模型供应商抢滩算力交易中心:37000P 算力、16 种芯片统一纳管,国产大模型告别"英伟达依赖"?

These representative cases reflect a profound industrial structural change: domestic chips are no longer alternative backups, but have become the primary mainstream choice for domestic large model training and deployment.


37,000P Computing Power & Unified Scheduling for 16 Chip Types: The End of Domestic Large Models’ Dependence on NVIDIA

July 2026 marked a pivotal turning point at WAIC 2026: domestic large AI models are undergoing a large-scale "chip replacement" revolution, gradually breaking the long-term monopoly of NVIDIA hardware.

Meituan open-sourced the 1.6-trillion-parameter LongCat-2.0, completing the entire process from pre-training to inference on more than 50,000 domestic AI chips. Zhipu GLM-5.2 achieved Day 0 inference adaptation with domestic computing platforms including Huawei Ascend and T-Head. DeepSeek-V4-Pro also realized full-stack in-depth compatibility with mainstream domestic chips such as Huawei Ascend and Cambricon.

These representative cases reflect a profound industrial structural change: domestic chips are no longer alternative backups, but have become the primary mainstream choice for domestic large model training and deployment.

Computing Fragmentation: The Most Underrated Industry Bottleneck

The domestic chip replacement trend is far more complex than simply replacing NVIDIA GPUs with domestic hardware.

According to SemiAnalysis calculations, the single-card performance of mainstream domestic AI chips is roughly one-third of NVIDIA’s Blackwell architecture, limited by advanced process technology constraints. However, the real industry bottleneck is not single-card performance gaps, but inefficient cluster scheduling and fragmented resource isolation.

Most AI enterprises hold mixed hardware resources: hundreds of A100s, dozens of H100s, and a variety of domestic test chips scattered across different data centers and procurement batches. The traditional operation mode leads to rigid resource division: A100s are dedicated to model training, H100s undertake inference tasks, and domestic chips are left idle or only responsible for low-value edge computing. Isolated clusters turn massive computing resources into static, unusable dead assets.

Xia Lixue, Co-founder and CEO of Infinicore, pointed out the core industry dilemma straightforwardly: "Computing demand is growing exponentially, while computing power expansion relying solely on linear production capacity cannot bridge the widening supply-demand gap. Simply stacking hardware chips can never keep up with the explosive growth of AI demand."

Computing Distribution Hub: Unified Scheduling of 16 Chip Types with Global Coordination

Infinicore’s core solution to fragmented computing resources is building a global computing distribution hub, which realizes unified aggregation, elastic scheduling and on-demand distribution of heterogeneous and scattered computing resources across the country.

At present, the platform’s cross-domain training system has deployed and covered over 37,000P of national computing power, supporting unified access and collaborative scheduling for 16 types of mainstream chips — covering NVIDIA, Huawei Ascend, Cambricon, Hygon and other hardware with different architectures, generations and manufacturers, all integrated into one unified scheduling network.

This cross-brand and cross-generation heterogeneous collaboration capability is supported by three self-developed core technologies: computing-communication pipeline orchestration, adaptive traffic scheduling, and full-dimensional heterogeneous computing scheduling.

The system has achieved outstanding verified results in multiple production-level scenarios:

• Completed ultra-long-distance cross-domain hybrid training spanning over 4,000 kilometers from Hami, Xinjiang to Shenzhen, Guangdong in cooperation with China Telecom, boosting cluster performance by 165%.

• Connected 4 GPU clusters of different generations to support collaborative training of nearly billion-parameter large models, improving resource collaboration efficiency by 41%.

• Realized hybrid computing interconnection between private clusters and public cloud resources, lifting overall system performance by 68%.

• Achieved one-week uninterrupted stable operation in cross-domain reinforcement learning training scenarios.

Xia Lixue revealed that the platform has scalable potential and can be further expanded to a 100,000-card-level super computing scale in the future.

PDD Architecture: Enable Heterogeneous Chips to Maximize Their Strengths

Admittedly, domestic chips have inherent gaps in comprehensive single-card performance, but they show outstanding advantages in specific segmented scenarios. Internal tests by Infinicore verified that a certain domestic E-series chip only reaches 81% of flagship GPU performance in compute-intensive Prefill stages, but surges to 210% of flagship GPU efficiency in memory-access-intensive Decode stages.

