The Shrinking Profit Margins of AI Inference Highlighted by GLM 5.2

GLM 5.2 is an open-weight AI model developed by Z.ai, showcasing performance on par with Anthropic’s Opus and OpenAI’s GPT models (the latest being GPT 5.5). However, its emergence does not just signify escalating competition in model development; it also sheds light on vulnerabilities inherent in the revenue structure of the AI industry. This article, based on Martin Anderson’s analysis, examines the economics of AI inference and the looming possibility of margin collapse in the sector.

The Essence of the DeepSeek Shock

In early 2025, the release of DeepSeek R1 sent shockwaves through the stock market. Reports of the model’s training costs, estimated at around $6 million, prompted growing concerns about excessive capital investments in the AI sector, causing Nvidia’s stock to plummet overnight. However, Anderson argues that this perspective fails to grasp the core of the AI economy.

Training is a fixed cost, incurred once and completed. In contrast, inference is a variable cost that fluctuates with demand. AI providers’ business models rely on making massive investments in training and then recovering these costs over the long term through the sale of inference APIs. According to Anderson, the market’s focus on training costs alone reflects a fundamental misunderstanding of this structure.

The Economic Structure of Training and Inference

While training costs are fixed, inference costs represent true marginal expenses. According to Anderson’s estimates, the computational resource costs associated with providing APIs, priced at around $25 per MTok by companies like Anthropic and OpenAI, amount to roughly one-tenth of that figure. This means the gross profit margin on API sales could reach as high as 90%.

However, leaked financial statements from OpenAI indicate an overall gross margin of approximately 60%, including expenses for support, payment processing, and other services. This discrepancy reveals how non-API-related costs erode overall profitability. Anderson states, “The essence of the AI lab business model lies in making significant capital investments in training and maintaining high profit margins through inference sales.” If companies can achieve sufficient volume with these margins, they can secure overall profitability.

The Capabilities and Limitations of GLM 5.2

Anderson describes GLM 5.2 as “the first open-weight model that genuinely rivals Opus and GPT.” In practice, he notes that any quality differences between GLM 5.2 and Opus are difficult to discern in daily use. However, GLM 5.2 comes with clear limitations.

First, its response time is slow. Due to the extensive processing involved in “thinking,” it is unsuitable for interactive tasks. While it performs adequately in non-interactive tasks like reviewing pull requests in the background, it struggles with real-time interactions. This sluggishness translates to higher token consumption, ultimately reducing cost efficiency.

Second, it lacks support for vision capabilities. Anderson explains that since the introduction of high-resolution vision in Opus 4.7, vision capabilities have become integral for tasks like reading image-based PDFs, screenshots, and design files. While earlier iterations of vision technology were rarely used due to low accuracy, they have now become indispensable. The absence of vision in GLM 5.2 represents a significant practical limitation.

Inference API Pricing and the Shrinking Margins

The emergence of competing models like GLM 5.2 is exerting downward pressure on prices in the inference API market. Open-weight models, which can be hosted by users themselves, add competitive forces that push API providers to lower their prices. Anderson warns that this process could trigger an “AI margin collapse.”

If the current high gross profit margins cannot be maintained, the profitability of AI companies will be significantly undermined. This shift would complicate the recovery of massive training investments and influence investment decisions across the entire industry. Notably, this change could also impact the revenue recovery plans of hyperscalers investing heavily in GPUs.

Anderson describes this as “the most underestimated transformation” and promises a more detailed analysis in the second part of his study.

Implications for the Industry

The arrival of GLM 5.2 serves as an indicator of the commoditization of AI models. As open-weight models begin to catch up with frontier models, AI companies will be forced to pivot from relying solely on the superiority of their models to differentiating themselves through services and ecosystems.

While falling API prices are good news for users, they spell revenue declines for providers. The focus will now shift to how AI companies navigate these challenges to maintain profitability. The structural changes in the industry are set to draw significant attention.

Editorial Opinion

In the short term, the release of GLM 5.2 is likely to intensify price competition in the inference API market, potentially reducing the profit margins of companies like OpenAI and Anthropic over the next six to twelve months. The increasing adoption of open-weight models for in-house use may further decrease API demand, presenting additional risks. AI companies will need to reassess their balance between model quality and API pricing.

In the long term, continued improvements in open-weight models may drive the marginal cost of AI inference closer to the cost of computational resources. Over the next one to three years, the revenue structures of AI companies may undergo fundamental changes, with a shift toward generating income through platforms, data, and customization services rather than solely through model sales. The smoothness of this transition will play a crucial role in determining the sustainability of the industry.

From the editorial perspective, the balance between profitability and technical advancement will be a key focus for AI companies moving forward. It’s clear that the current high profit margins are unsustainable, and the entire industry is at a crossroads where it must seek new business models. Close attention should be paid to trends in API pricing and advancements in open-weight models to understand the evolving market landscape.

References

• GLM 5.2 and the coming AI margin collapse — Published on 2026-07-06