AI Competition Restructured: Pressure on Open Weight Costs

The competitive landscape of the artificial intelligence (AI) market is divided into three axes: closed frontier models, open weight models, and application companies. Analysis suggests that the focus of competition has shifted from who can create the highest-performing model first to how corporate clients utilize AI in terms of costs and control conditions.

Saurabh Gupta, co-founder and managing partner of DST Global, likened the 2026 AI economy to a "three-body problem" in a column for Fortune. He observed that closed frontier labs like OpenAI and Anthropic, the China-centric open weight model, and application companies creating products on top of these two model groups are altering each other's trajectories.

Closed frontier models are characterized by developers controlling the models and services, with users accessing them via APIs or product forms. Until now, the prevailing structure has been that the company with the strongest model holds market dominance, but as AI spending increases, corporate clients have begun to consider return on investment and data control issues together.

The Fortune column highlighted recent changes such as increased demand for Anthropic, pressure on the return on investment for AI spending, and the launches of Meta's Muse Spark 1.1 and xAI's Grok 4.5. This suggests that while performance competition continues among closed models, cost and deployment flexibility have emerged as separate competitive axes.

The presence of the open weight camp has also grown. Moonshot presented 2.8 trillion parameters and a 1 million token context window during the official announcement of Kimi K3, while Z.ai revealed that GLM-5.2 has 753 billion parameters and public weights, along with an MIT license.

Open weight differs from open source, which means full source code disclosure. Even if access to model weights is permitted, not all learning data, code, and redistribution rights are freely available, so model-specific licenses and usage conditions must be checked separately.

This trend is also linked to the spread of low-cost open weight models from Chinese companies, as previously reported. In that analysis, Chinese AI companies were described as competing with American companies' closed models by emphasizing model efficiency and the overseas developer ecosystem amid limited semiconductor access.

American companies are also pursuing a public model strategy. Thinking Machines unveiled Inkling on July 15, and NVIDIA ($NVDA) presented its Nemotron 3 product line with three public models: Nano, Super, and Ultra.

NVIDIA stated in its announcement that it provides public models, data, and libraries to enable developers to create AI agents tailored to specific tasks. This is interpreted as a move aimed at targeting the demand for combining purpose-specific models rather than relying solely on closed frontier models for all tasks.

For application companies, costs are a direct variable. Posit AI added Kimi K3 and GLM-5.2 to its product lineup on August 10, explaining that the two models are offered at lower costs than closed models that provide similar perceived performance.

These options indicate a trend where app companies are not tied to a single closed model but are choosing models based on specific tasks. Application companies with customer touchpoints and operational data can compare not only model performance but also costs, latency, and security conditions.

The growth rate of frontier labs is also rapid. Reuters reported that Anthropic's annualized revenue exceeded $65 billion (approximately 90 trillion won) at the end of July, while Business Times reported that OpenAI's annualized revenue surpassed $40 billion (approximately 55 trillion won).

However, these figures are not audited annual revenues but indicators that annualize current revenue levels. While they can be seen as signals of increasing demand for high-performance AI models, they should be distinguished from actual annual performance or profitability metrics.

Policy risks also remain. Reuters reported that the Trump administration conveyed to AI companies that it would not include open weight AI models in voluntary safety testing. This aligns with the previously reported policy trend of voluntary pre-review discussions centered on closed frontier AI at the White House.

The same report distinguished between models with publicly available core components and closed models controlled by specific companies. As open weights proliferate, debates about who will bear the security verification and distribution responsibilities may intensify.

For Korean companies, this issue translates into the costs of AI adoption and data control. When selecting operational AI, it is no longer sufficient to compare only closed models like OpenAI or Anthropic; the structure now requires reviewing Chinese open weight models, American public models, and application services that combine these options.

Based on currently confirmed facts, the balance of the AI market has not been settled on one side. Closed models are increasing revenue and performance, open weight models are emphasizing cost and deployment flexibility, and application companies are responding to the market by combining both axes.

This content is provided for general informational purposes only and doesn't constitute financial, investment, legal, or tax advice. Any events, rewards, online promotions, or related information mentioned herein should not be considered a recommendation, solicitation, or invitation to purchase, sell, trade, or otherwise deal in any crypto assets. Crypto assets are highly volatile and may result in loss. The availability of WEEX services, products, and related events may vary by region. You are responsible for ensuring that your participation is in accordance with applicable local laws and regulations.