Who benefits as open models make AI cheaper to run?

Shaon Baqui: Closed models are your proprietary, off-the-shelf models where the code and the weights are largely fixed. So, think your [Chat]GPT, your Gemini, your Claude.

Open models, on the other hand, are highly customizable in the sense that you can download them on your own server, you can modify them, and really fine-tune them to your particular workload. So, think your Llama, your Qwen, and your DeepSeek.

Now, open models historically have been less performant –but cheaper – than closed models. But really, we’ve seen that performance gap narrow quite a bit in recent weeks and months. And in many cases, open models are quite good at running some of your standard workflows like search, summarization, and office productivity.

In recent weeks and months, we’ve really seen a sharp rise in open model adoption by enterprises as they seek to, A, drive more control and better customization, and B, really bring down the cost of inference and prevent vendor lock-in into one or two sort of frontier models.

So, the key implications here are that open models have been a real catalyst in unlocking enterprise AI adoption and really driving exponential token growth by bringing down the cost of inference.

So, from an AI infrastructure perspective, cheap AI is a huge positive. By bringing down the cost of inference, you’re generating more tokens, you’re running more agents, and you’re automating more workflows. So that’s hugely beneficial across the stack, from compute, networking, storage, and memory.

Now, we also think the hyperscalers can benefit here. By acting as the gatekeepers, these guys host these open models, and they can route workflows to the relevant model. And they can add value potentially no matter which model wins.

And then lastly, as from an AI lab perspective, while there is some concern that there’s some pricing pressure or potentially deflationary token pricing, we don’t see any evidence in that. In fact, the economics continue to benefit the AI labs. While the majority of token volume comes through open models, the majority of token revenue is actually being generated by closed or frontier-level models.

So, in terms of what to watch for, well, there’s been a lot of fears that open models could end up commoditizing the model layer. Right now, that’s just not the case. And we don’t think this is going to be a winner-takes-all market. In fact, we continue to believe that the most sophisticated workflows, like scientific research, advanced coding, and legal reviews, will be routed to the highest-value, high-end frontier models. Now, conversely, we think that some of the more basic workflows, like productivity tools, will be routed to open models. In other words, the right model for the right workload.

We also see that the AI labs are not standing still here, and some of the latest low- to mid-range offerings such as GPT, Terra, Luna, Opus 5.5, are quite compelling in terms of price and performance versus their open-model counterparts.

So, stepping back, we think both open and closed models will have a key role in advancing AI. The key question isn’t whether open models will cannibalize closed models; it’s just how that evolving mix of open versus closed really changes the equation on value creation in the overall AI ecosystem, and that’s something we’ll continue to watch and monitor going forward.