AI's Next Frontier: Cost, Control, and Compute Race

The artificial intelligence landscape is rapidly evolving, moving beyond the initial frontier of simply building the most powerful models. The next critical phase of this AI race, as discussed in a recent segment featuring Deirdre Bosa from CNBC, is increasingly centered on cost, control, and compute. This signals a significant shift in how AI will be deployed and utilized by businesses and developers alike.

The Post-Frontier Era of AI

The conversation highlights a transition into what's being termed the 'post-frontier era' of AI. In this new phase, the focus is not just on the capabilities of a single model, but on the entire system surrounding it. Aravind Srinivas, Co-founder & CEO of Perplexity, joined the discussion to elaborate on this critical evolution. Srinivas emphasized that a model alone is no longer the complete product; the true value will reside in the ability to effectively route, manage costs, and exert control over AI deployments.

This shift is driven by the realization that while frontier models offer impressive capabilities, their associated costs and lack of granular control can be prohibitive for widespread enterprise adoption. The trend is leaning towards companies building their own AI infrastructure and utilizing more specialized, cost-effective models for specific tasks. Srinivas noted that 'the model alone is no longer the product. It is the harness.' This 'harness' refers to the orchestration layer that enables efficient use of various models and resources.

The Rise of Orchestration and Open-Source

The concept of 'orchestration' is becoming paramount, allowing businesses to dynamically select and deploy the most appropriate AI model for a given task, rather than relying on a single, potentially expensive, general-purpose model. This approach also opens the door for greater utilization of open-source models, which are becoming increasingly competitive in performance while offering significant cost advantages.

Srinivas pointed out that '90-plus percent of the tokens will come out of open-weight models over the next 18 to 24 months, possibly even by the end of the year.' This prediction underscores the growing momentum behind open-source AI development and adoption. He further elaborated that for AI to be truly productive and beneficial to businesses, it needs to be 'a lot more affordable,' which is where open-source models and efficient orchestration play a crucial role.

Benchmarking and Real-World Value

The discussion also touched upon the limitations of traditional AI benchmarking. While benchmarks provide a useful starting point, Srinivas highlighted that they often don't fully capture the real-world performance and cost-effectiveness of models in enterprise applications. He cited Perplexity's own work in developing internal benchmarks, such as 'DashBench,' to evaluate AI code reviewers. Their findings indicated that a system leveraging multiple models, rather than a single high-performing one, achieved '1.7x higher weighted recall, with similar precision.' This demonstrates the power of tailored orchestration.

The shift in focus is from purely academic benchmarks to practical, enterprise-driven metrics. Companies are increasingly focused on the 'value of enterprise data and context,' ensuring that AI deployments are not only capable but also cost-efficient and controllable. This means that the value proposition of AI is moving beyond raw model performance to encompass the entire deployment lifecycle.

The Future of AI is Open and Orchestrated

The prevailing sentiment is that the future of AI for enterprises will be characterized by a hybrid approach, combining the power of frontier models with the flexibility and cost-effectiveness of open-source alternatives. The ability to orchestrate these different models, run them efficiently on local hardware, and maintain control over data and deployments will be key differentiators. As Srinivas stated, 'you don't use a Ferrari to go to the grocery store.' This analogy perfectly captures the need for task-specific AI solutions that balance performance with practicality.

Perplexity's own development of its new orchestrator model, built on China's Z.Ai GLM 5.2, exemplifies this trend. By leveraging open-source models and focusing on efficient deployment, companies can offer more accessible and cost-effective AI solutions. This approach allows businesses to leverage their own data and expertise, rather than relying solely on closed, proprietary systems. The future of AI is not just about bigger models, but smarter, more adaptable, and more democratized systems.