Goldman Sachs: What DeepSeek V4 Means for China's AI Industry
高盛：DeepSeek V4 对中国 AI 产业的意义
Goldman Sachs believes that DeepSeek V4's core significance lies in enabling more complex AI agent applications at significantly lower costs, opening up new possibilities for large-scale AI adoption. The model's architectural innovations have drastically reduced long-context inference costs, and its clear bet on Huawei's Ascend 950 chips is expected to drive further API price cuts in the second half of 2026. As competition in China's AI model market intensifies, programming capabilities and multimodal performance will become the key differentiators. Goldman Sachs maintains its "Outperform" rating on the cloud computing and data center sectors.
高盛认为，DeepSeek V4 的核心意义在于以更低成本支撑更复杂的智能体应用落地，为 AI 规模化应用打开了新空间。该模型通过架构创新大幅降低了长上下文推理成本，同时明确押注华为昇腾 950 芯片，预计将在 2026 年下半年推动 API 价格进一步下降。随着中国 AI 模型市场竞争加剧，编程能力和多模态表现将成为核心差异化因素。高盛维持对云计算和数据中心板块的 "首选" 评级。
Architectural Breakthroughs: Dramatic Efficiency Gains for Long Contexts
架构突破：长上下文效率大幅提升
DeepSeek V4 is released in two versions: Pro and Flash. The flagship Pro version has 1.6 trillion parameters (49 billion active parameters), while the lighter Flash version has 284 billion parameters (13 billion active parameters). Both models support a 1 million-token context window, matching the world's leading models, but require significantly less memory and KV cache.
DeepSeek V4 分为 Pro 和 Flash 两个版本。旗舰版 Pro 拥有 1.6 万亿参数（激活参数 490 亿），轻量版 Flash 拥有 2840 亿参数（激活参数 130 亿）。两款模型均支持 100 万词元的上下文窗口，与全球顶尖模型持平，但所需内存和 KV 缓存大幅减少。
According to Goldman Sachs, V4 Pro requires only 27% of the FLOPs and 10% of the KV cache per token compared to DeepSeek V3.2 in 1 million-context scenarios. The Flash version is even more efficient, with FLOPs reduced to 10% and KV cache to 7%.
根据高盛报告，在 100 万上下文场景下，V4 Pro 每个词元的浮点运算量仅为 DeepSeek V3.2 的 27%，KV 缓存占用仅为 10%；Flash 版本效率更高，浮点运算量降至 10%，KV 缓存压缩至 7%。
These efficiency gains are achieved through three key innovations: a hybrid attention architecture combining Compressed Sparse Attention (CSA) and Heavy Compression Attention (HCA), an mHC mechanism for improved training stability, and the use of the Muon optimizer alongside AdamW. These improvements are particularly valuable for long-duration agent tasks that require processing large amounts of context.
这些效率提升通过三项关键创新实现：结合压缩稀疏注意力 (CSA) 和重度压缩注意力 (HCA) 的混合注意力架构、增强训练稳定性的 mHC 机制，以及与 AdamW 配合使用的 Muon 优化器。这些改进对于需要处理大量上下文的长周期智能体任务尤为重要。
Huawei Ascend 950: Paving the Way for Lower API Prices
华为昇腾 950：为 API 降价铺路
A key signal from the V4 launch is DeepSeek's explicit integration of Huawei's Ascend 950 super nodes into its commercial roadmap. DeepSeek expects API prices for the V4 Pro to drop significantly once Huawei Ascend 950 enters mass production in the second half of 2026.
V4 发布的另一个重要信号是 DeepSeek 明确将华为昇腾 950 超级节点纳入其商业路线图。DeepSeek 预计，随着华为昇腾 950 在 2026 年下半年实现大规模供货，V4 Pro 的 API 定价将迎来显著下降。
This development has two important implications: first, it will further strengthen DeepSeek's cost competitiveness and enable broader application adoption; second, it confirms the trend of leading Chinese AI models migrating to domestic computing power amid ongoing chip restrictions.
这一进展具有双重意义：首先，它将进一步强化 DeepSeek 的成本竞争力，为更广泛的应用落地创造条件；其次，在芯片持续受限的背景下，它证实了中国顶尖 AI 模型向国产算力迁移的趋势。
China's AI Model Competition Enters Differentiation Phase
中国 AI 模型竞争进入差异化阶段
The open-source release of DeepSeek V4 has triggered a new wave of model updates in China. Recent launches include Kimi K2.6, Alibaba's Qwen3.6-Max, Tencent's Hy3 preview, and Xiaomi's V2.5, with MiniMax's M3/Hailuo expected in May.
DeepSeek V4 的开源发布引发了中国 AI 模型阵营的新一轮密集更新。近期推出新模型的厂商包括 Kimi K2.6、阿里巴巴 Qwen3.6-Max、腾讯 Hy3 预览版和小米 V2.5，MiniMax 的 M3/Hailuo 预计将于 5 月推出。
Goldman Sachs identifies two key differentiators that will determine pricing power: programming/task completion rates, where Zhipu's GLM models excel, and multimodal capabilities, where ByteDance, Alibaba, and MiniMax are leading the way.
