--- title: "Goldman Sachs: What Does DeepSeek V4 Mean for China's AI?" type: "News" locale: "en" url: "https://longbridge.com/en/news/284099030.md" description: "Goldman Sachs believes that DeepSeek V4, through architectural innovations such as mixed attention, significantly reduces the cost of long-context reasoning. Its core significance lies in supporting the deployment of complex agent applications at a lower cost. V4 has explicitly bet on the Huawei Ascend 950, and following its mass production in the second half of 2026, API pricing is expected to drop significantly. Domestic AI model competition is intensifying, with programming capabilities and multimodal features becoming key differentiators. Goldman Sachs maintains its top-rated recommendation for cloud computing and data centers" datetime: "2026-04-26T04:33:24.000Z" locales: - [zh-CN](https://longbridge.com/zh-CN/news/284099030.md) - [en](https://longbridge.com/en/news/284099030.md) - [zh-HK](https://longbridge.com/zh-HK/news/284099030.md) --- # Goldman Sachs: What Does DeepSeek V4 Mean for China's AI? Goldman Sachs believes that **the core significance of DeepSeek V4 lies in supporting the deployment of more complex agent applications at a lower cost, thereby opening up new space for the scaling of AI applications.** According to Trading Desk news, on April 24, the Ronald Keung team at Goldman Sachs released a research report stating that the newly open-sourced V4 model continues DeepSeek's efficiency-first, open-source strategy. **On the technical front**, V4 achieves significant cost reductions for long context windows through architectural upgrades and explicitly bets on domestic Chinese chips. **On the market front**, this launch accelerates the intensification of China's AI model competition, where programming capabilities, task completion rates, and multimodal features will become the core dividing lines for pricing power. Goldman Sachs maintains its recommended rating for the **cloud computing** and **data center** sectors. Continued improvements in compute cost efficiency will drive faster penetration of AI applications. The dual drive of enterprise-side AI agent growth and consumer-side AI assistants will support a continuous rise in cloud service pricing power. ## V4 Architectural Upgrade: Supporting Longer Contexts with Less Memory DeepSeek V4 was released in two versions: Pro and Flash. The Pro version is the flagship scale, with parameters reaching 1.6 trillion (49 billion activated parameters); the Flash version is relatively lightweight, with 284 billion parameters (13 billion activated parameters). **Both models support an ultra-long context window of 1 million tokens, matching top US models (SOTA), but with significantly compressed memory requirements and KV cache usage.** According to the Goldman Sachs report, in a 1-million-token context scenario, the floating-point operations (FLOPs) required per token inference for V4 Pro are only 27% of those for DeepSeek V3.2, and KV cache usage is reduced to 10%; V4 Flash is even more aggressive, reducing FLOPs to 10% and compressing KV cache to 7%. This leap in efficiency is achieved through three key architectural innovations: > - **Regarding mixed attention mechanisms**, V4 introduces a hybrid architecture combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA). CSA first compresses the KV cache along the sequence dimension before performing sparse attention calculations, while HCA adopts more aggressive compression but retains dense attention. **Together, they significantly reduce the temporary memory required for long inputs.** > - **Regarding training stability**, V4 introduces the mHC mechanism to enhance the stability of information transmission across multiple network layers; > - **It also adopts Muon as the primary training optimizer** (retaining AdamW for some modules) to adapt to more complex network architectures than V3 and improve convergence quality during training. Goldman Sachs points out that these efficiency gains are particularly significant for long-duration task scenarios, such as long-cycle agent tasks requiring the processing of large amounts of context. Notably, DeepSeek currently remains focused on foundational text models, whereas internet giants like Alibaba, ByteDance, MiniMax, and independent model players tend to favor multimodal/full-modal routes, showing clear divergence in their paths toward AGI exploration. ## Domestic Chips Accelerate Deployment; Huawei Ascend 950 Paves the Way for Price Reductions Another important signal from this V4 launch is that DeepSeek has explicitly included the mass production of Huawei Ascend 950 super-nodes in its commercial roadmap. DeepSeek expects **that as the Huawei Ascend 950 super-nodes achieve large-scale supply in the second half of 2026, the API pricing for the V4 Pro version will see a significant decline.** According to the Goldman Sachs report, this statement carries a dual meaning: > - **First**, DeepSeek's cost competitiveness will be further strengthened, creating conditions for broader application deployment; > - **Second**, amidst continued chip tightening, the trend of top Chinese AI models migrating to domestic compute infrastructure receives explicit endorsement from industry leaders. From current pricing perspectives, Goldman Sachs data shows that V4 Pro already possesses competitive pricing on mainstream API platforms, and **as domestic compute supply expands, this advantage is expected to widen further in the second half of 2026.