DeepSeek V4 is not a zero-sum shock, but a tailwind for China's LLM sector

Overview

This report discusses the impact of the release of DeepSeek V4 Preview on April 24, 2026, on China's artificial intelligence and LLM sector, especially Zhipu AI and MiniMax. The market viewed V4 as a competitive shock to listed pure-play LLM companies, leading to a notable pullback in related share prices; the report argues that this reaction misread the sector logic, and that V4 should be seen as a sector tailwind rather than a zero-sum shock. The core reasons are: the availability of domestic computing power has been validated, tiered pricing proves that high-value tasks can still be monetized at a premium, open-source technology and a downward shift in the cost curve will improve overall industry efficiency, and the relative model capability landscape has not been reset by DeepSeek V4.

Core views

The report reiterates its Overweight view on Zhipu AI and MiniMax, and recommends using the pullback to add positions ahead of the releases of GLM-5.5 and MiniMax M3 around June. DeepSeek V4 strengthens three pillars in the report's constructive framework: release of computing power supply, pricing discipline, and structural compression of the cost curve; the fourth pillar, relative competitive positioning, remains intense but has not been rebalanced. The V4 release also removed the largest known competitive catalyst risk for April-May, putting the next two months into a more positive catalyst window.

Analysis framework

The report uses a combined approach of event catalysts, model capability comparison, pricing system, computing power supply, and valuation framework. It first analyzes the industry implications of DeepSeek V4 for domestic GPU inference, token costs, and long-context efficiency; then compares the pricing, context length, and benchmark rankings of models such as DeepSeek, GLM, Kimi, Qwen, and MiniMax; next evaluates the impact of events such as index inclusion, Southbound, model releases, lock-up expiries, and financing on share price direction; and finally derives Dec-26 target prices using 2030E earnings, 30x P/E, and 15% WACC.

Methodology notes

• industry_frameworkFour-pillar constructive framework

  computing power supply, pricing discipline, cost curve, and relative competitive positioning

  DeepSeek V4 is used to test four key variables in China's LLM industry. The report believes V4 strengthens the first three and does not overturn the relative positioning of Zhipu and MiniMax in enterprise tasks.

• Valuation methods2030E P/E discounted valuation

  30x 2030E P/E + 15% WACC

  The target prices for both Zhipu and MiniMax are based on 2030E normalized earnings, a target P/E of 30x, discounted to December 2026 using a 15% WACC.

• benchmarkingModel capability and pricing comparison

  Artificial Analysis, LMArena Code Arena, token pricing, and context length

  The report compares models including DeepSeek V4, GLM-5.1, Kimi K2.6, Qwen3.6, and MiniMax M2.7, concluding that while DeepSeek V4 has made progress in general reasoning, it does not comprehensively lead in core B2B tasks such as coding and agents.

• event_analysisCatalyst path and lock-up expiry timeline

  index inclusion, Southbound, model releases, lock-up expiry, financing, and global AI IPOs

  The report maps major events from April 2026 to January 2027 to potential share price direction, and believes positive catalysts are concentrated in May-June, while lock-up expiry and financing pressure need attention after July.

Asset mapping & comparison

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

• Knowledge Atlas Technology Joint Stock Co. Ltd. (Zhipu AI) / 0100.HK

  Core recommended name; DeepSeek V4 is viewed as a sector tailwind rather than a competitive reset

  Strengths

  Has a commercialization foundation in enterprise ARR, on-premise deployment, GLM model iteration, agentic systems, tool-augmented reasoning, and developer-facing infrastructure; GLM-5.1 remains competitive in key B2B tasks.

  Weaknesses

  Depends on model capability remaining in the top tier; heavy R&D investment, computing power supply, and customer adoption uncertainty may weigh on earnings delivery.

  Comparison

  The report believes DeepSeek V4 has improved in general reasoning performance, but has not surpassed GLM-5.1 in coding and agent tasks; DeepSeek V4 Pro also has no obvious pricing advantage over GLM-5.1.

  Risks

  Export controls, geopolitics, entity list restrictions, intensifying competition, R&D execution risk, commercialization uncertainty, and dependence on computing infrastructure and external suppliers.

• MiniMax Group Inc-H / 2513.HK

  Core recommended name; benefits from declining industry cost curves and expansion of multimodal commercialization

  Strengths

  Has a full-spectrum model portfolio across text, voice, and video, supporting B2B/B2C dual-path monetization; a global orientation enhances scalability and earnings elasticity; historical R&D investment of about US$450mn supports high-frequency model iteration.

  Weaknesses

  The post-listing lock-up expiry ratio is relatively high: about 39.0% at the 6-month expiry in July and about 18.2% at the 9-month expiry in October, which may create supply pressure.

