Zhipu AI's 10x Rally Exposes Hong Kong's AI Narrative Premium Over MiniMax
A HK$400 billion valuation gulf between two near-identical Chinese AI startups is less a verdict on fundamentals than a masterclass in how narrative economics, reflexivity, and winner-take-all option pricing converge to distort markets in real time.
On May 29, 2026, shares of Zhipu AI whipsawed violently — surging as much as 23.18% intraday before closing down 1.42%, a single-session swing exceeding 20 percentage points. The episode crystallized a question that has been building since January: how can two Chinese AI companies that listed on the Hong Kong Stock Exchange on the same day, generated nearly identical 2025 revenues, and operate in overlapping markets diverge by more than HK$400 billion in market capitalization?
As of the May 29 close, Zhipu AI's market capitalization had exceeded HK$700 billion, representing a gain of more than 1,000% year-to-date. MiniMax closed at HK$840 per share with a market capitalization of HK$263.45 billion, leaving a valuation gap of more than HK$400 billion. Yet on the revenue line, the two companies are almost inseparable.
Financials Tell a Story of Near-Identical Twins
Zhipu AI reported full-year 2025 revenue of RMB 724 million (US$100.6 million), up 131.9% year-over-year. MiniMax generated US$79.04 million in revenue (RMB 569 million), with growth accelerating to 158.9% — faster than Zhipu's. The revenue ratio between the two stands at only 1.27-to-1, far narrower than their roughly 2.7-to-1 market cap gap.
Both companies are deeply loss-making. Zhipu AI recorded an adjusted net loss of RMB 3.182 billion (US$441.9 million) in 2025. MiniMax's adjusted net loss reached approximately US$251 million (roughly RMB 1.73 billion). Neither company is required under Hong Kong Stock Exchange rules to file quarterly reports, making management commentary at earnings calls the primary source of real-time operating data.
The most meaningful fundamental divergence between the two lies in gross margin. Zhipu AI posted a blended gross margin of 41% in 2025, versus MiniMax's 25.4% — a gap of approximately 16 percentage points. Zhipu's superior margin reflects a revenue mix weighted toward on-premise deployment, enterprise-grade Agent solutions, and general-purpose large model licensing, all of which carry higher per-unit economics. MiniMax derives a greater share of revenue from AI-native consumer products and international markets, where its open platform and developer ecosystem are scaling rapidly but had not yet reached revenue dominance by end-2025.
Diverging ARR Trajectories Sharpen the Valuation Debate
Because Hong Kong does not mandate quarterly disclosures, both companies have used earnings calls and operational updates as de facto guidance vehicles — and the numbers they have chosen to highlight reveal starkly different strategic postures.
At Zhipu AI's March 31 earnings call, CEO Zhang Peng disclosed that MaaS API annualized recurring revenue (ARR) had reached approximately RMB 1.7 billion (US$250 million) as of March 2026, with first-quarter API call pricing raised by a cumulative 83% while call volume still grew 400% — a simultaneous volume-and-price expansion that analysts describe as rare validation of genuine demand inelasticity.
MiniMax, in an operational update released May 28, reported that global enterprise and developer clients had exceeded one million — a fivefold increase in six months — with total global users at approximately 300 million. The company said ARR had more than doubled over the preceding two months. Using MiniMax's February disclosure of ARR exceeding US$150 million as a base, its current ARR is estimated to have surpassed US$300 million, which would actually place it ahead of Zhipu AI on this metric.
Recalculating valuations on an ARR basis produces a striking inversion. Zhipu AI's HK$700 billion market capitalization against US$250 million ARR (approximately HK$1.95 billion) implies a price-to-ARR multiple of roughly 360x. MiniMax, with a HK$263.45 billion market capitalization and an estimated ARR exceeding US$300 million (approximately HK$2.34 billion), trades at roughly 110x ARR. By this metric, MiniMax is significantly cheaper — yet investors continue to award the premium to Zhipu.
For reference, OpenAI's latest post-money valuation of US$852 billion against estimated 2025 revenue of US$13 billion implies a price-to-sales ratio of approximately 66x. Using an annualized revenue run rate exceeding US$20 billion, that multiple falls to roughly 43x. Even on the more generous annualized basis, both Chinese AI companies trade at valuation multiples that far exceed those of their most richly valued Western peer.
March 18 Marks the Inflection Point That Rewired Investor Psychology
The chart-level divergence between Zhipu and MiniMax can be traced to a single session: March 18, 2026. That day, Alibaba Cloud announced price increases of up to 34% on AI compute and storage products, with Baidu AI Cloud subsequently following suit. The market read this as price confirmation of tightening AI inference capacity and surging token demand. Hong Kong-listed AI names rallied broadly in the afternoon session; Zhipu AI closed up 19.47% and MiniMax up 19.85%.
But the day held asymmetric significance. MiniMax also launched its next-generation model M2.7 on March 18, with intraday gains briefly exceeding 28% and shares touching HK$1,330 — the company's all-time high since listing. Having catalyzed on both the macro tailwind and a product launch simultaneously, MiniMax's stock peaked that day and has not sustainably reclaimed that level since.