** ## Domestic AI Model Competition Enters a Differentiation Phase The open-source release of DeepSeek V4 quickly triggered a new wave of intensive follow-ups by China's AI model camp. According to Goldman Sachs' analysis, recent competitors launching new models include Kimi K2.6, Alibaba Qwen3.6-Max, Tencent Hy3 preview version, Xiaomi V2.5, and the upcoming MiniMax M3/Hailuo expected in May. In Goldman Sachs' view, the key differentiating factors determining pricing power for each model in the future will focus on two dimensions: > - **Programming/Task Completion Success Rates**: Zhipu's GLM models rank highly in code capability rankings; > - **Multimodal Capabilities**: ByteDance, Alibaba, and MiniMax have invested most deeply in this direction. The research report points out that the strengths and weaknesses of the two types of players are clear: > - **Independent AI Players** like MiniMax enjoy high organizational efficiency and short decision chains; even with extremely low base text API pricing, **they can still achieve gross margins of 40%**, according to Goldman Sachs' forecast. > - **Internet Giants** like ByteDance, Tencent, and Alibaba possess abundant cash flow from core businesses, making them better suited to layout AI infrastructure and cloud tracks. They need to set up independent AI teams with incentive plans to retain talent; for example, ByteDance's Doubao team already has independent incentives. Of particular note is that the Goldman Sachs report cites news reports stating that Tencent and Alibaba are negotiating to invest in DeepSeek with a valuation exceeding $20 billion, while Zhipu and MiniMax have latest valuations of approximately $53 billion and $31 billion respectively. **This potential transaction reflects the logic of tech giants competing for scarce top-tier AI capabilities.** ## Preferred Sector Logic Remains Unchanged: Cloud Computing and Data Centers Goldman Sachs maintains its judgment that cloud computing and data centers remain the preferred sub-sectors for China's internet industry, based on the following logic: > - Continuous growth in AI token demand will drive increased cloud service procurement; > - Growth in enterprise customers and AI agents is driving improvements in cloud/token pricing power; > - Continuous penetration of consumer-grade AI assistants contributes incremental demand. In the To-B enterprise cloud market, Alibaba leads in external AI cloud revenue scale; in the To-C consumer market, ByteDance is currently the platform with the largest daily average token usage for AI chatbots. Overall DAU for Chinese AIGC applications remains strong, with a month-on-month growth rate of 36% in March 2026. **Regarding key investment targets**, Goldman Sachs continues to emphasize four core holdings: GDS Holdings, 21Vianet, Alibaba, and Kingsoft Cloud, as the core allocation directions to capture the dividends of China's AI infrastructure expansion. Furthermore, **the second tier** consists of e-commerce and mobility sectors, **the third tier** comprises AI model-related targets, and **the fourth tier** includes gaming and entertainment sectors. ### Related Stocks - [GS.US](https://longbridge.com/en/quote/GS.US.md) - [GS-D.US](https://longbridge.com/en/quote/GS-D.US.md) - [GS-C.US](https://longbridge.com/en/quote/GS-C.US.md) - [GS-A.US](https://longbridge.com/en/quote/GS-A.US.md) - [DPSK.NA](https://longbridge.com/en/quote/DPSK.NA.md) - [HUAWEI.NA](https://longbridge.com/en/quote/HUAWEI.NA.md) - [BABA.US](https://longbridge.com/en/quote/BABA.US.md) - [09988.HK](https://longbridge.com/en/quote/09988.HK.md) - [00700.HK](https://longbridge.com/en/quote/00700.HK.md) - [01810.HK](https://longbridge.com/en/quote/01810.HK.md) - [00100.HK](https://longbridge.com/en/quote/00100.HK.md) - [02513.HK](https://longbridge.com/en/quote/02513.HK.md) - [BYTED.NA](https://longbridge.com/en/quote/BYTED.NA.md) - [GDS.US](https://longbridge.com/en/quote/GDS.US.md) - [09698.HK](https://longbridge.com/en/quote/09698.HK.md) - [VNET.US](https://longbridge.com/en/quote/VNET.US.md) - [KC.US](https://longbridge.com/en/quote/KC.US.md) - [03896.HK](https://longbridge.com/en/quote/03896.HK.md) - [W4VR.SG](https://longbridge.com/en/quote/W4VR.SG.md) - [89988.HK](https://longbridge.com/en/quote/89988.HK.md) - [HBBD.SG](https://longbridge.com/en/quote/HBBD.SG.md) - [TCEHY.US](https://longbridge.com/en/quote/TCEHY.US.md) - [80700.HK](https://longbridge.com/en/quote/80700.HK.md) - [HTCD.SG](https://longbridge.com/en/quote/HTCD.SG.md) - [TCTZF.US](https://longbridge.com/en/quote/TCTZF.US.md) - [XIACY.US](https://longbridge.com/en/quote/XIACY.US.md) - [81810.HK](https://longbridge.com/en/quote/81810.HK.md) - [HXXD.SG](https://longbridge.com/en/quote/HXXD.SG.md) ## Related News & Research - [Investment bank Lincoln International files for US IPO](https://longbridge.com/en/news/284057313.md) - [Goldman Invests $50 Million In BLP Digital As AI Finance Demand Grows](https://longbridge.com/en/news/283705657.md) - [AI smart glasses will help visually impaired runners take on the London Marathon](https://longbridge.com/en/news/284039660.md) - [08:02 ETChapsVision Unveils ChapsAgents: Enterprise Platform for Trustworthy Agentic AI](https://longbridge.com/en/news/283660375.md) - [What AI bubble? Why JPMorgan says AI stocks have momentum back.](https://longbridge.com/en/news/283517214.md)