I have been arguing for some time that open AI matters. Not because every model should be treated as risk-free, and not because openness magically solves questions of bias, safety, energy use or power. It matters because AI is becoming part of the basic infrastructure through which organisations work, learn and communicate. If that infrastructure is controlled by a small number of companies, accessed only as a rented service and governed through terms that can change without notice, then users have little real agency.

That is why a recent interview with Raffi Krikorian, Mozilla’s chief technology officer, caught my attention. His central argument is straightforward: policymakers should stop thinking of AI solely as a product and start thinking of it as infrastructure. In his view, the important question is not just which chatbot consumers use, but whether organisations can control their data, keep sensitive processes inside their own firewall, adapt systems to local needs and change supplier when necessary (Punit, I. S. 2026).

Mozilla’s reported figures help explain why this is becoming a serious argument rather than a hobbyist preference. The organisation says the performance gap between the leading open models and proprietary systems such as Claude and ChatGPT has narrowed to around 3%. The interview also notes that Alibaba’s Qwen was downloaded more times in February 2026 than the next eight models combined (Punit, I. S. 2026). Even allowing for the usual caveats about benchmarks and download statistics, the direction of travel is clear: open models are becoming credible infrastructure for real organisations.

However, we need to be precise with the language. Open source and open weight are not the same thing. The distinction is frequently blurred, sometimes carelessly and sometimes for marketing advantage.

Term	What is normally available	What is still missing in many releases	Why the distinction matters
Open weight	The trained model parameters, allowing a user to download, run and often fine-tune the model.	The full training data, data-processing pipeline, training code and evaluation process.	It enables local deployment and adaptation but usually does not permit a full independent reconstruction or audit of the model.
Open source AI	The ability to use, study, modify and share the system, alongside the information and materials needed to make meaningful modifications.	Nothing essential to understanding and recreating the process should be withheld.	It offers a stronger basis for transparency, reproducibility, collaborative improvement and independent scrutiny.

The Open Source Initiative’s definition sets a demanding standard. It says that the preferred form for modifying an AI system must include sufficiently detailed training-data information, complete code for training and running the system, and the model parameters themselves (Open Source Initiative, 2024).  Most of the high-profile models described as “open” today are more accurately called open weight. Their weights can be downloaded and adapted, but we do not know the full provenance of their training data, the details of filtering, or the complete pre-training and post-training process.

That does not make open-weight models unimportant. On the contrary, they can be a major practical gain. Krikorian argues that the ability to run and fine-tune models locally is what makes it possible for organisations and communities to develop AI around their own languages, values and use cases. This is particularly relevant outside the markets that proprietary suppliers are most likely to prioritise. But we should not confuse the opportunity to customise a model with complete openness and transparency (Punit, I. S. 2026).

The geopolitics here are increasingly striking. The dominant story of AI in the United States is still one of frontier companies, gigantic data centres and the race to produce the next most powerful proprietary model. Krikorian’s criticism is that Washington is too easily drawn towards backing the existing perceived winners. He contrasts this with the possibility of building an open layer of intelligence that companies and public bodies could control and customise. This is his political interpretation, but it identifies a real strategic tension 1.

It would be too simple, however, to say that the United States is opposed to open AI. The current US AI Action Plan explicitly includes a commitment to encourage open-source and open-weight models (The White House, 2025). Yet the wider plan is framed around American global AI dominance: exporting US systems, hardware and standards; countering Chinese influence in international governance; and strengthening controls on advanced compute and semiconductor exports 3. Open AI is therefore present in US policy, but within a broader strategy of national technological advantage and frontier leadership.

China has taken a rather different route in practice. Stanford HAI’s analysis describes a diverse Chinese ecosystem of open-weight models, built by companies including DeepSeek, Alibaba and Baidu. It concludes that Chinese open-weight models have caught up with and in some areas may have pulled ahead of global counterparts in capability and adoption. The brief highlights an emphasis on computationally efficient models designed for flexible downstream deployment, and notes that Chinese state support has played a substantial, though not exclusive, role in the ecosystem’s development (Stanford Institute for Human-Centered Artificial Intelligence, 2025).

This is not merely a debate about national pride. It is also about cost and the practical ability to deploy AI. On their current published API price lists, DeepSeek-V4-Pro’s standard rates are $0.44 per million non-cached input tokens and $1.32 per million output tokens. OpenAI’s GPT-5.6 Terra is listed at $2.00 and $12.00 respectively; even OpenAI’s lower-cost GPT-5.6 Luna is listed at $0.20 input and $1.20 output (Open AI, n.d.) (DeepSeek, n.d.).

These figures are not a measure of equivalent quality, and they will change. They also do not include the technical and computing costs of self-hosting an open-weight model. But they do show why “price-performance” has become part of the conversation. A small organisation that needs substantial volume may find that the economics, as well as the control of data, point it towards open-weight alternatives.

For education, public services and smaller organisations, the key issue is not to choose between an American closed model and a Chinese open-weight one as though that were the only choice. The important question is whether we can create and sustain an ecosystem that gives institutions genuine options: interoperable systems, public-interest infrastructure, locally adaptable models, transparent evaluation and the capacity to understand what is being deployed.

Mozilla’s analogy with the internet is useful here. The internet was not made open by the goodwill of the largest companies. It remained open because protocols, communities and alternative implementations created limits on any one company’s power. AI will need something similar. The growing open-weight ecosystem is not the whole answer, but it is an important part of the answer - and it deserves to be understood on its own terms, rather than dismissed as a lesser version of proprietary AI.

References

Mozilla’s argument, and the political and adoption data reported from its work, are discussed in the interview cited below. Definitions of openness, pricing and policy positions are drawn from the corresponding primary and independent sources.

Open Source Initiative. (2024). The Open Source AI Definition—1.0. https://opensource.org/ai/open-source-ai-definition

Stanford Institute for Human-Centered Artificial Intelligence. (2025, December 16). Beyond DeepSeek: China’s diverse open-weight AI ecosystem and its policy implications. https://hai.stanford.edu/policy/beyond-deepseek-chinas-diverse-open-weight-ai-ecosystem-and-its-policy-implications

DeepSeek. (2026). Models & pricing. https://api-docs.deepseek.com/quick_start/pricing/

OpenAI. (2026). Business pricing. https://openai.com/business/pricing/#api

Punit, I. S. (2026, August 13). Mozilla’s CTO thinks AI should be built like the internet. Rest of World. https://restofworld.org/2026/open-source-ai-infrastructure-mozilla/

The White House. (2025). America’s AI Action Plan. AI.gov. https://www.ai.gov/action-plan

About the Image

This print speaks to the ways OpenAI (and others) are forcing genAI into our lives, families and homes. It frames genAI as a power(ful) tool to control and influence behaviour. The piece is part of a larger art series titled “Power Tools: A critique of genAI and its toolmen".