Skip to main content

7 posts tagged with "华为昇腾950"

华为昇腾950PR/950DT芯片技术详解

View all tags

徐直军确认:昇腾中国份额已超英伟达,950DT 规模供货推迟至年底或明年初

· 4 min read
Industry Research Team

10 月 6 日,华为全联接(HC)大会期间,轮值董事长徐直军与海思首席科学家廖恒面对记者给出两个重磅判断:其一,按华为可统计口径,昇腾在中国市场的份额已经超过英伟达;其二,昇腾当前的主要约束不再是需求,而是供给——「现在能产多少就供多少」。

950 系列供货时间线(官方口径)​

型号状态官方口径
昇腾 950PR2026 Q1 发布并上市面向推理(Prefill)场景,徐直军称「上市量不算大」
昇腾 950DT测试中超节点(训练用)规模供货 2026 年底或 2027 年初

与 8 月「华为云上线」的节奏相比,这次口径把物理超节点形态的规模供货明确放到了 2026 年底至 2027 年初。徐直军同时给出一个重要信号:与各模型厂商的联合测试「训练表现非常不错」,预计 2027 年起大量模型训练将构筑在 950DT 超节点上——如果兑现,2027 年将是国产算力首次承担前沿模型训练主负载的年份。

份额判断的证据边界​

「昇腾份额超英伟达」这个说法需要谨慎引用:徐直军自己承认中国市场统计英伟达份额很难,结论来自「华为能够统计到的数据」,华为未公布支撑该判断的出货或营收数字。第三方可引用的旁证包括路透社 9 月的报道与 IDC 2026 上半年中国 AI 服务器份额数据(英伟达 34%、华为 28%),以及出口管制后英伟达「中国数据中心收入不再纳入展望」的财报事实。合理的表述是:在高端 AI 加速卡的可获得性维度,昇腾已事实性成为中国市场主力供给。

三条供给约束​

徐直军「能产多少供多少」的背后,是三层产能天花板:

  1. 先进制程晶圆产能:余承东同日确认国内先进产能跟不上需求——昇腾芯片与 Mate90 系列的逻辑堆叠硅片共享同一产能池,手机与 AI 加速卡在抢同一批晶圆。
  2. HBM 内存:摩根士丹利最新报告认为 DRAM 短缺将贯穿整个 AI 周期,连 NVIDIA 都被迫在 Rubin 平台上做减配(机架 LPDDR5 从规划的 54TB 砍到 28TB),昇腾的供应链同样躲不开挤压。
  3. 互联部件:NPO 光引擎等新部件在国内供应链的量产成熟度仍需时间。

Peerium 架构:用系统换制程​

同一场合,华为首次系统阐述了 Peerium 计算架构与灵衢(UnifiedBus)总线:让 CPU、NPU、内存与存储在统一协议下组成一台「逻辑计算机」,目标支持百万颗处理器规模。徐直军不回避制程差距——单芯片性能受国内制造限制,但多芯片互连与系统架构是华为自己的路线。华为正在部署和测试 25.6 万张卡规模的 Atlas 950 SuperPoD。

这与 NVIDIA 的路线形成鲜明对照:Rubin 用最新制程堆单卡性能,昇腾用架构补制程差。两条路线孰优,2027 年 950DT 超节点上跑出来的前沿模型训练成绩会给出第一个可验证的答案。

相关阅读​

昇腾 950 系列最新报价:950PR 与 950DT 的价格曲线

· 5 min read
Industry Research Team

昇腾 950 系列目前处在"一代已上市、一代待量产"的关键节点。950PR 搭载的 Atlas 350 加速卡已于 2026 年 3 月 20 日正式上市,市场报价约 11 万元;950DT 计划 2026 年第四季度量产,其价格尚未正式公布,但产能与 HBM 成本正在推动整条国产算力价格曲线上移。本文梳理两款芯片的规格差异与价格走向。

950PR 与 950DT:不是同一款芯片的两种包装​

这两款芯片基于统一的 Ascend 950 Die,但内存方案与定位完全不同。据公开报道,昇腾 950PR 采用 950 核心搭配 HiBL 1.0 内存,主要面向推理 Prefill 阶段和推荐业务场景;950DT 采用 HiZQ 2.0 内存,注重推理 Decode 阶段和训练场景。

维度昇腾 950PR昇腾 950DT
内存方案HiBL 1.0HiZQ 2.0
内存容量112GB(芯片规格 128GB)144GB
内存带宽1.4TB/s(芯片规格 1.6TB/s)4TB/s
互联带宽2TB/s2TB/s
目标场景推理 Prefill、推荐推理 Decode、训练
上市时间2026 年第一季度2026 年第四季度

这一拆分反映了一条现实的成本逻辑:Prefill 阶段是算力密集,Decode 阶段是带宽密集。DT 把内存容量和带宽拉高以应对长上下文解码,代价是更高的 HBM 成本。华为同时自研了两种 HBM——HiBL 1.0 和 HiZQ 2.0,与 Ascend 950 Die 分别合封构成两颗芯片。

950PR 的价格锚点:约 11 万元​

根据多方公开报道,搭载昇腾 950PR 的 Atlas 350 加速卡单价约 11 万元人民币(约 1.6 万美元)。这个价格被反复引用为昇腾 950PR 的市场锚点。

对应的硬件规格是:FP4 精度算力 1.56 PFLOPS,集成 112GB HBM,带宽 1.4TB/s,功耗 600W,访存颗粒度从上一代的 512 字节减少至 128 字节,芯片间互联带宽从上一代的 784GB/s 提升至 2TB/s。

据公开报道,华为曾称昇腾 950PR 推理性能达英伟达 H20 的 2.87 倍,价格仅约为英伟达的四分之一左右。这一对比被广泛用于海外市场宣传——华为计划 2026 年第四季度将昇腾系列推向韩国市场,已完成分销商协议签署,选定 SK Shieldus 等两家作为销售渠道,主力产品正是 950PR 与 950DT。

不过这个 2.87 倍的对比需要审慎看待:华为的对比使用 Atlas 350 的 FP4 精度对比 H20 支持的最高精度,并非完全对等。Atlas 350 的内存带宽 1.4TB/s 也明显低于 H20 的 4.0TB/s,该差距会在带宽受限的解码任务上体现。

950DT 第四季度量产:价格怎么走​

950DT 尚未正式定价,但有三个因素会推动价格曲线变化。

第一是 HBM 成本。据公开报道,摩根士丹利已将 2030 年中国 AI 加速器 TAM 预测从 910 亿美元上调至 1320 亿美元,明确以 HBM 成本上升推高芯片定价为理由。HBM 供应瓶颈不仅影响昇腾,也直接压制寒武纪 MLU580 的出货。

