What tech leaders should take from Mozilla’s inaugural open source AI report

The economics of running AI broke at Microsoft, Uber and Stripe this year, and each company responded differently.

Microsoft cancelled most of its Claude Code licenses at the end of June because the pay-per-use billing had eaten its division’s annual AI budget in a matter of months, according to a new report from Mozilla.

Uber blew through its 2026 AI coding budget in four months, the report says, with individual engineers running up bills between $500 and $2,000 apiece before the company put a cap on it.

Stripe cut its AI running costs by 73%, the report says, by moving 50 million daily AI requests off closed vendor APIs and onto open source models it runs on its own hardware. The payments company now handles that volume on a third of the computing hardware it needed before, and its costs are fixed and forecastable rather than climbing with usage.

Those three examples sit in the middle of Mozilla’s first State of Open Source AI report, published this week and built on a global survey of more than 950 developers. The report’s argument is that open source AI models have nearly matched the closed ones from OpenAI, Anthropic and Google on capability, the cost of running them has fallen roughly 50 times in three years, and the ground where the industry now competes has moved off the models themselves and onto the software that wraps them.

An open source AI model is one the vendor publishes in full, so anyone can download it and run it on their own computers. That includes the model’s weights, which are the billions of numbers the model learned during training and which do the actual work of generating an answer. A closed model, by contrast, stays on the vendor’s servers, and the customer pays every time they use it.

“Open source AI has reached a turning point,” said Raffi Krikorian, chief technology officer, Mozilla in a blog announcing the report. “It’s no longer about expanding access to models. It’s about who has the power to shape, audit, and improve them. Without investment in the infrastructure, tooling, and governance around open models, we risk locking in a system where only restrictive, closed AI can scale.”

Mozilla Foundation president Mark Surman put it more plainly in a post on LinkedIn. “The ‘is open good enough’ question is settled. What isn’t settled is who gets to shape what comes next,” he wrote. “This is the difference between owning your AI and renting it.”

Three findings from the report are important for Canadian technology leaders.

1. Open models have nearly closed the capability gap with closed ones, and the price of running them has collapsed. 
2. Open source AI is now a real commercial layer with paying customers and billions in revenue behind it, and the pay-per-use pricing model that closed vendors rely on is starting to break at enterprise scale.
3. Competition has moved one layer up from the model, into what the report calls the harness, and a new form of lock-in is forming there as the labs pull that layer in-house and tune it to their own models.

Ottawa is already moving on the strategic implications, with roughly $890 million behind a sovereign public AI supercomputer.

Open models have nearly closed the gap, and the cost of running them has collapsed 

The Chatbot Arena, a public leaderboard that scores AI models based on how users rate them in blind side-by-side tests, had the leading closed models 3.3 points ahead of the leading open weight models as of March, according to the report. At the start of 2024 that difference was more than eight points, and by early 2025 the two were roughly tied.

The picture has shifted again with Anthropic’s Claude Fable 5 release in June, which pushes the closed labs back out in front on the hardest reasoning and coding problems, though the broader trend the report is drawing attention to is the collapse of that lead over two years.

Inference costs, which is what an organization pays every time a model generates an answer, fell roughly 50 times over the last three years, according to the report. The price of running a model at the level of GPT-4 went from about $20 per million tokens to around 40 cents. A token is a fragment of text, and one million of them runs to several novels’ worth of words.

That drop is more than three times faster than the historical cost curve for personal computing over the same window, and more than two and a half times faster than the drop in bandwidth prices during the dotcom era, the report argues.

Mozilla says developers are moving to open models as the price drops. 

On OpenRouter, which lets developers switch between different AI models through a single connection, open weight models now handle roughly one-third of all the traffic, up from around 2% in late 2024. The platform processes 25 trillion tokens a week, five times what it did six months ago, and the single largest source of that traffic is an open model. 

On Hugging Face, the main library where developers publish and download AI models, there are now 2.5 million models available and 13 million users pulling from them, including one-third of the Fortune 500.

Open models are as good as closed ones at writing code, following instructions, and answering general knowledge questions, according to the report. Closed models still lead on reasoning through hard problems, on pulling information out of long documents, and on the kind of multi-step work where an AI agent has to call outside tools and take actions on the user’s behalf.

For a technology leader, the question has become which jobs are a fit for open models and which still need closed ones.

There’s also a difference between what developers want and what they can put into production. Of the developers surveyed, 79% said they use open models, but only 51% have put them into production, compared to 63% for closed models.

“This gap indicates that there is not an issue purely of model quality, but of missing infrastructure,” said Álvaro Ruiz Cubero of SlashData, the company that fielded the survey for Mozilla. “Deployment rates for open models barely increase with company size, highlighting a lack of mature tooling and support. At the same time, buyers are prioritising licensing terms and ownership, signalling a clear shift toward control and flexibility over raw capability.”

Open source AI also now has real revenue behind it. 

