DeepSeek V4 release strengthens long-context efficiency; cloud and data centers remain Goldman Sachs' top China internet sub-sector pick

Overview

This report discusses the impact of the DeepSeek V4 release on China's AI models, cloud services, and data centers. Goldman Sachs believes DeepSeek V4 is not an isolated "defining moment," but rather a continuation of DeepSeek's progress in computational efficiency and the open-source path. The core significance of V4 lies in supporting a 1M-token context with lower long-context memory and inference costs, and it may further reduce the price of the Pro version in 2H 2026 with the help of domestic compute supply.

Core views

Goldman Sachs' core views are: first, DeepSeek V4's efficiency gains in long-context scenarios could lower the cost of agents, long-duration tasks, and enterprise AI applications; second, competition among Chinese AI models will intensify further, with model pricing power depending more on coding ability, task completion rates, and multimodal capabilities; third, cloud and data centers remain the sub-sector within China's internet sector that benefits the most from expanding AI token demand; fourth, internet giants are better positioned to capture AI infrastructure and cloud opportunities thanks to cash flow from core businesses, but they need more independent incentive mechanisms for AI chip and model teams to retain top talent; fifth, independent AI companies have advantages over internet giants in organizational efficiency and speed of decision-making.

Analysis framework

The report analyzes the industry impact of DeepSeek V4 from dimensions including model technical parameters, API pricing, context length, inference FLOPs, KV cache ratio, OpenRouter token usage rankings, AIGC application DAU, and preferences across China internet sub-sectors, and compares DeepSeek with Kimi, Qwen, Tencent Hy3, Xiaomi MiMo, MiniMax, and U.S. SOTA models.

Methodology notes

• Model efficiency comparisonLong-context inference efficiency framework

  Use 1M context, inference FLOPs, and KV cache to measure a model's cost efficiency in long-text tasks.

  Through architectural upgrades such as CSA/HCA hybrid attention, DeepSeek V4 significantly reduces computation and cache requirements relative to V3.2 under a 1M context, making it more suitable for long-cycle tasks, agent workflows, and enterprise long-text applications.

• Industry preferenceChina internet sub-sector preference ranking

  Evaluate sub-sector attractiveness based on AI token demand, cloud pricing, enterprise AI agent growth, and consumer AI assistant growth.

  In its April 2026 update, Goldman Sachs continued to rank cloud and data centers as its preferred sub-sector, believing this segment benefits most directly from AI application diffusion and improving cloud/token pricing.

• Goldman Sachs disclosure frameworkGS Factor Profile

  Compare stock attributes across Growth, Financial Returns, Multiple, and Integrated dimensions.

  The disclosure appendix explains that Growth is based on forward sales, EBITDA, and EPS growth; Financial Returns are based on ROE, ROCE, and CROCI; Multiple is based on valuation metrics such as P/E, P/B, and EV/EBITDA; and Integrated combines growth, returns, and valuation.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• Cloud and data centers

  Goldman Sachs' top China internet sub-sector pick, directly benefiting from expanding AI token demand and improving cloud/token pricing.

  Strengths

  AI application diffusion, growth in enterprise AI agents, rising demand for consumer AI assistants, and declining long-context costs.

  Weaknesses

  Compute supply, price competition, and capital expenditure cadence may affect earnings realization.

  Comparison

  Compared with sub-sectors such as gaming, e-commerce, and local services, cloud and data centers ranked first in the April 2026 preference ranking.

  Risks

  Model prices falling too quickly, intensifying cloud service competition, compute build-out missing expectations, or insufficient utilization.

• GDS, VNET, Alibaba, Kingsoft Cloud

  Listed by the report as key ideas within the preferred cloud and data center sub-sector.

  Strengths

  Highly correlated with expanding demand for AI infrastructure, cloud resources, and data centers.

  Weaknesses

  Specific company financial and valuation details are disclosed only to a limited extent in this input.

  Comparison

  As direct recipients of AI token demand growth, they are closer to infrastructure demand than pure application-layer companies.

  Risks

  Timing of AI demand realization, pricing pressure, capital expenditure, and policy/supply-chain constraints.

• Tencent Holdings (0700.HK)

  An internet giant and participant in the AI model/cloud ecosystem; related research views Hy3 preview as an important step in Tencent's AI upgrade.

  Strengths

  Strong cash flow from core businesses provides the resource base to capture AI infrastructure and cloud opportunities.

  Weaknesses

  Needs independent incentive mechanisms to attract and retain top AI chip/model talent.

  Comparison

  Relative to independent AI companies, internet giants have stronger cash flow and ecosystems; relative to independent players such as DeepSeek, organizational speed and incentive mechanisms may be challenges.

  Risks

  AI model competition, talent incentives, returns on capital investment, and uncertainty over model pricing power.

• Xiaomi Corp. (1810.HK)

  A participant in China's AI model competition; the report mentions the release of the Xiaomi V2.5/MiMo-V2.5 series.

  Strengths

  Faster model update cadence may drive commercialization and expansion of AI applications.

  Weaknesses

  Compared with competitors such as DeepSeek, Alibaba, Tencent, and MiniMax, its model ecosystem and cloud infrastructure position still need continued validation.

  Comparison

  MiMo-V2-Pro ranks high on OpenRouter's rolling 30-day token usage leaderboard, indicating strong usage.