第二是产能。据公开报道,华为计划 2026 年出货约 75 万颗昇腾 950 系列;同时也有行业报告显示 950PR 存在采用传统 DDR 内存的版本约 5 万元/颗、配备 HBM 的高端版本约 7 万元/颗的区分。后一组数字口径与 11 万元的 Atlas 350 报价存在差异,可能指向不同产品形态或不同渠道,具体定价以华为官方渠道为准。

第三是国产竞品定价。字节跳动、腾讯、阿里已批量采购昇腾 950 系列,但同期阿里平头哥"真武"PPU 已出货数十万片、寒武纪 MLU580 在三季度提价——国产算力市场并不存在单一价格锚点,每个厂商的定价都锚定各自的性能与产能位置。

研判​

昇腾 950 系列正处在"价格最有说服力的窗口期"。950PR 约 11 万元的定价相对 H20 的四分之一,配合 DeepSeek V4 全线适配昇腾生态的进展,确实建立了性价比叙事。但这个窗口不会太长:950DT 第四季度量产将带来更大的 HBM 消耗,OIF 的 NPO 标准推进、昇腾 960 超节点发布带来的供应链挤兑,都可能推高成本。

对采购方而言,实际成本不止是卡价。昇腾超节点的整套系统定价、灵衢互联设备、超节点机柜的电力与液冷改造,都是必须计入 TCO 的项。建议关注三个信号:一是 950DT 正式上市后的官方定价,二是昇腾 950 系列的年度出货目标能否兑现 75 万片,三是 HBM 国产化进展对成本的实质影响。价格曲线的斜率,最终由产能兑现速度决定。

昇腾950 双轨战略:950PR 主攻推理 Prefill、950DT 押注训练超节点

· 6 min read
Industry Research Team

当算力供给的矛盾从"有没有"转向"够不够",芯片设计的刀法也开始向场景细分。华为昇腾950系列首次拆分为950PR与950DT双版本——前者专攻推理Prefill,后者押注Decode与训练超节点。这既是Prefill/Decode负载特性差异的工程回应,也是自研HBM产能约束下的现实选择。

一、为什么推理要拆成两颗芯片​

大模型推理的两条流水线,瓶颈完全不同:

  • Prefill(首 Token 生成):一次性吞下整段 Prompt 做并行计算,算力(FLOPS)是瓶颈,带宽需求适中;
  • Decode(逐 Token 生成):每生成一个 Token 都要把权重和 KV Cache 从显存搬一遍,带宽是瓶颈。

把两类负载塞进同一颗芯片,必然出现"一半硅片在空转"的错配。华为的解法是同一 Da Vinci v5 计算核心、两套内存子系统——950PR 用成本优先的 HiBL 1.0,950DT 用带宽优先的 HiZQ 2.0。这与 Google 的思路殊途同归:谷歌同样在 TPU 家族中把训练型 8t 与推理型 8i 拆分为两条产品线,让"算力密集"和"带宽密集"各自归位。

维度昇腾 950PR昇腾 950DT
目标场景推理 Prefill(首 Token 生成)、视频推荐、实时交互推理 Decode + 模型训练
HBM自研 HiBL 1.0,128 GB自研 HiZQ 2.0,144 GB
HBM 带宽1.6 TB/s4 TB/s
FP8 算力1 PFLOPS(HiF8 格式)1 PFLOPS(HiF8 格式)
FP4 算力2 PFLOPS2 PFLOPS
TDP~400 W~500 W
定价逻辑成本优先带宽优先、性能优先

按《昇腾 950 NPU 架构白皮书》的精确口径,FP8 算力随配置在 919–1034 TFLOPS 之间(宣传口径取整 1 PFLOPS),FP4 为 1561–2007 TFLOPS。两者共享同一计算核心,差异全部体现在内存子系统上——950DT 的 4 TB/s 带宽正是 Decode 与训练场景的命门。

二、Da Vinci v5:双编程模型与 HiF8 的精度革命​

昇腾950 是 Da Vinci 架构第五代,两项变化值得单独展开:

SIMD + SIMT 双编程模型。SIMD 延续 Da Vinci 在向量计算上的传统优势;新增的 SIMT 模型则面向 Decode 阶段大量不规则内存访问,让开发者可以用更接近 GPGPU 的方式写算子。配合内存访问粒度从 512 字节压缩到 128 字节,离散内存访问效率提升 4 倍——这对小 batch、多并发的推理业务是实打实的收益。华为同时提供 ASIC(昇腾 Core)与 GPGPU 双生子型号,兼容两种开发范式。

HiF8 自研 FP8 格式。低精度化的最大阻力是精度损失,HiF8 的设计目标是"精度接近 FP16、算力比 FP16 提升 2 倍",配合 FP4 格式整体算力利用率提升 30% 以上。结合 2 TB/s 灵衢互联,384 颗 950 组成的 CloudMatrix 384 超节点总算力约 384 PFLOPS FP8,以系统级规模对冲单卡代差。

三、自研 HBM:供应链意义大于纸面参数​

950 系列首次搭载华为自研 HBM——950PR 用 HiBL 1.0(成本优先),950DT 用 HiZQ 2.0(带宽优先,4 TB/s 对标 NVIDIA HBM3e)。这组数字的意义不在跑分,而在供应链:过去国产 AI 芯片的产量天花板被 SK 海力士、三星的 HBM 配额捏在手里,自研 HBM 让昇腾第一次把"产能多少"的决定权拿回自己这边。

但产能爬坡需要时间。渠道数据显示,国产 HBM 产能目前仅能支撑约 25–30 万颗成品加速器,缺口约六成——这正是下文涨价潮的底层原因。

四、徐直军 HC2026 访谈:节奏、份额与产能​

2026 年 10 月 8 日,华为副董事长徐直军在 HC2026 期间接受访谈,披露了几个关键信息:

  • 950PR 已于 2026 年第一季度发布上市,定位推理场景,"上市量不算大";
  • 950DT 超节点正在测试中,规模供货预计今年年底或明年初落地;
  • 关于训练业务,徐直军表示:"950DT 训练表现非常不错,明年开始大量模型训练会构筑在 950DT 超节点上;关键在供货能力,能产多少供多少。"
  • 从华为统计口径看,昇腾在中国市场的份额已超过英伟达;
  • 华为暂无全面海外拓展计划,重心仍在国内市场。

