For the past two years, the race for artificial intelligence has been easy: bigger models, better benchmarks and any company can claim the lead, at least until the next launch.
That scorecard is starting to look incomplete.
As companies move from testing AI to implementing it in real products and workflows, it’s no longer about tapping the best model, but reaching the one best suited for a given task, at the right cost, with the required data and in the chosen location.
That shift opens the door to a new kind of AI competition, one that focuses less on model size and more on routing, cost, control and computation.
“The model alone is no longer the product,” Perplexity CEO Aravind Srinivas told CNBC. “Harness, a singing system that puts the model inside a very capable harness and pairs the model with a lot of tools.”
That means AI products are becoming systems that can decide which model to use, when to use it and which external tools or company data sources are needed. A customer service job may not require an expensive model. A complex coding problem may. A typical internal workflow may work with a cheaper open model. A difficult step can be upgraded to a more powerful one.
“The answer is always use whatever is best for the job,” Srinivas said.
The emergence of other models comes as the American company tightens its belt in AI spending, and presents another challenge to OpenAI and Anthropic, which has flourished in the past few years by selling cutting-edge technology.
Aravind Srinivas, CEO of Perplexity AI.
CNBC
Confusion this week previewed a new version of its desktop product built around GLM 5.2, an unlocked model from China’s Z.ai. The system is designed to allow the cheaper model to handle most of the work while the more expensive model is only used when needed.
That trend reflects a broader shift in the market. Open weight models, which can be downloaded, tuned and driven by the companies themselves, are becoming more and more powerful. It’s also cheaper to run than premium proprietary models from the big AI labs.
Benchmark general partner Peter Fenton said the change could be significant.
“The perhaps counterintuitive view that is beginning to be agreed upon is our belief that 90-plus percent of the tokens created will come out of open-weighted models within the next 18 to 24 months, maybe even by the end of the year,” Fenton told CNBC.
Tokens are units of process and AI models of data.
“The inference margins produced by the border model companies, I think, will be under pressure if you can run those without the markup they offer, if you have enough models from the open weights,” Fenton said.
Fenton said the move to open models isn’t just about saving money. In some cases, smaller models tuned for a specific task can be faster and perform better than larger general-purpose models.
‘Where it runs and how it runs’
This is one of the reasons Benchmark invested in Ollama, a company that makes it easy for developers and businesses to download, run and manage open source models.
“One thing is where the model comes from and where it was created and trained,” said Ollama CEO Jeff Morgan. “But the most important thing about these businesses we’re talking to is where they work and how they’re doing.”
Morgan said Ollama is accepted by more than 85% of the Fortune 500, including companies in regulated industries such as aviation, insurance and health care. He said many companies start with small models that work closely with their data, then move to larger open models as they get comfortable.
The rise of open-source models also poses a strategic challenge to the US Many of the most competitive open-source models come from Chinese labs, including Z.ai and DeepSeek. That has made open source AI a business issue, a policy issue and a national competitiveness issue.
Srinivas said the US should support open models because they make AI affordable and accessible.
“If you want the benefits of AI to be widely distributed to small businesses in America and in the United States, you really need AI to be more affordable,” Srinivas said. “And open source is the only way to do that.”
The shift may also affect the massive data center construction that continues throughout the technology industry. The current AI boom assumes that demand will continue to flow to large cloud data centers filled with high-end chips. Srinivas says some AI work could end up running locally, on devices owned by consumers or businesses.
That wouldn’t eliminate the need for data centers, but it would create a hybrid AI system, with routine tasks performed locally and more complex tasks sent to a more powerful model in the cloud.
For investors, the question is whether the biggest AI labs can maintain their pricing power as open models improve and companies become more selective about who uses them.
WATCH: OpenAI’s Sam Altman says China’s open source models are going very well
