More Than Just Code: OpenAI’s Return to Open Weights and the US Quest for AI Dominance

In August 2025, OpenAI released, for the first time in five years, its open-weight large language models to the public. This marked the end of a long silence since the release of GPT-2 in 2019. This event can be seen not only as a technical development but also as a strategic and political move.

On January 21, 2025, one of the largest AI infrastructure projects in US history was launched. OpenAI, together with Oracle and SoftBank, started this massive project, which aims to build a nationwide network of high-capacity data centers with $500 billion investment over four years. Known as the Stargate Project, it is considered not only a technological initiative but also a strategic and geopolitical effort. Its goals are to secure America’s leadership in AI, create hundreds of thousands of jobs, and contribute to the global economy.

Since taking office, Donald Trump has tried to support this vision by easing environmental regulations in the energy sector and facilitating the sale of American AI technologies. In the document titled “Winning the Race: America’s AI Action”, AI is described as a critical national security and economic necessity. The ultimate goal is clear: to establish undisputed and unrivaled global technological dominance.

Trump has also taken concrete steps internationally to support this vision. On May 15, 2025, he completed a tour of the Gulf countries with agreements worth trillions of dollars in investment and technology. Among these agreements, the most notable were:

• A deal to give the UAE easier access to AI semiconductor chips.
• A plan to build a $20 billion AI data center in Abu Dhabi.

These initiatives show that AI is no longer just a technology. It has become a matter of sovereignty, security, and economic power. The Trump administration is pursuing a proactive strategy, building domestic infrastructure and strengthening international partnerships. In the coming years, the AI competition between the US and China is expected to turn into a state-level power race.

In this context, these strategic efforts are not just political statements. They were realized technically with the release of OpenAI’s GPT-OSS models in August 2025. The models gpt-oss-120b and gpt-oss-20b serve as the technical manifestation of this vision: strengthening the US’s AI leadership and gaining an advantage over China in the open model race.

Generated by artificial intelligence.

GPT-OSS Series

Open-weight models mean that the trained AI model’s weights are publicly available. Developers can download these models, run them on their own systems, and fine-tune them according to their own applications. Even though these models are not fully open-source, sharing the weights of such powerful models allows for inspection and customization.

OpenAI offers the models in two versions: gpt-oss-120b and gpt-oss-20b. Both models use the Mixture-of-Experts (MoE) architecture.

OpenAI

The gpt-oss-120b model is technically advanced. It has a total of 117 billion parameters, but only 5.1 billion parameters are active per token. This design allows high performance while keeping computing costs optimized. Developers can set reasoning levels during inference as low, medium, or high, balancing speed and accuracy based on the application.

The model provides transparent chain-of-thought (CoT) reasoning, which helps with debugging, trust analysis, and model insight development. gpt-oss-120b is quantized in MXFP4 format and optimized to run on a single NVIDIA H100 GPU. Released under the Apache 2.0 license, it is suitable for both commercial and research use.

The gpt-oss-20b model is lighter and more accessible. It has 21 billion total parameters, with 3.6 billion active per token. Thanks to its MoE structure, it maintains high accuracy and speed even on smaller devices.

Its main advantage is that it can run with 16 GB of RAM, making local inference on laptops possible. Like the larger model, the reasoning level is adjustable, CoT reasoning is transparent, it is licensed under Apache 2.0, and fine-tuning is allowed.

When compared to other OpenAI models such as o3i, o3-mini, and o4-mini, gpt-oss-120b outperforms o3-mini and performs almost equal to o4-mini in coding (Codeforces), general problem-solving (MMLU and HLE), and tool calls (TauBench). In health-related queries (HealthBench) and math competitions (AIME 2024 and 2025), it even performs better than o4-mini. The gpt-oss-20b model is close to or slightly better than o3-mini in these comparisons.

It has been created with artificial intelligence based on data from OpenAI’s official website.

You can access the model’s GitHub repository **here**.

You can access the Hugging Face gpt-oss-120b repository **here**.

You can access the Hugging Face gpt-oss-20b repository **here**.

References

Guldogan, Diyar. “Trump announces artificial intelligence infrastructure project.” Anadolu Agency. https://www.aa.com.tr/en/americas/trump-announces-artificial-intelligence-infrastructure-project/3458842#.

Hernandez, Michael. “Trump signs 3 executive orders to bolster US standing in global AI race.” Anadolu Agency. https://www.aa.com.tr/en/americas/trump-signs-3-executive-orders-to-bolster-us-standing-in-global-ai-race/3640325.

Okay, Dilara Zengin. “ABD ile BAE, 200 milyar doları aşan yeni ticaret anlaşmaları yaptı.” Anadolu Agency. https://www.aa.com.tr/tr/ekonomi/abd-ile-bae-200-milyar-dolari-asan-yeni-ticaret-anlasmalari-yapti-/3569863.

OpenAI. “Introduction gpt-oss.” https://openai.com/index/introducing-gpt-oss.

OpenAI. “Open models by OpenAI.” https://openai.com/open-models/.