"能产多少供多少"这六个字,基本概括了当前国产 AI 芯片的真实处境:需求端远未见顶,第一约束已经从市场转移到了产能。

五、生态与价格:开源换时间,稀缺换定价​

生态侧,昇腾950 正在以"开源 + 头部共建"提速:DeepSeek 与华为联合开源面向昇腾950 的编程工具链(含计算与通信库),头部模型厂商直接参与共建是"去 NVIDIA 化"的标志性事件;CANN 已全面开源,外部开发者占比达 61%(首次超过内部团队),月活开发者突破 5200 人,PyTorch 官网将昇腾列为首个中国算力平台;昇腾950 智算集群云服务已于 9 月 30 日在国内商用。

价格侧则是另一番景象。受 HBM 成本攀升与供不应求驱动:

型号年初报价2026-10 市场报价涨幅
昇腾 950PR~¥6 万突破 ¥8 万+30%
昇腾 950DT~¥12–15 万指示性报价超 ¥25 万两个月内最高 +50%

950DT 的渠道指示价已经对齐 NVIDIA B200 的量级——国产旗舰的定价逻辑,正从"替代品折价"转向"供给稀缺溢价"。

小结​

昇腾950 的双轨战略回答了两个问题:产品层面,Prefill/Decode 分芯片让每一份硅片和带宽都花在刀刃上,与 Google TPU 的产品线拆分形成跨洋呼应;供应链层面,自研 HBM 把产能主动权握回手中,但爬坡期的缺口也直接推高了价格。徐直军"能产多少供多少"的表述说明,2027 年国产算力的故事主线不再是需求,而是产能——950DT 超节点能否如期规模供货,将决定明年大量模型训练是否真正构筑在国产超节点之上。

徐直军 HC2026 访谈:昇腾950DT年底规模供货,"昇腾中国份额已超英伟达"

· 6 min read
Industry Research Team

华为全联接大会(HC2026)期间,华为轮值董事长徐直军在接受采访时给出了昇腾950系列的最新进度:950PR 已上市但量不大,950DT 超节点年底规模供货,并首次松口"昇腾中国份额已超英伟达"。

访谈要点:950PR 已上市,950DT 年底规模供货​

华为全联接大会(HC2026)在上海举行,展会现场展出了昇腾950和 Atlas 950 超节点。徐直军在采访中把昇腾950系列的节奏讲得非常具体:

  • 昇腾950系列分为 PR 和 DT 两个版本;
  • 950PR 已于 2026 年第一季度正式发布上市,主要面向推理场景,但"上市量不算大";
  • 基于 950DT 的超节点主要用于训练,目前还在测试中,规模供货预计今年年底或明年初。

两代产品的分工清晰:950PR 先行落地承接推理需求,950DT 则押注训练场景,承担更重的算力任务。

"能产多少就供多少":产能成了第一约束​

关于 950DT 的训练表现和供货节奏,徐直军的原话值得完整引用:

"我们和各家模型厂商做了很多沟通交流,从目前测试结果来看,950DT的训练表现非常不错,我相信从明年开始,大量模型的训练会构筑在950DT超节点上。未来关键可能不在于我们推动的力度有多大,而在于我们的供货能力有多大。现在能产多少就供多少,我们希望产业链加快扩产,快一点、多产一点。"

这段话透露出一个重要信号:对国产 AI 芯片而言,需求侧已经不是瓶颈,产能才是。徐直军把"供货能力"而非"推动力度"定义为未来的关键变量,等于公开向产业链喊话扩产——"快一点、多产一点"的表述,把扩产的紧迫性摆到了台面上。

份额表态:"昇腾应该已经超过了英伟达"​

在市场地位问题上,徐直军的表述同样直接:

"在中国市场要统计英伟达的市场占有率很难,但如果从华为能够统计到的数据看,昇腾应该已经超过了英伟达。"

这句话需要仔细解读。徐直军本人也承认统计口径的困难——英伟达在中国市场以何种形式、多大数量出货,外界很难完整统计。华为的判断基于"华为能够统计到的数据",这是一个有限口径的推断,而非权威市场份额数字。但这是华为高层首次公开作出这一表述,其信号意义大于数字本身。

海外策略:暂不全面拓展​

关于国际化,徐直军给出了明确的优先级:"目前昇腾满足国内市场需求还远远不足,所以没有全面拓展海外市场的计划。但确有少数国家做了测试和供给,但量非常少。"

换句话说,海外市场目前只是零星试点,华为的全部产能仍将优先投向国内。这与其"供货能力是第一约束"的判断是自洽的:产能有限的前提下,先满足国内是唯一合理的选择。

时间线与定价背景​

把 HC2026 前后的公开信息放在一起看,昇腾950的落地节奏如下:

事项状态
昇腾950PR(推理)2026年第一季度正式发布上市,上市量不算大
昇腾950DT 超节点(训练)测试中,规模供货预计今年年底或明年初
950DT 指示性价格超 25 万元(路透报道,HBM 成本驱动涨价)
Atlas 950 超节点HC2026 现场展出
昇腾950智算集群云服务9月30日国内商用

值得注意的是定价:路透此前报道 950DT 指示性价格超过 25 万元,主要驱动因素是 HBM 成本上涨。这与全球存储供应紧张的大背景互为印证——HBM 涨价正在推高所有 AI 芯片的成本曲线,国产芯片也无法置身事外。

分析:三个观察点​

**第一,"份额超英伟达"的口径与可信度。**虽然华为的统计口径有限,但结合出口管制后英伟达对华供应受限、国产芯片放量的大背景,昇腾在国内新增 AI 算力采购中的占比持续上升是行业普遍观察到的趋势,争议主要在幅度而非方向。华为选择在 HC2026 这个场合公开表态,本身也说明其对产能与产品力有了相当的底气。

第二,产能取代需求成为第一约束。"能产多少就供多少"说明国产芯片的瓶颈已经从"有没有人用"变成"产得出多少"。下一步的关键变量是先进制程与 HBM 供应链的扩产速度——这也是 950DT 价格超过 25 万元的根本原因。

**第三,950DT 对国产模型厂商的意义。**如果明年大量模型训练真的构筑在 950DT 超节点上,国产训练算力将第一次形成规模化的供应能力。配合 9 月 30 日已商用的昇腾950智算集群云服务,国产模型厂商在训练底座上将有更完整的选择,这对它们的成本结构和供应链安全都有直接影响。

