9 posts tagged with "DeepSeek"

长期以来，质疑国产 AI 芯片的理由从来不是算力峰值，而是「软件栈到底能不能打」。SemiAnalysis 这份报告的价值在于：第一次用系统性的第三方测试回答了这个问题——答案是能，而且比想象中好。

报告核心数据​

SemiAnalysis 在《The Huawei-DeepSeek Symbiosis》中对昇腾 950PR 给出独立评估：

• GEMM 利用率 99.7%——在华为自研内核下。作为对照，这个数字意味着矩阵运算单元几乎不再有闲置周期，是内核工程成熟度的直接体现
• 绝对 GEMM 性能「介于 A100 与 MI300X 之间」——考虑 950PR 是一张推理定位卡，这个位置符合预期
• 小消息规格下 HBM 带宽与 H100 具竞争力——大消息规格下仍有差距，但推理负载的关键路径（小 batch、高并发）恰好落在有利区间

DeepSeek 的软件贡献：五个开源组件​

共生关系的另一半是 DeepSeek 的软件开源——已发布昇腾支持的组件清单：

这套组合的意义在于软件栈的去 CANN 化：当 TileLang 这样的社区 DSL 能在部分形状上打赢厂商官方栈，说明昇腾的内核开发门槛已经低到社区可以参与——这正是 CUDA 生态当年起飞的路径。

战略判断：1-2 年的窗口​

SemiAnalysis 给出的结论相当大胆：中国本土 AI 软件生态的成熟速度，可能在 1-2 年内把 AMD/NVIDIA 挤出中国云部署。

支撑证据链：

1. 硬件侧：950PR 单卡性能达到可用区间（A100–MI300X 之间），超节点弥补集群规模
2. 软件侧：99.7% GEMM 利用率 + DeepSeek 五组件证明工具链可用
3. 需求侧：出口管制让英伟达中国数据中心收入占比不足 1%，采购转向是被动但确定的
4. 标杆案例：智谱 GLM-5.3-Flash 全部线上流量已由超过 10 万张国产芯片集群承载，其 Infra Agent 两周内把端到端吞吐提升到基线的约 3 倍

需要标注的边界​

• 99.7% 利用率基于华为自研内核——华为自己写的内核跑自己的测试形状，与第三方负载下的表现是两回事；SemiAnalysis 亦仅给出了部分形状的评估
• 「1-2 年替换」是分析机构判断，且主要指推理云部署；训练侧对互联与稳定性的要求更高，替换节奏会慢得多
• 产能约束未解除：950PR 全年产能规划约 75 万颗，DeepSeek 内蒙古 16 万张 950DT 的规划（彭博社 9 月报道，未获双方完整确认）本身就会挤占其他客户的配额

结论​

对采购方，这份报告的实际意义是给「国产替代」提供了第一个有分量的第三方技术背书：选型评估中可以把昇腾推理栈从「观望」上调为「可测」。对行业观察者，TileLang vs CANN 的社区生态演化值得持续跟踪——软件栈的护城河从来不是官方文档写的，是社区代码堆出来的。

数据来源与口径说明​

• 99.7% GEMM 利用率、A100–MI300X 区间、五组件清单、1-2 年判断：SemiAnalysis《The Huawei-DeepSeek Symbiosis》（2026-10），AMD News Daily 摘要转述
• 智谱 10 万张集群与 Infra Agent 3 倍吞吐：唐杰 2026-09-17 披露，官方口径
• DeepSeek 内蒙古 16 万张 950DT：彭博社 2026-09 报道，未获双方完整公开确认
• 950PR 产能规划：产业链口径，非华为官方披露

10 月 21 日，DeepSeek 把支撑自家生产服务的推理框架 DeepInfer 全面开源（Apache 2.0）。权重开放是上一阶段的国产 AI 叙事，这一步把「推理运行时」也交了出去——同一套框架跨英伟达与昇腾部署。

一、DeepInfer 技术栈​

上线一天 GitHub Star 破 1.2 万。注意它是离线推理框架：KV Cache 系统、Prefill/Decode 分离、副本间负载均衡等压箱底组件未开源——开源的是可复现的生产方法论，不是全部护城河。

二、为什么「双底座」比开源本身更重要​

DeepInfer 对英伟达 SM100 与昇腾 950 的同时适配，是国产推理栈的一个分水岭：

1. 迁移成本降维：昇腾侧的适配不再依赖华为单方面推动，头部模型厂商自己的运行时成为「跨硬件抽象层」
2. 推理侧议价结构改变：当同一框架可以跑在两种硬件上，「换底座」的工程成本从数月缩到数周——呼应 TorchTPU/TorchNeuron 的全球趋势：软件栈中立正在成为所有阵营的共识武器
3. 国产芯片的生态新路径：与其等 CUDA 兼容，不如等头部模型框架原生适配——昇腾 950 获得的不是兼容层，而是一线生产验证

