Woofun AI reports that Goldman Sachs analyst Ronald Keung has published a comprehensive assessment of China's artificial intelligence sector, identifying a structural shift where domestic open-source models like DeepSeek and Zhiyu GLM are achieving performance parity with global proprietary leaders through superior cost efficiency. The 50-page analysis, which also evaluates ByteDance, MiniMax, and Kuaishou, argues that the industry is moving beyond simple parameter scaling toward a competitive framework defined by pricing power, gross margin advantages, and financial resilience, positioning Chinese firms to capture significant share in the global enterprise market.

The core driver of this competitive advantage lies in architectural innovation rather than brute-force compute expansion. Goldman Sachs notes that Chinese open-source models typically operate within a parameter range of 200 billion to 1.6 trillion, representing only 2% to 10% of the scale of the world's largest proprietary models, a constraint largely imposed by limited access to high-end computing hardware. To compensate, developers have adopted mixture of experts architecture (MoE) and sparse attention mechanisms, ensuring that actual activated parameters account for merely 3% to 5% of the total model size. This structural efficiency drastically reduces both training and inference costs. Specific examples include DeepSeek V4 Pro with 1.6 trillion parameters, Zhiyu GLM5.2 with 0.7 trillion, and MiniMax M3 with 0.4 trillion. These figures illustrate a strategic pivot toward maximizing output per unit of compute, allowing Chinese firms to deliver comparable intelligence at a fraction of the resource expenditure required by their Western counterparts.

Performance enhancements have been further accelerated by localized infrastructure breakthroughs and advanced post-training techniques. On June 27, DeepSeek introduced the DSpark speculative decoding framework, which was immediately deployed in its V4-Flash and V4 Pro online services. This innovation increased user generation speed by 60% to 85% for V4-Flash and 57% to 78% for V4 Pro, all without altering model weights or compromising output quality. Simultaneously, the localization of hardware stacks has reached a critical milestone with the release of Meituan's LongCat 2.0 on June 30. Goldman Sachs highlights this as China's first open-source MoE model with 1.6 trillion parameters, trained and deployed entirely on 50,000 domestically produced computing cards. This achievement demonstrates the viability of a self-sufficient hardware ecosystem during the compute-intensive pre-training phase, significantly reducing reliance on foreign high-end chips and securing the long-term scalability of Chinese AI development.

The market structure is rapidly polarizing into distinct high-end and low-end tiers, each with unique pricing dynamics and margin profiles. In the premium segment, models such as Zhiyu GLM5.2 and Alibaba Qwen3.7 Max are priced at approximately $1 per million tokens, which is five times the rate of low-end offerings and yields an estimated gross margin of 10% to 20%. While this price point is only 10% to 25% of the $4 to $8 per million tokens charged by top U.S. models, Chinese firms maintain positive margins due to their lower parameter activation ratios. Conversely, the low-end market targets price-sensitive global small and medium enterprises and individual users, with agent-focused models priced as low as $0.06 to $0.20 per million tokens. MiniMax derives 60% to 70% of its revenue from these overseas markets.

Notably, DeepSeek announced that starting mid-July, it would implement a peak and off-peak pricing mechanism for its V4 series, with peak rates double the off-peak rates, resulting in mixed pricing of about $0.35 per million tokens for V4 Pro and $0.12 for V4 Flash.

Woofun AI reports that this pricing stratification supports aggressive revenue growth projections, with API and subscription income for Chinese AI models expected to surge from an estimated 35 billion RMB in 2026 to 879 billion RMB by 2030. This financial expansion correlates with a massive increase in daily token consumption, projected to rise from 350 trillion to 4,600 trillion, representing a 25-fold increase over the period. The sheer volume of data processing is expected to create a powerful flywheel effect, where increased usage drives further model iteration and efficiency gains, reinforcing the competitive moat of leading players. This trajectory suggests that the Chinese AI sector is not merely catching up but is establishing a new standard for cost-effective intelligence delivery, capable of sustaining high growth rates even in a capital-constrained environment.

The widespread adoption of open-source and open-weight strategies is central to this growth, though monetization paths remain complex. Models such as the Alibaba Qwen series, DeepSeek, Zhiyu GLM, and MiniMax M3 utilize open-source or open-weight distributions to maximize deployment flexibility and accelerate iteration through community feedback. ByteDance's Seed model stands as a notable exception, adhering to a completely closed-source proprietary route.

