Michael Gerstenhaber, Product VP at Google Cloud, leads the development of Vertex, the company’s unified platform for enterprise AI deployment. His role gives him a unique vantage point into how organizations are adopting AI, particularly in the emerging domain of agentic systems. In a recent conversation, he introduced a compelling framework for understanding the evolving capabilities of AI models: they are advancing along three interconnected frontiers simultaneously. The first is raw intelligence—measured by a model’s ability to solve complex tasks, such as writing high-quality code or generating nuanced content. Models like Gemini Pro exemplify this frontier, where performance is prioritized regardless of speed. The second frontier is response time, critical in real-time applications like customer support, where even the most accurate answer is useless if it takes minutes to deliver. Here, the goal is to find the most intelligent model that operates within strict latency constraints, ensuring user engagement isn’t lost. The third frontier, often overlooked, is cost efficiency at scale. For platforms like Reddit or Meta, which must moderate vast, unpredictable volumes of content, the challenge isn’t just intelligence or speed—it’s affordability and scalability. They need models that remain effective and affordable even under extreme, variable loads, making cost a key determinant of deployment viability. Gerstenhaber attributes his move to Google partly to the company’s rare vertical integration: from custom chips and data centers to proprietary models, inference infrastructure, agent frameworks, and end-user interfaces like Gemini Enterprise. This full-stack control enables faster iteration and more reliable deployment—something he sees as a competitive edge. Despite the apparent parity in raw model capabilities across major AI labs, he argues the real competition lies in mastering these three frontiers in tandem. The delay in widespread adoption of agentic AI, he believes, isn’t due to lack of capability but to immature infrastructure. There are still no standardized patterns for auditing agent behavior, managing data authorization, or ensuring compliance at scale. These missing pieces make production deployment risky and complex. While agentic systems have seen rapid progress in software engineering—where development environments allow safe experimentation and human review acts as a safety net—other domains lack similar guardrails. Until such operational patterns are established across industries, the full potential of agentic AI will remain unrealized. Industry observers note that Gerstenhaber’s three-frontier model reflects a maturing industry focus: from pure performance to practical deployment. Google’s investment in end-to-end control positions it well to lead in scalable, responsible AI. Meanwhile, companies like Anthropic and Meta are racing to close the infrastructure gap. The next phase of AI innovation won’t be about bigger models, but about smarter, safer, and more scalable systems—precisely the challenge Gerstenhaber is tackling at Vertex.