AI’s Next Race: Cost, Control, and Compute

The transcript presents a comprehensive discussion of the shifting AI landscape from a frontier model-dominated paradigm to a multi-model ecosystem centered on cost, control, and orchestration. Arvin Hashemi from Perplexity emphasizes that "the model alone is no longer the product. It is the harness"—referring to the system architecture that routes requests across multiple models based on task requirements. This shift is exemplified by Door Dash's internal testing, which found that no single model excels at all tasks, and that an orchestrated system using multiple models delivers superior results compared to relying on benchmarks or single-model deployments.

Peter Fenton from Benchmark makes a striking prediction: over the next 18-24 months, potentially even by year-end, more than 90% of AI tokens will be generated by open-weight models. This projection is driven by two forces: (1) cost advantages—open models can be run without the 3-5x markup that frontier labs charge, and (2) performance benefits—companies like Sierra find that fine-tuned open models deliver lower latency and higher performance for bounded tasks. Chinese models like GLM and DeepSeek have captured dominant market share in token usage, with GLM particularly noted as achieving near-Opus-level performance at one-third the cost when post-trained properly.

Hashemi discusses Perplexity's new orchestrator system, which starts with a cheaper open model and escalates to premium models only when necessary—analogous to not using a Ferrari for grocery shopping. He advocates for "intelligence per watt" as a defining metric, arguing that power constraints will be the bottleneck limiting AI adoption more than demand. Local compute solutions like Nvidia's DGX Spark, which unifies CPU/GPU memory, could enable efficient edge deployment, reducing reliance on centralized data centers.

A critical theme is data sovereignty and control. Both Hashemi and Fenton argue that enterprises should maintain ownership of proprietary workflow data rather than feeding it into frontier model companies. This creates economic incentive for adopting open-weight models that can be self-hosted and fine-tuned on internal data. Jeff Morgan from Ollama reports that 85% of Fortune 500 companies now use Ollama to run open models locally, citing trust, control, and regulatory compliance (particularly in aviation, insurance, and healthcare) as primary drivers.

The conversation challenges common misconceptions: open models aren't chosen primarily for cost but for performance when fine-tuned to specific tasks; Chinese open models can be securely deployed in US/EU data centers, addressing security concerns; and open-weight models can be monetized through API access and specialized services. The discussion includes policy considerations, with Fenton warning against regulatory strategies that block Chinese model access, arguing such restrictions would disadvantage the US ecosystem globally.