  Risks

  Risks related to model capability, open-source cadence, commercialization, and price competition.

• MiniMax Group (0100.HK)

  An independent AI company; the report mentions its multi-/omni-modal strategy and potential M3/Hailuo release.

  Strengths

  Strong organizational efficiency and model design/inference efficiency; the report estimates its base text API channel can still achieve 40% GPM under competitive pricing.

  Weaknesses

  Valuation-sensitive, with a disclosed rating of Neutral.

  Comparison

  Relative to internet giants, independent players have advantages in organizational efficiency and decision speed; relative to open-source models such as DeepSeek, they need to maintain differentiation in capability and commercialization.

  Risks

  Valuation, model release cadence, API price competition, and uncertainty in multimodal commercialization.

• DeepSeek V4 ecosystem

  As an open-source model and efficiency benchmark, it creates spillover effects on demand for China's AI models, cloud, and data centers.

  Strengths

  1M context, lower KV cache and inference FLOPs, open-source path, and potential domestic compute support.

  Weaknesses

  The report states DeepSeek still leans toward base text models, while internet giants and some independent model players are more focused on multimodal/omni-modal approaches.

  Comparison

  Among open-source models, it is competitive on price and long-context efficiency, but it still needs ongoing comparison with U.S. SOTA models and other Chinese models in coding, task completion, and multimodality.

  Risks

  Rapid catch-up by competing models, price declines compressing profitability, and risks around the pace and performance validation of domestic compute supply.

Key data

• DeepSeek V4 release date2026-04-24The report states that DeepSeek released the open-source DeepSeek-V4 Preview on April 24, 2026.
• DeepSeek V4 Pro parameter scale1.6T total parameters, 49B activated parametersFlagship-scale model supporting a 1M-token context.
• DeepSeek V4 Flash parameter scale284B total parameters, 13B activated parametersA relatively smaller version that also supports a 1M-token context.
• V4-Pro long-context efficiencyPer-token inference FLOPs are 27% of V3.2, and KV cache is 10% of V3.2Measured in a 1M-context scenario.
• V4-Flash long-context efficiencyPer-token inference FLOPs are 10% of V3.2, and KV cache is 7% of V3.2The report believes it has efficiency advantages for long-text workloads.
• Top China internet sub-sector pickCloud & Data Centers #1Goldman Sachs lists GDS, VNET, Alibaba, and Kingsoft Cloud among its key ideas.
• MiniMax base text API channel gross margin40% GPM, GSeThe report uses this as an example of efficient model design and inference capability at an independent AI company.
• OpenRouter token usage changesMiMo-V2-Pro rose from 6.87T to 9.13T; Qwen 3.6 Plus was 6.27T; DeepSeek V3.2 was 5.37TComparing the rolling 30-day rankings on April 1 and April 22, 2026.
• AIGC To-C/Chatbot engagement+36% month over month in March 2026The chart title shows AIGC To-C/Chatbot engagement increased +36% mom in Mar.
• Relevant ratings and target pricesMiniMax Group Neutral HK$858.50; Tencent Holdings Buy HK$495.20; Xiaomi Corp. Buy HK$31.18Company-specific rating and price information from the disclosure page.

Impact & implications

If DeepSeek V4's efficiency gains and domestic compute supply materialize, AI model inference costs may continue to decline, thereby expanding the trial and commercialization scope for enterprise AI agents, long-text tasks, consumer AI assistants, and AIGC applications. In terms of asset impact, the most direct beneficiaries are cloud and data centers as well as internet giants with cloud infrastructure capabilities; for model companies, open-source and low pricing will intensify competition, making differentiation in model capabilities, API gross margins, and multimodal capabilities critical.

Risks

• The accelerating release cadence of AI models may intensify price competition and weaken pricing power for model services.
• The spread of open-source models may pressure API pricing lower, weighing on commercialization and gross margins.
• If domestic compute supply and the mass production of Huawei Ascend 950 supernodes fall short of expectations, the path to price cuts for DeepSeek V4 Pro may be delayed.
• The long-context efficiency advantage mainly applies to long-cycle tasks, while demand elasticity in ordinary short-prompt scenarios may be smaller.
• If internet giants lack independent incentive mechanisms, they may lag independent AI companies in AI talent retention and model iteration speed.
• While cloud and data centers benefit from AI token demand, they also face uncertainty around capital expenditure, utilization, cloud price competition, and regulation/supply chains.

What to watch

• The magnitude of DeepSeek V4 Pro price cuts in 2H 2026 and progress in the mass production of Huawei Ascend 950 supernodes.
• The release cadence and capability comparisons of models such as Kimi K2.6, Alibaba Qwen3.6-Max, Tencent Hy3 preview, Xiaomi V2.5, and MiniMax M3/Hailuo.
• Whether coding ability, real task completion rates, and multimodal capability become the core differentiators of model pricing power.
• Whether Chinese models continue to gain share in API token usage rankings such as OpenRouter.
• Incremental cloud/token demand driven by enterprise AI agents and consumer AI assistants.
• AIGC To-C/Chatbot DAU and user engagement trends, especially share changes among apps such as Doubao, DeepSeek, Kimi, Yuanbao, and Qianwen.
• The actual financial realization by cloud and data center companies in AI demand, pricing, utilization, and capital expenditure.