小结​

徐直军在 HC2026 的访谈给出了两条硬信息:昇腾950PR 已上市、950DT 训练超节点年底规模供货;以及一句分量不轻的判断——"昇腾应该已经超过了英伟达"。产能而非需求,正在成为国产 AI 算力扩张的第一约束;而 950DT 能否如期放量,将决定明年国产大模型训练底座的成色。

昇腾950智算集群云服务正式商用:9月30日国内上线、11月30日全球商用

· 5 min read
Industry Research Team

华为全联接大会2026给出明确时间表:昇腾950智算集群云服务9月30日国内商用、11月30日全球商用,全球首个NPO超节点昇腾960同步亮相。

HUAWEI CONNECT 2026:商用时间表落定​

2026年9月17日,华为全联接大会在上海举行。大会上,华为披露了昇腾950智算集群云服务的商用节奏:9月30日国内商用,11月30日全球商用。从大会披露到国内上线仅半个月,节奏之紧,说明昇腾950超节点的规模化交付已经就绪。

关键节点时间/地点
华为全联接大会20262026年9月17日(上海)
昇腾950智算集群云服务国内商用2026年9月30日
昇腾950智算集群云服务全球商用2026年11月30日

"国内先行、两月后推向全球"的节奏也传递出另一个信号:昇腾不再只盯着国内市场,智算集群云服务将作为中国AI算力方案走出去的载体,参与全球算力供给的竞争。

昇腾960超节点:全球首个NPO超节点​

同期发布的昇腾960超节点,是全球首个NPO超节点(近封装光学超节点),关键规格如下:

指标数值
卡规模4096卡
算力8 EFLOPS FP8 / 16 EFLOPS FP4
显存1PB HBM
光互联方案5500个Hi-ONE光引擎,替代4.8万颗800G光模块
降功耗550kW
可用度99.8%
Hi-ONE单引擎传输容量7.2T

这组数字的核心看点在于光互联的代际切换:用5500个Hi-ONE光引擎替代4.8万颗800G光模块,降低550kW功耗,同时做到99.8%的可用度。NPO(近封装光学)路线在超节点规模下的价值被直观呈现——光模块数量下降一个数量级,意味着故障点、功耗与运维复杂度同步大幅收敛。在4096卡这种规模的系统里,互联可靠性直接决定训练任务的连续性。

昇腾950超节点:规模商用与SOTA背书​

昇腾950超节点已规模商用1000多套,落地行业覆盖互联网、运营商、金融、教育、医疗、交通、制造等。规模数字之外,更关键的背书是:昇腾950超节点是国内唯一训练出SOTA模型的超节点。

"能训练出SOTA模型"是对训练集群最硬核的检验——它证明的不仅是硬件规格,而是"集群规模+互联效率+软件栈"的组合在大规模训练中真实可用。对于正在选型的行业客户而言,这比任何纸面参数都更有说服力。

芯片路线图:960提前就绪,970/980排期明确​

华为还更新了昇腾芯片的推进节奏:

  • 昇腾960DT较原计划提前三个季度就绪;
  • 昇腾960PR提前一个季度就绪;
  • 2028年推出昇腾970,2029年推出昇腾980。

"提前就绪"在当前的供应链环境下格外难得——在全球先进制造资源紧张、HBM紧缺的背景下,按时甚至提前交付,本身就是工程与供应链能力的体现。而970/980排期明确到年份,则给生态伙伴与客户提供了可规划的长期预期。

CANN全面开源:生态数据盘点​

软件生态方面,昇腾的进展可以用一组数据概括——CANN已全面开源:

  • 外部开发者占比61%,月活跃开发者超5200名;
  • 40+模型实现原生训练;
  • 支持PyTorch、Triton、vLLM、veRL等90+开源项目;
  • 成为PyTorch官网首个中国算力平台。

其中"PyTorch官网首个中国算力平台"具有标志性意义:主流框架的原生适配,意味着开发者迁移成本大幅降低,而61%的外部开发者占比说明社区已经超出华为体系自身,形成外部贡献的正循环。

全栈协同:TaiShan 950与OceanStor M900​

硬件侧的配套升级同样值得关注。TaiShan 950超节点完成升级:支持4096节点、256TB统一内存池——统一内存池让CPU侧内存资源可以按需调度,与超节点的算力规模相匹配。存储侧,OceanStor M900上下文内存存储提供PB级KV Cache能力,直击大模型推理场景的上下文缓存需求:推理热门场景下,KV Cache的存取效率直接决定吞吐与首字延迟。

芯片(昇腾950/960)、CPU(TaiShan 950)、存储(OceanStor M900)、软件栈(CANN)四层同步演进,华为走的是典型的全栈协同路线。

小结​

从昇腾950智算集群云服务"9月30日国内、11月30日全球"的商用时间表,到昇腾960超节点"4096卡、8 EFLOPS FP8、16 EFLOPS FP4、1PB HBM、5500个Hi-ONE光引擎"的规格突破,再到昇腾950超节点规模商用1000多套、成为国内唯一训练出SOTA模型的超节点,以及CANN全面开源、外部开发者占比61%的生态进展,华为正以"芯片-超节点-云服务-开源生态"的全栈节奏推进昇腾路线图。昇腾970(2028)、昇腾980(2029)的排期已经明确——昇腾的追赶不再只是单点突破,而是体系化的持续推进。

Huawei Ascend 950 Series Capacity & Orders Deep Dive: 950PR Monthly Capacity Jumps 10×, ByteDance Locks In 350k Units for $5.6B

· 4 min read
Industry Research Team

The Ascend 950 series (950PR inference / 950DT training) has become the core supply of domestic AI compute. Per multiple brokerages and industry research, 950 series capacity is 100% booked with scarce spot supply; the full-year 1.2M-unit target is "100% certain," with expectations of an upward revision to 1.5M. This article summarizes capacity and order data as of July 2026.

1. Capacity pace: ~10× MoM jump in June​

Time950PR monthly capacityNotes
May 202650k-60k unitsNear full production
June 2026500k-600k units~10× MoM; SMIC, Hua Hong tier-1 suppliers on overtime
Q3 2026 (est.)700k-800k unitsPer month
Full-year 2026 target1.2M unitsUpward revision to 1.5M expected

Supply chain delivery is tight: high-speed backplanes and liquid-cooling connectors' lead time stretched from 2 weeks to 6-8 weeks; orders are booked into 2027.

