In a market where AI demand still outstrips inference supply, a price cut is not a gesture of goodwill but a self-certification of diminished competitiveness.
JPMorgan Chase published a research report on June 12, using this logic as its core to deliver starkly contrasting rating judgments on two listed Chinese AI model companies—maintaining an "Overweight" on KNOWLEDGE ATLAS while downgrading MiniMax to "Neutral". The divergence in the companies' paths is attributed to a single variable: pricing power.
The Catalyst for Downgrade
The immediate trigger for this rating adjustment was a pricing move by MiniMax on June 8.
The launch price for MiniMax's flagship M3 model was approximately double that of its predecessor, M2.7. However, after only about a week, the company announced a permanent 50% price reduction, bringing it back to a level close to M2.7. JPMorgan Chase interprets this as a clear signal that the intelligence gains from M3 failed to secure market acceptance of the intended premium.
In contrast, KNOWLEDGE ATLAS has charted a path in the opposite direction.
Since the start of the year, KNOWLEDGE ATLAS has doubled its API prices and has maintained this level despite continued growth in usage volume. JPMorgan Chase views this as a classic characteristic of a "price maker," consistent with its rhythm of consistently refreshing the domestic SOTA (state-of-the-art) benchmark through GLM 5 and 5.1.
Beyond Ratings: A Valuation Framework
The significance of this report extends beyond the rating changes for the two companies; it provides an actionable framework for valuing AI model companies. A premium valuation must pass three concurrent tests: repeated delivery of SOTA models, validated pricing power, and sustainable adoption in workflows.
As monetization paths increasingly converge on enterprise workflows, API consumption, and coding agents today, leadership in model capability has become highly correlated with pricing power.
Price as the Ultimate Arbiter
JPMorgan Chase's report notes that the two dimensions for evaluating a model's cost-effectiveness are intelligence and price, with price being the more easily observable and harder-to-fake signal. Benchmarks are refreshed monthly and can be optimized, while listed prices are continuous, public, and set autonomously by the party with the best understanding of its own demand curve.
The report's core logic is: in a phase where AI usage demand still exceeds inference supply, no developer would voluntarily cut prices when demand is in surplus.
If a model company quickly retreats from a premium price after launching a new model, it is essentially using its pricing action to admit that the market does not accept that premium. JPMorgan Chase terms this rapid retreat from premium pricing as "the developer's own admission that the intelligence gains did not justify the intended premium to the market."
From this logic, the bank constructed a SOTA recognition framework centered on token pricing, supplemented by third-party benchmarks (like Artificial Analysis), LMArena real-user preferences, and actual adoption by developers and enterprise workflows for cross-verification.
The report emphasizes that, for investors, the most compelling evidence is the convergence of the above four dimensions: a combination of premium pricing and resilience, strong benchmark performance, positive LMArena preference, and observable sustained adoption in real workflows.
A Tale of Two Experiments
JPMorgan Chase explicitly wrote in the report: "KNOWLEDGE ATLAS and MiniMax conducted the same experiment, but with opposite results."
KNOWLEDGE ATLAS's experiment resulted in: continued usage growth after a price increase.
Since the start of the year, KNOWLEDGE ATLAS has doubled its API price, but clients have not churned, indicating that downstream workflow reliance on the GLM series is sufficient to support the hike.
JPMorgan Chase believes this combination—sustained SOTA delivery coupled with market-validated pricing power—is the strongest evidence for evaluating an AI foundational model company. Even after the releases of Kimi's K2.6 and DeepSeek's V4, GLM-5.1 still ranks among the top domestic models in Code Arena and WebDev Arena, demonstrating consistency in delivering cutting-edge capabilities.
MiniMax's experiment yielded the opposite result. The M3 launched at roughly double the price of M2.7, but announced a permanent 50% price cut within a week.
JPMorgan Chase interprets this as the market rejecting M3's premium expectations through its actions. Furthermore, since the release of M2, MiniMax has not re-established a domestic SOTA position in subsequent iterations, while competitors like KNOWLEDGE ATLAS (GLM-5/5.1), Kimi (K2.6), and DeepSeek (V4) have continued to refresh the frontier. From a pure model capability perspective, JPMorgan Chase views MiniMax as still in a catch-up phase.
This contrast directly determined the valuation assigned to each company: JPMorgan Chase gives KNOWLEDGE ATLAS a premium valuation corresponding to a 57x 2027 expected price-to-sales ratio, while MiniMax's target price corresponds to 29x, "in line with providers priced on an anchor basis."
DeepSeek Resets the Market Clearing Price
The competitive pressure on MiniMax is compounded by a systemic pricing reset from DeepSeek.
