The AI race is getting more complicated.
Despite the hype around the latest and biggest frontier models, companies are finding they don’t always need the most powerful approach for every task. As open-weight models get better, cheaper, and easier to customize, there’s more pressure on frontier model companies like OpenAI and Anthropic to develop features that retain customers.
That’s creating a new set of tradeoffs for businesses between cost, control, and security.
In an interview with Business Insider, Aaron Jacobson, a partner at New Enterprise Associates whose portfolio companies include Databricks, Factory, and Together, said that he sees this debate at the center of the next phase of the AI industry. He expects frontier models to remain important for the hardest, most complex problems, but for many everyday tasks, cheaper models may be good enough.
The result could be a shift away from the idea that bigger is always better — and toward a world where companies take a more curated approach to model selection.
This interview has been edited and condensed for clarity.
Can you define the difference between open-source, open-weight, and closed AI?
Open source is where you can actually recreate the entire model yourself. You have all the data, all the data pipelines, and you can basically train it from scratch and get the weights.
Open weight is where you just have the weights, but you can’t actually create the model from scratch because you don’t have all the data and all the data pipelines. You don’t actually know the data the model was trained on or how it was trained.
With closed frontier models, you don’t have access to either of those. You can do some fine-tuning of the weights through APIs that Anthropic and OpenAI provide, but you don’t actually have access to the weights. You can’t run the models yourself.
Is open versus closed really a debate for you? Or are frontier models so superior that there isn’t even a question of whether open source can catch up?
I think the frontier models are, in terms of the broad set of problems— and even for coding — better. But here’s the thing: That doesn’t mean they’re the most useful. They’re better, but they’re also more expensive.
The debate for me, and for a lot of the industry, is: Do you need frontier intelligence for everything?
Open weight is getting good enough for a lot of use cases, and that’s where it begins to challenge the frontier business model. You don’t need to send all complex tasks to frontier intelligence.
What do you think are the best use cases for frontier models?
I think very hard, complex problems. Or open-ended problems that are hard to solve or require longer time horizons.
We’re really shifting away from the chatbot era to the agentic era. If you ask a model to do something complex — like plan an entire vacation and book flights, or, as a developer, construct a highly complex banking application that must meet a range of security requirements — that kind of open-ended, long-horizon problem is a great use case for a frontier model. It can handle the planning and some of the more complex token generation, while routing simpler tasks — such as parts of the code that don’t require as much sophistication — to lower-cost models.
It sounds like there are also a lot of use cases for open-weight models.
I think this is the debate: What are the use cases for the open-weight models available in the US today given that the most advanced open models are of Chinese origin?
Individual developers are using open-weight models. Some companies are OK using Chinese open-weight models to generate their code or perform other work. A lot of enterprises are not, and they don’t permit Chinese models to be used to generate code or touch their data.
For the broad swath of the market — a lot of which is enterprise workloads — the most advanced open-weight models can technically do a cheaper and maybe good-enough job. But because these models are of Chinese origin and not truly open source, there can be security concerns in using them that will hold back deployment. So, this work will either be sent to less performant, non-Chinese open models or the more expensive closed frontier models.
Certainly, the more sensitive the data, the bigger the concern. Coding applications, like source code, are super sensitive. Anything that touches finance or customer data is sensitive.
Can you talk a little bit more specifically about the pricing pressure that open-source or open-weight alternatives are putting on closed models?
I think you see it twofold. The Financial Times reported that Anthropic’s lower-cost Opus 5 model has overtaken Fable in business spending since launching in late July, according to Ramp.
I also think that’s because not only is Opus good enough, but for a wide swath of startups and some enterprise use cases, open models are good enough for some of the code generation. And open models are cheaper because you don’t have to pay the tax on the frontier-model company that runs it itself.
So that cost for tokens that are just as good or good enough puts pricing pressure on frontier models, which price them more expensively to recoup the cost of training.
Where do you see the market heading?
I’m pretty bullish on open weight capturing more share.
If you characterize the last 10 years in AI, there was this idea that bigger is better: more data, more compute, bigger models. I think the next decade is going to be defined by efficient use of models. Not everything needs to be frontier intelligence.
Read the original article on Business Insider
The post An NEA partner says not every AI task needs frontier intelligence appeared first on Business Insider.
