J.P. Morgan Splits China AI, Upgrades Zhipu AI to HK$2,000 on Open-Weight Monetization
In a July 7, 2026 research note on China's artificial intelligence sector, J.P. Morgan analysts laid out a framework that cuts to the heart of one of the industry's most debated questions: does releasing open-weight models cannibalize revenue, or does it create a wider monetization funnel? The bank's answer is neither simple nor uniformly bullish — and it has direct, divergent consequences for two of Hong Kong's most closely watched AI listings.
The report, authored by analysts Olivia Xu, Alex Yao, and Daniel Chen at J.P. Morgan Securities (China), raises the target price on Zhipu AI to HK$2,000 (from HK$1,800) with an Overweight rating, while cutting MiniMax to HK$300 (from HK$400), maintaining a Neutral stance. The divergence is not arbitrary — it reflects a structural argument about which companies can actually convert open distribution into durable, premium-priced usage.
Open Weights Are Not a Monetization Leak — Unless Your Model Is Weak
The conventional read on open-weight releases is that they bleed revenue: once weights are public, cloud service providers (CSPs), API aggregators, and enterprise IT teams can self-host, bypassing the model provider's official API entirely. J.P. Morgan acknowledges this risk but reframes the debate around model quality as the decisive variable.
"For competitive models, open weights can scale usage across external GPU capacity — CSPs, inference platforms, private deployments — rather than relying solely on the provider's own compute stack," the analysts write. The key insight is that open-weight releases represent a published checkpoint, not a finished product. Official APIs continue to evolve post-launch through instruction-tuning updates, caching optimizations, latency improvements, and enterprise-grade SLA enhancements — most of which are never pushed back into the public weight package.
The practical implication: two endpoints running nominally the same model can deliver meaningfully different user experiences. For coding agents, long-context workloads, and multi-step agentic tasks — where users pay for task completion rather than raw token throughput — that gap matters enormously.
The bank's data makes this concrete. For DeepSeek V4 Pro, official API pricing combined with aggressive prompt caching yields an effective monthly cost of roughly US$24–41 for a standardized 100 million input token workload, versus US$85–196 across various third-party providers. For MiniMax M3, the official path wins not on sticker price but on cache hit rates, speed, and usage concentration.
Zhipu: Optionality Value, Not Guaranteed Revenue
J.P. Morgan's bull case on Zhipu is carefully qualified. The analysts argue that GLM-5.2's competitive positioning — holding top rankings on WebDev Arena even after Kimi K2.6 and DeepSeek V4 launches — validates the open-weight monetization thesis for frontier-class models. By releasing under MIT licensing while keeping GLM-Turbo variants on managed API channels, Zhipu has structured a distribution strategy that expands developer reach without fully commoditizing its premium endpoints.
The bank raises its 2026–2028 revenue estimates by 3–9%, reflecting improved visibility into global expansion via open-source distribution. Adjusted net loss forecasts shift to RMB 3.71 billion yuan (US$473 million) in 2026 and RMB 3.14 billion yuan in 2027, with a return to profitability projected at RMB 2.37 billion in 2028. The HK$2,000 target is based on 30x 2030 estimated earnings, discounted at 15% WACC.
Critically, the analysts stress that this is option value, not locked-in upside. "The key test is whether GLM-5.2 represents a step-change in the company's own capabilities relative to Kimi K3 and DeepSeek V4.1, and whether GLM-5.5/6 can widen the gap." If model leadership slips, the open-weight distribution advantage evaporates rapidly.
MiniMax: Wider Access, Faster Price Comparison
The logic cuts the other way for MiniMax. While the company's M3 model offers a 1-million-token context window, native multimodality, and an improving product narrative through MiniMax Code, J.P. Morgan sees a critical gap: M3 has not demonstrated the kind of differentiation that commands pricing power over domestic competitors.
The tell, according to the analysts, is M3's permanent 50% price discount — a signal that the model cannot yet extract a capability premium in a crowded field. "When a model lacks clear differentiation, broader access makes comparison, routing, and substitution easier," the report states. Open weights that could be an asset for Zhipu become a liability for MiniMax, accelerating traffic diversion rather than expanding monetization pathways.
Revenue estimates for MiniMax are trimmed 2–8% for 2027–2030, with adjusted net losses widening to US$948 million in 2027 and US$1.003 billion in 2028. The revised HK$300 target, also based on 30x 2030 earnings at 15% WACC, implies roughly 7% downside from the stock's July 7 close of HK$323.80.
The Funding Overhang Neither Company Can Ignore
Both companies remain in capital-intensive phases, and J.P. Morgan flags this explicitly. The bank projects J.P. Morgan models two additional funding rounds for each company in 2026 and 2027. For Zhipu, R&D expenditure (primarily training costs) is forecast at RMB 5.7 billion yuan (US$727 million) in 2026, rising to RMB 9.6 billion and RMB 14.5 billion in subsequent years. For MiniMax, the equivalent figures run at US$809 million, US$1.2 billion, and US$1.5 billion. Operating cash outflows are expected to persist through 2027 for Zhipu and 2028 for MiniMax under base-case assumptions.
The bottom line from J.P. Morgan is blunt: open-weight commercialization is becoming a winner-take-most dynamic. Frontier models convert distribution into premium revenue; everything else gets commoditized faster.
Related Coverage:
Zhipu AI Eyes RMB 15B STAR Market Raise in China’s First Pure-Play LLM Listing Bid