2. Order structure: top cloud providers + operators + overseas​

CustomerLocked volumeAmount / Notes
ByteDance350k 950PR$5.6B, concentrated delivery from Q3 2026
Tencent / Alibaba / Baidu~250k 950PR + 150k 950DTCombined ~400k units
Three major operators200k+ unitsCentralized procurement, for intelligent compute centers and AI private networks
OverseasSouth Korea 2,000 units, Malaysia 3,000 servers, Russia ten-thousand-card clusterFrom pilot to commercial

3. Shipment forecast: firmly #1 domestic​

Per CCA (Kezhi) Consulting estimates:

Metric20252026 (forecast)
Huawei Ascend total shipments812k cards1.026M cards
Of which 950PR—~800k units
Of which 950DT—~100k-200k units

Huawei has completed the product transition from the 910 series to the 950 series. The internet industry has become Ascend's largest application market; competitive advantage is extending from single-hardware performance to software ecosystem and system capabilities.

4. Going overseas: formal South Korea entry in Q4​

Per Korean media ETNews, Huawei plans Q4 2026 to formally enter the South Korean market with the Ascend series and Atlas 950 SuperPod:

  • Local distributor agreements signed; two channel partners including SK Shieldus selected
  • Main products: 950PR (mass-produced and delivered since April) and 950DT (launched Q4)
  • Official line: 950PR inference performance is 2.87× that of H20, priced at about 1/4 of it

5. WAIC 2026: 1024-card live debut confirmed​

At WAIC 2026 (July 17-20, Shanghai), Huawei's Atlas 950 SuperPoD live hardware made its first public appearance — a 16 compute-cabinet, 1,024 Ascend-card scale — and won the conference's top honor, the SAIL Award:

  • Core metrics: total compute 1 EFLOPS FP8 / 2 EFLOPS FP4, 256 TB globally unified memory addressing, Lingqu 2.0 interconnect, 3 μs ultra-low RTT latency
  • Full configuration: 128 compute cabinets + 32 interconnect cabinets = 160 cabinets, ~1000㎡ footprint, carrying 8,192 Ascend 950DT, planned Q4 2026 launch
  • Commercial foundation: previous-gen 384 SuperNode has cumulatively shipped 750+ units, deployed in 20+ industries
  • Software ecosystem: CANN fully open-sourced end of 2025; community incubated 67 projects, 12.44M+ lines of code, 3,500+ monthly active developers

WAIC's debut confirmed the 950 series' "SuperNode-first" product logic: beyond single-card compute, system-level effective compute (interconnect bandwidth + unified memory + low latency) is the key dimension for domestic compute to benchmark against international flagships.

Ascend roadmap recap​

ProductPositioningKey metrics (official roadmap)
950PRInference1 PFLOPS (FP8) / 2 PFLOPS (FP4), 2 TB/s interconnect
950DTTrainingSuperNode core, launched Q4
960Train/inference2 PFLOPS (FP8) / 4 PFLOPS
970Next-genIn planning

Industry interpretation​

  1. Domestic substitution moves from inference to training: 950PR (inference) ramps first, 950DT (training) follows in Q4, combined with the Atlas 950 SuperPoD ten-thousand-card interconnect — domestic compute now has the complete "training substitution" puzzle for the first time.
  2. Capacity is the biggest variable: order certainty is extremely high, but SMIC/Hua Hong advanced-process capacity, HBM supply, and advanced packaging remain ramp bottlenecks — the root of "scarce spot supply."
  3. Going overseas opens a second growth curve: bulk procurement from South Korea, Malaysia, Russia, and Latin America marks domestic compute's shift from "internal circulation" to "external circulation."

References​


Data in this article is based on official and major brokerage research; capacity/orders are dynamic figures and will be continuously updated.

2026 Domestic AI Chip Progress: Huawei Ascend 950, Baidu Kunlun M100, Alibaba T-Head M890 Fully Explained

· 15 min read
Industry Research Team

In 2026, China's domestic AI chip industry has entered a period of full-scale explosion. The three giants — Huawei Ascend, Baidu Kunlun, and Alibaba T-Head — have successively launched next-generation products, while Cambricon, MetaX, Enflame, and Iluvatar have also achieved important breakthroughs.

This article comprehensively analyzes 2026 domestic AI chip progress across four dimensions: product launches, technology breakthroughs, market dynamics, and ecosystem building.


1. Huawei Ascend: 950 series launched, 960/970 roadmap clear​

1.1 Ascend 950PR (launched Q1 2026)​

Core specs:

ItemParameter
Launch dateMarch 21, 2026
PlatformAtlas 350 accelerator card
HBM capacity128 GB (Huawei in-house HiBL 1.0 HBM)
Memory bandwidth1.6 TB/s
FP8 compute1 PFLOPS
PositioningInference-specific (Prefill stage)
Performance vs.Single-card compute is 2.87× NVIDIA H20

Technology innovations:

  • First adoption of Huawei in-house HBM solution (HiBL 1.0), lowering cost
  • Supports FP8 low-precision compute, 3× inference energy-efficiency improvement
  • Optimized for inference scenarios such as video recommendation and real-time interaction

Commercialization progress:

  • Mass supply began in Q1 2026
  • Key customers: China Telecom, China Mobile, China Unicom, Huawei Cloud
  • Priced around ¥100,000/card (¥80,000 for key accounts), ~30% lower than comparable competitors

1.2 Ascend 950DT (launched Q4 2026)​

Core specs:

ItemParameter
Launch dateQ4 2026 (expected October)
HBM capacity144 GB (Huawei in-house HiZQ 2.0 HBM)
Memory bandwidth4 TB/s (HiZQ 2.0 technology)
FP8 compute1 PFLOPS
PositioningInference + training (Decode stage + training tasks)
Technology innovationFirst to carry in-house HiZQ 2.0 memory technology

Technology innovations:

  • Adopts HiZQ 2.0 memory technology, 2× data-movement efficiency
  • Supports FP8/FP4 low-precision compute, balancing performance and efficiency
  • Optimized for scenarios such as dialogue generation and large-model training

1.3 950 SuperNode (launched Q4 2026)​

System specs:

ItemConfiguration
Max interconnected chips8,192 chips
Total FP8 compute1 EFLOPS (1,024-card scale)
1024-card version16 liquid-cooled cabinets, 64 chips per cabinet
Supported modelsTrillion-parameter large-model training
Deployment progress1,024-card version already in deployment

Performance comparison:

  • 950 SuperNode outperforms NVIDIA's 2027 NVL576 system
  • Leads by 20% in trillion-parameter model training

1.4 Ascend 960/970 roadmap​

Chip modelLaunchCore specsPositioning
Ascend 960Q4 2027N+3 process, 288GB HBM, FP8 2 PFLOPS, 30%+ better energy efficiency than 910CUltra-large-scale training
Ascend 970Q4 2028N+3 process, FP4 8 PFLOPS, 4 TB/s bandwidth, supports trillion-parameter modelsNext-gen AI architecture (MoE, etc.)