JPMorgan Chase's report points out that DeepSeek V4, through its lower-cost Flash version and more aggressive caching pricing, has significantly lowered the market clearing price for "sufficiently intelligent" models—for tasks DeepSeek can adequately handle, this price anchor has been pulled downward.
The report shows that among major domestic LLM providers, comprehensive token prices (calculated at an 80% cache hit rate and a 10:1 input/output ratio) present a clear hierarchy: Qwen3.7-Max is about ¥7.2/million tokens, GLM-5.1 is ¥5.45, MiniMax M3 post-permanent-cut is about ¥1.45, DeepSeek V4 Pro is about ¥1.11, and V4 Flash is only ¥0.38.
Post-cut, MiniMax is in the same price tier as the DeepSeek series, meaning it has chosen to compete on price in the "layer where DeepSeek sets the price."
JPMorgan Chase believes this is most challenging for models primarily positioned on cost-effectiveness—they face pressure from both ends: low-cost providers (DeepSeek, large platforms) pressuring on price, and SOTA model providers pressuring on the completion of high-value tasks.
Workloads like routine text generation, low-risk coding assistance, and simple tool calls will face more intense price compression, while complex workflows with high failure costs and strong reliability requirements can still support the premium pricing of SOTA models.
The Narrowing Path to Monetization
JPMorgan Chase's report posits that current AI industry monetization paths are highly convergent—both domestically and internationally, for independent model companies and large platforms alike, the clearest monetization layers are concentrating on enterprise workflows, API consumption, coding, and agent deployment.
Alibaba, Tencent, and ByteDance are all deploying in the same direction, meaning independent model companies like KNOWLEDGE ATLAS, MiniMax, and Kimi now operate in a more direct competitive environment, competing not only with each other but also with large platforms possessing model capabilities, distribution channels, cloud infrastructure, and stronger balance sheets.
In this landscape, the compression of model iteration cycles further elevates the strategic value of sustained SOTA delivery.
JPMorgan Chase notes that release cycles have compressed from a relatively relaxed rhythm of about 3 to 6 months to a shorter competitive window, raising the cost of falling behind. A single strong release can boost usage, but sustained leadership in coding, reasoning, agent execution, and enterprise reliability is what truly underpins revenue quality.
The bank also points out that for most independent Chinese model companies, switching costs remain low—developers can test multiple models and allocate traffic via aggregators, and enterprises can benchmark multiple models within the same workflow.
When models are not deeply integrated with proprietary tools, product workflows, or data flywheels, the durability of API revenue relies more on continuous model leadership than on transient usage scale.
Contrasting Financial Forecast Adjustments
At the financial forecast level, the adjustment directions for the two companies also form a strong contrast.
For KNOWLEDGE ATLAS, JPMorgan Chase raised its expected revenue for 2026-2030 by 26% to 42%, reflecting improved visibility for quality revenue growth supported by a robust model iteration cycle.
Adjusted net loss forecasts for 2026-2028 narrowed, and the target price was raised from HK$950 to HK$1,400, corresponding to a 30x 2030 expected P/E ratio, discounted at a 15% weighted average cost of capital.
For MiniMax, the bank raised its 2026-2027 revenue forecasts by 34% to 74% (based on the industry still being compute-constrained and MiniMax having some flexibility in compute procurement) but lowered its 2028-2030 revenue forecasts by 5% to 21%, citing reduced visibility into long-term monetization for non-SOTA LLM suppliers.
The permanent 50% price cut for M3 led to a significant downward revision in margin expectations, with adjusted net loss forecasts for 2026-2028 widening from $309 million, $596 million, and $512 million to $432 million, $940 million, and $972 million, respectively. The target price was slashed from HK$1,100 to HK$400.
Potential Triggers to Reverse the Calls
JPMorgan Chase specified concrete conditions under which these rating adjustments could be overturned, demonstrating the framework's operational nature.
For MiniMax's "Neutral" rating, a return to an "Overweight" is possible if: MiniMax launches a next-generation flagship product at a premium and maintains that price for a full quarter; or achieves a reset of domestic frontier capabilities validated by both third-party benchmarks and user data; or clear revenue evidence emerges for multimodal monetization paths in specific scenarios (marketing automation, game content production, video production, education, etc.); or API pricing stabilizes, retention improves, and gross margins become attractive.
The downside trigger for KNOWLEDGE ATLAS's "Overweight" rating would be: DeepSeek's next frontier release forcing a sudden price cut for the GLM premium tier, or usage churn exposing the demand elasticity of the doubled price.
JPMorgan Chase notes that short-term catalysts are concentrated around the release windows for GLM and the M series, as well as monthly updated third-party leaderboards.
Comments