高盛认为，未来决定各模型定价权的两个关键差异化因素是：编程 / 任务完成率（智谱的 GLM 模型在代码能力方面名列前茅）和多模态能力（字节跳动、阿里巴巴和 MiniMax 在这一方向投入最为深入）。
Independent AI players like MiniMax benefit from high organizational efficiency and can achieve 40% gross margins even with low API prices. Internet giants like ByteDance, Tencent, and Alibaba have strong cash flows from core businesses, making them better positioned to invest in AI infrastructure. Notably, Tencent and Alibaba are reportedly in talks to invest in DeepSeek at a valuation of over $20 billion.
独立 AI 厂商如 MiniMax 组织效率高、决策链路短，即使基础文本 API 定价极低，仍能实现 40% 的毛利率。字节、腾讯、阿里等互联网大厂核心业务现金流充沛，更适合布局 AI 基础设施。值得关注的是，有报道称腾讯和阿里巴巴正洽谈以逾 200 亿美元估值投资 DeepSeek。
Investment Outlook: Cloud Computing and Data Centers Remain Top Picks
投资展望：云计算与数据中心仍是首选
Goldman Sachs maintains its view that cloud computing and data centers are the top sub-sectors in China's internet industry, driven by growing token demand, improving cloud pricing power, and increasing penetration of consumer AI assistants.
高盛维持云计算与数据中心为中国互联网首选子赛道的判断，驱动因素包括词元需求持续增长、云服务定价能力改善以及消费级 AI 助手渗透率提升。
In the enterprise cloud market, Alibaba leads with the largest external AI cloud revenue. In the consumer market, ByteDance currently has the highest daily token usage for AI chatbots, with China's overall AIGC app DAUs growing 36% month-on-month in March 2026.
在企业云市场，阿里巴巴以最大的外部 AI 云营收规模领跑；在消费市场，字节跳动目前是 AI 聊天机器人日均词元用量最大的平台，中国 AIGC 应用整体日活用户在 2026 年 3 月环比增长高达 36%。
Goldman Sachs recommends four core stocks to capture China's AI infrastructure expansion: GDS Holdings, 21Vianet, Alibaba, and Kingsoft Cloud. E-commerce and mobility are the second-tier picks, followed by AI model players and then gaming and entertainment.
高盛推荐四只核心标的以把握中国 AI 基础设施扩张红利：万国数据、世纪互联、阿里巴巴和金山云。电商与出行是第二梯队，第三梯队是 AI 模型相关标的，第四梯队是游戏与娱乐赛道。
以下是本群专属福利，可按需自取：
【海外券商】
❥ 复星证券专属公开渠道：BPMQTY
❥ 复星证券公开页面：
https://h5.fotechwealth.com/pages/startAccount.html?channel=030003&aeCode=B2&invitationCode=BPMQTY&langType=zhCn
❥ 致富 专属公开渠道：DQ3638
❥ 致富 公开页面：https://www.chiefgroup.com.hk/hk/open?rl=ECR&PromoCode=DQ3638
❥ 盈立（香港）公开渠道码：261773
公开页面：https://hk.usmartbroker.com/webapp/marketing/jump-register.html?register-ct=ic®ister-cid=5998&ICode=261773#/register
❥ 盈立（新加坡）nyw2w4
❥公开页面：https://m.usmartsg.com/promo/overseas/sg-register.html?langType=1&HCode=nyw2w4#/marketing-register
❥ 盈透证券IBKR公开页面：
------------------------------------
【银行】
❥ 中银香港专属公开渠道：M23756872072
❥ ZA众安银行公开渠道：VUJ474
❥ 恒生银行公开渠道: GXHDSQ
❥信银国际（中信银行国际）公开渠道：1G2T4Y2
❥ 工银亚洲专属公开渠道：26042152716854
❥ 天星银行公开页面https://app.airstarbank.com/apph5/friendsInvite_appLink.html#/detail
❥ 汇立银行公开渠道：FPJYXH
❥美国华美银行公开页面 （300红包+103美金（你自己付））https://buy.italkbb.com/chs/cu/PkM?q=874266
❥ 英国Ifast公开页面https://www.ifastgb.com/tellafriend/yuanyuanc6480
❥ 德国N26公开渠道：yuanyuac44482c
公开页面https://n26.com/r/yuanyuac44482c?cid=0JY&lang=en
❥瑞士杜高斯贝银行公开页面：https://www.dukascopy.bank/swiss/open-mca-account/?ref=&lang=zh
❥ 澳门蚂蚁银行LPE22RM9
https://render.alipay.com/p/c/180020570000149153/index.html?invite-code=LPE22RM9
❥ Wise邀请链接:https://wise.com/invite/arhn/4v9di89
------------------------------------
【数字货币】
❥ B安专属公开渠道：CPA_00N0KYMEP1
❥ B安公开页面：https://www.bsmkweb.cc/activity/referral-entry/CPA?ref=CPA_00N0KYMEP1&utm_source=referral_entrance
❥ O意专属公开渠道：67834863
❥ O易公开页面：https://www.nzdcvhpxqwy.com/join/67834863
❥ H币公开页面：https://www.htx.com.ph/invite/zh-hk/1f?invite_code=48qfd223
❥OSL HK公开页面：https://trade-hk.osl.com/invite/activities?invitecode=7g7xF
❥OSLGlobal公开页面：https://trade-hk.osl.com/register?channel=cjySdK
欢迎加博主绿泡泡：NBWcat ，也欢迎进群交流更多用卡攻略：
