When a stock surges five to eight times within three months of its IPO, the immediate question is whether it is still a buy. J.P. Morgan's response is affirmative, but with a caveat: investors must first understand what is already priced into the stock. KNOWLEDGE ATLAS and MiniMax listed on the Hong Kong Stock Exchange in January 2026, and their share prices have skyrocketed since, with cumulative gains of approximately 500% to 800% by April 23. KNOWLEDGE ATLAS, for instance, rose from its IPO price of HKD 116.20 to HKD 1,001. They are currently the world's only two publicly traded companies focused purely on frontier AI large language models, and this scarcity inherently commands a premium. According to market sources, J.P. Morgan maintained an "Overweight" rating on both companies in a research report dated April 22, with a target price of HKD 950 for KNOWLEDGE ATLAS and HKD 1,100 for MiniMax. However, the report's focus is not merely on whether to buy, but on systematically addressing a more complex question: what factors have already been discounted in this rally, what could drive further gains, and what potential risks could lead to a decline.
The share prices embed an Anthropic-like growth narrative. The current market capitalizations of the two companies range between $40 billion and $55 billion. What does this figure imply? J.P. Morgan's analysis suggests the market is already pricing in an expectation that KNOWLEDGE ATLAS will achieve an Annual Recurring Revenue of approximately $1 billion by the end of 2026, with MiniMax reaching around $700 million. But this is just the first layer. The more critical, second layer of implied expectation is that the market is not only betting on them reaching the $1 billion mark but also on their ability to follow an Anthropic-like growth trajectory after surpassing it. The Anthropic benchmark is as follows: in March 2025, its ARR was approximately $1.4 billion with a valuation of about $61.5 billion; by August 2025, its ARR had surpassed $5 billion—a 3.5x increase in just six months. Following this logic, the current valuation of KNOWLEDGE ATLAS and MiniMap presupposes that they can roughly replicate this growth curve after crossing the $1 billion ARR threshold.
J.P. Morgan directly pointed out: "Such expectations are high, especially in the more competitive Chinese market." When Anthropic grew from $1 billion to $5 billion, it faced limited direct competition. The situation in China is截然不同—Alibaba, ByteDance, and Tencent are all increasing their B2B efforts. Tencent has launched WorkBuddy, while Alibaba and ByteDance are scaling up Qoder and TRAE respectively, with ByteDance bundling its Doubao model into enterprise suites. If the same growth opportunity is distributed among 6-8 companies, the growth rate for each will naturally be slower.
ARR growth is not linear; it is a step function. Despite the high expectations, J.P. Morgan maintains a structurally optimistic stance, partly because current ARR figures may underestimate true demand. KNOWLEDGE ATLAS disclosed that as of March 31, 2026, its API ARR had exceeded $250 million, representing a 60-fold year-on-year increase and a 6.4-fold increase since the start of the year. Each new release of the GLM model (from 4.5 to 4.6, 4.7, 5, and 5.1) has triggered a jump in ARR—the growth curve became "almost vertical" after the launch of GLM-5. MiniMax exhibited a similar pattern following the release of its M2.2 to M2.7 series. This indicates that ARR growth is not a steady climb but a step function tied to model iterations. The next catalyst—the anticipated releases of GLM 5.5 and MiniMax M3 around June—is currently the most predictable single ARR driver. Demand-side support is also evident. According to People's Daily data, China's daily token usage exceeded 140 trillion in March 2026, a sharp increase from just 100 billion at the beginning of 2024 and 100 trillion at the end of 2025—the acceleration slope is extremely steep.
The problem lies in computing power. All major Chinese LLM providers have indicated that inference computing capacity cannot keep pace with token demand growth. This leads to a counterintuitive conclusion: the current observed ARR growth rate represents a floor, not a ceiling. Once computing bottlenecks are alleviated, pent-up demand can be directly converted into recognized revenue. Alibaba Cloud announced a 34% price increase for AI computing on April 18, and KNOWLEDGE ATLAS's API prices have nearly doubled since the start of the year—demand remains robust despite the price hikes, indicating a "significantly more solid" pricing environment. Pricing improvement involves more than just hikes; three mechanisms are working in tandem. Pricing improvement is one of the most significant industry changes in 2026. J.P. Morgan identified three concurrently acting mechanisms: First, direct repricing tied to model upgrades. From GLM-4.7 to GLM-5 and GLM-5.1, KNOWLEDGE ATLAS significantly increased its API pricing, with effective token prices nearly doubling since the year's start. The absence of significant demand contraction alongside price increases suggests rising corporate willingness to pay. Second, a migration from subsidized packages to pay-as-you-go models. Alibaba Cloud stopped accepting new users for its Lite coding package from March 20, making the Pro version the main entry point; KNOWLEDGE ATLAS employs a similar separation strategy—coding package quotas are restricted to its coding tools, while API calls are billed separately. Discounted packages serve to reduce adoption friction, but the real revenue opportunity comes from customer consumption beyond the bundled content. Third, KV cache repricing. This is the most easily overlooked mechanism. Compared to M2.5, MiniMax's M2.7 maintains the same overall input/output pricing, but the price for prompt cache reads increased from $0.03 to $0.06 per million tokens. In coding and agent workloads, cached tokens constitute a significant portion of billing. J.P. Morgan estimates the effective price of M2.7 is over 30% higher than M2.5—even though the headline price remained unchanged.
