//php echo do_shortcode(‘[responsivevoice_button voice=”US English Male” buttontext=”Listen to Post”]’) ?>
As AI systems evolve beyond capital-intensive LLMs controlled by a handful of corporations, small language models (SLMs) are gaining ground with open-weight systems narrowing the performance gap with closed platforms.
In a conversation with EE Times, Mark Surman, president of the Mozilla Foundation, said that the high training and inference costs required to train and run large AI models create incentives to innovate in ways that lower cost and energy use.
“The cost structure of open-source AI is very different from traditional open-source software because of the compute, energy, and infrastructure,” Surman said. “This year, there will likely be more focus on small language models designed for specific use cases. There is also work on improving efficiency and using unused compute cycles through distributed training.”
Highlighting the release of Llama by Meta as a turning point, Surman said that before Llama, it looked as if LLMs would exist without any open source because everything was locked up by OpenAI and Anthropic.
By Bluetooth SIG 03.10.2026
By GigaDevice 03.10.2026
By GigaDevice 03.10.2026
“For most people, AI is a chatbot or a social media feed that guesses what they want to read. AI is all around us, but under the hood, it is made up of different components, like Lego blocks,” Surman said. “When someone uses a chatbot or a social media feed, a company such as OpenAI, Meta, or Alibaba has snapped those blocks together in its own way and closed the box. Open source means having access to all those Lego blocks,”
Breaking open the box
Arguing that meaningful openness in AI must be assessed by examining the different components of a system, Surman pointed to several freely available models that developers can build on. These include Gemma and GPT-OSS from major U.S. players, Qwen and Kimi from China, and models from France’s Mistral AI.
It’s important to distinguish between open-weight AI models and fully open-source AI models. With open-weight models, pre-training happens in a black box. Developers can fine-tune and adapt them, but they can’t see how the models were originally built. Fully open-source models allow developers to see what data went into them and the pre-training checkpoints. They exist today, but they still lag behind open weight and closed models in performance. Surman said more investment is needed to improve fully open-source models.
However, open-source AI isn’t just about the models. It includes open-source data, open ways to orchestrate compute infrastructure, and open ways to design and orchestrate systems built from AI models. “The techniques used in AI, including transformers, large language models, reinforcement learning, and agentic frameworks, are widely known,” Surman said. “Open weight and proprietary models are becoming similar in capability, giving users what they need for most purposes.”
Academic research is beginning to quantify the economic impact of this shift.
The billion-dollar undercount
In November 2025, Frank Nagle, a research scientist at the MIT Initiative on the Digital Economy and chief scientist at the Linux Foundation, along with Daniel Yue, an assistant professor in the Information Technology Management area at Georgia Tech’s Scheller College of Business, published a working paper that puts empirical weight behind what Surman described. Studying the market for LLM inference—the process by which a trained model generates responses to user prompts, and where most day-to-day AI spending takes place—they found that open-weight models are dramatically underused relative to what their price and performance would justify.
Using data from OpenRouter, a platform that routes API requests across dozens of inference providers and captures between 0.3% and 1.06% of the total inference market, Nagle and Yue tracked daily token usage, prices, and model availability from May to September 2025. Their findings revealed a striking paradox at the heart of the AI economy. Closed models from companies such as OpenAI, Anthropic, and Google, accessible only through proprietary APIs, account for around 80% of tokens processed and over 95% of revenue on the platform. Open-weight models, whose weights are publicly released and can be hosted by any inference provider, account for roughly 20% of tokens but only about 4% of revenue.
The price gap is significant. On average, open-weight models cost only 15.66% of the price of closed models, making closed models approximately 6× more expensive. Surman said the difference lies in freedoms. “With open source, you are not paying high rents, you have flexibility, and you can build on top of it,” he said.
The reason open models are cheaper, the paper argued, is structural. Since anyone can host an open model, they attract many competing inference providers, which drives prices toward marginal cost. Closed models, by contrast, are typically served by only one or two providers—the original company and perhaps a cloud partner—which allows them to maintain higher markups.
Yet the performance gap is comparatively modest. Nagle and Yue theorized in their research that open-weight models achieve about 90% of closed-model performance on leading benchmarks, including GPQA, MMLU Pro, LiveCodeBench, and LM Arena, with 89.6% on the GPQA graduate-level reasoning benchmark specifically. The time it takes for a leading open model to match the performance of the best available closed model has fallen from an average of 27 weeks in the first half of 2024, to 17 weeks in the second half of 2024, to just 13 weeks in the first half of 2025, as per the research paper.
Even after controlling price and benchmark performance in regression analysis, open models receive 63% to 88% less usage than comparable closed models. The authors identified numerous cases where closed models are both more expensive and lower performing than available open alternatives, yet users continue to choose them. Simulating a switch from these observably dominated closed models to superior open alternatives, the paper estimated annual savings of between $104 million and $146 million on OpenRouter alone. Extrapolated to the full inference market using three independent methods, the potential unrealized savings range from $20.1 billion to $48.3 billion per year, with a preferred estimate of $24.8 billion.
Trust and switching costs in AI adoption
Importantly, Nagle and Yue don’t argue that users are behaving irrationally. Their findings suggest that switching costs, brand trust, security concerns, and the possibility that standardized benchmarks fail to capture real-world performance differences all help explain the preference for closed models. The economic magnitude of these hidden factors, they argued, is far larger than previously recognized.
Surman cited research commissioned by the U.S. Department of Commerce that found no marginal risk difference between open-weight models and closed models. “Both can be jailbroken,” he said. “Misinformation and deepfakes depend on who is deploying the system and how. At Mozilla, we are involved in ROOST, the Robust Open Online Safety Tools foundation, which builds AI-driven trust and safety software that platforms can use to detect and remove abusive material.”
He explained, “Business models [around open-source AI] include charging for services, charging for support, and pooling resources among users who rely on the same software. Open source has traditionally been strongly known for auditability and transparency. Anyone can see the code. When a problem appears, many people can try to fix it. Proprietary systems are also largely unregulated [when it comes to AI]. The question is who you trust.”
Surman said the technology industry already runs on open source. He cited cloud providers such as Amazon Web Services, Microsoft Azure, and Google Cloud Platform, as well as companies such as Meta and X, as being built on Linux and other open-source software.
Sharing Mozilla’s perspective on the ongoing debate around licensing, Surman said, “Permissive licenses remain the right tool. Open source uses copyright law through licenses that grant rights to modify and share what is created. For responsible AI use, we need guardrail technologies, business processes, and national laws. Changing open-source licenses is not the path,” he opined.
As the global order reorganizes, countries are increasingly exploring approaches that enable broader participation, creating new space for open-source AI. With AI touted as the next era of digital technologies, Surman walked through the computer era, the personal computer era, the web era, and the internet era, and pointed out how open source has been the counterforce to centralization.
“If open-source AI succeeds, it means more people can participate in creating this era of society, the economy, and technology,” Surman said. “Five years from now, I would look for greater diversity in where AI comes from and more of us being creators, not just consumers.”