Pricing has bottomed out and is now rising, while model parameter sizes are surging, signaling a fundamental shift in the competitive landscape of China's large language model (LLM) industry.

For a long time, the Chinese LLM market was defined by a "price war," but a Morgan Stanley research report released on August 9th indicates this assessment is becoming outdated. The report highlights that the industry is moving from price competition to commercialization driven by intelligence, with three major structural changes occurring simultaneously: pricing becoming more rational, open-source licenses tightening, and model parameter sizes increasing.

This shift was directly triggered by DeepSeek's pricing action. On August 8, 2026, DeepSeek announced a significant increase in its API prices. For a company long known for its "high performance and extremely low pricing," this move carries implications far beyond a simple price adjustment. Morgan Stanley analysts Gary Yu and Lydia Lin interpret this as a positive signal of improving pricing discipline across the industry.

Behind the pricing pivot lies a deeper reshaping of the business logic. The analysts argue that model intelligence, not price, is the ultimate barrier to competition in the LLM space. Top-tier firms can launch lower-priced, lightweight versions, but mid-tier players face immense difficulty in reverse-engineering their way to leading models. This asymmetry makes pure price competition an unsustainable strategy. The true winners will be those who can maintain the positive flywheel of "stronger models leading to more revenue, leading to greater investment, leading to even better models."

Pricing Has Bottomed Out: DeepSeek's Price Hike is a Signal, Not an Anomaly

For an extended period, the average API prices for Chinese LLMs have been steadily declining, fueling market concerns that a "price war" would drag down the entire industry's profitability. However, data suggests a turning point may have arrived. According to the Morgan Stanley report, based on a sample including ByteDance, Alibaba, Baidu, Tencent, MiniMax, Zhipu AI, Moonshot AI, and DeepSeek, the average API output price for Chinese LLMs has recovered from about 12.2 yuan per million tokens in Q1 2025 to 21.9 yuan per million tokens in Q2 2026.

Analysts believe DeepSeek's price increase is driven by three factors: first, strong demand for its V4 model supports greater pricing power; second, firms need to balance market share with gross margins to sustain investment in cutting-edge model development; and third, increased use of domestic chips may lead to higher inference costs. The analysts conclude that a price war is not a sustainable competitive strategy. As model intelligence and scale continue to improve, the pricing of China's next-generation LLMs is expected to trend upward, with the competitive dimension increasingly shifting from price to capability and capacity.

This trend is visually evident: the higher a model's intelligence index, the higher its API output pricing. Models like K3 and Qwen3.8-Max sit in the high-price, high-intelligence quadrant of the chart, while V4-Flash is in the low-price, low-intelligence zone, illustrating a growing positive correlation between price and capability.

Licensing Tightens: From "Free to Use" to "Negotiate for Large-Scale Use"

Open-source models are another variable. A common market fear is that the widespread distribution of open-source weight models will further depress demand for AI infrastructure and intensify commoditization. Morgan Stanley holds a different view. The analysts point out that cheaper, more efficient open-source models should, through the Jevons Paradox, drive higher usage rates – lower inference costs will create greater demand, thereby accelerating AI penetration. Survey data shows that 63% of respondent companies use both open-source and closed-source models.

A more significant change lies in the evolution of licensing terms. Chinese LLM companies previously widely adopted permissive open-source licenses like Apache 2.0 or MIT (L1 level), allowing third-party free deployment and commercialization. However, the report shows that companies are accelerating their migration to stricter L2/L3 licenses. Specific examples include Moonshot AI's K3 license, which requires MaaS providers with annual revenue exceeding $20 million to sign a separate commercial agreement, and reports suggesting Alibaba may introduce revenue-sharing requirements for users who deploy its future Qwen open-source models on a large scale. This shift means the monetization path for LLM companies is expanding from simple API direct sales (1P) to revenue sharing (3P) with cloud service providers and model aggregation platforms, thereby enlarging the addressable ARR pool.

Parameter Leap: Chinese LLMs Enter the Trillion-Parameter Era

In the first half of 2026, the parameter sizes of mainstream Chinese open-source LLMs were generally below 1 trillion. In July, Moonshot AI's Kimi K3 broke this ceiling with 2.8 trillion parameters, becoming the first Chinese mainstream open-source LLM to cross the 2-trillion parameter threshold. Following closely, Alibaba's Qwen3.8-Max reached 2.4 trillion parameters. According to analysis, the 2-3 trillion parameter range will gradually become the baseline for China's frontier models in the second half of 2026, with several heavyweight products on the horizon: MiniMax M3 Pro (expected September-October, 2.7 trillion parameters), Zhipu AI's next-generation GLM model (expected October), and a potential release of Alibaba's Qwen4. Reports from LatePost on August 6, 2026, suggest ByteDance is even exploring models with over 5 trillion parameters.

The significance of increasing parameter size goes beyond improving performance metrics; it systematically raises the competitive barrier. The analysts point out that training large-parameter models requires not only more capital and computing power but also places higher demands on inference efficiency, infrastructure optimization, engineering execution capability, and financing ability. This mirrors the dynamics of capital-intensive industries: the higher the tier, the higher the entry barrier, and the stronger players become even stronger. The analysts believe that as the scale of frontier models continues to expand, the industry will further consolidate towards resource-rich leading players.

