Whoever lowers prices is weaker! JPMorgan: Zhipu and Minimax conducted the same experiment, but got opposite results.

In a market environment where AI demand still exceeds inference supply, price reductions are not a gesture of goodwill, but are a self-acknowledgment of insufficient competitiveness.

According to Chase Wind Trading Desk, on June 12, JPMorgan released a research report using this logic as the core and made diametrically opposite rating decisions for two Chinese listed AI model companies—maintaining an “overweight” rating for Zhipu, and downgrading MiniMax to “neutral.” The reason the two companies went down different paths boils down to the same variable: pricing power.

The immediate trigger for this rating adjustment was a pricing action taken by MiniMax on June 8.

The price for MiniMax’s flagship model M3 was roughly twice that of the previous generation M2.7 at launch, but just about a week later, it announced a permanent 50% price reduction, dropping close to the M2.7 level. JPMorgan interpreted this as a clear signal—the intelligence improvement brought by M3 failed to gain market recognition for its intended premium.

In contrast, Zhipu has taken the opposite path.

Since the beginning of the year, Zhipu has doubled its API prices, and maintained this price level even with continual growth in usage. JPMorgan believes this is typical “pricer behavior,” which aligns with its consistently setting domestic SOTA (state-of-the-art models) via GLM 5 and 5.1 releases.

The significance of this report goes beyond the upgrade and downgrade of company ratings; it also provides an actionable valuation framework for AI model companies: premium valuation must pass three tests—repeated SOTA delivery, verified pricing power, and sustainable workflow adoption.

As monetization paths increasingly converge on enterprise workflows, API consumption, and coding agents, leading model capabilities are now highly linked to pricing power.

Price reduction is a verdict: Pricing is a more reliable market signal than benchmarking

JPMorgan points out in the report that the two dimensions for evaluating model cost performance are intelligence and price, and price is the easier to observe and harder to fake signal. Benchmarks are updated monthly and may be optimized for, but pricing is continuous, public, and set by those most familiar with their own demand curve.

The core logic of the report is: at a stage where AI usage demand still exceeds inference supply, no developer will proactively reduce prices in such an oversupply situation.

If a model company quickly retracts a premium after launching a new model, it essentially admits via its pricing behavior—the market does not accept this premium. JPMorgan calls this rapid retreat from premium pricing an admission from the developer that the intelligence improvement did not justify the intended premium for the market.

Based on this logic, JPMorgan built an SOTA recognition framework centered on token pricing, complemented by third-party benchmarking (Artificial Analysis, etc.), LMArena real user preferences, and actual developer and enterprise workflow adoption as cross-verification.

The report emphasizes that for investors, the strongest evidence is convergence of the above four dimensions: high-end pricing with resilience, strong benchmark performance, positive LMArena preference, and sustained adoption in real workflows.

Same experiment, two radically different results

JPMorgan clearly writes in the report: “Zhipu and MiniMax did the same experiment, but the results were the opposite.”

Zhipu’s experiment result: after the price increase, usage continued to grow.

Since the beginning of the year, Zhipu’s API price has doubled, but customers did not leave, meaning downstream workflows are dependent enough on GLM series to support price hikes.

JPMorgan believes this combination—continuous SOTA delivery plus market-validated pricing power—is the strongest evidence for evaluating AI foundational model companies. Even after the successive releases of Kimi K2.6 and DeepSeek V4, GLM-5.1 still ranks at the top of Code Arena and WebDev Arena among domestic models, demonstrating consistency in delivering cutting-edge capabilities.

MiniMax’s experiment result was the opposite. M3’s launch price was about twice that of M2.7, but within a week a permanent 50% price reduction was announced.

JPMorgan interprets this as the market firmly rejecting M3’s premium pricing. Moreover, since M2, MiniMax has not reclaimed the SOTA position in China with subsequent iterations, while competitors continuously refresh the leading edge via GLM-5/5.1, Kimi K2.6, and DeepSeek V4. JPMorgan believes that from a pure model capability perspective, MiniMax is still in the catch-up phase.

This comparison determines the valuation status of the two companies: JPMorgan gives Zhipu a premium valuation of 57x expected P/S in 2027, while MiniMax’s target price corresponds to 29x, “in line with providers pricing at anchor points.”

DeepSeek drives down the market clearing price for 'good enough'

The competitive pressure MiniMax faces is compounded by DeepSeek’s systematic pricing reset.

JPMorgan’s report points out that DeepSeek V4, through its lower-cost Flash version and aggressive caching price, significantly lowers the market clearing price for “sufficiently intelligent” tasks—wherever DeepSeek can fully handle a task, that price anchor is pulled lower.