  Comparison

  The report believes MiniMax's multimodal portfolio and global expansion capability are relatively scarce, and the upcoming release of MiniMax M3 may again influence model rankings and market expectations.

  Risks

  Litigation proceedings with U.S. studios, intensifying competition, execution and profitability pressure from high R&D spending, commercialization and customer adoption uncertainty, and risks in computing power and external supply chains.

• DeepSeek V4

  Industry catalyst and competitive test variable

  Strengths

  Validates the feasibility of domestic-chip inference; Pro/Flash tiered pricing proves task-based monetization; token compression and DSA significantly reduce long-context costs.

  Weaknesses

  Has not yet comprehensively led domestic top-tier models in core enterprise tasks such as coding and agents.

  Comparison

  Versus DeepSeek V3.2, V4 shows significant improvements in FLOPs and KV cache efficiency; versus GLM-5.1 and Kimi K2.6, it still lags on some B2B metrics.

  Risks

  If subsequent price cuts are too rapid or capability iteration exceeds expectations, it may again compress industry pricing and the relative competitive landscape.

Key data

• DeepSeek V4 release date2026-04-24After the release, Zhipu and MiniMax shares fell 9%, while the HSI rose 0.2% over the same period.
• DeepSeek V4 Pro price$1.74/M input tokens; $3.48/M output tokensCreates an approximately 12x output price gap versus DeepSeek V4 Flash.
• DeepSeek V4 Flash price$0.14/M input tokens; $0.28/M output tokensUsed for low-cost, high-throughput tasks.
• GLM-5.1 price$1.05/M input tokens; $3.50/M output tokensThe report believes DeepSeek V4 Pro has no obvious pricing advantage versus GLM-5.1.
• MiniMax M2.7 price$0.30/M input tokens; $1.20/M output tokensMiniMax maintains lower pricing and has multimodal commercialization potential.
• DeepSeek V4 Pro long-context efficiencysingle-token FLOPs约比DeepSeek V3.2低3.7x，KV cache约小9.5xFrom official DeepSeek charts, showing a significant decline in long-context inference costs.
• DeepSeek V4 Flash long-context efficiencysingle-token FLOPs约比DeepSeek V3.2低9.8x，KV cache约小13.7xReinforces the logic of cost curve compression.
• Code Arena rankingDeepSeek V4约第14名Ranks behind GLM-5.1, Kimi K2.6, and Qwen3.6 plus.
• Zhipu AI target priceHK$950.00Based on 2030E revenue of RMB98,832mn, adjusted net profit of RMB20,360mn, a target P/E of 30x, and a 15% WACC.
• MiniMax target priceHK$1,100.00Based on 2030E revenue of US$9,136mn, adjusted net profit of US$2,322mn, a target P/E of 30x, and a 15% WACC.

Impact & implications

The investment implication of the report is that DeepSeek V4 does not weaken the core investment case for Zhipu and MiniMax; instead, it validates three industry themes: domestic computing power, tiered commercialization, and declining costs. If GLM-5.5 and MiniMax M3 are launched on schedule around June and remain relatively competitive, ARR conversion, institutional fund inflows, and valuation support may all improve simultaneously. However, lock-up expiries, financing needs, and fluctuations in relative model rankings in July and beyond may still create periodic pressure.

Risks

• DeepSeek or other unlisted/internet giant model iterations exceed expectations, causing Zhipu and MiniMax to lose relative model leadership.
• Industry pricing discipline weakens, forcing foundation model services toward commoditized economics.
• The capability gap between Chinese and U.S. models widens, affecting the global competitiveness and valuation framework of Chinese LLM companies.
• ARR conversion falls short of expectations, and token demand cannot be effectively converted into recognizable revenue.
• The release cadence or performance of models such as GLM-5.5 and MiniMax M3 falls below market expectations.
• Share lock-up expiries from July onward, potential equity financing, and high R&D spending may weigh on share prices.
• Export controls, geopolitics, entity list restrictions, and computing power supply constraints may affect training and inference costs.
• MiniMax faces litigation risk related to U.S. studios.

What to watch

• The outcome of the Hang Seng Composite review on May 22, 2026.
• Whether Hang Seng Composite / TECH inclusion becomes effective on June 8, 2026.
• Whether Zhipu enters Southbound in the first week of June 2026.
• The release progress, performance rankings, and pricing of GLM-5.5 and MiniMax M3 around June 2026.
• Whether DeepSeek V4 Pro pricing declines alongside 2H26 Ascend 950 supernode capacity expansion.
• Actual selling pressure after the 6-month lock-up expiry for Zhipu and MiniMax in July 2026.
• ARR trajectory, pricing resilience, and margin trends in 1H26 results in August 2026.
• The impact of potential IPOs by Anthropic and OpenAI on global AI valuation anchors.
• Whether potential listings by Kimi, StepFun, and others dilute the scarcity premium of listed LLM companies.