If you have been following AI news for the past several months, you could be forgiven for thinking the entire story is GPT-4, ChatGPT Enterprise, and whatever Sam Altman said most recently. The coverage is loud, the valuations are eye-watering, and the product launches come with the energy of a stadium concert.
Meanwhile, something quietly significant has been building in the background — in open repositories, on Hugging Face, in university research groups, and inside some engineering teams at companies you may not have heard of yet. And in mid-July 2023, it got considerably less quiet.
Meta Llama 2: The Moment the Open-Source Conversation Changed
On July 18, 2023 — just over few weeks ago as this is written — Meta released Llama 2. This matters for a reason that goes beyond the model’s capabilities: for the first time, Meta made a large language model available under a commercial licence, free to use for organisations with fewer than 700 million monthly active users.
To put that in context: Llama 1, released in February 2023, was research-only. Weights leaked almost immediately, which created an awkward, semi-underground ecosystem of developers fine-tuning a model they technically weren’t supposed to have. Llama 2 resolved that tension formally. Microsoft partnered on distribution. Hugging Face hosted it. The enterprise world sat up.
I keep thinking about what that shift actually means. It’s not just a licensing change. It’s a signal that a major technology company — not a scrappy startup, not an academic lab — decided that open availability was a strategic choice, not a concession. That decision has consequences for everyone else in the ecosystem.
The Ecosystem That Was Already Forming
Llama 2 didn’t arrive in isolation. By the time it launched, a genuine open-source AI ecosystem had been assembling piece by piece.
Falcon 40B, released in May 2023 by the Technology Innovation Institute in Abu Dhabi, briefly topped the Hugging Face Open LLM Leaderboard and was released for commercial use without restriction — a genuinely remarkable move from an organisation outside the traditional Silicon Valley orbit. Databricks released Dolly 2.0 in April 2023 — the first open-source, commercially usable instruction-following language model built on a fully open dataset. MosaicML’s MPT-7B followed in May with training transparency that proprietary providers simply don’t offer.
What Hugging Face has quietly become in all of this is worth pausing on. It’s part model hub, part GitHub for AI, part community — a single destination where a researcher in Bangalore, a startup in Lagos, and an enterprise team in Frankfurt can all access, compare, and fine-tune the same models. The infrastructure for open AI development has arrived faster than most people expected, and it looks considerably less chaotic than early open-source software ecosystems did.
The Honest Tradeoffs — Because There Are Some
This is the part where I think the open-source AI conversation sometimes oversimplifies, and it’s worth being direct about the tradeoffs.
Proprietary models — GPT-4 particularly — are still meaningfully more capable on complex reasoning, instruction following, and nuanced generation tasks. The gap is real, not a marketing fiction. They come with support, safety filtering, reliability guarantees, and continuous improvement that an open model does not automatically provide. For many enterprise applications, those things matter enormously.
Open-source models, on the other hand, offer something proprietary models cannot: control. You can run them on your own infrastructure, which means your data never leaves your environment — a consideration that is not hypothetical for regulated industries. You can fine-tune them on your specific domain, which in many narrow tasks produces better results than a general-purpose proprietary model. You have no vendor dependency, no API pricing risk, no terms-of-service change that suddenly affects your product roadmap.
The interesting practical observation is that organisations aren’t really choosing between open and proprietary. They are — often somewhat messily — combining both: using open-source models for high-volume, lower-complexity tasks and proprietary models where capability justifies cost. The engineering teams figuring out that orchestration layer are doing genuinely novel work right now.
The Competitive Pressure Is the Real Story
Here is the dynamic I find most fascinating, and it doesn’t get talked about enough.
Open-source AI doesn’t just help organisations that use it. It changes the behaviour of the organisations it competes with. When a capable open model exists that can be deployed for free, it puts a floor on what proprietary model providers can charge. It forces investment in differentiation rather than moat-building. It accelerates the entire ecosystem — because when the baseline is free and accessible, innovation has to happen at a higher layer.
This is roughly what happened with Linux and enterprise software. Linux didn’t replace proprietary operating systems in every context. But its existence materially changed how those systems were priced, positioned, and developed. The AI equivalent of that dynamic appears to be underway. Whether it plays out at the same pace or faster — the infrastructure is already different from anything the open-source software movement had in its early years.
The Lens Worth Applying
The question worth sitting with is not “open source vs. proprietary” — that framing produces a false binary that doesn’t reflect how most organisations will actually operate. The more useful question is: for each AI use case we’re building, what does the make-or-buy decision look like when “make” suddenly became dramatically cheaper?
That reframe changes the architecture conversation, the vendor negotiation, the build-versus-integrate calculus, and the data strategy simultaneously. The organisations thinking through it carefully now — rather than defaulting to a single provider because it’s the path of least resistance — are the ones who will have real optionality as the ecosystem matures further.
Whether the models within that infrastructure are open or proprietary is, in some ways, the less consequential decision. How they’re governed, how their outputs are managed, and how human oversight is maintained — those questions hold across the entire ecosystem.
I’m genuinely curious about what’s happening in practice: are teams you know making deliberate open-versus-proprietary decisions, or is it mostly defaulting to whichever API the first engineer tried? The gap between the strategic conversation and the actual implementation path seems wider than most organisations want to admit.
Let’s keep learning — together.
harish.balasubramanian 
Share your thoughts