The Political Economy of Open Source AI: Who Benefits and Why

In February 2023, Meta released LLaMA — a set of large language models ranging from 7 billion to 65 billion parameters — to the research community. In July 2023, Llama 2 was released for both research and commercial use. Llama 3 followed in April 2024. Each release generated significant goodwill in the open-source AI community and substantial coverage about Meta’s commitment to open, accessible AI development.

The coverage mostly missed the point.

The Competitive Logic Is Not Complicated

Meta does not have a significant AI API business. OpenAI charges customers per token. Anthropic charges customers per token. Google charges customers per token. Meta’s revenue comes from advertising on Facebook, Instagram, and WhatsApp. The AI race, as it’s currently configured, does not directly threaten Meta’s advertising business — but it does present a specific risk: if AI assistants become the primary interface for information discovery, and if those assistants are built by OpenAI or Google, Meta could lose distribution.

Against this backdrop, consider what Llama does for Meta strategically. Every enterprise that deploys Llama instead of GPT-4 or Claude is an enterprise that isn’t paying OpenAI or Anthropic. Every developer who builds a product on Llama instead of the OpenAI API is a developer who isn’t helping OpenAI’s flywheel. Every researcher who fine-tunes Llama rather than building on Google’s foundation models is a researcher who isn’t improving Google’s systems.

Commoditization is a well-understood competitive weapon. If you don’t control the premium product but you can make the premium product’s differentiating features free, you shift competition to a domain where you have advantages. Meta’s advantages are data (the social graph, user interaction data from billions of accounts) and compute (world-class infrastructure built for ad serving at Facebook scale). If model weights become a commodity — if anybody can download a competitive model for free — then the advantage shifts toward whoever has the best data and the most compute. That is Meta.

This is not a new strategy. It’s what Microsoft did with Internet Explorer (distributed free to commoditize Netscape’s browser business), what Google did with Android (distributed free to commoditize Apple’s mobile OS advantage), and what Amazon did with AWS pricing in the early years (priced to commodity to establish volume while incumbents couldn’t respond at cost). Every major platform has done this. The unusual thing about the open-source AI case is that people seem to have forgotten the pattern.

The TensorFlow Parallel

Google open sourced TensorFlow in November 2015. TensorFlow was, at the time, one of Google’s key tools for machine learning research and production deployment. Making it open source looked like remarkable corporate generosity — sharing internal infrastructure with the world.

The strategic rationale was straightforward: Google’s competitive advantage in AI was not the deep learning framework. It was the data (Google Search queries, YouTube watch patterns, Gmail usage), the compute (custom TPUs), and the research talent. TensorFlow was how Google organized its ML work, but it wasn’t the source of the advantage. Open sourcing it cost Google essentially nothing in terms of competitive position while generating substantial benefits: developer goodwill, an enormous community that contributed bug fixes and improvements, de facto standardization on Google’s approach to ML (making it easier to hire people already trained on Google tools), and a credible claim to be driving open development in a field where regulatory attention was increasing.

Facebook followed with PyTorch in 2017. By 2022, PyTorch had largely displaced TensorFlow as the dominant research framework — partly because its dynamic computation graph was easier to work with, partly because the research community strongly preferred it. The competition between TensorFlow and PyTorch was real. What both companies had in common: neither gave away the actual source of their competitive advantage.

Google kept its TPUs proprietary and its training data locked. Meta kept its social graph proprietary and its advertising algorithms locked. What they open sourced was the infrastructure layer where commoditization served their interests — where making the tool free reduced competitors’ ability to charge for equivalent tools.

The Llama releases follow this pattern precisely. Meta is not open sourcing its recommendation algorithm, its ad auction system, or the specific training data derived from billions of social interactions. It’s open sourcing the weights of a model trained on largely public text — meaningful for the broader ecosystem, not the source of Meta’s actual moat.

What Open Source Actually Means When a Trillion-Dollar Company Does It

The term “open source” carries moral freight from its origins. Richard Stallman’s GNU Project, Linus Torvalds’ Linux kernel, the Free Software Foundation — these emerged from a genuine ideological commitment to software freedom: the idea that users should have the right to run, study, modify, and distribute software. This tradition has produced enormous value and deserves its positive reputation.

