The artificial intelligence industry's measure of progress is shifting from model size to model selection, as companies build systems that route each task to the cheapest capable model. Perplexity CEO Aravind Srinivas said the real product is no longer the model itself but the orchestration layer that decides which model to use, when to escalate, and what tools or data sources to tap. The change could squeeze the pricing power of frontier model providers and reshape the data center buildout currently underway.
"The model alone is no longer the product," Srinivas said. "It is the harness, the orchestration system that puts the model inside a very capable harness and pairs the model with a lot of tools." That system might handle a routine customer service query with a cheap open model, then call a more powerful proprietary model only for a complex coding problem or a high-stakes step.
Open-weight models gain ground
Open-weight models - which companies can download, fine-tune, and run on their own infrastructure - are becoming capable enough to handle a growing share of enterprise AI workloads. Benchmark general partner Peter Fenton said the shift could be dramatic. "A maybe contrarian view that is becoming consensus is our belief that 90-plus percent of the tokens created will come out of open-weight models over the next 18 to 24 months, possibly even by the end of the year," Fenton said. Tokens are the units of data AI models process and generate.
Fenton added that the inference margins generated by frontier model companies will come under pressure when enterprises can run open-weight models without the markup those providers charge. In some cases, smaller models tuned for a specific task can be faster and perform better than larger general-purpose models. Perplexity this week previewed a computer-use system built around GLM 5.2, an open model from China's Z.ai. The system lets a cheaper model handle most of the work and escalates to a stronger one only when needed.
Where the model runs matters as much as where it was built
Benchmark invested in Ollama, a company that helps developers and enterprises download, run, and manage open models. Ollama CEO Jeff Morgan said more than 85% of the Fortune 500 have adopted the platform, including firms in regulated industries such as aviation, insurance, and health care. Many start with smaller models running close to their own data, then expand to larger open models as they gain confidence.
"One thing is where the model's from and where it was created and trained," Morgan said. "But the more important thing to these businesses we speak to is where it runs and how it runs." That preference for local or on-device execution could eventually reduce the volume of AI workloads sent to large cloud data centers, creating a more hybrid infrastructure. Routine tasks might run on local devices, while only the hardest problems get routed to a powerful model in the cloud.
A strategic challenge for the U.S.
Many of the most competitive open-weight models are coming from Chinese labs, including Z.ai and DeepSeek. That makes open-source AI a business issue, a policy issue, and a national competitiveness issue. Srinivas argued the U.S. should support open models because they make AI more affordable and accessible. "If you want the benefits of AI to be widely distributed to small businesses in America and American allied countries, then you really need AI to be a lot more affordable," he said. "And open source is the only way to do that."
For investors, the open question is whether the largest AI labs can maintain their pricing power as open models improve and companies become more selective about which model they use for each job. The massive data center expansion premised on insatiable demand for high-end cloud inference could face a reset if a significant fraction of AI work moves to cheaper, locally run models.
Why this matters for IT, development, and research professionals
The move toward model routing and open-weight systems directly affects how teams choose, deploy, and pay for AI. Instead of defaulting to a single expensive API, you can evaluate a mix of models - open and proprietary - and assign each task to the right one based on cost, latency, and data sensitivity. Tools like Ollama lower the barrier to running open models on your own hardware, which can cut inference costs and keep sensitive data in-house. For researchers and developers, the growing viability of fine-tuned open models means you can often get better task-specific performance without paying per-token premiums. The shift also means architecture decisions - where a model runs, not just which model - will become a standard part of AI system design.
Your membership also unlocks:
