The 'Winner-Takes-More' Thesis

According to JPMorgan, the Chinese LLM market is consolidating rapidly. The bank explicitly describes the period as one where "amid the open-source wave, China's AI enters a 'winner-takes-more' phase" . This is not a winner-takes-all scenario, but one where the top-tier providers capture a disproportionately large share of the value .

Alex Yao, JPMorgan's head of China equity research, argues that the winners will be decided by enterprise value conversion, not by whose model is the smartest on leaderboards . The focus has shifted to monetization through enterprise workflows, API consumption, coding tools, and agents . JPMorgan forecasts that the annual recurring revenue (ARR) from China's major LLMs will grow by approximately 4x to 7x in 2026 .

The Open-Source Structural Divide

A central pillar of JPMorgan's analysis is the structural divide created by open-source strategies . The bank's July 2026 report argues that firms with consistently state-of-the-art (SOTA) open-weight models can generate "significant optionality value" through commercialization . In contrast, lagging models get commoditized and struggle to command any pricing power .

This creates a self-reinforcing cycle: top-tier models attract more users and developers, which generates more data and revenue, which funds further model improvements. Weaker models, even if open-source, get stuck in a low-value trap where they are used but not monetized effectively .

DeepSeek V4 Pro: The Pricing Shockwave

DeepSeek's V4 Pro is the clearest example of the extreme cost pressure reshaping the market. Its pricing advantages are stark:

• Permanent 75% price cut: In May 2026, DeepSeek made permanent a 75% discount on V4 Pro, pricing output tokens at $0.87 per million tokens .
• Radical cost advantage: This pricing makes V4 Pro roughly 12x cheaper than GPT-5.5 and 7x cheaper than Claude Opus 4.8 at comparable intelligence .
• Near-identical performance: V4 Pro scores 80.6% on SWE-bench Verified, within 0.2 points of Claude Opus 4.6, which costs $25 per million output tokens—a 7x price gap .
• Structural cost advantage: JPMorgan noted in a May 2026 report that only DeepSeek's proprietary infrastructure can operate V4 with cutting-edge economic efficiency. The disparity in cache hit input between DeepSeek's official API and third-party pipelines is approximately 40x, highlighting structural first-party advantages .

According to JPMorgan, the market initially misinterpreted V4 as a competitive threat to other Chinese AI companies, but the bank argued it actually strengthened three of the four key pillars supporting domestic LLM monetization .

JPMorgan's Divergent Ratings: Zhipu AI vs. MiniMax

The most concrete expression of the "winner-takes-more" thesis is JPMorgan's starkly divergent treatment of two leading Chinese AI companies: Zhipu AI and MiniMax .

Zhipu AI (02513.HK) — Overweight, Target Raised to HK$2,000

JPMorgan raised Zhipu's target price three times in rapid succession:

• June 15, 2026: Raised to HK$1,400 from HK$950, citing model visibility and pricing power
• June 22, 2026: Raised further to HK$1,800 (30x forward P/E based on 2030 earnings)
• July 8, 2026: Raised to HK$2,000 for December 2026, citing GLM-5.2 as reinforcing the thesis that "open-weight commercialization can generate significant optionality value for leading model providers"

The bank also raised Zhipu's revenue forecasts by 26% to 42% for FY2026–2030 and lowered adjusted net loss projections . Zhipu shares surged as much as 48% on the initial upgrade .

MiniMax (00100.HK) — Neutral, Target Cut to HK$300

Simultaneously, JPMorgan downgraded MiniMax:

• June 14, 2026: Downgraded from "Overweight" to "Neutral", target cut from HK$1,100 to HK$400
• July 8, 2026: Target further slashed to HK$300, reflecting "limited pricing power and weak commercialization prospects"

The bank's rationale: MiniMax has not launched a new domestic SOTA model since its M2 model, and in terms of pure model capability it is falling behind peers . Its M3 model (released June 1) ranked #4 on Code Arena WebDev but failed to close the gap with leading models . JPMorgan cited "weak distribution and brand recognition" outside of MiniMax's narrow entertainment use case .

The contrasting treatment is clear: Zhipu's consistent model iteration (especially GLM-5.2) earned it pricing power and an "Overweight" rating, while MiniMax's inability to keep pace with leading SOTA models resulted in a "Neutral" rating and a target cut of approximately 73% in just one month .

Broader Context: Surging Global Chinese AI Adoption

The JPMorgan analysis is set against a backdrop of rapidly accelerating adoption of Chinese AI models globally.

OpenRouter Traffic Dominance: Chinese models accounted for over 45% of OpenRouter traffic by April 2026, capturing the majority of token consumption on the world's largest AI aggregation platform, according to data highlighted by JPMorgan Asset Management strategist Michael Cembalest . By late May 2026, Chinese AI models were surging in global usage rankings at an unprecedented pace .

Cost Advantage vs. U.S. Frontier Models: Chinese models offer 60–90% lower costs than U.S. frontier models while approaching comparable performance . DeepSeek V4 Pro alone undercuts GPT-5.5 by roughly 12x on input pricing . The economic gap is widest on reasoning-heavy enterprise workloads .

Alibaba's Qwen: The Adoption-Revenue Paradox: Alibaba's Qwen became the world's most downloaded open-source AI system by January 2026 . However, JPMorgan notes that Qwen faces significant revenue conversion struggles — illustrating the core tension of the "winner-takes-more" dynamic where open-source adoption does not automatically translate into pricing power or sustainable revenue . The broader Chinese AI ecosystem has rapidly built 538 registered LLMs (up from 14 in October 2023), but much of this capacity has been channeled into open-source, low-cost models offered free or near-free to maximize adoption .

Key Caveats

• JPMorgan's analysis is explicitly about commercialization and monetization, not technical leadership — the bank warns that model capability alone is insufficient for investment outperformance .
• DeepSeek itself is not publicly listed, so JPMorgan's ratings apply to other players (Zhipu, MiniMax, and indirectly Alibaba/Baidu) .
• The MiniMax target of HK$300 represents a ~73% cut from the initial HK$1,100 target in just one month, reflecting how rapidly JPMorgan is adjusting to the open-source commoditization dynamic .

The 'Winner-Takes-More' Thesis

According to JPMorgan, the Chinese LLM market is consolidating rapidly. The bank explicitly describes the period as one where "amid the open-source wave, China's AI enters a 'winner-takes-more' phase" . This is not a winner-takes-all scenario, but one where the top-tier providers capture a disproportionately large share of the value .

Alex Yao, JPMorgan's head of China equity research, argues that the winners will be decided by enterprise value conversion, not by whose model is the smartest on leaderboards . The focus has shifted to monetization through enterprise workflows, API consumption, coding tools, and agents . JPMorgan forecasts that the annual recurring revenue (ARR) from China's major LLMs will grow by approximately 4x to 7x in 2026 .

The Open-Source Structural Divide

A central pillar of JPMorgan's analysis is the structural divide created by open-source strategies . The bank's July 2026 report argues that firms with consistently state-of-the-art (SOTA) open-weight models can generate "significant optionality value" through commercialization . In contrast, lagging models get commoditized and struggle to command any pricing power .

This creates a self-reinforcing cycle: top-tier models attract more users and developers, which generates more data and revenue, which funds further model improvements. Weaker models, even if open-source, get stuck in a low-value trap where they are used but not monetized effectively .

DeepSeek V4 Pro: The Pricing Shockwave

DeepSeek's V4 Pro is the clearest example of the extreme cost pressure reshaping the market. Its pricing advantages are stark:

• Permanent 75% price cut: In May 2026, DeepSeek made permanent a 75% discount on V4 Pro, pricing output tokens at $0.87 per million tokens .
• Radical cost advantage: This pricing makes V4 Pro roughly 12x cheaper than GPT-5.5 and 7x cheaper than Claude Opus 4.8 at comparable intelligence .
• Near-identical performance: V4 Pro scores 80.6% on SWE-bench Verified, within 0.2 points of Claude Opus 4.6, which costs $25 per million output tokens—a 7x price gap .
• Structural cost advantage: JPMorgan noted in a May 2026 report that only DeepSeek's proprietary infrastructure can operate V4 with cutting-edge economic efficiency. The disparity in cache hit input between DeepSeek's official API and third-party pipelines is approximately 40x, highlighting structural first-party advantages .

