Perplexity CEO Aravind Srinivas said AI market competition is shifting from model size to operationa...

Perplexity CEO Aravind Srinivas said AI market competition is shifting from model size to operational efficiency. In a July 12 CNBC interview, Srinivas argued that the key competitive differentiator is no longer building the single best frontier model, but orchestrating task-appropriate models — deploying cheap models for customer support and internal automation while reserving high-performance models for complex coding and reasoning. Perplexity is testing this architecture in its newest computer-use product by integrating Zhipu AI's (智谱AI) open-weight GLM 5.2 model as the primary worker, calling a more capable model only for high-difficulty reasoning steps.

Why it matters: This marks the emergence of a new pattern we have tracked as orchestration-as-moat, where an AI service's defensibility lies not in owning the largest parameter count but in routing queries across heterogeneous models to optimize cost and latency. The interview explicitly validates our longstanding debate frame about whether best-in-class model performance will remain the dominant axis of competition. Peter Fenton, general partner at Benchmark, adds weight to the thesis: he predicts over 90% of AI inference tokens will come from open-weight models within 18–24 months, putting structural pressure on frontier model providers' inference-service margins.

Srinivas's remarks and Perplexity's concrete product moves accelerate a capital-compression dynamic already visible across the industry. As open-weight model quality rises — Zhipu AI's GLM 5.2 being a case in point — enterprises gain the option to escape hyperscaler-distribution moats and retain data control while cutting inference spend. The winner in this new regime may not be the lab with the largest cluster, but the system builder with the most sophisticated routing and orchestration layer.