J.P. Morgan believes DeepSeek V4 is a tailwind for China’s large language model sector, rather than a zero-sum shock to Zhipu AI and MiniMax

Overview

This report evaluates the impact of the DeepSeek V4 release on China’s large language model sector and on Zhipu AI and MiniMax. J.P. Morgan believes V4 is not a zero-sum shock to listed pure-play large-model companies, but rather confirms three positive sector variables: usable domestic compute supply, still-effective task-based tiered pricing, and a continued long-term decline in the inference cost curve. The report argues that the biggest known competitive catalyst from April to May has already materialized and been absorbed by the market, while catalysts over the next two months are more positive, including index inclusion, Southbound Stock Connect eligibility, and the launches of GLM-5.5 and MiniMax M3.

Core views

The core view is to add to Zhipu AI and MiniMax on share-price pullbacks. DeepSeek V4 validates the infrastructure feasibility of China’s large-model sector, but does not reset the relative positioning of Zhipu AI and MiniMax in enterprise-grade model capabilities, coding tasks, and agent workflows. The report emphasizes that pricing, compute, model iteration, and ARR conversion remain the main drivers of investment returns, rather than short-term sentiment shocks caused by a single model release.

Analysis framework

The report assesses the impact of DeepSeek V4 through four frameworks: whether compute supply is unlocked, whether pricing discipline is damaged, whether the cost curve structurally shifts downward, and whether the relative competitive landscape is reordered. It then combines model pricing tables, Code Arena rankings, Artificial Analysis metrics, event timelines, and company valuation models to judge the impact on the share prices, fundamentals, and catalyst paths of Zhipu AI and MiniMax.

Methodology notes

• Industry competition frameworkFour-pillar evaluation framework

  Assess the impact of DeepSeek V4 from four angles: compute supply, pricing discipline, cost curve, and relative model capability.

  The report believes DeepSeek V4 strengthens the first three positive variables, while the fourth—relative competitive positioning—remains tight, with no adverse reordering for Zhipu AI and MiniMax.

• Valuation methodologyForward earnings discount valuation

  Based on 2030E earnings, applying a 30x P/E and discounting to the end-2026 target price using a 15% WACC.

  The target prices for both Zhipu AI and MiniMax are based on normalized 2030 earnings, a target P/E of 30x, and a 15% discount rate, with the valuation premium reflecting expectations of revenue CAGR above 100% from 2026 to 2030.

• Catalyst event analysisEvent path and share-price direction assessment

  Evaluate the impact of model releases, index inclusion, Southbound inclusion, lock-up expiries, earnings, and potential financing on share-price direction along a timeline.

  The report believes the short-term negatives related to DeepSeek V4 have largely cleared, and the index, Southbound, and new-model releases around June create a more positively skewed window.

Asset mapping & comparison

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

• Zhipu AI (2513.HK)

  Core beneficiary

  Strengths

  GLM-5.1 remains competitive in enterprise metrics, coding, and agent tasks; has a relatively strong foundation in on-premise deployment; cloud API and developer ecosystem have expansion potential.

  Weaknesses

  Highly dependent on continued iteration in model capability and compute supply, while commercialization and ARR conversion still need validation.

  Comparison

  The report believes DeepSeek V4 does not reset Zhipu AI’s relative competitive position, and the release of GLM-5.5 could again change ranking dynamics.

  Risks

  Export controls, geopolitics, entity-list risk, intensifying competition, persistently high R&D spending, commercialization uncertainty, and external compute supply risk.

• MiniMax Group Inc- H (0100.HK)

  Core beneficiary

  Strengths

  Has full-spectrum model capabilities across text, voice, and video, with both B2B and B2C commercialization paths and global expansion potential.

  Weaknesses

  Profitability is still affected by high R&D investment and commercialization progress, and some model and application paths remain in the early stage.

  Comparison

  The report believes the launch of MiniMax M3 will be a key positive catalyst, and that the technology diffusion from DeepSeek V4 can lower sector costs rather than uniquely hurting MiniMax.

  Risks

  Litigation with U.S. film studios, intensifying competition, R&D execution risk, uncertainty in customer adoption, and compute supply cost risk.

• DeepSeek V4

  Sector catalyst and competitive benchmark

  Strengths

  Validates that domestic chips can support inference for a 1.6T-parameter model, while reducing long-context costs through token compression and DSA.

  Weaknesses

  Has not yet shown clear leadership over GLM-5.1 and Kimi K2.6 in key enterprise tasks such as coding and agents.

  Comparison

  Compared with GLM-5.1 and Kimi K2.6, DeepSeek V4 has made significant progress in general reasoning, but still has not established an overwhelming advantage in enterprise core tasks.

  Risks

  If pricing falls rapidly later or capabilities continue to leap forward, it could reignite sector-wide pricing and competitive pressure.

Key data

• Share-price reaction after DeepSeek V4 releaseZhipu AI/MiniMax about -9%, Hang Seng Index about +0.2%The report considers this an overreaction.
• Output price gap between DeepSeek V4 Pro and Flashabout 12xUsed to support the tiered pricing logic that premium enterprise tasks can sustain a premium.
• DeepSeek V4 Pro pricinginput $1.74/M token, output $3.48/M tokenCompared with GLM-5.1 and Kimi K2.6, it does not show a clear pricing advantage.
• DeepSeek V4 Flash pricinginput $0.14/M token, output $0.28/M tokenGeared toward more commoditized and high-throughput tasks.
• Code Arena rankingsDeepSeek V4 about No.14; GLM-5.1 No.5; Kimi K2.6 No.6Based on this, the report believes V4 has not yet led in enterprise core tasks such as coding and agents.
• Zhipu AI target priceHK$950.00Based on 2030E revenue of Rmb98,832mn, adjusted net profit of Rmb20,360mn, 30x P/E, and 15% WACC.
• MiniMax target priceHK$1,100.00Based on 2030E revenue of US$9,136mn, adjusted net profit of US$2,322mn, 30x P/E, and 15% WACC.

Impact & implications

In terms of investment implications, the report believes DeepSeek V4 enhances the investability of China’s large-model sector rather than weakening Zhipu AI and MiniMax. Validation of domestic compute helps ease inference cost and supply bottlenecks, pricing tiers indicate that high-value enterprise tasks can still maintain a premium, and the diffusion of open-source technology is likely to be absorbed by the industry within one to two model cycles. After the short-term share-price pullback, investors should pay more attention to June new-model launches, index and Southbound capital catalysts, and whether ARR conversion accelerates as compute improves.

Risks

• A widening China-U.S. technology capability gap or escalation in export controls.
• Model iteration by competitors such as DeepSeek, Tencent, Alibaba, and Moonshot exceeds expectations.
• High R&D spending continues to suppress profitability and creates financing pressure.
• ARR conversion, customer adoption, and enterprise budget release fall short of expectations.
• Lock-up expiries, potential equity financing, and an expanding free float put pressure on share prices.
• MiniMax-related litigation or compliance risks affect globalization and multimodal commercialization.

What to watch

• The release timing and performance of GLM-5.5 and MiniMax M3 around June.
• Whether Zhipu AI and MiniMax are included in indices and the associated passive fund inflows.
• Potential Southbound inclusion for Zhipu AI and the resulting improvement in mainland investor demand and liquidity.
• The extent of price adjustments for DeepSeek V4 Pro after capacity expansion of the Ascend 950 supernode.
• Validation of ARR, pricing discipline, gross margin, and customer growth in 1H26 results.
• The impact of July and subsequent lock-up expiry windows on share prices and the supply-demand structure.