However, Goldman Sachs warns that reported ARR figures for open-source companies likely underestimate actual deployment scale and revenue potential. For instance, while Zhiyu targets an ARR of $1 billion by the end of 2026, the global deployment of GLM5.2 far exceeds the token volume generated through Zhiyu's own API channels. Platforms like Alibaba Cloud's Bailian MaaS can host the GLM5.2 open-source model directly without paying fees to Zhiyu, highlighting the disconnect between model popularity and direct revenue capture.

To address these monetization limitations, the industry is shifting from pure open-source models under MIT licenses to an 'open-weight + community license' framework, where commercial usage requires revenue-sharing agreements. MiniMax has already adopted this model for its M series. Goldman Sachs anticipates this transition will significantly improve unit economics, as model developers can benefit from revenue shares with hosting platforms like AWS Bedrock and Alibaba Cloud Bailian without bearing the inference computing costs themselves. This structural change allows model creators to capture value from the broader ecosystem while offloading the operational burden of serving large-scale traffic, creating a more sustainable business model for independent AI firms.

Global expansion is identified as the most critical upward space for Chinese AI models, particularly in non-U.S. markets where price sensitivity is higher. Goldman Sachs' U.S. research team estimates that by 2030, agent AI will drive global token consumption to grow 24 times, reaching 120 trillion tokens per month. Within this volume, enterprise agents are expected to contribute a 55-fold increase, while consumer agents will see a 12-fold rise. Chinese models have already gained significant token share in these global markets due to their performance improvements and price advantages. The usage paradigm among global enterprises is undergoing a fundamental shift from 'token maximization' to 'ROI priority.' The former approach, prevalent from late 2025 to early 2026, equated high token consumption with organizational productivity. The emerging ROI-focused model emphasizes clear task boundaries, daily active agent counts, backend process automation, and actual output. A Jellyfish AI engineering trend study supports this shift, showing that heavy AI users in enterprises consumed 10 times the tokens but only doubled their output, indicating diminishing returns on pure volume.

Strategic partnerships with major cloud providers are facilitating this global penetration. Alphabet's Gemini Enterprise Agent Platform and Amazon's AWS Bedrock have both integrated hosting services for Chinese models including DeepSeek, MiniMax, Moonshot, GLM, and Qwen. According to the Wall Street Journal, Microsoft is also considering hosting a version of DeepSeek on Copilot as a low-cost option, with CEO statements emphasizing that the model would run within Microsoft's cloud ecosystem to ensure customer data remains within Azure. These collaborations validate the technical quality of Chinese models while providing them with the distribution channels necessary to scale globally, effectively bypassing some of the geopolitical barriers that might otherwise limit their reach.

Goldman Sachs has constructed a three-dimensional competitive positioning framework to quantitatively assess long-term winning probabilities, using the formula: ARR scale × gross margin advantage + financial strength. The pricing capability dimension evaluates market entry speed, LMArena arena scores based on blind user evaluations, and mixed pricing levels. The cost advantage dimension assesses throughput, cache hit rate, parameter activation ratio, and inference gross margin. The financial strength dimension examines cash on hand, net cash as a percentage of total assets, and valuation multiples. In the foundational text model field, Zhiyu (neutral rating, target valuation of $110 billion) and DeepSeek (unlisted) are identified as the strongest positioned players, excelling in both pricing and cost metrics. The overall implied valuation of independent AI model companies exceeds $200 billion, reflecting strong investor confidence in their scalable business models.

In the multimodal and video generation sector, ByteDance leads with Seedance, which LatePost and 36Kr report has a gross margin as high as 70% and an ARR run rate exceeding $2 billion. Kuaishou's Keling and MiniMax's Hailuo and upcoming H3 model are also viewed positively, expected to benefit in the second half of 2026 from functional breakthroughs in video generation and LLM integration, as well as healthy pricing due to supply constraints. Goldman Sachs maintains a buy rating on MiniMax with a target price of HKD 860, reasoning that its M3 model sits in the ARR maximizing quadrant of high token volume and attractive pricing. With a current valuation of only 13 times the ARR at the end of 2026, MiniMax presents a significant discount compared to global peers, offering an asymmetric risk-reward profile.