Google Cloud VP on the three frontiers of AI model capability

Michael Gerstenhaber, the product VP at Google Cloud who oversees Vertex AI—Google’s unified platform for deploying enterprise AI—argues that AI models are pushing against three distinct frontiers. According to Gerstenhaber, these boundaries are raw intelligence, response time, and cost. While raw intelligence is critical for complex tasks like writing code, other enterprise applications require balancing capability against latency budgets. Latency, which refers to the response time of an AI model, is a constraint in customer-facing scenarios. For instance, Gerstenhaber notes that in customer support, a user will hang up if an answer takes 45 minutes, making a model tuned for raw intelligence less useful if it cannot deliver within a required time frame. “I see three boundaries. Models like Gemini Pro are tuned for raw intelligence,” Gerstenhaber said. Conversely, the cost frontier is important for platforms like Reddit or Meta, which are potential users of scalable AI models that might need to moderate the internet. These platforms must restrict their budgets to models at the highest intelligence they can afford, but in a scalable way.

Gerstenhaber, who has about two years of experience in AI—including a year and a half at Anthropic and almost half a year at Google—presents Google’s vertical integration as a strategic strength. He states that he thinks Google is uniquely vertically integrated in the world, controlling everything from infrastructure and chips to the model, the inference layer, and the agentic layer. Inference refers to the process of running a trained AI model. This vertical integration, which extends to consumer chat interfaces like Gemini enterprise and Gemini chat, allows the company to manage the trade-offs between intelligence, latency, and cost for customers like Shopify and Thomson Reuters.

Despite interest in agentic AI—which refers to AI systems capable of autonomous action—these systems are taking time to catch on. Gerstenhaber asserts that this slow adoption is due to missing infrastructure, particularly the lack of established patterns for auditing what agents do and authorizing their access to data. Because the technology is basically two years old, production implementations are still catching up to what the models can do. To deploy these systems, enterprises must establish human-in-the-loop processes. For example, Google manages risk in its own software development by requiring two people to audit and approve code before it is put into production.

Why it matters

Gerstenhaber’s framework provides a practical lens for enterprise leaders to evaluate AI models beyond raw capability, focusing on the trade-offs required for real-world deployment.