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

Industry Context

The AI industry is shifting from a "model capability competition" to a "cost and efficiency competition." In the past few years, the industry focus was on model parameter counts and benchmark scores, but as applications have been deployed, companies are beginning to focus on the economics and controllability of actual deployment. This CNBC interview clearly illustrates this shift: Perplexity is building a new orchestrator model and has chosen to develop it based on Chinese open-source models. This choice in itself reflects the maturity of the open-source ecosystem and the pragmatic considerations under cost pressures from closed-source models.

Market Impact

The "token value per watt" proposed by Srinivas brings compute efficiency to the core of competition. This metric directly links model output value to energy consumption, meaning that the competitiveness of AI companies in the future will depend not only on model quality but also on economic benefits per unit of power consumption. For customers, this means lower inference costs and more flexible deployment options; for investors, it requires re-evaluating AI companies that rely solely on scale expansion without efficiency advantages.

Competitive Landscape

The rise of open-source models is reshaping the market landscape. Ollama's growth shows that companies are increasingly inclined to run downloadable, controllable models, and this shift in preference puts direct pressure on closed-source API providers. Benchmark's Peter Fenton views this trend as a fundamental change in the industry structure—companies are no longer willing to hand over core data and model control to third parties. Perplexity's choice to build on Chinese open-source models also indicates that, against the backdrop of geopolitical tensions, the open-source ecosystem has become important infrastructure for cross-regional innovation.

Enterprise Implications

For enterprise decision-makers, the key takeaway is that the initiative in AI deployment is returning to the enterprise itself. Through open-source models and local operation, companies can better control data privacy, customization needs, and cost structures. At the same time, however, they need to build corresponding capabilities in model operations and governance. Tools like Ollama, with their low barriers to entry, allow companies to experiment quickly, but production-grade deployment still requires rigorous evaluation.

Outlook

Over the next 12–24 months, we are likely to see more companies adopt a hybrid AI architecture: using controllable open-source models for core tasks while still calling on top-tier closed-source APIs for complex tasks. Meanwhile, "token value per watt" will become a key metric for measuring the return on investment in AI infrastructure. Over a three-year horizon, open-source and closed-source models will coexist in the long term, but control and cost efficiency will determine the ultimate positions of various players in the value chain.