cash flow.

Early Zero-Commission Benefit. Exclusive zero-platform-commission privileges for early settled suppliers drastically reduce the industry entry threshold. SMEs and individual developers can enter the market with zero additional cost and monetize idle computing resources efficiently.

Dynamic API Bidding Mechanism. Abolishing unilateral platform fixed pricing, API calling prices float intelligently with real-time market supply and demand. Individual developers and enterprises can enjoy ultra-low calling prices during off-peak idle periods to cut R&D costs; suppliers can gain higher profit margins during peak demand to maximize resource value, realizing a win-win for both supply and demand.

The greatest value of the Unexhub model lies in inclusive empowerment. It does not require participants to have the billions of capital strength to build self-built data centers like Zhipu, nor to own top CAS-level Infra technical teams. As long as there are idle computing resources or high-quality model capabilities, participants can join the market circulation and obtain stable profits.

From Self-Built Closed Loop to Open Market Circulation: The Reshaping of Industrial Rules

Zhipu’s development path is a typical giant closed-loop layout: building exclusive super computing bases and acquiring top underlying technical teams to realize self-sufficient supply and efficient utilization of internal computing resources, which is a competitive strategy for leading enterprises to build moats.

The computing trading center represented by Unexhub takes an open circulation path: breaking resource barriers, gathering global multi-level AI resources, and building a fair trading market where all suppliers and demanders can participate freely. Resources are priced by market supply and demand, and funds are settled in real time.

Although the two paths are different, they point to the same industrial evolution direction: AI computing is accelerating its transformation from a scarce resource monopolized by giants to a universal, tradable standardized commodity.

Zhipu’s 1GW data center proves the feasibility of large-scale construction of domestic chip computing infrastructure; the acquisition of Zhongke Jiahe proves the feasibility of high-efficiency utilization of heterogeneous domestic chips; and Unexhub undertakes the third key proof: domestic computing can achieve large-scale efficient circulation and universal commercialization.

The competition logic of the AI industry has undergone a fundamental change. The era of "winning by occupying scarce resources" is over, and the era of "winning by efficient resource circulation" has arrived. Leading enterprises represented by Zhipu are consolidating their own advantages, while open trading platforms represented by Unexhub are creating fair competition opportunities for all small and medium-sized practitioners in the industry.

Unexhub Core Advantages & Exclusive Empowerment Benefits

Unexhub comprehensively reshapes the global token distribution system and builds an integrated token aggregation, trading and distribution center. The platform gathers full-spectrum global AI suppliers, covering top-tier large models such as Gemini, ChatGPT, Claude, DeepSeek and GLM, as well as high-quality niche models from SMEs and individual developers.

Relying on millisecond-level intelligent routing technology, massive AI resources participate in open market bidding, matching user demands accurately and delivering ultra-stable and ultra-low-cost calling services.

Platform Exclusive Benefits: Developers pay strictly based on actual token consumption, with the lowest discount as low as 30% of the official market price. Suppliers can access the trading channel with one click and enjoy real-time settlement of calling revenue, realizing zero-threshold monetization of idle AI and computing resources.


1GW国产芯片数据中心+中科院Infra收购:智谱算力大棋,算力交易中心成终极拼图

2026年7月21日,智谱AI接连落地两大重磅战略动作,一举补齐国产AI算力赛道的核心竞争力,重塑行业竞争格局。

第一,落地1GW级全国产AI算力数据中心,全域搭载国产化AI芯片,彻底摆脱海外硬件依赖。1GW超级算力的供电规模十分可观,单园区峰值供电能力,可同时满足75万户普通家庭的日常用电需求,为大模型训练与推理提供超大规模、高稳定的自主算力支撑。

第二,正式完成对中科加禾的全资收购。这家硬核技术企业脱胎于中科院计算所编译实验室,核心团队深度参与过龙芯、神威、寒武纪、华为昇腾等全系主流国产芯片的编译器核心研发,手握国内顶尖的底层算力基础设施技术。

两大重磅落地消息直接引爆资本市场,智谱股价盘中暴涨超35%,收盘涨幅超36%,总市值回升至5676亿港元。资本用真金白银给出了最明确的判断:智谱一举补齐了AI产业最核心的两大能力——算力规模化供给能力、算力高效率释放能力。

自建规模&优化效率:彻底根治算力工程短板

自GLM-5发布以来,智谱长期存在的算力工程短板被彻底放大,成为制约业务爆发的核心瓶颈。

今年1月,算力扩容速度跟不上用户与业务增长速度,智谱被迫将GLM Coding Plan的每日用户购买额度,压缩至正常水平的20%。3月GLM-5正式上线后,智能编码代理Coding Agent日调用量飙升至数亿级,大量用户反馈复杂任务场景下模型输出异常。团队复盘后确认,核心症结在于高并发场景下,推理架构与KV Cache缓存管理存在明显工程缺陷,算力资源无法高效释放。

所有问题最终指向同一个核心:算力资源不再稀缺,但算力利用效率严重不足。

智谱1GW全国产芯片数据中心,精准解决「算力规模」问题。在国产优质算力结构性紧缺的行业背景下,这套自研自建的超级算力基地,让智谱彻底掌握算力供给自主权,摆脱对外采购的资源约束与价格波动,筑牢企业核心竞争壁垒。

而收购中科加禾,精准解决「算力效率」问题。其自研的编译器、运行时框架、推理引擎等底层工具链,能够全方位提升异构国产芯片的资源利用率,大幅降低模型推理成本,极速迭代大模型部署效率。业内共识是:自建数据中心解决「有没有算力」的规模难题,底层Infra技术优化解决「好不好用算力」的效率难题,二者相辅相成,缺一不可。

中科加禾的行业稀缺性无可替代。公司2023年成立,核心技术完全承袭中科院计算所编译实验室底蕴,产品覆盖大模型推理优化工具、算子自动生成系统、异构算力统一调度平台。区别于通用AI基础设施厂商,它是国内极少数深耕国产异构芯片编译适配、效率优化的专精技术团队。

这场收购,让智谱手握一项核心稀缺能力:让架构、代际、性能各异的国产芯片,实现底层语言统一、调度协同统一,彻底打破异构芯片的适配壁垒。

算力交易中心:打通国产算力从自建闭环到全域流通的最后一公里

1GW超级算力基地+中科院顶级Infra技术,只解决了智谱单一企业的算力供给与效率问题。但整个国产AI产业的崛起,不能依赖头部企业的单点自建闭环,更需要全行业算力资源的全域盘活、自由流通、普惠共享。

而算力交易中心的核心价值,就是重构行业资源配置逻辑,把海量碎片化、闲置化的算力「死资源」,转化为可交易、可变现、可复用的产业「活资产」。

公开数据显示,国内算力生态已实现15个省市分平台联动组网,接入130余家算力服务商、200余款基础大模型及垂直行业模型。京算Token工厂搭建起PD分离智能调度、全栈推理加速、异构算力统一调度、国产模型全域适配等七大核心能力;东南Token运营中心扎根浙南、闽北、赣东,成为区域核心算力与词元服务枢纽。

各类区域算力平台的核心价值高度统一:整合全网零散算力资源,封装为标准化、低门槛的AI服务能力。广大中小模型厂商、开发者无需重资产投入自建算力集群,可按需采购、灵活复用优质国产算力,极大降低创业与研发门槛。

其中,Unexhub实现了行业模式的颠覆性升级,将传统词元「固定标价、被动售卖」的超市模式,迭代为「供需定价、自由竞价」的市场化交易模式,依托三大核心机制,全方位赋能供需两端,普惠全行业中小从业者:

全程即时清算结算 彻底破除传统算力行业账期冗长、回款滞后、现金流承压的痛点。平台搭载专业金融清算体系,供应商可根据实时算力供给量全额回款、即时提现,清零企业财务压力,让闲置GPU、模型资源稳定转化为经营性现金流。

早期入驻零佣金特权 面向首批入驻的算力、模型供应商开放永久零平台佣金权益,零成本入驻、零门槛变现,大幅拉低行业入局门槛,让中小团队无需承担平台服务费即可参与市场交易。

API动态智能竞价 告别平台单向固定定价模式,API调用价格随市场供需实时浮动、自主调节。开发者可在业务低谷闲时薅取极致低价算力,压缩研发成本;供应商可在需求高峰时段抬升收益,最大化盘活闲置资源,实现供需双向共赢。

这套机制的核心普惠价值在于:无需具备智谱级别的百亿融资能力、无需搭建中科院级别的底层技术团队,只要拥有闲置算力或优质AI模型资源,即可入驻平台参与全域流通、高效变现。

从巨头自建闭环,到全行业普惠流通

智谱的发展路径,是头部巨头的闭环深耕:自建超级算力基地、收购顶级技术团队,实现内部算力自给自足、效率最优,打造专属竞争护城河。

以Unexhub为代表的算力交易中心,是全行业的开放共生:打破资源垄断与技术壁垒,搭建全域AI资源交易市场,让所有供需方自由交易、市场自主定价、资金实时到账。

两条路径看似迥异,实则同源:AI算力的产业属性正在彻底变革,从巨头垄断的稀缺战略资源,逐步转变为全民可参与、可交易、可普惠的标准化商品。

智谱用1GW自建算力证明了「国产算力可规模化落地」;中科加禾的技术能力证明了「国产异构算力可高效利用」;而Unexhub正在完成行业最后一块拼图——证明「国产算力可市场化高效流通」。

AI行业竞争规则已然彻底改写:告别「谁手握稀缺资源谁胜出」的存量竞争,迈入「谁高效流通、谁极致降本谁领跑」的增量时代。巨头深耕自有壁垒,而Unexhub为全行业中小玩家提供了平等入局、公平竞争的机会。

Unexhub平台核心权益,普惠全球开发者与供应商

Unexhub重构全球Token分发底层逻辑,打造全球化Token聚合、分发、交易一体化算力枢纽。平台汇聚Gemini、ChatGPT、Claude、DeepSeek、智谱GLM等全球头部大模型,同时吸纳海量中小企业、个人开发者的优质特色模型资源。依托毫秒级智能路由技术,全网AI资源公开竞价匹配,为用户提供极速稳定、极致低价的调用服务。

平台专属硬核福利:开发者全程按Token实际消耗计费,最低可享官方市场价3折优惠;供应商一键入驻开通变现通道,调用分成实时结算,零门槛盘活闲置AI算力与模型资产,轻松实现稳定盈利。