Mistral, the Paris-based AI startup, is at $400 million in annual recurring revenue, up twentyfold in 12 months, and is reportedly in talks to raise €3 billion at a €20 billion valuation. DeepSeek raised $7.4 billion in its first external funding round in June at a valuation over $50 billion. Databricks, the American data platform that acquired MosaicML in 2023 to move into open model training, is running at a $5.4 billion revenue rate, the Mozilla report says.

Six of the largest U.S. technology companies, Microsoft, Amazon, Nvidia, Google, IBM and Meta, have taken positions across the open stack, either by investing in labs like Mistral and Hugging Face, or by shipping their own open weight models. Nvidia alone has stakes in Mistral, Cohere, Together, Fireworks, Baseten, Hugging Face and Replicate.

The stories from Microsoft, Uber and Stripe are not one-offs. The Mozilla report describes a pattern: pay-per-use pricing works until the tool becomes so useful that people stop rationing how much they use it, and by then the vendor’s bill is what limits what the customer can build next.

According to the Mozilla report, Microsoft was exploring Azure-hosted DeepSeek V4 for its heaviest Copilot workload by June, cutting out the pay-per-use bill it would otherwise owe to Anthropic. And the companies that hit the ceiling first are giving themselves an out by setting up an open-weight model behind the same software they already use, so they can switch to it if the vendor’s price climbs.

The report also flags a risk beyond price. In June, Anthropic’s newly released Claude Fable 5 model went dark globally for 19 days after the U.S. government issued an export control order, a sovereignty story Digital Journal covered at the time. Every business that had built its work on top of that model lost access with no notice.

The next AI moat is being built above the model

Mozilla’s third finding is about the software that sits above the model, which it calls “the harness.” It’s a term that many leaders may not have encountered yet, but one they will need to understand. If the model is the engine, the harness is everything that turns that engine into a working vehicle. It’s the software that sits between a person and the model and does the practical work: giving the AI a memory, connecting it to outside tools like email and databases, running it safely, and deciding what it’s allowed to do on someone’s behalf. This is the hardest part to get working inside a business, and it’s where the next round of vendor lock-in is forming.

The same model can perform very differently depending on the harness wrapped around it. On a coding test called Terminal-Bench in May, an independent harness scored far higher using Anthropic’s model than Anthropic’s own tool did with the identical model underneath. The wrapper was doing more than the engine, according to the report. Within two months, Anthropic and its rivals had rebuilt their own wrappers to close that difference, and now each company’s harness works best with its own model and worse with a competitor’s. The report calls it a moat in formation, one being built one release at a time.

For tech leaders buying AI, it means the tool and the model are becoming a single package that is hard to leave. A company that wants to switch to a cheaper or better model from a rival can’t just swap the model out. Because the harness is tuned to work with one company’s model, moving to a new model also means replacing the tool the team has built its daily work around, and doing both at once, so the cost of leaving climbs the longer a company stays.

The report also argues that the technology is being adopted faster than the safeguards around it. 

The Model Context Protocol, an open standard for connecting AI agents to outside tools like email, calendars and databases, went from about 2 million monthly downloads at its late-2024 launch to roughly 97 million by early 2026, and 28% of the Fortune 500 now runs it in production. 

The safeguards around it have not kept pace, and security researchers filed more than 30 vulnerabilities against the protocol in the first eight weeks of this year. Only about 21% of companies say they have proper oversight of the AI agents they are running, and people approve those agents’ requests by default as often as 93% of the time, a habit the report calls consent fatigue.

The report names Canada among the governments treating open-weight models as national strategy. One of the six pillars of Ottawa’s AI for All plan is what the government calls a Sovereign Foundation, aimed at keeping Canadian data and intellectual property in the country and cutting its dependence on foreign technology suppliers.

Canadian companies are building for this.

In May, Toronto-based Cohere released Command A+, a powerful model whose weights it published openly so any organization can run it on its own systems, even fully disconnected from the internet. Cohere built it for governments and regulated industries that need to keep sensitive data in the country, and followed it in June with a smaller model aimed at developers. The company calls the goal a sovereign developer ecosystem, where a Canadian business owns its AI tools outright instead of renting them from American vendors.

“AI can go the way the internet and mobile phones did, a mostly disempowering tech. Or it can empower the people that use it. We are working towards that second one,” Cohere co-founder Nick Frosst said in the Mozilla report.

Mozilla’s recommendation to enterprise leaders is to be owners, not renters. That means holding operational data behind the company’s own firewall, keeping a second model ready to run against open interfaces, and paying attention to how permission and identity get defined in the agent layer.

The report uses Uber to make the point. Uber kept fares cheap for years while riders gave up their cars and built their daily lives around the low price. Once that dependence set in, fares rose roughly 92%, and riders had no easy way back. Closed AI vendors, the report warns, are running the same play. The low price today builds the dependence and the bill comes once leaving is hard.