The AI race is no longer just about who has the biggest model
Companies are quietly moving away from the most powerful AI and toward cheaper, smarter systems that pick the right tool for each job. Here is what that means for you.
Key points
- Perplexity CEO Aravind Srinivas said in May 2025 that "the model alone is no longer the product" as AI companies shift focus to orchestration systems.
- Perplexity this week previewed a new system built around GLM 5.2, an open-weight model from China's Z.ai, designed to use cheaper AI for routine steps and stronger AI only when needed.
- Benchmark general partner Peter Fenton predicted that more than 90 percent of AI-generated tokens, the units of data a model processes and produces, could come from open-weight models within 18 to 24 months.
- Ollama, a company that helps businesses run AI models on their own computers, says more than 85 percent of Fortune 500 companies have adopted its software.
For the past two years, the AI race had a simple scoreboard: biggest model, best benchmark. That scoreboard's getting complicated.
Businesses aren't just testing AI any more. They're plugging it into real workflows, and suddenly the question isn't which AI is most powerful but which AI is right for this specific job at a price that makes sense.
Think of it like hiring contractors. You wouldn't bring in a specialist architect to hang a shelf. A customer service chatbot doesn't need the same firepower as a tool debugging complex software. The new generation of AI products is designed to make that call automatically, routing each task to the cheapest model that can handle it.
Perplexity CEO Aravind Srinivas put it plainly to CNBC: "The model alone is no longer the product. It is the orchestration system that puts the model inside a very capable harness and pairs the model with a lot of tools."
This week Perplexity previewed exactly that kind of system, built around GLM 5.2, an open-weight model from Chinese lab Z.ai. Open-weight means the model's underlying code can be downloaded, customised and run on a company's own servers, no monthly subscription to a big AI lab required.
That matters because open-weight models are getting genuinely good. Peter Fenton, a general partner at Benchmark, told CNBC he believes more than 90 percent of all AI-generated output could come from open-weight models within 18 to 24 months, possibly even by the end of 2025. That would squeeze the profit margins of premium AI providers like OpenAI and Anthropic, which currently charge a mark-up for access to their top-tier models. Our earlier story on DeepSeek's 75 percent price cut found that cheaper models haven't yet made AI businesses more profitable, largely because usage grows to fill any savings.
What does this mean for ordinary people and small businesses?
Cheaper AI tools, probably sooner than expected. Jeff Morgan, CEO of Ollama, which makes it easier to run AI models on your own hardware rather than in someone else's cloud, says more than 85 percent of Fortune 500 companies already use his software. That includes heavily regulated sectors like aviation, insurance and healthcare, where keeping data in-house isn't optional.
If AI runs locally on a business's own computers, sensitive customer data never has to leave the building. That's a genuine privacy win over sending everything to a third-party server.
The catch is geopolitics. Several of the strongest open-weight models come from Chinese labs, a tension we looked at in depth when Chinese open-weight models took 41 percent of Hugging Face downloads this spring. Open-source AI is now a policy question as much as a technology one.
Watch the orchestration layer, not the model leaderboards. The builders who know how to stitch models together cleverly are pulling ahead of the labs chasing the largest parameter counts, and that gap is widening faster than most people realise.