三、与 DeepInfer 同日的开源生态​

• llm-d 1.0（谷歌云+Red Hat+IBM）：Kubernetes 推理编排，把多个 vLLM/SGLang 实例池化
• TorchNeuron/TorchTPU：硬件后端原生开放

一个共同方向正在成形：推理栈的每一层（硬件后端、内核、框架、编排）都在独立开源，专有锁定只剩硬件本身——而硬件的差异化正快速集中到成本与能效上。

延伸阅读：昇腾 950PR 卡片 · 完整对比表

「大模型参数已经到头了」的说法可以收回——壁仞科技 AI 框架架构副总裁丁云帆在工博会国芯展分论坛给出的数字是：Kimi K3 已达 2.8 万亿参数，明年大模型参数或将迈入 5 万亿级别。同时，代码与 Agent 场景正驱动 1M 级超长上下文成为刚需。

一、两个数字的分量​

• 2.8T → 5T：参数规模一年内再翻近一倍。训练算力需求按近似比例膨胀——这不是「效率提升可以抵消」的量级
• 1M 上下文：长上下文对 KV Cache 显存的要求超线性增长，推理成本结构中显存容量的权重骤增

二、模型结构的演进方向​

丁云帆同时指出结构层面的迁移：业界正从 Full Attention 走向线性/混合 Attention。这条路线的逻辑链是：

Full Attention：KV Cache 随上下文线性爆炸 → 1M 上下文不可承受
线性/混合 Attention：KV Cache 压缩到近常数级 → 长上下文成本可控

DeepInfer 开源栈里的 SWA Bounded Replay、MLA 低秩注意力，都是同一路线的工程实现。模型结构的每一次瘦身，都是为更大参数量腾出算力预算——总算力需求不减反增。

三、5 万亿参数需要什么算力​

粗算账：2.8T 参数训练按 20T token 量级、MFU 40% 计，需要的 GPU 时已属「超节点专属」；5T 参数意味着：

• 单柜算力密度不足：传统八卡服务器的跨节点通信在 MoE 训练中是灾难，超节点成为必然工程解
• 柜间扩展：华为 Atlas 950（8192 卡）级集群、NVL288/NVL576 演进路线的目标场景正在这里
• 显存墙：2.8T 参数的 MoE 权重即使 FP8 也远超单柜显存——统一内存编址（超节点核心特征）不是优化项而是必需品

四、对算力供应商的含义​

1. 「5 万亿参数」是超节点最强的需求叙事：2027 年超节点渗透率的最大变量是头部模型厂商的实际训练计划
2. 推理侧同步受益：1M 上下文的 Agent 应用把大显存推理卡（MI450 432GB、Rubin 288GB、C600 144GB）的需求拉到新高度
3. 国产模型的国产算力匹配：Kimi K3 若以国产超节点集群训练，将是国产算力「训得动前沿模型」的标志性验证

延伸阅读：DeepSeek DeepInfer 开源 · 机柜即产品 · 完整对比表

每当 DeepSeek 发布新模型，国产芯片厂商的适配速度就成了软件栈成熟度的公开考试。从 0day 支持到性能调优，各家交出的答卷差异不小。

适配路径：两条技术路线​

国产芯片跑 DeepSeek 目前有两条主流路径。第一条是接入开源推理框架：通过 vLLM、SGLang 的硬件插件机制适配，昇腾的 vllm-ascend、寒武纪的 vLLM-MLU 都走这条路，好处是模型覆盖随上游社区自动扩展。第二条是自研推理栈深度调优：在自己引擎内做算子融合与并行策略定制，性能上限更高但适配周期更长。头部厂商普遍双线并行。

各家进展盘点​

以 2025 年 9 月底 DeepSeek-V3.2-Exp 发布为例：昇腾实现 0day 支持并开源全部推理代码，寒武纪同步完成适配并开源 vLLM-MLU 源码，海光宣布无缝适配加深度调优。三天内的响应速度说明头部厂商的模型结构级适配能力已经工程化、流水线化，不再是运动式攻关。

性能口碑怎么看​

公开数据要分清口径。摩尔线程联合硅基流动公布的 DeepSeek-V3 671B 单卡 Prefill 超 4000 tokens/s、Decode 超 1000 tokens/s，是集群级部署下分摊到单卡的分片吞吐，不能与单卡跑全量模型的场景混比。昇腾路线的优势在于生产案例密度：据社区年度回顾，互联网、银行、运营商等用户已基于 vLLM 加昇腾插件上线 DeepSeek V3.2、Qwen 系列、GLM 系列等十余个模型，覆盖智算平台与 MaaS 形态。寒武纪的叙事重点是把 V3.2 的稀疏 Attention 机制与自家算子结合，压低长序列场景的训推成本。

实操建议​

1. 选型前明确目标模型清单，让厂商在真实负载下跑 POC，重点看连续 72 小时压测的吞吐衰减与故障恢复。
2. 优先选有开源插件、活跃社区与公开性能报告的路线，闭源黑盒调优承诺要打折扣看。
3. 关注稀疏 Attention、FP8 等新模型特性的跟进速度，这比跑老模型快慢更能预示未来半年的体验。
4. 部署形态上多卡分片跑 MoE 大模型是当前主流，网络互联与调度软件的成熟度往往比单卡算力更影响实际效果。

2026 年 4 月 24 日，DeepSeek V4 开源并宣布全面适配华为昇腾，成为全球首个在国产 AI 芯片上完成全栈推理适配的万亿参数级开源模型。一场"开源模型拉动国产算力、国产算力反哺开源生态"的飞轮，开始加速转动。

Day 0 适配：从后补迁移到原生生长​

过去国产大模型普遍在海外 GPU 上开发训练，再向国产芯片做后补式迁移，成本高、周期长、性能损耗显著。DeepSeek V4 改写了这一模式：据公开报道，其团队联合华为、寒武纪从底层框架开展重构，基于 CANN 与 NeuWare 软件栈重写数十万量级的核心算子，针对稀疏 Attention、GroupGemm 等关键内核深度优化，对昇腾、寒武纪最新旗舰芯片实现发布即适配。

适配潮随即扩散：寒武纪依托 vLLM 框架完成 V4 双版本适配并将代码开源至 GitHub；摩尔线程联合智源 FlagOS 社区在旗舰 GPU 上完成深度优化部署；百度千帆、阿里国际等平台实现 Day 0 接入。多家券商研报预计，2026 年国产算力芯片出货量将实现翻倍增长。

飞轮的四个咬合齿轮​

这条飞轮的运转逻辑可以拆成四步：

两个补充齿轮同样关键。其一，昇腾已成为 PyTorch 官方支持的首个中国算力底座，覆盖 90 多个主流第三方社区，全球开发者无需繁琐流程即可调用昇腾算力；其二，智源研究院牵头的开源统一软件栈 FlagOS 已支持英伟达、摩尔线程、华为、海光、天数等多种芯片，直指"换一家芯片就要重学一遍"的行业痛点——软件碎片化曾是国产算力最大的隐性成本。

推理成本下降的正循环​

模型与芯片的深度协同直接压低了推理边际成本。DeepSeek 方面表示，受高端算力供给限制，V4-Pro 服务吞吐量暂时有限，待下半年昇腾 950 超节点量产后，定价有望大幅下调——算力供给的改善会直接传导为应用调用降价，进而加速 AI 向实体经济渗透。硬件形态上，国产厂商也在集体升级：在世界人工智能大会上，壁仞、沐曦、燧原、摩尔线程相继发布超节点产品，阿里云推出可公共云租用的超节点实例，"超节点"已成为国产算力的行业共识。此外，有报道指出，新一代模型在昇腾上完成混合专家并行方案验证后，推理任务实现了 1.5 至 1.73 倍的加速。

结语：飞轮转起来之后​

DeepSeek 效应的真正启示在于：生态壁垒的攻破不靠单点性能对标，而靠"模型免费开放、算力成本可控、软件迁移低摩擦"三者共振。飞轮一旦转起来就会自我强化——更多应用选择国产芯片，芯片出货放大摊薄软件投入，软件更成熟又吸引更多模型原生适配。从"能用"到"好用"仍有距离，工具链成熟度与高端产能仍是短板，但方向已经清晰：国产算力的竞争逻辑，正在从追赶参数转向经营生态。

This article is based on market analysis published by Bloomberg Intelligence and Bernstein in September 2026 and public reporting; market share and revenue figures are analyst estimates, not official disclosures.

On September 28, two Wall Street reports simultaneously painted a dramatic picture: the balance of market share in China's AI compute market completed an almost total flip within 12 months.

1. The Share Flip: 40% to 8%, Ascend to 50%​

Bernstein's latest forecast:

• NVIDIA's share of the China AI chip market will fall from about 40% to about 8% by the end of this year
• Huawei Ascend rises to nearly 50%
• The immediate trigger: China's September ban on new H20 orders from NVIDIA, which shut the last channel for NVIDIA's return to the Chinese market
• ByteDance, Alibaba, and Tencent have all placed large Ascend orders

On the timeline, this was not a sudden shock but the cumulative result of three years of tightening export controls: NVIDIA's share of the China AI chip market once peaked at 95%; as high-end chips were cut off, Chinese hyperscale cloud providers pivoted wholesale to Huawei and domestic alternatives. Even when Washington allowed H200 sales to China in January 2026, the gears of that shift had already meshed — there was no going back.

2. Ascend 950PR: The "Chinese H200" in Analysts' Eyes​

What underpins this share data is product strength itself:

• The Ascend 950PR entered mass production in March this year, with FP4 compute of up to 2 PFLOPS and 128GB of domestic HBM
• Analysts rate it as roughly on par with the NVIDIA H200
• Huawei's AI chip revenue is expected to reach $12 billion this year, up 60% from $7.5 billion in 2025

For the full specifications of the 950 series, see our earlier in-depth articles: Ascend 950PR/950DT Dual-Configuration Analysis and Ascend 960 Official Launch — the 960 was ready three quarters ahead of schedule, confirming a one-generation-per-year cadence.

3. The Model Side: Chinese Open Models Take the OpenRouter Token Pie​

While chip shares flipped, the model usage data is equally striking (OpenRouter platform statistics):

Compute and models form a positive feedback loop: cards that cannot be bought force domestic compute to scale up; scaled-up domestic compute produces cheap tokens; and cheap tokens let Chinese open-source models swallow the bulk of global inference traffic.

4. But the Stock Market Isn't Buying It​

Bloomberg Intelligence pointed out a contradiction in the same period: the valuation discount of the China Tech 8 relative to the US Magnificent Seven has widened to more than 50%, the widest of the year, and BI believes only a "true AI breakthrough" can close it. Year-to-date stock performance: Alibaba about -7%, Tencent about -23%, while US AI infrastructure leaders broadly gained more than 15%.

There are two readings of this divergence: either the market is underestimating China's AI fundamentals, or earnings execution (Alibaba's EPS missed expectations by 17.3% last quarter, Baidu's by 36%) has dragged down the delivery of the AI narrative. For industry observers, the notable point is: the compute-side data (share, revenue, shipments) has moved ahead, while application-side monetization is still on the way.

5. Three Takeaways for Compute Buyers​

1. For deployments in China, the Ascend ecosystem is already the default option. The cost of migrating from CUDA to CANN is one-time, while supply uncertainty is persistent; after the September ban, there is no compliant path to procuring new NVIDIA cards in the Chinese market.
2. Domestic compute "matching the H200" is a watershed. The implicit assumption of domestic substitution used to be "discounted performance, discounted price"; when single-card performance catches up to the H200 generation, procurement decisions return to pure TCO and supply-security calculations — we recommend using the TCO Calculator to put electricity, depreciation, and utilization into the same table.
3. The token cost curve determines the application landscape. Chinese models' 61% share of OpenRouter tokens shows that the chain of compute self-sufficiency leading to token price drops leading to application prosperity has completed its first leg; the next question is whether inference costs can keep falling.

Summary​

From 40% to 8% is not a slogan but a market reshuffle driven jointly by the ban, product strength, and the model ecosystem. For the Chinese market, the mass-production cadence of Ascend 950/960 and the maturity of the CANN ecosystem will determine whether this share can be held; for the global market, the combination of China's compute "internal circulation + open-source models going overseas" is rewriting the geographic distribution of inference traffic.

(Market share and revenue forecasts come from Bernstein / Bloomberg Intelligence analyst reports; actual figures are subject to each company's financial reports.)

On September 22, 2026, multiple media outlets reported that DeepSeek's CEO explicitly stated the company is betting heavily on Huawei chips, expects to receive a new batch of Huawei chips for large-model training, and stressed that "this choice must succeed." Following earlier reports that "DeepSeek planned to procure 160,000 Ascend 950DT units for inference," this is the most significant alignment signal yet in the domestic compute ecosystem — an upgrade from inference procurement to a training bet.

1. The Signal Chain: Four Steps to a "Training Bet"​

Piecing together the public information from the past six months, the DeepSeek–Ascend collaboration shows a clear escalation path:

The key change is the workload tier: inference deployment means "cutting costs with off-the-shelf compute," while a training bet means "staking the existence of next-generation models on domestic chips" — training clusters demand an order of magnitude more in stability, interconnect efficiency, and software-stack maturity.

2. Huawei's Ability to Deliver​

DeepSeek's willingness to commit rests on a series of verifiable progress points on Huawei's supply side (all official figures):

• One generation per year, delivered: the 950PR is in mass production, the 950DT ramps in 2026 Q4, and the 960DT moved three quarters earlier than its original 2027 Q4 plan to ready in 2027 Q1 / launch in Q2 — the roadmap's credibility validated twice in a row;
• Deployment scale: Ascend supernodes have been commercially deployed at scale in over 1,000 sets, covering internet, finance, healthcare, and manufacturing;
• Ecosystem maturity: CANN has entered routine open-source operation, with external developers exceeding 61% for the first time and 5,200 monthly active developers; there are over 40 Ascend-native training models, making it the only domestic technology route supporting pretraining;
• Training evidence chain: China Telecom's Xing 4.0-29B-A4B agentic MoE model, open-sourced in September, was announced as trained end-to-end on Ascend (company claim) — "Ascend can train large models" is no longer just Huawei's self-attestation.

See the Ascend 950DT and Ascend 960 spec pages for details; for supernode analysis, see Ascend 960 Official Launch.

3. Why DeepSeek?​

DeepSeek's choice has strong structural drivers:

1. Supply certainty: under export-control constraints, NVIDIA's flagship supply to China keeps tightening; domestic compute is the only plannable large-scale training supply;
2. Cost structure: DeepSeek has always been known for extreme engineering efficiency (the V3 training cost set the industry benchmark), and domestic compute plus supernode system efficiency fits its approach;
3. Betting on ecosystem dividends: CANN open-sourcing plus deep binding with leading model vendors means model vendors can participate in shaping the toolchain's direction — something impossible within CUDA's closed system;
4. Self-fulfilling demonstration effect: a leading lab's public commitment pulls back on Huawei's production scheduling and upstream HBM and system investment, making "must succeed" a rational commitment rather than a slogan.

4. Risks and Open Questions​

Viewed coolly, this route still has three items that need time to verify:

• Actual training scale: the quantity, model type (950DT or 960DT), and delivery schedule of the new batch of chips are all undisclosed;
• MFU methodology: the MFU / latency gains Huawei cites all come from Markov-lab simulations, with no independent third-party measurements yet; the real effective compute of a training cluster depends on long-term data from large-scale production environments;
• Per-card generation gap: the 960DT's 4 PFLOPS (FP4) versus Rubin R200's 50 PFLOPS (FP4) — the per-card gap objectively exists, and training efficiency depends on whether supernode scale and software optimization can compensate; that is precisely the decisive battleground of "system-level competition."

5. Summary​

• DeepSeek confirms betting on Huawei chips for training large models — the first training-grade commitment from a leading lab in domestic compute;
• Signal chain: CUDA→CANN migration → 160,000-unit 950DT inference procurement rumor → 960 ready ahead of schedule → training bet;
• Supporting factors: one-generation-per-year delivery, 1,000+ supernode sets, and 61% external developers in the CANN open-source ecosystem;
• What to watch: actual arrival of the new chips and training-cluster scale, third-party MFU data, and training-compute disclosure in DeepSeek's next release.

Further Reading​

• Ascend 960 spec page / Ascend 950DT spec page (this site)
• Ascend 960 Official Launch: FP4 Compute Doubled, World's First NPO Supernode
• Domestic Top Three in H2 2026: Localization Rate Breaks 40%, Heading Toward 60%
• Alibaba Zhenwu V900 Unveiled at Apsara Conference

References​

• Toutiao Tech Morning Report: DeepSeek confirms betting on Huawei chips for large-model training (2026-09-22)
• HUAWEI CONNECT 2026 official announcements (2026-09-17)
• Earlier report: DeepSeek plans to procure 160,000 Ascend 950DT units (2026-08)

This article is compiled from public reports and vendors' official statements. Details such as chip quantities and training-cluster scale are subject to subsequent disclosures from DeepSeek and Huawei.

本周（2026 年 9 月 1 日–5 日）AI 算力行业的题眼，是两条相反方向的成本重构：海外，英伟达以史上最大并购吞下开发者生态入口，同时给旗舰减配显存——"内存太贵"倒逼硬件从单卡堆料走向系统级互联；国内，DeepSeek 16 万颗昇腾订单把"国产替代"从口号变成头部实验室的资产负债表决策，字节近 300 亿美元融资为算力扩张装上杠杆。

1. DeepSeek 拟购 16 万颗昇腾 950DT：国产算力的"标志性订单"​

彭博社 9 月 4 日报道，DeepSeek 计划在内蒙古乌兰察布新建的约 1GW 数据中心部署至少 16 万颗华为昇腾 950DT，主要用于推理而非训练。按每颗约 11.1 万元的市场价估算，订单总额约 178 亿元人民币（约 25.6 亿美元）——这是迄今已知规模最大的昇腾集群，是半年前深圳首个万卡级集群的 16 倍。

三个关键读数：

• 主动选择，而非被迫替代：DeepSeek 是公认最会"榨干"算力的实验室。创始人梁文锋 7 月曾直言：华为超节点能完成 GB300 的任务、延迟没有明显差别，但约需 4 颗昇腾才抵 1 颗英伟达、技术上落后约两年。选择昇腾跑推理，是在综合成本、供应链安全与本土化后的理性决策——推理对软件生态依赖较浅，正是国产芯片的突破口；
• 瓶颈在供给端：受高端内存（HBM）短缺制约，昇腾 950DT 今年产量仅数十万颗（2026 年全部昇腾 die 计划约 160 万颗），DeepSeek 希望加购更多但产能所限，整单交付或需一年以上——国产 HBM 与先进封装成为整条链的胜负手；
• 过渡性押注：DeepSeek 同时在与中芯国际合作开发自研推理芯片，并于 6 月完成约 500 亿元融资、继续洽谈数十亿美元基建融资。国产算力的终局是多元自主，而非单一依赖。

单颗 950DT：144GB HBM、4.0TB/s 带宽、2.0TB/s 互联，原生支持 FP8/FP4/HiF8，详见昇腾 950DT 规格页。

2. 英伟达 129.3 亿美元收购 Hugging Face：买下"铲子的交易所"​

当地时间 9 月 3 日，英伟达宣布以 129.3 亿美元收购全球最大开源 AI 平台 Hugging Face，超过 2020 年 69 亿美元收购 Mellanox，成为其史上最大并购。交易含约 119 亿美元股权款与最高约 10 亿美元员工留任计划，预计 2027 年上半年完成，待监管批准。

Hugging Face 托管超 300 万个模型、50 万个数据集，服务超 1800 万名开发者。黄仁勋承诺平台继续开放中立运营：不强制绑定英伟达资源、开源属性完整保留。

解读：算力霸主的护城河从"芯片 + 网络 + 软件栈"一路修到了"模型分发 + 开发者生态"——卖铲人开始收购铲子的交易所。平台的"中立性"承诺能否兑现，将成为全球监管与竞争性云厂商持续盯防的焦点；对国内产业而言，模型托管、权重分发这类"轻资产 AI 基础设施"与芯片一样，正在成为大国科技博弈的卡点。

3. Rubin Ultra 显存减配：内存占 TCO 40% 后的算术题​

SemiAnalysis 最新报告（科创板日报 9 月 3 日转引）显示，英伟达已将旗舰 Rubin Ultra 的 HBM 配置从 HBM4E 12-Hi（384GB 档）下调至 HBM4 8-Hi（192GB），三星正配合开发 8 层产品。

• 动因：HBM/DRAM 涨价后，内存已占整机 TCO 约 40%；减配后 HBM 成本降超 50%，即便计入 2026 年 HBM 涨价预期，内存占总资本开支比例也将从 40% 压至 28%；
• 钱去了哪：转向 Scale-up 纵向扩展网络——以 NVL576 NPO 方案测算，光模块取代机架间互连后，Scale-up 网络占机架总支出比例从 4% 升至 12%；
• 连锁反应：TrendForce 显示英伟达自 2026 Q3 起并行评估 HBM4E 8-Hi / 12-Hi / HBM4 8-Hi 多套方案，部分云厂商也在考虑下调下一代自研 ASIC 的 HBM 容量。

连定价权最强的英伟达都给旗舰"减配求量"，等于官宣：当前 AI 硬件最紧的约束是内存（美光高管称新增内存供应 2028 年前难有实质放量），单卡堆料的军备竞赛告一段落，硬件价值重心正迁移到光互联、CPO、高速交换与 PTFE 背板。华泰测算 2027 年存储供需缺口将从约 -7% 收窄至 -3%——上行周期未逆转，但从"普涨"走向"结构性紧缺"。

详见本站已同步更新的 Rubin Ultra 规格页与HBM4 量产竞速分析。

4. 推理 ASIC 军备提速：谷歌"一年两款"，Jalapeño 规格落地​

• 谷歌 TPU 迭代周期从两年一代压缩至"每年两款"（华创证券 9 月 4 日研报）：第八代已拆分为训练（8t）与推理（8i）双架构，8i 把内存/算力配比拉到 8t 的 1.65 倍，专为推理放量设计；
• OpenAI Jalapeño 规格披露：6 堆栈 HBM4 共 216GB、带宽 15.4TB/s、700W，围绕投机解码设计；OpenAI 称在 DeepSeek R1 负载下 tokens/kW 达 GB300 的 1.7 倍；RTL 冻结到流片 9 个月，首批硅片后约 10 周承载 ChatGPT 流量。详见规格页（本站已更新）；
• workload-specific 时代开场：训练、推理、推荐各自长出专用芯片，上游 HBM/封装/光互联的供应节奏必须跟上"半年一代"。

