For the past two years, the artificial intelligence race has been easily scored on bigger models, better benchmarks, and which companies can take the lead, at least until the next launch.
That scorecard is starting to look incomplete.
As companies move from testing AI to using it in real products and workflows, it’s no longer about having the best model, it’s about accessing the best model for a specific job, at the right cost, with the data you need, and in the environment of your choice.
This change opens the door to a new kind of AI competition that focuses on routing, cost, control, and computing rather than model size.
“The model is no longer the product,” Perplexity CEO Aravind Srinivas told CNBC. “It’s a harness, it’s an orchestration system that puts the model inside a very high-performance harness and combines the model with a lot of tools.”
In other words, AI products are becoming systems that can decide which models to use, when to use them, and what external tools and enterprise data sources are needed. Your customer service tasks may not require the most expensive model. Possibly a complex coding problem. Routine internal workflows can be performed in a cheaper open model. More difficult steps can be escalated to more powerful steps.
“The answer is always to use what’s best for the task,” says Srinivas.
The emergence of alternative models comes as American companies ramp up spending on AI, posing new challenges for OpenAI and Anthropic, which have thrived over the past few years by selling cutting-edge technology.
Perplexity AI CEO Aravind Srinivas said:
CNBC
Perplexity this week previewed a new system for its computer products built on China’s Z.ai’s open model GLM 5.2. The system is designed to handle more work with cheaper models, while calling more powerful models only when needed.
This approach reflects broader changes in the market. Openweight models, which companies can download, adjust, and run themselves, are becoming more powerful. It also costs less to run than premium proprietary models offered by the largest AI labs.
Benchmark general partner Peter Fenton said the change could be dramatic.
“The contrarian view that is probably becoming the consensus is our belief that over 90 percent of the tokens created will come from an open weight model over the next 18 to 24 months, probably by the end of the year,” Fenton told CNBC.
A token is a unit of data that an AI model processes and produces.
“I think the inference margins generated by Frontier Models companies are going to come under pressure once they can run them without the markup that Frontier Models companies provide and we get enough models from open weight,” Fenton said.
Fenton said the move to an open model is not just about saving money. In some cases, a smaller model tailored for a specific task can be faster and perform better than a larger, general-purpose model.
“Where and how do we run?”
That’s one of the reasons Benchmark invested in Ollama, a company that makes it easy for developers and enterprises to download, run, and manage open models.
“One is where did the model come from? Where was it created and trained?” said Ollama CEO Jeff Morgan. “But what’s more important to these companies we talk to is where and how they operate.”
Morgan said Orama is used by more than 85% of Fortune 500 companies, including companies in regulated industries such as aviation, insurance and health care. He said many companies start by running small models that are close to their own data and expand to larger open models as they become comfortable.
The rise of open models also creates strategic challenges for the United States. Many of the most competitive open weight models come from Chinese labs such as Z.ai and DeepSeek. As such, open source AI has become a business problem, a policy problem, and a national competitiveness problem.
Srinivas said the U.S. should support an open model to make AI more affordable and accessible.
“If we want the benefits of AI to be widely available to small and medium-sized businesses in the United States and our allies, we need to make it more affordable,” Srinivas said. “And open source is the only way to do that.”
This change could also impact the massive data center construction underway across the technology industry. The current AI boom assumes that demand will continue to flow to large cloud data centers equipped with high-end chips. Srinivas said some AI work could eventually be performed locally on devices owned by consumers and businesses.
While the need for data centers will not disappear, more hybrid AI systems may be built where routine tasks are performed locally and the most difficult work is sent to more powerful models in the cloud.
The question for investors is whether the biggest AI labs can maintain pricing power as open models improve and companies become more selective about what they use.
WATCH: OpenAI’s Sam Altman says China’s open source model is getting much better