Traditional integrated cluster scheduling forces all chips to undertake both Prefill and Decode tasks simultaneously. The scenario-specific strengths of "specialized" heterogeneous chips cannot be exerted, and their inherent advantages are wasted, dragging down the overall cluster efficiency.

To solve this pain point, Infinicore pioneered the cross-cluster heterogeneous PD Separation Architecture (PDD). It splits LLM inference into two independent links, deploying Prefill tasks on high-computing chips and Decode tasks on high-bandwidth chips, and interconnects homogeneous clusters in different regions via WAN Ethernet.

Actual production tests show that the PDD architecture reduces the first token latency by 51.5% and cuts single-token inference costs by 37.5%.

For model suppliers, underlying chip differences are completely transparent. The platform intelligently matches tasks and hardware automatically, eliminating manual adaptation work and enabling zero-threshold efficient use of mixed domestic and NVIDIA computing resources.

Agent Task Stratification: The Inflection Point for Domestic Chip Rise

The substantial improvement in domestic chip adaptation efficiency over the past year stems from two core driving factors. Beyond the technical breakthrough of cross-cluster heterogeneous scheduling, the more critical reason is task stratification brought by the explosive growth of Agent applications.

As of July 2026, the daily token calling volume on the Infinicore platform has surged 40 times compared with last December, withover 95% of inference tasks originating from Agent business scenarios.

Agent operation logic features hierarchical decomposition: a single complex intelligent task will generate numerous sub-tasks with different complexity and computing requirements. Not all tasks require flagship high-end NVIDIA chips. Cost-effective domestic chips can undertake matching lightweight tasks, giving full play to their scenario advantages without bearing ultra-high-intensity training loads.

Task stratification makes hierarchical and refined scheduling of heterogeneous chips feasible, creating a huge incremental market and application space for domestic chips.

Computing Trading Center: Complete the Upgrade From Efficient Production to Efficient Circulation

Infinicore’s heterogeneous scheduling technology solves the problem of how to maximize the utilization efficiency of existing computing resources. To fundamentally help domestic large models get rid of NVIDIA dependence, the industry also needs to solve the core problem of efficient circulation and monetization of computing resources — the core value of professional computing trading centers.

Public data shows that domestic computing platforms have connected 15 provincial and municipal sub-platforms, accessing more than 130 computing service providers and over 200 AI models. The Southeast Token Operation Center serves as a core token service hub covering southern Zhejiang, northern Fujian and eastern Jiangxi.

Unexhub further upgrades the industrial model, transforming token trading from the traditional fixed-price "supermarket model" to a fully market-oriented "futures trading market model". It empowers small and medium-sized suppliers and developers through three core mechanisms:

Real-time Settlement: Completely eliminate long account periods and slow capital recovery in the industry. The professional financial clearing system enables suppliers to withdraw funds in real time based on actual supply, zeroing out corporate financial pressure and converting idle computing assets into stable cash flow.

Early Zero-Commission Benefit: Early settled suppliers enjoy permanent zero-platform-commission privileges, drastically lowering industry entry barriers and enabling zero-cost resource monetization.

Dynamic API Bidding: API calling prices float intelligently with real-time market supply and demand, realizing autonomous market pricing instead of unilateral platform pricing.

In short, Infinicore’s computing distribution hub acts as the token production workshop that activates heterogeneous computing resources efficiently, while Unexhub serves as the token circulation market that realizes free trading, dynamic pricing and instant monetization of computing assets.

Two Paths, One Ultimate Industrial Trend

The unified scheduling of 37,000P computing power and 16 chip types solves the core pain point of the industry: shifting from "whether computing power is available" to "whether every computing resource can be fully utilized". Agent task stratification finds precise application scenarios for domestic chips, while computing trading centers transform tokens from manufacturer-priced scarce resources to market-priced tradable commodities.

Meituan LongCat-2.0’s full domestic chip training, DeepSeek’s Day 0 adaptation on Ascend chips, and Zhipu’s domestic computing data center deployment are not isolated individual cases, but the starting point of an irreversible industrial trend.

The day when domestic large models completely bid farewell to NVIDIA dependence is coming faster than expected.