5. 国内动态：字节 296 亿美元加杠杆，摩尔线程 Token 超节点投产​

• 字节跳动获约 296 亿美元融资安排（彭博 9 月 3 日），较最初约 200 亿美元目标大幅上调，用于数据中心与 AI 基建；另据产业报道正洽谈在内蒙古新增 5-6GW 算力产能——中国 AI 公司迄今最大规模基建融资之一，算力正被当作可融资、可证券化的重资产经营；
• 摩尔线程 × 趋境科技 "Token 超节点"投产（光明网 9 月 4 日）：以 MTT S500 承担 Prefill 与 KV Cache 生成、高带宽 GPU 专注 Decode 的 PD 异构方案，实测平均生成速度超 50 TPS、KV Cache 命中率超 90%、稳定性 99.9%，已承接头部模型厂商官方业务流量——超节点竞争维度从"单卡参数"转向"单位 Token 生产成本"；
• OpenAI 发布 GPT-6 Astra（9 月 4 日）：超 10 万颗 GPU 在 Stargate 集群完成训练，推理放量与超大规模集群仍是全球算力叙事主线；
• SemiAnalysis 基准：AMD MI355X 新提交在 AgentX 基准低交互区间tokens/$ TCO 击败 B300——vLLM + LMCache 软件栈的贡献首次被独立机构量化认可。

6. 市场：中美算力资产一涨一调​

美东 9 月 4 日，美国 8 月非农仅增 8.9 万人（预期 16 万），10 年期美债收益率回落至 3.78%，成长股走强：科技板块 XLK 周涨 4.2%，英伟达周涨 8.7% 收于 230.36 美元。A股 9 月 4 日反向回调：AI 算力芯片板块 -1.99%、服务器 -2.51%、超节点 -2.86%，浪潮信息跌停、寒武纪 -2.54%——基本面无恶化（博通、戴尔、中际旭创订单与财报持续验证景气），更多是交易层面获利兑现。

下周起三连催化密集：CIOE 光博会（9/9-11）→ 华为全联接大会（9/17-19）→ 云栖大会（9/22-24）。国产超节点从"发布会 PPT"到"批量交付"的成色，将迎来集中检验。

本周一句话​

内存太贵改变了所有人的算法：英伟达给旗舰减配显存、把钱投给互联；DeepSeek 用 16 万颗昇腾买推理确定性；谷歌把 TPU 迭代压到半年一代。2026 年 Q4 起，"每兆瓦/每美元 token 数"将取代"单卡 PFLOPS"，成为算力采购的第一指标。

相关链接​

• 华为昇腾 950DT 规格页
• NVIDIA Rubin Ultra 规格页
• OpenAI Jalapeño 规格页
• NVIDIA Rubin 规格页

参考资料​

• 彭博社（转引）：DeepSeek 拟部署 16 万颗昇腾 950DT（2026-09-04）
• cnBeta/今日头条：DeepSeek 据称采购 16 万颗华为昇腾 950DT 芯片（2026-09-05）
• 科创板日报：SemiAnalysis 拆解 Rubin Ultra 减配（2026-09-03）
• Natural News/DailyAI：Nvidia To Acquire Hugging Face For $12.9B（2026-09-05）
• 光明网：摩尔线程 × 趋境科技 PD 异构推理方案投产（2026-09-04）
• AGI Hunt：AI News Daily 2026-09-05

本文基于彭博社、科创板日报、SemiAnalysis、TrendForce、华创证券研报及光明网等公开报道整理。订单金额与市场数据为媒体/机构预估口径，实际以相关公司正式披露为准。

On June 5, 2026, Shenzhen announced a major piece of news: Shenzhen Hetao College, together with HIT (Shenzhen) and Huawei, used 1,000 Huawei Ascend 910C chips to successfully complete full-parameter post-training of the 1.6-trillion-parameter DeepSeek-V4-Pro large model.

This was no tentative attempt, but a milestone technological breakthrough. It proved with irrefutable engineering results that: domestic AI chips are fully capable of supporting world-class, super-large-parameter model training.

Why this matters​

The two thresholds of AI chips: "inference" and "training"​

• Inference: using an existing model to chat, write copy. Domestic chips could already do this
• Training: adjusting model parameters to learn new capabilities. Full-parameter training adjusts all 1.6 trillion parameters at once — maximum difficulty

Previously, full-parameter training of trillion-scale models was monopolized by NVIDIA H100/H200. Domestic chips could only do inference, not large-scale training.