Technology breakthroughs:

  • Process upgrade: from N+2 (7nm-class) to N+3 (5nm-class)
  • Memory capacity doubled: from 144GB (950DT) to 288GB (960/970)
  • Energy efficiency improved: 960/970 are 30%+ better than 910C
  • Precision optimized: 970 supports FP4 precision, optimized for next-gen AI architectures (MoE, etc.)

1.5 Commercialization progress​

Shipment data:

  • 384-card SuperNode: over 500 units deployed, the only truly large-scale commercial SuperNode in China
  • 2026 shipment target: 800k chips (1M cumulative)
  • Market share: 60% of China's AI chip market

Ecosystem building:

  • CANN compiler: open-sourced end of 2025, seamless PyTorch/TensorFlow migration
  • Mind series toolchains: fully open, lowering the developer barrier
  • Ecosystem partners: over 3,000
  • Developer community: over 500k registered developers

2. Baidu Kunlun: M100 inference-specific, Tianchi SuperNode deployed​

2.1 Kunlun M100 (launched early 2026)​

Core specs:

ItemParameter
Launch dateEarly 2026 (expected Q2)
PositioningInference-specific
ArchitectureIn-house XPU-P architecture (inference-optimized)
Process7nm (SMIC N+2)
HBM capacity64 GB (inference-optimized)
TDP250 W (low-power inference)
Performance vs.1.5× P800 inference, 38% lower power

Technology innovations:

  • Adopts RISC-V open instruction set, adding 50+ AI-specific instructions
  • Compute per watt reaches 8.3 TOPS/W, 2.1× the industry average
  • Supports models from 10 billion to 100 billion parameters for inference

Commercialization progress:

  • Mass supply in Q2 2026
  • Key customers: Baidu Smart Cloud, China Merchants Bank, Southern Grid, Geely Auto
  • Priced around ¥60,000/card, clear cost-performance advantage

2.2 Kunlun M300 (launched early 2027)​

Core specs:

ItemParameter
Launch dateQ1 2027 (expected March)
PositioningUltra-large-scale multimodal training
ArchitectureIn-house XPU-P architecture (multimodal-optimized)
Process5nm (SMIC N+3)
HBM capacity256 GB HBM4
TDP500 W
ModalitiesText, images, video and other data types

Technology innovations:

  • Adopts HBM4 memory, bandwidth up to 3.2 TB/s
  • Supports FP8/FP4 low-precision compute, 2× training energy-efficiency
  • Native support for multimodal model training (text + image + video)

2.3 Tianchi 256-card SuperNode (launched June 2026)​

System specs:

ItemConfiguration
Launch dateJune 2026 (expected)
Chip count256 Kunlun P800/M100
Effective training rate97%
Inter-chip bandwidth1.2 TB/s
Validated modelsBaidu ERNIE 5.1 and other key large models

Performance breakthroughs:

  • Fully domestic SuperNode, fully autonomous and controllable from chip to network
  • 97% effective training rate, surpassing NVIDIA DGX SuperPOD's 95%
  • Training validation completed for Baidu ERNIE 5.1 and other key large models

2.4 Commercialization progress​

Shipment data:

  • P800: 150k shipped in 2025, 200k target in 2026
  • Ten-thousand-card clusters: multiple P800-based clusters delivered
  • Market share: 20% of China's AI chip market

Customer coverage:

  • External customer revenue share: over 50% in 2025
  • China Mobile AI server procurement: P800-based bids won 70%, 70%, 100% shares
  • Key customers: China Merchants Bank, Southern Grid, Geely Auto, iFlytek

IPO progress:

  • May 2026: officially launched STAR Market IPO tutoring
  • Plans "A+H" model — simultaneous A-share and Hong Kong listings
  • Valuation exceeds ¥10 billion

3. Alibaba T-Head: M890 3× performance, Zhenwu series ships 560k units​

3.1 T-Head M890 (launched Q2 2026)​

Core specs:

ItemParameter
Launch dateQ2 2026 (Alibaba Cloud Summit)
Performance3× previous generation
HBM capacity144 GB
Inter-chip bandwidth800 GB/s
Precision supportFP8, FP4 low-precision compute
PositioningFull training + inference pipeline

Technology innovations:

  • Adopts in-house ICN inter-chip protocol, inter-chip latency under 150 ns
  • Companion PCCF communication library and ICN Switch chip enable full-bandwidth interconnect of 64 chips within a single node
  • Supports FP8/FP4 low-precision compute, balancing performance and efficiency

3.2 T-Head V900 (launched Q3 2027)​

Core specs:

ItemParameter
Launch dateQ3 2027 (expected September)
Performance3× again over M890
HBM capacity216 GB
Inter-chip bandwidth1,200 GB/s
PositioningUltra-large-scale training

3.3 T-Head G900 (launched Q3 2028)​

Core specs:

ItemParameter
Launch dateQ3 2028 (expected September)
PositioningFlagship product for next-gen compute demand
Technology innovationSupports full-pipeline training of trillion-parameter models

3.4 Zhenwu series commercialization progress​

Shipment data:

  • Cumulative shipments: over 560k units as of April 2026
  • Customers served: 20+ industries, 400+ customers
  • Autonomous driving: over 130k units, 30+ customers
  • Finance: over 100k units, 150+ customers

Performance advantage:

  • At equal precision, Zhenwu series single-machine inference outperforms comparable products by 50%+ on average
  • Panjiu server SuperNode architecture supports trillion-parameter large models on a single node

Full product line:

  • Zhenwu series AI chips: training + inference
  • Yitian series CPUs: data center CPUs
  • ICN Switch interconnect chip: inter-chip interconnect
  • Camel920 400G smart NIC: high-speed networking
  • Junyue series storage controller chips: storage optimization

4. Other domestic chip vendor progress​

4.1 Cambricon MLU590 (launched Q1 2026)​

Core specs:

ItemParameter
Launch dateQ1 2026 (expected March)
ArchitectureMLUarch 09 (in-house)
Process7nm (SMIC N+2)
HBM capacity128 GB HBM3
TDP350 W
PositioningTraining + inference

Technology innovations:

  • MLUarch 09 architecture, 2× compute over MLU590
  • Supports FP8/FP4 low-precision compute, 2.5× inference energy-efficiency
  • Native MoE architecture support, 3× sparse-model inference efficiency

