Perplexity Pushes Cheaper Open AI Models as 90% of Tokens May Shift Within 24 Months
Updated
Updated · CNBC · Jul 10
Perplexity Pushes Cheaper Open AI Models as 90% of Tokens May Shift Within 24 Months
1 articles · Updated · CNBC · Jul 10
Summary
Perplexity this week previewed a computer-use system built around Z.ai’s open GLM 5.2 model, letting a cheaper model do most tasks and escalating only harder work to stronger models.
That setup reflects a broader enterprise shift from chasing the biggest model to routing each job by cost, speed, control and data needs as companies tighten AI spending.
Benchmark’s Peter Fenton said open-weight models could generate more than 90% of AI tokens within 18 to 24 months, pressuring inference margins at OpenAI, Anthropic and other frontier-model providers.
Ollama CEO Jeff Morgan said more than 85% of the Fortune 500 already use its tools to run open models, including firms in regulated sectors that want models closer to their own data.
The move toward open models—many from Chinese labs such as Z.ai and DeepSeek—also raises U.S. competitiveness questions and could shift some AI workloads from giant cloud data centers to local devices.
With AI costs spiraling, can open-source models truly democratize AI or just shift expenses from APIs to internal operations?
As AI models become commodities, will the 'orchestration layer' become the new battleground for tech supremacy?
Open-Source AI and Orchestration in 2026: The Shift Driving Enterprise Adoption, Price Wars, and New Risks
Overview
As of July 2026, the AI industry is experiencing a major transformation, moving away from relying only on large, proprietary models to a more distributed and cost-effective approach using open-weight models and advanced orchestration systems. This shift is changing how AI is deployed and managed, leading to hybrid systems where some AI tasks run locally on user devices while more complex work is handled in the cloud. This new model aims to optimize both performance and efficiency, impacting the ongoing data center buildout but not removing the need for data centers entirely.