Meta’s Llama releases are not this tradition. They’re open weights — you get the trained parameters, which lets you run the model and fine-tune it, but not necessarily the training code, training data, or the full technical details of how the model was built. (Some details are published in the accompanying research paper; others are not.) The license is not a traditional open-source license: Llama’s license prohibits use by organizations with more than 700 million monthly active users, which conveniently includes Meta’s direct competitors (Google, Microsoft, potentially Amazon) while including effectively every entity that might create goodwill for Meta by using its model.

This is not criticism. It is observation. Meta is making a calculated business decision about what to release under what terms. The decision is defensible on business grounds. What it isn’t is a principled commitment to open-source values. Calling it “open source AI” is generous to the point of inaccuracy, and the imprecision matters because it shapes how people reason about who benefits and why.

Who Actually Benefits from Open Source AI

Despite the strategic motivations of the companies doing the releasing, open-weights AI models have produced real benefits that are worth acknowledging without attributing them to corporate altruism.

Academic researchers at institutions without GPU budgets can now run experiments on models that would have been inaccessible two years ago. A graduate student at a university in India or Brazil can fine-tune Llama on domain-specific data for a master’s thesis — something that previously would have required either proprietary API access (expensive, data privacy issues) or the computational resources of a large research institution. The democratization here is real, even if the corporate motivation is not idealistic.

Organizations in regulated industries — healthcare, finance, government — that cannot send data to third-party APIs for compliance reasons can now run capable models in their own infrastructure. A hospital that couldn’t use GPT-4 because of HIPAA concerns around patient data can run Llama on its own servers with full control over data flows. This is significant.

Researchers working on AI safety, interpretability, and alignment have access to models they can actually inspect, modify, and probe — rather than being limited to black-box API access. This matters for the field’s ability to understand how these systems work.

None of this required Meta’s motivation to be altruistic for the benefits to be real. The open question — which won’t be answered for several years — is whether the long-term consequence of commoditizing model weights is net positive for the field and for society, or whether it primarily accelerates deployment of capable models in contexts where the harm potential is high and the safety work hasn’t been done.

The Limits of What Open Source Means for AI

Here’s the uncomfortable version of the open-source AI question: what does it mean to “democratize” access to AI if the training data, the compute, and the expertise required to build frontier systems are controlled by a small number of trillion-dollar companies?

You can download Llama 3 70B for free. You can run it on your own hardware, if your hardware is capable (which requires a high-end GPU setup that costs $5,000-$15,000 at minimum to do seriously). You can fine-tune it on your own data. You cannot train a model of comparable quality yourself, because training a model of Llama 3 70B’s quality required an estimated $50-100 million in compute, a dataset of roughly 15 trillion tokens curated by a team of data scientists, and a research team of dozens of senior ML engineers. Those resources are not democratized by open-weighting the result.

The analogy is instructive: imagine that Airbus published the full engineering specifications for the A380 and called it “open aviation.” Anyone could study the blueprints. What they couldn’t do is build an A380, because doing so requires a factory, materials, a supply chain, and decades of aerospace engineering expertise. Publishing the blueprints democratizes the knowledge of how the thing was built while doing almost nothing to democratize the ability to build it.

Open-source AI at the current frontier is closer to publishing blueprints than to enabling independent manufacturing. It’s valuable — genuinely valuable. It is not what the discourse around it often implies.

The Line Always Falls in the Same Place

There’s a reliable heuristic for predicting what any large tech company will and won’t open source: the line falls at “this doesn’t give away our actual competitive advantage.”

Google open sourced the Google File System paper in 2003, the MapReduce paper in 2004, the Bigtable paper in 2006. These papers were instrumental in creating Hadoop, which powered the big data industry. Google open sourced the research while keeping the actual operational systems proprietary — and by the time competitors had built Hadoop, Google had moved on to systems that were a generation ahead.

The same logic applies to AI. The models being open sourced are the ones where the advantage isn’t in the weights — it’s in the data flywheel, the inference infrastructure, the continuous fine-tuning from user feedback, and the distribution relationships with enterprise customers. Release the weights; keep the advantages that matter.

This is a coherent and rational strategy for the companies pursuing it. It produces real benefits for the ecosystem. It is not, and was never designed to be, a transfer of power from large technology companies to anyone else.