According to JPMorgan, the market initially misinterpreted V4 as a competitive threat to other Chinese AI companies, but the bank argued it actually strengthened three of the four key pillars supporting domestic LLM monetization .

JPMorgan's Divergent Ratings: Zhipu AI vs. MiniMax

The most concrete expression of the "winner-takes-more" thesis is JPMorgan's starkly divergent treatment of two leading Chinese AI companies: Zhipu AI and MiniMax .

Zhipu AI (02513.HK) — Overweight, Target Raised to HK$2,000

JPMorgan raised Zhipu's target price three times in rapid succession:

• June 15, 2026: Raised to HK$1,400 from HK$950, citing model visibility and pricing power
• June 22, 2026: Raised further to HK$1,800 (30x forward P/E based on 2030 earnings)
• July 8, 2026: Raised to HK$2,000 for December 2026, citing GLM-5.2 as reinforcing the thesis that "open-weight commercialization can generate significant optionality value for leading model providers"

The bank also raised Zhipu's revenue forecasts by 26% to 42% for FY2026–2030 and lowered adjusted net loss projections . Zhipu shares surged as much as 48% on the initial upgrade .

MiniMax (00100.HK) — Neutral, Target Cut to HK$300

Simultaneously, JPMorgan downgraded MiniMax:

• June 14, 2026: Downgraded from "Overweight" to "Neutral", target cut from HK$1,100 to HK$400
• July 8, 2026: Target further slashed to HK$300, reflecting "limited pricing power and weak commercialization prospects"

The bank's rationale: MiniMax has not launched a new domestic SOTA model since its M2 model, and in terms of pure model capability it is falling behind peers . Its M3 model (released June 1) ranked #4 on Code Arena WebDev but failed to close the gap with leading models . JPMorgan cited "weak distribution and brand recognition" outside of MiniMax's narrow entertainment use case .

The contrasting treatment is clear: Zhipu's consistent model iteration (especially GLM-5.2) earned it pricing power and an "Overweight" rating, while MiniMax's inability to keep pace with leading SOTA models resulted in a "Neutral" rating and a target cut of approximately 73% in just one month .

Broader Context: Surging Global Chinese AI Adoption

The JPMorgan analysis is set against a backdrop of rapidly accelerating adoption of Chinese AI models globally.

OpenRouter Traffic Dominance: Chinese models accounted for over 45% of OpenRouter traffic by April 2026, capturing the majority of token consumption on the world's largest AI aggregation platform, according to data highlighted by JPMorgan Asset Management strategist Michael Cembalest . By late May 2026, Chinese AI models were surging in global usage rankings at an unprecedented pace .

Cost Advantage vs. U.S. Frontier Models: Chinese models offer 60–90% lower costs than U.S. frontier models while approaching comparable performance . DeepSeek V4 Pro alone undercuts GPT-5.5 by roughly 12x on input pricing . The economic gap is widest on reasoning-heavy enterprise workloads .

Alibaba's Qwen: The Adoption-Revenue Paradox: Alibaba's Qwen became the world's most downloaded open-source AI system by January 2026 . However, JPMorgan notes that Qwen faces significant revenue conversion struggles — illustrating the core tension of the "winner-takes-more" dynamic where open-source adoption does not automatically translate into pricing power or sustainable revenue . The broader Chinese AI ecosystem has rapidly built 538 registered LLMs (up from 14 in October 2023), but much of this capacity has been channeled into open-source, low-cost models offered free or near-free to maximize adoption .

Key Caveats

• JPMorgan's analysis is explicitly about commercialization and monetization, not technical leadership — the bank warns that model capability alone is insufficient for investment outperformance .
• DeepSeek itself is not publicly listed, so JPMorgan's ratings apply to other players (Zhipu, MiniMax, and indirectly Alibaba/Baidu) .
• The MiniMax target of HK$300 represents a ~73% cut from the initial HK$1,100 target in just one month, reflecting how rapidly JPMorgan is adjusting to the open-source commoditization dynamic .