The Flywheel Logic: Commercialization and AGI are Not Opposing Forces

The analysts argue that commercialization and AGI are complementary, not competing goals. Stronger commercialization provides the funding to support frontier model research, while stronger frontier intelligence, in turn, drives commercial monetization. This creates a virtuous cycle: better models lead to stronger user adoption and monetization, which funds more training and computing power investment, leading to continuous model iteration. The analysts believe the true moat is not a temporary top ranking, but the ability to sustain this flywheel. Top-tier firms can easily launch lower-priced lightweight models to cover the mass market, but the difficulty for mid-tier firms to break through to the top tier is on a completely different level.

Z.AI (02513): The Flywheel is Forming, Target Price Raised to HK$1,700

Morgan Stanley has significantly raised its target price for Z.AI by 72%, from HK$990 to HK$1,700, and has increased its 2026 ARR forecast from $1 billion to $2 billion, corresponding to a 2027 price-to-sales ratio of approximately 42 times. The flywheel for Z.AI is accelerating, supported by three dimensions:

Model Performance: GLM-5.2 has joined the ranks of the world's top large models, performing well not only in benchmark tests but also receiving positive feedback from the global developer community. According to the report, Zhipu AI's Chief Scientist, Jie Tang, posted on the X platform on June 16, 2026, stating that the company would release a product by Q1 2027 that matches the current best US large model. The analysts expect Z.AI to achieve this goal ahead of schedule, by October 2026.

Computing Power Supply: Inference computing power, especially overseas capacity, has been a bottleneck limiting Z.AI's ARR growth. Zhipu AI is currently in negotiations with AWS, and the GLM-5.2 API has been integrated with the AWS Marketplace. Morgan Stanley expects this 3P partnership model to expand further to other overseas cloud service providers and API platforms.

Financing Cycle: On July 13, 2026, Z.AI completed an H-share rights issue, raising HK$31.4 billion (approximately $4 billion), with a dilution rate of 4.25%. The company plans to use 55% of this (about $2.2 billion) for model training and computing power expansion. Additionally, a potential A-share listing could provide a further approximately RMB 15 billion in funds.

MiniMax: Short-Term Pressure, but Bullish Reasons Remain

Morgan Stanley has slightly lowered its target price for MiniMax by 18%, from HK$1,100 to HK$900, while maintaining its 2026 ARR forecast of $1 billion (consistent with company guidance), corresponding to a 2027 price-to-sales ratio of approximately 32 times. MiniMax's stock price has recently come under significant pressure: following the June 1st release of the M3 model, negative feedback from the developer community – due to its relatively small parameter size (428 billion total parameters, only 23 billion active), a price double that of M2.7, and relatively weaker coding capabilities – caused the stock price to fall up to 77% from its June 1st high. On July 8th, the lock-up period for cornerstone investors and pre-IPO investors expired, instantly increasing the floating share ratio from 5.44% to 54.38% of total shares, further amplifying selling pressure.

However, the analysts state they are "not overly pessimistic" about MiniMax, citing four reasons:

M3 Pro is the True Catalyst: The analysts view the current M3 as an experimental version testing a new architecture (MiniMax's sparse attention mechanism), not the final product. MiniMax has announced that M3 Pro, with 2.7 trillion total parameters, will be launched in September-October. It is expected to be the largest Chinese LLM by parameter size at that time and demonstrate significant pricing power.

LLMs are More Than Just Coding: The analysts argue that while coding is the first area in AI to cross the technology and commercialization inflection point, it is not the whole picture. Future LLMs will penetrate broader knowledge worker markets like law, finance, consulting, and healthcare. MiniMax has consistently focused on vertical industry training data and launched a "10x Team" initiative in May 2026 to recruit experts from various fields.

Multimodality is Underestimated: The commercial value of the video model market is substantial, with relatively lower competition compared to the LLM market itself. The H3 video model released by MiniMax in July 2026 has surpassed global competitors and received positive feedback, potentially accelerating MiniMax's multimodal ARR growth.

Financing Support: On July 10, 2026, MiniMax announced a rights issue to raise HK$9.5 billion (about $1.2 billion) and a one-year convertible bond issuance to raise HK$6.5 billion (about $800 million), with a conversion price of HK$335. The company plans to use 80% of the funds from the rights issue and convertible bonds (approximately $1.6 billion) for model training and computing power expansion, and 10% for the global commercialization of its Harness product. A potential A-share listing could further broaden its financing channels. MiniMax founder and CTO, Dr. Yan Junjie, stated in an internal letter on July 10th that he would forgo his salary until the company achieves AGI, and would donate 4% of his personal company shares over the next four years for employee incentives, plus an additional 1% to establish a special fund supporting the open-source community.

R&D Spending Significantly Revised Upward: The Inevitable Cost of a High-Barrier Era

As model parameter sizes surge, R&D spending forecasts for both companies have been substantially revised upward. The analysts have raised their 2027 R&D expense forecasts for both Z.AI and MiniMax to over $1 billion, reflecting the higher investment required to train large-parameter models. Specifically, revenue forecasts for Z.AI for 2026-2028 have been raised by 37.3%, 53.1%, and 56.7%, respectively, while those for MiniMax for the same period have been raised by 104.2%, 71.1%, and 99.2%. Both companies are currently in a loss-making phase. Morgan Stanley expects Z.AI to achieve breakeven as early as 2028, when the rapid expansion of its cloud deployment business revenue outpaces the growth in R&D spending.