Zhipu AI, by contrast, had no company-specific catalyst on March 18. Its subsequent rally was driven by a series of fundamental developments: the March 31 earnings release, API metrics showing simultaneous growth in both pricing and usage volume, and a target-price upgrade from China International Capital Corporation (CICC) to HK$900 per share, citing stronger-than-expected API ARR and accelerating demand for Agentic AI solutions. On April 1, Zhipu AI shares surged another 31.94%, briefly topping HK$1,000 intraday. By May 29, the stock had climbed past HK$1,500.
Benchmark Wars Expose the Limits of Model-Capability Pricing
The market's willingness to assign a premium to Zhipu AI rests substantially on a coding-capability narrative, with Anthropic — which completed a US$65 billion Series H financing round and reached a post-money valuation of US$96.5 billion, surpassing OpenAI's prior US$85.2 billion record — cited as the aspirational peer.
On SWE-bench Pro, the industry's most widely cited agentic coding benchmark, Zhipu AI's GLM-5.1 scored 58.4%, edging out GPT-5.4 at 57.7% and Claude Opus 4.6 at 57.3%. MiniMax's M2.7 scored 56.22% on the same benchmark, with strong results on VIBE-Pro and Terminal Bench 2. The capability gap between the two Chinese models is narrow.
The benchmark itself, however, may be losing its discriminating power. In April 2026, OpenAI publicly acknowledged in an official blog post that SWE-bench Verified can no longer effectively differentiate frontier model capabilities.
On May 27, San Francisco-based data company Datacurve released DeepSWE, a new benchmark designed to eliminate data contamination. Its 113 problems were written from scratch by engineers, span 91 repositories and five programming languages, and have never appeared in any public code base. Reference solutions average 5.5 times the code volume of SWE-bench Pro problems, precluding pattern-matching shortcuts.
The DeepSWE results are sobering for Chinese AI developers. GPT-5.5 scored 70% ±4%, establishing a commanding lead. Claude Opus 4.7 reached 54% ±5%. Chinese models clustered in a much lower band: Moonshot AI's Kimi K2.6 at 24%, Xiaomi's MiMo V2.5 Pro at 19%, Zhipu AI's GLM-5.1 at 18%, and DeepSeek V4 Pro at 8%. MiniMax's M2.7 participated in the evaluation but did not appear among the top 12 of 16 models listed.
The results imply a three-to-four-fold capability gap between GPT-5.5 and Chinese frontier models on long-horizon, zero-contamination engineering tasks. Analysts note important caveats: the 113-question sample generates error bands of ±4-5% that materially affect rankings in the lower score range; all models were evaluated using a standardized mini-swe-agent framework rather than proprietary toolchains, potentially compressing each model's ceiling; and Datacurve's closer commercial relationship with OpenAI than with Anthropic introduces a conflict-of-interest dimension that the research community has flagged.
Narrative Economics and Reflexivity Explain What Fundamentals Cannot
A senior secondary-market analyst framed the divergence bluntly: "The market's primary focus right now is coding capability." That framing — regardless of whether it is empirically justified — has become self-fulfilling.
Four financial frameworks help explain the HK$400 billion gap. Yale economist Robert Shiller's narrative economics posits that asset prices are frequently driven by transmissible stories rather than numbers. Zhipu AI's story maps cleanly onto the global Coding Agent and Agentic AI trade. MiniMax's story — overseas consumer growth, multimodal products, developer ecosystem — resists compression into a single investable theme.
John Maynard Keynes's beauty-contest model describes investors buying what they expect others to buy. Once Zhipu AI was designated the leading candidate for China's Agentic AI crown, capital self-reinforcingly concentrated there.
George Soros's reflexivity theory explains the feedback loop: rising share price prompted CICC to raise its target to HK$900, generating research coverage that attracted additional buyers, pushing the stock higher — while the elevated valuation itself improved Zhipu's fundamentals through cheaper capital access, enhanced talent recruitment, and increased enterprise client confidence.
Finally, at a price-to-sales ratio of approximately 890x on trailing revenue — or 360x on ARR — the market is effectively purchasing a winner-take-all call option. In a market that has selected a single narrative winner, the option premium accrues entirely to that winner; all others are priced as ordinary businesses.
For context on how extreme this repricing has been: Zhipu AI's market cap on May 29 surpassed that of Xiaomi, which closed May 27 at at approximately HK$28.40 per share, implying a market capitalization of roughly HK$740 billion on full-year 2025 revenue of RMB 457.3 billion — roughly 630 times Zhipu's revenue base. Alibaba trades at approximately HK$121.8 per share, with a market capitalization of roughly HK$2.3 trillion, implying a price-to-sales ratio below 2x on fiscal 2025 revenue of RMB 996.35 billion. Qwen and MiMo, the large language models embedded within Alibaba and Xiaomi respectively, are treated by the market as value-neutral or negative — a stark illustration of how the same asset class commands radically different multiples depending on its corporate wrapper.
Whether the HK$400 billion premium represents rational option pricing or a Shiller-style narrative bubble can only be resolved by sustained cash flow generation over a multi-year horizon. What is clear is that two companies with near-identical revenue scales, overlapping business models, and comparable model capabilities have been assigned fundamentally different identities by capital markets — and that identity, not data, is currently doing the pricing.
Related Coverage: