Deep Dive: The Open vs. Closed AI Race

Open-weight models have come within roughly four months of the best publicly evaluated closed frontier models, and the gap between them has only gotten more volatile.

That is turning the open vs. closed AI race into an increasingly important commercial question. If open models can approach frontier performance while giving companies more control over their data, infrastructure, and customization, what are businesses still paying the frontier labs for? And where does the value ultimately accrue?

To understand the landscape, it helps to start with what we mean by open and closed models. Openness exists on a spectrum, from fully open models that users can download and freely modify to open-weight models with varying restrictions. Closed models keep their weights private and control access.

For businesses deciding which models to build on, the differences go beyond performance and price.

Palantir CEO Alex Karp has warned that enterprises risk surrendering their “alpha”, the proprietary knowledge and processes that distinguish their businesses, to frontier model providers.

Satya Nadella made a similar point in an essay on X: “You essentially pay for intelligence twice, once with money, and again with something even more valuable: the proprietary knowledge you must reveal to make that intelligence useful. The better you want the model to perform, the more of that knowledge you have to feed it!”

Both are warnings about strategic dependence and help explain why more enterprises are exploring open models.

But despite concerns about giving up proprietary knowledge, the evidence shows that companies are still willing to pay for frontier performance, even when the best open-weight models trail by only a few months. Frontier lab revenues continue to accelerate, even as token spending spreads across a much wider range of models.

Leading companies are using both, choosing open models for customization and control, and closed frontier models when they need the highest available capability. Some have reported several-fold efficiency gains and dramatic cost savings by matching different models to the workflows they suit best. We cover a few of these examples in the Deep Dive below.

Some model makers are also taking both approaches. Google has Gemini and Gemma, while Meta has Muse Spark and Llama. Why keep one model closed and give another away? And if open models are free to download, what strings are attached? How do their developers compete financially with closed labs?

Our 99-page Deep Dive examines the economics behind these choices and how they could shape who captures AI profits and which businesses can preserve pricing power as AI becomes cheaper and cheaper to use.

The full report also includes:

Whether frontier labs can sustain the cost of staying ahead as Chinese competitors push prices lower.
The five factors models compete on, and the trade-offs each requires.
The three places a model can run, and how the industry splits between them.
A case study reporting up to 12x greater engineering efficiency and over 20x cost savings.
Why NVIDIA and Samsung are investing in open weights, and where the returns could come from.