Upcoming risk: In July, 39% of MiniMax's shares will be released from lock-up. Beyond fundamentals, liquidity supply is the most direct technical pressure in the second half of 2026. Both companies had minimal free floats at IPO: KNOWLEDGE ATLAS around 3.9%, MiniMax around 5.4%. Lock-up shares will be released in batches— KNOWLEDGE ATLAS: 5.8% released in early July 2026, 90.3% released in January 2027. Near-term pressure is relatively mild. MiniMax: 39.0% released in early July 2026, approximately 18.2% released around October 2026, and 37.4% released in January 2027. Supply pressure in H2 2026 is noticeably greater for MiniMax. Pre-IPO shareholders have extremely low cost bases, with paper gains of 5-7 times, making profit-taking pressure post-lock-up non-negligible.
J.P. Morgan referenced the lock-up expiration histories of Kuaishou, SenseTime, and Horizon Robotics: Kuaishou's share price fell from HKD 250 to HKD 120 around its 6-month lock-up expiry; SenseTime declined significantly after its lock-up. However, there are differences—KNOWLEDGE ATLAS and MiniMax currently have average daily trading volumes of approximately HKD 1.2-1.5 billion, indicating significantly better liquidity than the aforementioned cases. Pre-expiry turnover rates are also higher, suggesting stronger market absorption capacity. More importantly, the lock-up expiry window highly overlaps with the June model release cycle. If the releases of GLM 5.5 and MiniMax M3 are strong, some selling pressure could be hedged; if the model performances are mediocre, share prices could be more vulnerable to increased supply. Hedging factors: KNOWLEDGE ATLAS and MiniMax are expected to be included in the Hang Seng Composite Index and the Hang Seng Tech Index on June 8, bringing passive fund inflows; KNOWLEDGE ATLAS is expected to be included in the Southbound Stock Connect in the first week of June, while MiniMax, due to its weighted voting rights structure, is anticipated to join in the first week of August. The scarcity premium has a window, estimated at 6-12 months. KNOWLEDGE ATLAS and MiniMax are currently the world's only two pure-play frontier AI LLM public companies. Scarcity is a key source of their valuation premium, but this window will not remain open indefinitely. J.P. Morgan cited a direct precedent: Cambricon was once the only pure-play AI chip listed company in the A-share market, with its share price around RMB 1,500 in late November 2025. Subsequently, Moore Threads, MetaX, and Biren Technology listed between December 2025 and January 2026. Even though Cambricon's 12-month forward revenue estimates were raised by 17% and profit forecasts by 26%, its share price has fallen approximately 2% year-to-date, with valuation multiples contracting about 25-30%. If Kimi (Moonshot AI) and StepFun were to list, the impact on KNOWLEDGE ATLAS and MiniMax would be similar. In the short term, IPO activity brings capital inflows and benefits sector sentiment; but each company's scarcity premium would structurally decline. J.P. Morgan judges the window for being the "only way to gain exposure to Chinese AI" to last approximately 6-12 months.
The US situation: IPOs of Anthropic and OpenAI could alter valuation frameworks. Anthropic's ARR surpassed $30 billion on April 6, up from $9 billion at the end of 2025—a 3.3x increase in roughly 3.5 months; OpenAI is projected to achieve revenue of approximately $29 billion in 2026. Their combined ARR exceeds $59 billion. This has a two-fold impact on Chinese LLM companies: Positive aspect: The escalating monetization scale of US frontier models continuously resets the market's perception of the potential market ceiling. Chinese token consumption is also experiencing explosive growth. If China's adoption curve catches up, KNOWLEDGE ATLAS's year-end $1 billion target might be conservative. The combined market cap of KNOWLEDGE ATLAS and MiniMax is approximately $80 billion, representing about 5% of their US peers' valuation—the relative discount is actually widening. Risk aspect: Once Anthropic or OpenAI goes public, the valuation framework might shift from "ARR multiples" to a public market logic that places greater emphasis on gross margins, operating leverage, and free cash flow. At that point, companies with strong revenue growth but lower medium-term profit visibility could face greater valuation pressure. J.P. Morgan's warning signals. While maintaining "Overweight" ratings, J.P. Morgan clearly outlined five signals that could prompt a shift to caution: These include losing relative model leadership in coding, agent, and enterprise workloads; internet or cloud giants establishing clear technical superiority, weakening the premium pricing power of independent model suppliers; a resurgence of API price competition, compressing the link between "superior models" and "higher pricing"; a renewed widening of the US-China model capability gap; and strong usage growth coupled with persistently weak ARR conversion. "As long as these conditions remain favorable, technical pullbacks should be viewed as windows for adding positions." Currently, J.P. Morgan estimates the gap between US and Chinese frontier models at about 9-12 months, with China rapidly catching up. This is one of the core supporting assumptions for the current valuation.
Comments