AI Price War: How Open Models and Cheap Inference Are Erasing Giants' Superprofits

The End of the High-Margin Era

The generative artificial intelligence industry has entered a phase of fierce price competition. While proprietary labs led by OpenAI and Anthropic previously commanded high profit margins due to a lack of viable alternatives, the situation has now changed dramatically. Chinese open-source projects—Qwen, GLM, DeepSeek—have entered the arena, matching flagship models across most practical tasks at a fraction of the cost.

Lowering inference costs has become the primary battleground. In response to open-source pressure, OpenAI resorted to tactical price cuts, reducing rates for GPT 5.6 Sol until November 2026 and expanding its lineup of lightweight models like gpt-5.6-luna. Such moves confirm a global shift: "pure intelligence" is losing its status as an elite premium product.

"We see a future where intelligence becomes a utility, like electricity or water, and people will buy it from us by the meter," OpenAI CEO Sam Altman stated in March 2026.

Architectural Tricks: How Qwen Drives Down Computing Costs

While closed-source labs cut prices at the expense of their margins, open-source developers are optimizing the underlying mathematics of neural networks. A prime example is the release of Qwen 3.8-Flash-Next. The model is built on a Mixture of Experts (MoE) architecture with a total of 125 billion parameters, of which only 6 billion (a6B) are activated per token.

This approach delivers outstanding economic efficiency:

• Compute power and RAM requirements correspond to a compact 6-billion-parameter LLM.
• Context window breadth, reasoning depth, and knowledge capacity remain at the level of a large 125-billion-parameter system.
• The primary hurdle of open source—the high cost of deploying heavy model weights on commodity servers—is eliminated.

Open Weights as a Strategy for Infrastructure Dominance

Another lever applied against the commercial sector is the distribution model. The recent stealth release by Z.ai—the Ox Alpha model, which turned out to be GLM-5.3-Flash—showcased impressive specifications: a 1-million-token context window, generation of up to 131,000 tokens per response, multimodality, and high performance in code generation.

The company made testing free and subsequently announced the open-sourcing of its weights. With API revenue rapidly shrinking, making flagship models publicly accessible allows companies to set market standards, build a software ecosystem around them, and undercut competitors who rely on paid access to proprietary solutions.

The Battle for Hardware, Energy, and Software

The race for the cheapest token extends far beyond algorithms and impacts the entire supply chain:

• Custom Silicon: The development of OpenAI's Jalapeño chip aims to reduce reliance on expensive Nvidia Blackwell GPUs, cutting infrastructure expenses at the silicon level. A similar focus on specialized NPUs can be seen in Apple's M5/M6 Ultra processors.
• Mergers and Acquisitions: Nvidia's $13 billion acquisition of Hugging Face and AWS's purchase of DuckDB are designed to secure control over the software stack and data retrieval speeds in RAG systems.
• The Energy Factor: Serving millions of requests runs up against physical power grid and cooling constraints, leading US authorities to protect large-scale data center construction projects despite local community pushback.

The Outcome of the Transformation

The AI market is transforming: model developers' margins are melting away, and profits are shifting toward infrastructure providers, custom hardware manufacturers, and data center operators. The winner in this new era will not be the lab that builds the smartest model, but the one that constructs the most cost-effective, vertically integrated pipeline to deliver tokens to the end customer.