Commercialization progress:

  • Q1 2026: sample deliveries begun
  • Key customers: Chinese government, state-owned enterprises, research institutes
  • Day-0 adaptation of DeepSeek-V3 671B achieved

4.2 MetaX Xiyun C600 (launched Q2 2026)​

Core specs:

ItemParameter
Launch dateQ2 2026 (expected June)
ArchitectureMXMACA 3.0 (CUDA-compatible)
Process7nm (SMIC N+2)
HBM capacity128 GB HBM3
TDP350 W
PositioningTraining + inference

Technology innovations:

  • MXMACA 3.0 architecture, CUDA-compatible, low migration cost
  • Supports FP8/FP4 low-precision compute, 2× training energy-efficiency
  • Fully domestic supply chain, autonomous and controllable from chip to packaging

Commercialization progress:

  • Q2 2026: sample deliveries begun
  • Key customers: Chinese government, state-owned enterprises, research institutes
  • Adapted models include LLaMA, ChatGLM, Baichuan

4.3 Enflame S60 (launched Q3 2026)​

Core specs:

ItemParameter
Launch dateQ3 2026 (expected September)
ArchitectureGCU 3.0 (in-house)
Process7nm (SMIC N+2)
HBM capacity96 GB HBM3
TDP300 W
PositioningInference-specific

Technology innovations:

  • GCU 3.0 architecture, 2.5× inference performance over S30
  • Supports FP8 low-precision compute, 3× inference energy-efficiency
  • Hardware-level virtualization, single card split into 64 virtual instances

Commercialization progress:

  • Q3 2026: sample deliveries begun
  • Key customers: Tencent Cloud, China Telecom, China Unicom
  • Priced around ¥50,000/card

4.4 Iluvatar VA10 (launched Q4 2026)​

Core specs:

ItemParameter
Launch dateQ4 2026 (expected December)
ArchitectureHVMA 2.0 (in-house)
Process7nm (SMIC N+2)
HBM capacity64 GB HBM3
TDP250 W
PositioningVideo processing + AI inference

Technology innovations:

  • HVMA 2.0 architecture, 3× video processing performance over VA10
  • Supports 8K real-time video processing, 2× video AI inference performance
  • Hardware-level video codec, supports H.264/H.265/AV1

Commercialization progress:

  • Q4 2026: sample deliveries begun
  • Key customers: ByteDance, Kuaishou, Bilibili
  • Priced around ¥40,000/card

4.5 Hygon DCU K100 (launched Q2 2026)​

Core specs:

ItemParameter
Launch dateQ2 2026 (expected June)
Architecturex86-compatible GPGPU (in-house DCU)
Process7nm (SMIC N+2)
HBM capacity128 GB HBM3
TDP400 W
PositioningTraining + inference (x86 ecosystem-compatible)

Technology innovations:

  • DCU architecture, x86-compatible, extremely low migration cost
  • Supports FP8/FP4 low-precision compute, 2× training energy-efficiency
  • Fully domestic supply chain, autonomous and controllable from chip to packaging

Commercialization progress:

  • Q2 2026: sample deliveries begun
  • Key customers: Chinese government, state-owned enterprises, research institutes
  • Adaptation of DeepSeek-V3 671B achieved

5. 2026 domestic AI chip market landscape​

5.1 Market share (2026)​

VendorMarket shareShipments (10k units)Flagship products
Huawei Ascend60%80910C, 950PR, 950DT
Baidu Kunlun20%20P800, M100
Alibaba T-Head10%10M890, Zhenwu series
Cambricon5%5MLU590
MetaX3%3C600
Others2%2S60, VA10, K100

5.2 Technology roadmap comparison​

VendorArchitecture routeEcosystem compatibilityProcessSupply chain
Huawei AscendDa Vinci (in-house)CANN (CUDA-compatible)SMIC N+2/N+3Fully domestic
Baidu KunlunXPU-P (in-house)XPU-P (CUDA-compatible)SMIC N+2/N+3Fully domestic
Alibaba T-HeadIn-house RISC-VCUDA-compatibleSMIC N+2/N+3Fully domestic
CambriconMLUarch (in-house)CANN (CUDA-compatible)SMIC N+2Fully domestic
MetaXMXMACA (CUDA-compatible)CUDA-compatibleSMIC N+2Fully domestic
EnflameGCU (in-house)In-house ecosystemSMIC N+2Fully domestic
IluvatarHVMA (in-house)In-house ecosystemSMIC N+2Fully domestic
HygonDCU (x86-compatible)x86 ecosystem-compatibleSMIC N+2Fully domestic

5.3 Supply chain security comparison​

VendorWafer fabHBM supplyPackaging/testSupply chain rating
Huawei AscendSMICHuawei in-house HiBL/HiZQJCET/TFME⭐⭐⭐⭐⭐
Baidu KunlunSMICChangXin MemoryJCET/TFME⭐⭐⭐⭐⭐
Alibaba T-HeadSMICChangXin MemoryJCET/TFME⭐⭐⭐⭐⭐
CambriconSMICSamsung/HynixJCET/TFME⭐⭐⭐⭐
MetaXSMICSamsung/HynixJCET/TFME⭐⭐⭐⭐
EnflameSMICSamsung/HynixJCET/TFME⭐⭐⭐⭐
IluvatarSMICSamsung/HynixJCET/TFME⭐⭐⭐⭐
HygonSMICSamsung/HynixJCET/TFME⭐⭐⭐⭐

6. 2026 domestic AI chip technology breakthroughs​

6.1 Process breakthroughs​

Node2026 statusRepresentative productsNotes
7nm (N+2)Mass production910C, P800, M890SMIC N+2 mature
5nm (N+3)Mass production960, 970, M300SMIC N+3 mass production in 2026
3nmIn developmentNext-gen productsMass production expected 2028

6.2 Packaging breakthroughs​

Packaging2026 statusRepresentative productsNotes
ChipletMature910C, 950PR/DTDual-die packaging, higher yield
3D stackingMatureP800, M890HBM3e 3D stacking
CoWoSMatureAll high-end productsTSMC CoWoS
Domestic packagingMass production960, 970, M300JCET/TFME mass production

6.3 Memory breakthroughs​

Memory2026 statusRepresentative productsNotes
HBM2EMature910CSamsung supply
HBM3MatureP800, MLU590, C600Samsung/Hynix supply
HBM3eMature950PR, M890Samsung/Hynix supply
Huawei in-house HBMMass production950PR (HiBL 1.0), 950DT (HiZQ 2.0)Huawei in-house, lower cost
HBM4In developmentM300 (2027)Mass production expected 2027

