Goldman Sachs: What Does DeepSeek V4 Mean for Chinese AI?

Goldman Sachs believes that the core significance of DeepSeek V4 lies in supporting more complex intelligent agent applications at lower cost, thereby opening up new space for large-scale AI deployment. According to Zhuifeng Trading Desk news, on April 24, the Goldman Sachs Ronald Keung team released a research report stating that the newly open-sourced V4 model is a continuation of DeepSeek’s efficiency-first, open-source strategy. On the technical level, V4 achieves a substantial reduction in long-context window costs via architectural upgrades and explicitly bets on domestically produced Huawei chips. On the market level, this release accelerates the feverish competition among China’s AI models; programming capability, task completion rates, and multimodality will become the critical dividing lines for pricing power. Goldman Sachs maintains its recommendation for the cloud computing and data center sectors, believing the continued improvement in compute cost efficiency will accelerate AI application penetration. The dual engines of enterprise AI agent growth and consumer AI assistants will support the continued strengthening of cloud services pricing power. V4 Architectural Upgrade: Longer Context with Less Memory DeepSeek V4 is released in two versions: Pro and Flash. The Pro version is flagship scale, with 1.6 trillion parameters (49 billion activated parameters); the Flash version is comparatively 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 memory and KV cache requirements are greatly compressed. According to Goldman Sachs, in the 1 million context scenario, V4 Pro requires only 27% of the FLOPs per token compared to DeepSeek V3.2, and just 10% of KV cache; V4 Flash is even more aggressive, reducing FLOPs to 10% and KV cache to 7%. This efficiency leap is achieved via three key architectural innovations: - Hybrid Attention Mechanism: V4 introduces a hybrid architecture of Compressed Sparse Attention (CSA) and Highly Compressed Attention (HCA). CSA first compresses the KV cache along the sequence dimension and then performs sparse attention calculations; HCA uses more aggressive compression while retaining dense attention. Together, they greatly reduce temporary memory requirements for long inputs. - Training Stability: V4 introduces the mHC mechanism, enhancing stability of information transfer across multi-layer networks. - Uses Muon as the main training optimizer (AdamW is retained for some modules), to adapt to a more complex network architecture than V3 and enhance convergence during training. Goldman Sachs notes that these efficiency gains are especially significant in long-duration task scenarios, such as agent tasks requiring large amounts of context. Notably, DeepSeek currently focuses on foundational text models, while major internet giants such as Alibaba, ByteDance, MiniMax, and independent model players favor the multimodal/all-modal route, showing clear differentiation in their AGI exploration paths. Domestic Chip Accelerates Landing: Huawei Ascend 950 Paves the Way for Price Reduction Another important signal from this V4 release is DeepSeek’s clear inclusion of mass production of Huawei Ascend 950 super nodes into its commercial roadmap. DeepSeek expects that, as Huawei Ascend 950 super nodes are widely supplied in the second half of 2026, the API pricing for V4 Pro will see a significant decrease. Goldman Sachs reports that this statement has dual implications: 1) DeepSeek's cost competitiveness will further strengthen, creating conditions for broader application. 2) Against the backdrop of tightening chip supplies, the migration of China’s top AI models to domestic compute power receives explicit endorsement from leading players. From current pricing, Goldman Sachs data shows V4 Pro's API pricing on mainstream platforms is already competitive, and as domestic compute supply expands, this advantage is expected to grow further in the second half of 2026. Domestic AI Model Competition Enters Differentiation Stage DeepSeek V4’s open-source release rapidly triggered a new round of dense follow-ups within the Chinese AI model camp. Goldman Sachs notes recent entrants with new models include: Kimi K2.6, Alibaba Qwen3.6-Max, Tencent Hy3 Preview, Xiaomi V2.5, and the soon-to-be launched MiniMax M3/Hailuo in May. According to Goldman Sachs, future key differentiators deciding pricing power will be concentrated in two dimensions: - Programming/task completion success rate: Zhipu’s GLM model ranks among the top in code capabilities. - Multimodal capabilities: ByteDance, Alibaba, and MiniMax have invested most deeply in this direction. The report notes that the strengths and weaknesses of the two types of players are clear: - Independent AI players such as MiniMax are highly efficient in organization and have short decision chains; even with very low foundational text API prices, they can still achieve 40% gross margins, per Goldman Sachs forecasts. - Large internet companies like ByteDance, Tencent, and Alibaba have strong cash flow for their core businesses, making them more suitable to deploy AI infrastructure and cloud tracks. They require independent AI team incentive plans to retain talent—for example, ByteDance's Doubao team already has independent incentives. Goldman Sachs reports, citing news sources, that Tencent and Alibaba are in talks to invest in DeepSeek at a valuation exceeding $20 billion, while Zhipu and MiniMax’s latest valuations are about $53 billion and $31 billion respectively. This potential deal reflects the logic of giants competing for scarce top-level AI capabilities. Top Track Logic Unchanged: Cloud Computing and Data Centers Goldman Sachs maintains that cloud computing and data centers are the preferred sub-sectors for China’s internet, based on the logic that: - Continued growth in AI token demand will drive up cloud service orders; - Growth in enterprise clients and AI agents is improving pricing power for cloud/token services; - Ongoing penetration by consumer-grade AI assistants contributes incremental demand. In the To-B enterprise cloud market, Alibaba leads with the largest external AI cloud revenue scale; in the To-C consumer market, ByteDance currently has the highest daily token usage for AI chatbots. China's AIGC applications’ DAU remains strong, with a month-on-month growth rate of 36% in March 2026. Goldman Sachs continually emphasizes four core recommended stocks as key allocations to capture the dividend from China’s AI infrastructure expansion: GDS Holdings, 21Vianet, Alibaba, and Kingsoft Cloud. Additionally, the second tier consists of e-commerce and mobility tracks, the third tier is AI model-related stocks, and the fourth tier is the gaming and entertainment tracks. ~~~~~~~~~~~~~~~~~~~~~~~~ The above highlights are from Zhuifeng Trading Desk. For more in-depth interpretations, including real-time analysis and frontline research, please join [Zhuifeng Trading Desk Annual Membership]. Risk Disclosure and Disclaimer The market has risks; investing must be cautious. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions contained herein are suitable for their particular circumstances. Invest accordingly at your own risk.