For decades, Mozilla has been one of the strongest advocates for an open internet. Today, as artificial intelligence becomes increasingly central to how businesses and governments operate, the organization is making a similar case for open AI.
In a report published last month, Mozilla, a free software community and nonprofit project, said the performance gap between top open-source models and proprietary systems like Claude and ChatGPT has narrowed to just 3%. The report highlighted the growing popularity of open models, noting that in February 2026, Alibaba’s open-source model Qwen was downloaded more times than the next eight models combined.
The debate over open AI has gained fresh momentum after Meta released a new open-weight model on August 10, and promised more would come soon.
Raffi Krikorian, Mozilla’s chief technology officer, believes organizations need alternatives to proprietary AI systems to gain greater control over their data and infrastructure.
Throughout his career — from leading engineering teams at X (then Twitter) to working on autonomous vehicles at Uber and now shaping Mozilla’s vision — Krikorian has worked on breakthrough technologies and what they mean for consumers. In a conversation with Rest of World, he discussed why policymakers need to view AI as infrastructure rather than as a product, why businesses are increasingly experimenting with open and open-weight models, and what role he believes the technology will play in the next generation of products and services.
The conversation has been edited for length and clarity. A recording of the live Rest of World event is available here.
You’ve said that open-source AI is no longer a “hobbyist ecosystem” and that it has a multi-hundred-billion-dollar commercial layer. What have you based this finding on?
When you talk about open source in certain circles, usually governmental policy circles, they would say things like, “Oh, it’s just some kid in the basement, right?” This ecosystem is not that. These businesses are building huge economies, infrastructure, products, and technology.
It’s like Linux, which started with just a kernel, not being taken very seriously. But Google was able to take Linux and make Android, and now, for all practical purposes, every computer on the planet runs on Linux. Every single new vending machine probably is running on Linux.
We are seeing the same thing in the open-source, open-weight AI world. It is because of the marketing blitz of the big frontier labs that we all view AI just as something that Anthropic, OpenAI, and a few other companies are doing. There’s actually a burgeoning ecosystem on the open-source, open-weight side that a lot of people actually do know about.
It’s potentially one of the best well-kept secrets right now when it comes to technology.
Why are more people not familiar with open-source tools?
There’s a difference between what consumers are doing and what businesses are starting to do.
It is true that an individual — like a secretary or another white-collar worker — is probably using ChatGPT and Claude. But the workflows that an IT or HR team is starting to put together are migrating toward open models because of price-performance reasons or potentially because of regulated industries, which want to have full control of their data and how their data flows. They want to self-host and have things stay within their firewall. They’re looking for an ability to deploy things under their control.
I don’t have a good answer as to why people aren’t shouting from their rooftops that they’re saving all this money.
Are people not talking about open-source models openly because many of them are Chinese and there are regulatory or trust issues with these tools?
We don’t have concrete proof that it’s true, but I think you can probably infer that to be true.
Open-source AI models are complicated to understand, and I think that is also part of the issue. When you look at these models, it’s like downloading a model you want to run, like a file, onto your computer. So, I can understand how people might — if they don’t truly understand how this technology works — be worried about just downloading malware that could be giving their data over in some places.
That’s not how the systems actually work. But the simplest explanation is the one that usually ends up winning.
And I also do believe that something about just China in general is confounding the equation a little bit. The confounding thing should be why aren’t more countries and organizations trying to fill that demand? If a large amount of traffic is going to a model that people want to be open, then you can imagine a world where American companies or Indian companies or European companies should also be trying to fill that demand.
In the context of the concerns with Chinese software, do you think governments — especially Washington — understand that open source is getting so popular?
I do think that the government understands what is happening. But they’re caught between “do we want to let our current existing standards bearers win” or “do we want to potentially take the risk and try to be a transformative thing to the ecosystem.”
I think it’s a simple thing to comprehend that we have two companies that are doing incredibly well, that are racing toward public listings either end of this year or the beginning of next year, and we want them to win. And the government is being lobbied very heavily in that regard.
So the simple thing to do would be to back the people they perceive to be the winners already.
But I would say that Americans have generally filed intelligence as a product. Products are by definition things that you can rent or turn off, but it should be filed as infrastructure, which is what the rest of the world seems to want to buy. That categorization error is actually the problem. We should be trying to fill that demand in some way, and I’m frustrated that my government can’t see that distinction right now, and they’re doing the simple thing.
Is open source really open for everyone?
I will admit that we are glossing over the open-source, open-weights conversation. There is the strictest definition of open source, which would come down to understanding all the data that went into training it on, all the evaluation systems, and what pre-training and post-training look like. I think that’s incredibly important, and we need to invest more in making that happen. That is what will lead to trust and transparency, and we can ethically feel really good about them.
Right now, a lot of what we’ve been talking about is open-weight systems. So we don’t have a good sense of all the data — pre-training and post-training. But what we do have is the ability to download, run, execute, and then modify and fine-tune. That fine-tune is where the rest of the world comes into play. The pathway to have really local models and local AI that reflect the value of the users is to take these big open-weight models and fine-tune them to our use cases in terms of the language we speak or the values of a community. The technology there still needs work. So, if you’re a computer scientist, you could probably do it, but if you’re a community leader, you probably can’t.
I’m not faulting what the big companies are doing, but they are going to optimize for the places where they can make the most money. So they’re going to optimize for the Western world and maybe parts of Asia, but they’re not going to optimize for the rest of the world. The rest of the world only becomes part of this AI transformative conversation through the open-source and open-weight ecosystem.
Mozilla has championed the open web for decades. Yet, Google has 90% of search, Meta has captured social media, and Amazon runs most of the cloud. Why should anyone believe that open-source AI would end differently?
The existence of things like Firefox has moved the entire industry. Someone like Google can’t build a closed version of the web because the vocal community of Firefox users will rise up and say you just can’t do that. Those types of numbers are enough to force and keep the internet in an open-protocol state and one that doesn’t benefit any particular company at that base infrastructure level.
I want the exact same dynamic to occur on the AI internet. You could make an argument that we get a small amount of traffic on the AI internet through the open models, but that would be enough to force a conversation around model choice, agentic harness choice, user agency, and making sure that memory stays on my side, not the user side. You can make the argument that it only takes a small amount of traffic to pull that off, but I’m also saying a large amount of traffic is already moving to the open model.
What do you think the biggest AI company of 2031 would be doing? Building models, memory, developer tools, agents, infrastructure, or something entirely different?
I think there will be a diffusion. We are going to find that the biggest provider is the one that provides an open version of intelligence that every single company can use and customize — similar to how we have a Google Drive subscription or a Microsoft Office 365 solution right now. It will be a services provider that can walk into a company and help them actually get over that hump and use AI.
Facts Only
* Mozilla has been an advocate for an open internet for decades.
* A report indicated the performance gap between top open-source models and proprietary systems like Claude and ChatGPT narrowed to 3%.
* In February 2026, Alibaba’s open-source model Qwen was downloaded more times than the next eight models combined.
* Meta released a new open-weight model on August 10, and promised more would follow.
* Raffi Krikorian believes organizations need alternatives to proprietary AI systems for greater control over data and infrastructure.
* The author suggests that governments should view AI as infrastructure rather than a product.
* The pathway to local models involves taking open-weight models and fine-tuning them to specific use cases and community values.
* The existence of technologies like Firefox moved the internet by creating an open protocol, which can be analogized to the potential for an open AI internet.
* The largest provider in 2031 is predicted to be a service provider offering an open version of intelligence that companies can customize.
Executive Summary
Mozilla advocates for an open internet and argues for open AI, citing a report indicating that the performance gap between open-source models and proprietary systems like Claude and ChatGPT has narrowed to 3%. The document points to the growing popularity of open models, exemplified by Alibaba’s Qwen model outselling other models in February 2026. This momentum is fueled by a burgeoning ecosystem on the open-source, open-weight side, despite public perception that it remains a niche activity.
The shift toward open models is driven by business needs; IT and HR teams are migrating to open models for reasons such as price-performance or the need for greater data control, favoring self-hosting within firewalls. The discussion touches on the difficulty in understanding these systems, suggesting that public awareness is limited, partly due to complexity, and potentially complicated by geopolitical factors concerning Chinese entities.
The piece suggests a fundamental shift from viewing AI as a product to viewing it as infrastructure, analogous to the open nature of the internet's foundational protocols like Linux. The author posits that this paradigm shift requires systemic change, advocating for greater transparency regarding training data and evaluation systems to build trust. Ultimately, true localization and value for users depend on fine-tuning these large models with community-specific values.
Full Take
The narrative presents a tension between the operational reality of proprietary AI systems and the potential structure offered by open-weight ecosystems. The core conflict lies in the perceived gap between what consumers observe (using ChatGPT) and the underlying technological and commercial structures (the multi-hundred-billion-dollar open ecosystem). This gap is maintained through marketing that frames AI as a proprietary product rather than shared infrastructure.
The pattern observed is one of institutional inertia resisting infrastructural change. The reluctance to adopt open models widely appears rooted less in technical feasibility—given the demonstrated capabilities of fine-tuning and execution—and more in established power structures, particularly concerning data ownership and regulatory capture. The author points to the historical precedent of Linux and the internet as models for how open systems can achieve dominance through foundational infrastructure adoption.
The implication is that the current trajectory risks concentrating transformative AI capabilities within entities that control the highest levels of access and fine-tuning. If demand continues to flow toward open models, the system may eventually be forced to accommodate an infrastructure-first model. The failure to recognize this potential due to concerns over trust and complexity—especially regarding data provenance—acts as a brake on broader adoption, potentially allowing incumbents to maintain control by framing any transition solely around feature sets rather than foundational governance.
Bridge Questions: If the focus shifts entirely from model performance (the 3% gap) to verifiable transparency in training data and evaluation methods, what specific regulatory mechanisms would need to be established globally to enforce this? How can community leaders effectively bridge the technical complexity of fine-tuning with the systemic demands for true open infrastructure ownership? What are the concrete political or economic levers required to shift government categorization from viewing AI as a product to viewing it as essential public utility infrastructure?
Sentinel — Human
The text appears to be a human-authored reflection built around expert opinion, successfully synthesizing technical trends with philosophical and geopolitical implications regarding open-source AI infrastructure.