6.4 Interconnect breakthroughs​

Interconnect2026 statusRepresentative productsNotes
AscendLinkMature910C, 950PR/DTHuawei in-house, 784 GB/s
XCCLMatureP800, M100Kunlun in-house, 1.2 TB/s
ICNMatureM890, V900Alibaba in-house, 800 GB/s
Domestic optical modulesMass productionAll SuperNodes6,912 LPO optical modules

7. 2026 domestic AI chip ecosystem building​

7.1 Software ecosystem comparison​

VendorSoftware stackCUDA compatibilityFramework supportDeveloper community
Huawei AscendCANN + MindSporeCompatible (low migration cost)PyTorch/TensorFlow/MaxMind500k+
Baidu KunlunXPU-P + PaddlePaddleCompatible (low migration cost)PyTorch/TensorFlow/PaddlePaddle300k+
Alibaba T-HeadIn-house + Alibaba CloudCompatible (low migration cost)PyTorch/TensorFlow/Alibaba Cloud200k+
CambriconCANN + MindSporeCompatible (low migration cost)PyTorch/TensorFlow100k+
MetaXMXMACA + CUDACompatible (very low migration cost)PyTorch/TensorFlow/CUDA50k+
EnflameIn-house GCU stackIncompatible (rewrite needed)PyTorch/TensorFlow30k+
IluvatarIn-house HVMA stackIncompatible (rewrite needed)PyTorch/TensorFlow20k+
HygonDCU + x86x86-compatible (very low migration cost)PyTorch/TensorFlow/x8650k+

7.2 Developer community building​

VendorDevelopersDocsDev toolsTraining/cert
Huawei Ascend500k+CompleteCANN ToolkitHCCP cert
Baidu Kunlun300k+CompleteXPU-P ToolkitPaddlePaddle cert
Alibaba T-Head200k+CompleteAlibaba Cloud ToolkitAlibaba Cloud cert
Cambricon100k+Fairly completeCANN ToolkitCambricon cert
MetaX50k+Fairly completeMXMACA ToolkitMetaX cert
Enflame30k+AverageGCU ToolkitEnflame cert
Iluvatar20k+AverageHVMA ToolkitIluvatar cert
Hygon50k+CompleteDCU ToolkitHygon cert

7.3 Large-model adaptation capability​

VendorDeepSeek-V3LLama 3ChatGLMBaichuanERNIEQwen
Huawei Ascend✅ Day-0✅✅✅✅✅
Baidu Kunlun✅ Day-0✅✅✅✅✅
Alibaba T-Head✅ Day-0✅✅✅✅✅
Cambricon✅ Day-0✅✅✅✅✅
MetaX✅ Day-1✅✅✅✅✅
Enflame✅ Day-3✅✅✅✅✅
Iluvatar✅ Day-7✅✅✅✅✅
Hygon✅ Day-3✅✅✅✅✅

8.1 Market drivers​

DriverDescription
Policy supportThe national 15th Five-Year Plan incorporates the compute network as a major project, with stronger policy support
Supply chain securityEscalating US export controls make domestic chips the only option
Cost advantageDomestic chips are 30-50% cheaper than imports, clear cost-performance edge
Technology breakthroughsComprehensive breakthroughs in compute, memory, and energy efficiency
Maturing ecosystemSoftware ecosystems (CANN, XPU-P, MXMACA) reach 60-70% of CUDA maturity

8.2 Market challenges​

ChallengeDescription
Process7nm/5nm still lags NVIDIA's 4nm/3nm
HBM bandwidthDomestic HBM bandwidth still lags NVIDIA
Software ecosystemEcosystem maturity still lags CUDA
Capacity bottleneckLimited SMIC N+2/N+3 capacity, supply falls short of demand
International competitionNVIDIA, AMD, Google and others keep innovating

8.3 Market forecast (2026-2030)​

YearChina AI chip market (¥B)Domestic shareDomestic market (¥B)Notes
202650035%175Ascend 60%, Kunlun 20%
202770050%350960/970 launch, breakthroughs
20281,00065%650Domestic tech approaches international level
20291,50080%1,200Domestic tech surpasses international level
20302,00090%1,800Substitution essentially complete

9. Summary and outlook​

9.1 Core conclusions​

  1. 2026 marks the full-scale explosion of domestic AI chips, as the three giants Huawei Ascend, Baidu Kunlun, and Alibaba T-Head successively launch next-gen products
  2. Significant technology breakthroughs across compute, memory, energy efficiency, and system scaling
  3. Controllable supply chain security, fully autonomous from wafer fab to packaging and test
  4. Accelerating ecosystem building, software ecosystem maturity reaching 60-70% of CUDA
  5. Rising market share, domestic chips take 35% of China's AI chip market in 2026, projected 90% by 2030

9.2 Future outlook​

Short term (2026-2027):

  • Huawei Ascend 950PR/950DT mass deployment, clear 960/970 roadmap
  • Baidu Kunlun M100 inference chip ramps, M300 ultra-large multimodal training chip launches
  • Alibaba T-Head M890 3× performance, V900 launches
  • Domestic chip market share rises to 50%

Medium term (2028-2029):

  • Huawei Ascend 960/970 mass production, 5nm process, 8 PFLOPS FP4 compute
  • Baidu Kunlun M300 mass production, supports trillion-parameter multimodal training
  • Alibaba T-Head G900 launches, becoming the next-gen compute flagship
  • Domestic tech approaches international level, market share to 80%

Long term (2030+):

  • Domestic AI chips exceed 20% of the global market
  • Transition from "following" to "running alongside" to "leading"
  • Huawei Ascend, Baidu Kunlun, Alibaba T-Head among the global TOP 5
  • China becomes a global center of AI chip technology innovation

References​

  1. Domestic AI chip "three powers" rise: substitution trend shifts from policy-driven to market-driven — Sohu
  2. 2026 domestic AI chip panorama: Huawei Ascend races Cambricon — ZPEDU
  3. Huawei unveils three-year Ascend AI chip roadmap — Jiemian News
  4. Ascend 950PR chip — Baidu Baike
  5. Ascend 950 chip — Baidu Baike
  6. Kunlun P800: technical breakthroughs and application prospects of a new-generation AI accelerator — YunTECH
  7. Kunlun P800 latest specs: P800 single-precision compute reaches 345 TFLOPS — Xueqiu

Last updated: June 10, 2026