Market Panic or Prime Opportunity? Why J.P. Morgan Says DEEPSEEK V4 is a Massive Tailwind for China’s AI Sector

JPMorgan reiterates "Overweight" ratings on Zhipu AI and MiniMax after shares tumbled 9% on Thursday following DeepSeek's V4 preview launch, arguing the market misread the release as competitive threat when it actually strengthens three of four pillars supporting domestic LLM monetization.

Chinese large language model (LLM) specialists Zhipu AI and MiniMax saw shares decline 9% on April 24, 2026, as DeepSeek unveiled its V4 preview model—a selloff JPMorgan analysts now characterize as a "knee-jerk overreaction" that obscures fundamental industry tailwinds rather than competitive displacement.

In a research note published Saturday, JPMorgan's equity research team led by Olivia Xu argued that DeepSeek V4's debut removes the "single largest near-term competitive uncertainty" facing Zhipu and MiniMax ahead of their own next-generation model cycles expected in June 2026. Rather than eroding competitive positioning, the V4 release validates domestic compute infrastructure, reinforces task-based pricing discipline, and compresses industry-wide cost curves—three structural positives that outweigh any immediate performance benchmarking concerns.

"The market interpreted V4 as a zero-sum competitive shock," JPMorgan wrote. "We see it as an industry positive that strengthens compute supply release, pricing discipline, and structural cost curve compression while leaving competitive positioning contested but not rebalanced."

Huawei Ascend Validation Eases Compute Bottleneck

The most underappreciated aspect of DeepSeek V4, according to JPMorgan, is its demonstration of inference viability on Huawei's Ascend chips—a critical de-risking event for China's LLM ecosystem given export controls that restrict access to NVIDIA's advanced GPUs. DeepSeek confirmed that V4 Pro's throughput is currently constrained by Huawei compute availability but pledged pricing reductions once Ascend 950 supernodes achieve mass production in the second half of 2026.

"This validates that domestic chips are not only technically feasible but potentially cost-competitive versus international alternatives," JPMorgan stated. "For Zhipu and MiniMax, whose ARR conversion has been historically limited by compute scarcity, broader Ascend adoption opens new pathways to convert token demand into confirmed revenue."

China's LLM industry has been throttled by compute constraints since U.S. export restrictions tightened in 2023. While companies like Zhipu secured access to limited inventories of NVIDIA A100/H100 chips, scaling inference capacity remained a binding constraint on revenue growth. DeepSeek's successful deployment of a 1.6-trillion-parameter model on Huawei silicon marks the first large-scale validation that domestic alternatives can support frontier AI workloads.

Pricing Strategy Mirrors Task-Based Monetization Frameworks

DeepSeek's product lineup—with V4 Pro priced at 1.74 per million input tokens and 3.48 per million output tokens versus V4 Flash at 0.14/0.28—creates an approximately 12x output price differential within its own portfolio. JPMorgan interprets this tiering as implicit validation of the task-based monetization logic that Zhipu and MiniMax have already operationalized.

"DeepSeek's own pricing validates what the industry has been building toward: high pricing for coding and agentic workflows, lower pricing for high-throughput simple tasks," the analysts wrote. "This structure aligns with how Zhipu GLM-5.1 and MiniMax already operate."

Contrary to bearish narratives that DeepSeek's aggressive pricing forces "commodity economics" across the sector, JPMorgan's token price comparison table shows V4 Pro's pricing is not materially cheaper than Zhipu's GLM-5.1 (1.05/3.50) or Moonshot AI's Kimi K2.6 (0.74/4.66) when adjusted for context window and task complexity. Even if V4 Pro prices decline with Ascend 950 availability, Zhipu and MiniMax will likely have released their next-generation models (GLM-5.5 and M3, respectively) by then, rendering current pricing comparisons moot.

Technical Advances Are Industry-Wide, Not Moat-Building

DeepSeek V4's architectural innovations—including token compression mechanisms and DeepSeek Sparse Attention (DSA)—reduce memory and compute costs for long-context inference, a critical capability for coding agents and enterprise knowledge management. While these advances set new industry benchmarks, JPMorgan argues they represent diffusible technology rather than durable competitive moats.

"China's LLM ecosystem has historically absorbed DeepSeek's architectural breakthroughs within 1-2 model cycles," the note observed, citing how Mixture-of-Experts (MoE) routing mechanisms from V3 were integrated by Alibaba's Qwen, Zhipu's GLM, and Moonshot's Kimi within approximately 4-6 months.

DeepSeek's open-source release of V4 accelerates this knowledge transfer, benefiting the entire industry rather than conferring lasting advantage to DeepSeek alone. The report emphasizes that Zhipu and MiniMax will need to incorporate token compression and sparse attention into future models to remain competitive—but frame this as table stakes rather than existential catch-up.

Performance Remains Competitive on Enterprise-Critical Tasks

While DeepSeek V4 advanced on general reasoning benchmarks, JPMorgan highlighted that it has not surpassed domestic leaders Zhipu GLM-5.1 and Moonshot Kimi K2.6 on enterprise-critical tasks like coding efficiency and agentic workflow reliability.

On LMArena's Code Arena leaderboard as of April 24, 2026, DeepSeek V4 Pro Thinking ranked 14th globally with an ELO score of 1,456—trailing Zhipu GLM-5.1 (5th, ELO 1,534) and Kimi K2.6 (6th, ELO 1,529). Similarly, Artificial Analysis intelligence indices show V4 improved markedly versus V3.2 but remains below top-tier domestic models on specialized tasks that drive B2B ARR conversion.

"What drives enterprise ARR are B2B metrics where DeepSeek V4 does not impact GLM-5.1's positioning," JPMorgan concluded. "Leaderboard rotations reflect normal industry patterns as new releases incorporate latest training techniques."

Catalyst Pathway Tilts Asymmetrically Positive

With V4 now released and digested, JPMorgan sees a two-month window before July 2026 lock-up expirations where catalysts skew constructively for Zhipu and MiniMax:

• Index Inclusion: Potential additions to Hang Seng indices in May review (effective June) would trigger passive inflows.
• Stock Connect: Zhipu's anticipated inclusion in Hong Kong Stock Connect in early June broadens mainland investor access.
• Model Refresh Cycle: GLM-5.5 and MiniMax M3 releases expected around June historically drive ARR step-functions.

"From V4 uncertainty removal to July lock-up expiry, the intervening two-month corridor presents an asymmetric positive setup," the analysts wrote. "Index inclusion provides passive flows; Stock Connect opens mainland demand; model iteration cycles produce operational catalysts that have historically lifted ARR."

JPMorgan maintains "Overweight" ratings with 12-month price targets of HK 950 for ZhipuAI and HK 1100 for MiniMax, both derived from 30x 2030E P/E multiples discounted at 15% WACC. The premium valuations versus mainstream Chinese internet peers reflect expected 2026-2030 revenue CAGRs exceeding 100%.

Related Coverage：

China’s Anthropic? UBS Sees 30% Upside in Zhipu AI

Kimi's K2.6 Opens Door for China's Domestic Chip Integration