The report shows that among major domestic LLM providers, composite token prices (calculated at 80% cache hit rate, 10:1 input-output ratio) are clearly stratified:

Qwen3.7-Max is about ¥7.2 per million tokens, GLM-5.1 ¥5.45, MiniMax M3 about ¥1.45 after permanent price cut, DeepSeek V4 Pro about ¥1.11, V4 Flash only ¥0.38.

After MiniMax’s price cut, it is now in the same price zone as DeepSeek’s series, meaning it has chosen to compete at the price level set by DeepSeek.

JPMorgan considers this most challenging for models positioned primarily on cost-performance—they are squeezed at both ends: low-cost providers (DeepSeek, big platforms) pressure the pricing end, and SOTA providers pressure the high-value task completion end.

Routine text generation, low-risk coding assist, simple tool calls, and similar workloads will face more drastic price compression, while complex workflows with high failure costs and strong reliability requirements can still support high-end pricing for SOTA models.

Monetization paths narrow, SOTA cadence determines valuation premium

JPMorgan’s report notes that the monetization path of the AI industry is now highly convergent—whether domestic or international, whether independent model companies or large platforms, the clearest monetization layers focus on enterprise workflows, API consumption, coding, and agent deployment.

Alibaba, Tencent, ByteDance are all pursuing the same direction, meaning independent model companies like Zhipu, MiniMax, and Kimi are now in a more direct competitive environment—competing both with each other, and with large platforms with model capabilities, distribution channels, cloud infrastructure, and stronger balance sheets.

In this landscape, the acceleration of model iteration cycles increases the strategic value of continuous SOTA delivery.

JPMorgan points out that release cycles have now shortened from relatively relaxed 3–6 months to a much tighter competitive window, raising the cost of falling behind. A strong release can drive usage, but only continuous leadership in coding, inference, agent execution, and enterprise reliability can truly sustain revenue quality.

JPMorgan also notes that switching costs are still low for most Chinese independent model companies—developers can try multiple models, allocate traffic via aggregators, and enterprises can benchmark multiple models within the same workflow.

When models are not deeply bound to proprietary tools, product workflows, or data closed-loops, the durability of API revenue relies more on model leadership than on usage scale at any point in time.

Forecasts adjusted sharply: Zhipu up, MiniMax down

At the financial forecast level, the adjustment directions for the two companies are also strongly contrasted.

For Zhipu, JPMorgan raises expected revenue from 2026 to 2030 by 26%–42%, reflecting improved visibility of high-quality revenue growth under steady model iteration cycles.

Predicted net loss for 2026–2028 is slightly narrowed after adjustment, target price raised from HK$950 to HK$1,400, corresponding to 30x 2030 expected P/E, discounted at a 15% weighted average cost of capital.

For MiniMax, JPMorgan raises revenue forecasts for 2026–2027 by 34%–74% (based on the industry still being constrained by compute supply and MiniMax’s flexible compute procurement), but lowers 2028–2030 revenue forecasts by 5%–21%, citing declining visibility in long-term monetization for non-SOTA LLM providers.

The permanent 50% price cut for M3 leads to sharply lowered profit margin expectations, with adjusted net loss forecasts for 2026–2028 expanding from US$309m/596m/512m to US$432m/940m/972m. Target price is cut sharply from HK$1,100 to HK$400.

Falsifiable triggers for rating adjustments

JPMorgan specifies concrete conditions under which this rating adjustment may be overturned, highlighting the framework’s operability.

For MiniMax’s “neutral” rating, several conditions could restore “overweight”:

MiniMax launches its next-generation flagship at a premium and maintains that price for an entire quarter; or resets domestic leading capability with both third-party benchmarks and user data;

or multimodal monetization paths show clear income evidence in specific scenarios (marketing automation, game content production, video creation, education, etc.);

or API pricing remains stable, retention improves, and gross margin is attractive.

For Zhipu’s “overweight” rating, downside triggers are: DeepSeek’s next frontier release causes sharp price reduction in GLM premium tier, or usage loss exposes demand elasticity at doubled prices.

JPMorgan notes short-term catalysts center on GLM and M series release windows, as well as monthly updated third-party leaderboards.

~~~~~~~~~~~~~~~~~~~~~~~~

The above brilliant content is from Chase Wind Trading Desk.

For more detailed interpretation, including real-time and frontline research, please join [Chase Wind Trading Desk▪Annual Membership]

Risk Warning and DisclaimerThe market has risks, and investment requires caution. This article does not constitute personal investment advice, nor does it take into account the individual user’s particular investment objectives, financial situation or needs. Users should consider whether any opinions, views, or conclusions herein fit their specific situation. Investment based on this, responsibility lies entirely with the user.