Unexhub Platform Core Strengths & Exclusive Benefits

Unexhub reshapes the global token distribution rules and builds a leading integrated token aggregation, distribution and trading 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, bringing enterprise users and developers ultra-low-cost and ultra-stable calling services.

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


37000P算力、16种芯片统一调度:国产大模型告别英伟达依赖的时代已至

2026年7月,世界人工智能大会WAIC 2026释放出明确的行业信号:国产大模型正在开启大规模「换芯突围」,持续打破英伟达硬件长期垄断的行业格局。

美团开源1.6万亿参数大模型LongCat-2.0,依托超5万张国产算力芯片,完成从预训练到推理的全流程国产化训练;智谱GLM-5.2实现与华为昇腾、平头哥等国产算力平台的首日适配(Day 0);DeepSeek-V4-Pro也完成对华为昇腾、寒武纪等主流国产AI芯片的全栈深度兼容。

一系列标杆案例背后,是AI产业正在发生的结构性变革:国产芯片彻底摆脱「备选替补」的定位,正式成为国产大模型训练、部署的核心主力硬件。

算力碎片化:被行业严重低估的核心发展壁垒

国产芯片替代绝非简单的「拔卡换卡」,底层技术适配与资源调度的复杂度远超行业认知。

据海外机构SemiAnalysis测算,受先进制程限制,单张国产AI芯片的原生性能仅为英伟达布莱克韦尔架构的三分之一。但真正制约国产算力规模化落地的核心瓶颈,并非单卡性能差距,而是算力碎片化、集群调度低效、资源相互割裂。

当下多数AI企业均存在算力混杂、分散部署的问题:手头既有数百张A100、数十张H100高端英伟达显卡,也配套各类国产测试芯片,设备分散在不同机房、不同采购批次中。传统运维模式下资源分工僵化:英伟达显卡专属训练、高端卡承载推理、国产芯片长期闲置或仅承接边缘轻量任务。算力集群彻底沦为信息孤岛,海量硬件资源变成无法复用、无法流转的「死资产」。

无问芯穹联合创始人兼CEO夏立雪精准点破行业痛点:「AI计算需求呈指数级增长,单纯依靠芯片产能线性堆叠,已经无法填补持续扩大的供需缺口。只堆硬件、不优调度,永远跟不上AI需求的爆发式增长。」

算力集散中心:16类异构芯片,全域统一调度一盘棋

针对行业算力碎片化难题,无问芯穹打造「算力集散中心」核心方案,实现全域分散异构算力的统一汇聚、弹性调度、按需分发,彻底盘活存量算力资源。

目前,平台跨域训练系统已实现全国布局,覆盖算力规模超37000P,全面兼容16种市面主流AI芯片,囊括英伟达、华为昇腾、寒武纪、海光等全品类、跨架构、跨代际硬件,所有异构算力统一接入一张智能调度网络。

这套打破硬件壁垒的协同能力,依托三大自研核心技术支撑:计算通信流水编排技术、自适应通信调度技术、全维度异构算力调度技术。

经过多场景生产级落地验证,技术成效十分亮眼:

• 联合中国电信完成新疆哈密—广东深圳超4000公里跨域混训,打破广域网延迟限制,集群整体性能提升165%;

• 打通4个不同代际异构GPU集群,支撑近百亿参数大模型联合训练,资源协同效率提升41%;

• 实现私有集群与公有云混合算力安全互通、协同运算,整体算力性能提升68%;

• 跨域强化学习训练场景实现连续7天零中断稳定运行,适配高强度生产需求。

夏立雪透露,平台具备极强的扩容潜力,未来可升级拓展至十万卡级超级算力规模,全面适配超大模型训练需求。

PDD分离架构:让每一款芯片都能发挥极致长板

不可否认,国产芯片综合单卡性能存在短板,但在细分场景具备极致爆发优势。无问芯穹实测数据显示:某国产E系列芯片在计算密集的Prefill预填充阶段,性能仅为旗舰GPU的81%;但在访存密集的Decode解码阶段,性能可飙升至旗舰GPU的210%,场景优势极其突出。

传统集群调度模式存在致命缺陷:所有芯片统一承载预填充、解码全流程任务,无法适配异构芯片的差异化特性,国产芯片的场景长板被彻底埋没,短板反而拖累整体集群效率。

基于此,无问芯穹首创跨集群异构PD分离架构(PDD),将大模型推理的两大核心阶段拆分独立调度:把Prefill算力任务分配给算力更强的芯片,将Decode访存任务分配给带宽更高的芯片,再通过广域网以太网串联跨地域同构集群。

落地实测效果显著:模型首Token延迟降低51.5%,单Token推理成本直接下降37.5%,实现性能与成本的双向优化。

对模型供应商而言,底层芯片的品牌、架构、性能差异完全透明,无需人工适配调试。平台智能按需调度、最优匹配,真正实现异构算力无感使用、高效落地。

Agent任务分层:国产算力迎来规模化逆袭拐点

过去一年,国产芯片适配效率实现质的飞跃。夏立雪拆解核心原因:五成归功于跨集群异构调度的技术突破,更核心的五成来自Agent智能体业务爆发带来的「任务分层变革」。

截至2026年7月,无问芯穹平台单日Token调用量较去年12月暴涨40倍,其中95%以上的推理任务均来自Agent场景。

Agent智能体任务具备典型的分层拆解特性:一项复杂的智能交互任务,会衍生出大量复杂度各异、算力需求不同的子任务。并非所有子任务都需要高端旗舰芯片支撑,性价比更高的国产芯片可精准承接轻量化、常态化任务,无需强行承载超高强度的大模型训练负载。

任务分层落地,让异构芯片精细化、分层化调度成为行业常态,彻底激活国产芯片的场景价值与增量空间。

算力交易中心:补齐产业短板,从高效生产到高效流通

无问芯穹的异构调度技术,解决了「算力如何高效利用」的生产端难题,盘活了存量异构算力资源。而想要彻底摆脱英伟达依赖、完善国产算力生态,还需解决「算力如何高效流转变现」的流通端难题——这正是专业算力交易中心的核心价值。

公开信息显示,国内算力生态已完成15个省市分平台联动,接入130余家算力服务商、200多款优质AI模型;东南Token运营中心正式落地,成为辐射浙南、闽北、赣东的算力与词元服务核心枢纽。

而Unexhub进一步迭代产业模式,将传统词元「超市式固定售卖」升级为「市场化动态交易」,依托三大核心机制彻底革新行业流通规则:

全程即时结算 彻底打破行业账期冗长、回款滞后的痛点,专业金融清算系统支持供应商按实时供给量全额回款、即时提现,清零中小团队资金压力,让闲置算力资产转化为稳定现金流。

早期零佣金特权 首批入驻供应商可享受平台永久零佣金权益,零成本入驻、零门槛变现,大幅降低行业入局门槛。

API动态竞价机制 摒弃平台固定定价模式,调用价格随市场供需实时浮动,由市场自主定价,供需双向共赢。

如果说无问芯穹算力集散中心是AI词元的「生产车间」,负责协同异构算力、高效产出词元资源;那么Unexhub就是词元的「流通市场」,实现算力资产自由交易、动态定价、即时变现。

双轨协同,奔赴国产算力自主终局

37000P全域算力覆盖、16类芯片统一调度,解决的不是简单的「算力有无问题」,而是「每一份算力能否最大化利用」的效率问题。Agent任务分层为国产芯片找准了场景定位,算力交易市场化则将Token从巨头垄断的稀缺资源,转化为全民可流通的标准化商品。

美团LongCat-2.0全国产芯片训练、DeepSeek昇腾首日适配、智谱国产算力中心落地……一系列案例绝非个例,而是国产算力突围、摆脱海外依赖的产业序幕。

国产大模型彻底告别英伟达垄断依赖的时代,已然加速到来。

Unexhub平台核心权益,普惠全行业开发者与供应商

Unexhub重构全球Token分发底层逻辑,搭建全球化Token聚合分发算力交易中心。平台汇聚Gemini、ChatGPT、Claude、DeepSeek、GLM等全球顶级大模型,同时吸纳海量中小企业、个人开发者的优质AI模型资源。依托毫秒级智能路由技术,全网AI资源公开竞价流通,为用户提供极致低价、极速稳定的调用服务。

平台专属硬核福利:开发者全程按Token实际消耗计费,最低可享市场官方价格3折极致低价;供应商一键入驻开通变现渠道,调用分成实时结算,零门槛盘活闲置AI算力资产,轻松实现高效盈利。