The significance of this breakthrough: domestic compute leapt from "usable" to "useful", from "inference" to "training".

Technical details​

Training configuration​

Performance metrics​

💡 About 30% compute utilization: many feel 30% is low, but in large-model training this is already a very respectable industrial-grade level. Even with the most advanced overseas chips, many teams' actual utilization is around 40%.

Ascend 910C detailed specs​

Ascend 910C is Huawei's AI training/inference chip announced at the Huawei Analyst Conference (April 24, 2024), with a theoretical peak of 800 TFLOPS (BF16), in the same class as NVIDIA H100.

Key upgrades:

• ✅ 3× compute: from 910B's 256 TFLOPS to 800 TFLOPS
• ✅ Mature software ecosystem: torch_npu adapts PyTorch, MindSpore framework mature
• ✅ Cluster stability: 1,000 cards trained continuously with no failures (the biggest breakthrough)

Technical challenges and solutions​

Challenge 1: Memory demand of trillion-scale models​

A 1.6-trillion-parameter model needs, just for model parameters:

• FP16: 1.6T × 2 bytes = 3.2 TB
• Plus gradients and optimizer states: at least 10 TB of memory

Huawei's solution:

• Model Parallel: distribute the model across 1,000 910C chips
• ZeRO optimizer: optimize memory footprint
• Gradient accumulation: update parameters in stages

Challenge 2: Communication efficiency of thousand-card clusters​

Training with 1,000 chips, inter-chip communication becomes the bottleneck. MoE models need All-to-All communication (each expert may need to communicate with all others).

Huawei's solution:

• HCCS (Huawei Collective Communication Scheduler): in-house high-speed interconnect protocol
• Layered communication: intra-node NVLink + inter-node HCCS
• Communication-compute overlap: data transfer concurrent with computation

Challenge 3: Training stability​

Trillion-scale model training takes weeks or months; any single card failure can interrupt the entire training.

Huawei's solution:

• Fault detection and auto-recovery: real-time monitoring of card status, auto-restart and recovery on failure
• Checkpoint optimization: high-frequency training-state saves (every N steps)
• Ascend cluster management software: designed specifically for enterprise training

Competitive comparison​

Conclusion: Ascend 910C has caught up to H100 in hardware performance, still lags in software ecosystem, but this training success proves engineering feasibility.

Industry impact​

1. The "Zunyi Conference" of domestic compute​

This breakthrough is called the "Zunyi Conference" of domestic compute — from passive defense to strategic counteroffensive.

Specific impact:

• ✅ Breaks the bias that "domestic chips can only do inference"
• ✅ Proves domestic chips can train frontier models
• ✅ Provides compute foundation for domestic large models (e.g., DeepSeek-V4, ERNIE 5.0)

2. Impact on NVIDIA​

Huawei Ascend 910C completing trillion-scale training means China's AI industry is less dependent on NVIDIA.

3. Boost to the domestic chip industry​

This breakthrough will drive the entire domestic AI chip supply chain:

• Chip design: Cambricon, MetaX, Moore Threads accelerate iteration
• Wafer manufacturing: SMIC, Hua Hong get more orders
• Packaging/test: JCET, TFME benefit

Huawei Ascend roadmap (2025-2028)​

Training lessons shared​

The Shenzhen Hetao College team accumulated valuable experience:

✅ Successes​

1. Progressive training: start from small models (7B), gradually scale to 1.6T
2. Mixed-precision training: BF16 main + FP32 gradient accumulation
3. Communication optimization: All-to-All overlap with computation
4. Fault recovery: save checkpoint every 1,000 steps

⚠️ Challenges encountered​

1. Memory fragmentation: severe fragmentation over long training, needs periodic cleanup
2. Communication bottleneck: MoE All-to-All takes 30%+ of training time
3. Software bugs: torch_npu occasional memory leak, needs training process restart

Related chips​

• Huawei Ascend 910C - the chip used in this training
• Huawei Ascend 910B - previous-gen chip
• Huawei Ascend 950PR - next-gen inference chip
• NVIDIA H100 - benchmark product
• NVIDIA H200 - benchmark product (HBM3e upgrade)

References​

• Sina Finance: Huawei Ascend 910C completes DS-V4-Pro training
• Tencent News: Milestone! Ascend 910C completes 1.6-trillion-param model training
• 21IC: Huawei chip completes full-parameter post-training of 1.6T AI model

This article is compiled from public reports. Salute to the teams at Shenzhen Hetao College, HIT (Shenzhen), and Huawei — you proved the feasibility of China's AI compute with engineering results.