Michael Gerstenhaber, a product vice president at Google Cloud, primarily focuses on Vertex AI, the company's comprehensive platform for enterprise AI deployment. This role provides him with broad insight into how businesses are implementing AI models and what advancements are still needed to fully realize the potential of agentic AI. In a recent conversation, Gerstenhaber introduced a compelling framework that reframes how we evaluate AI models. He described them as advancing along three simultaneous frontiers: raw intelligence, response speed, and a third, cost-related dimension—specifically, whether a model can be deployed affordably enough to operate at vast, unpredictable scales. This perspective offers a fresh and highly useful lens for anyone aiming to steer the development of cutting-edge models in innovative directions.

The following discussion has been condensed and clarified for readability.

**Could you begin by outlining your background in AI and your current responsibilities at Google?**

My journey in AI spans about two years. I spent a year and a half at Anthropic and have been at Google for nearly six months. I lead Vertex AI, Google's developer platform. Our primary users are engineers creating custom applications. They are seeking access to agentic frameworks, a robust agentic platform, and the inference capabilities of the world's most advanced models. My role is to supply these foundational tools, not the end-user applications themselves. Developing those specific applications falls to our diverse customers, such as Shopify and Thomson Reuters, within their respective fields.

**What attracted you to Google?**

Google stands out globally due to its remarkable vertical integration. The company operates across the entire stack, from user interfaces down to the underlying infrastructure. We design and construct data centers, manage power procurement and generation, develop our own semiconductor chips, and build our proprietary models. We control the inference layer, the agentic layer, and offer APIs for functionalities like memory and interleaved code generation. On top of this, we have an agent engine that enforces compliance and governance. We even provide the end-user chat interfaces with Gemini for both enterprise and consumer use. This unparalleled integration across all layers was a key factor in my decision to join, as I view it as a significant strategic strength.

**Despite the competitive landscape, the leading AI labs often seem closely matched in capabilities. Is the race purely about achieving greater intelligence, or are there more nuanced factors at play?**

I see three distinct frontiers models are pushing against. First, there's raw intelligence, exemplified by models like Gemini Pro. Consider a task like code generation: the paramount concern is obtaining the highest quality output, even if the process takes considerable time, because that code must be maintained and deployed. Here, sheer capability is the priority.

Second, there's the frontier of latency. In scenarios like customer support, where an agent needs to interpret and apply a policy—whether processing a return or a seat upgrade—intelligence is required, but speed is critical. An impeccably correct answer is useless if it arrives 45 minutes later, after the customer has already disconnected. For these use cases, you need the smartest possible model that can operate within a strict latency budget.

The third frontier involves cost at massive scale. A company like Reddit or Meta, aiming to moderate content across the entire internet, operates with large budgets but cannot assume unlimited enterprise risk. The volume of problematic content is unpredictable. Therefore, they must select a model that offers the highest intelligence they can afford, but in a manner that is scalable to a potentially infinite number of queries. In this domain, cost efficiency becomes absolutely paramount.

**One question I've been considering is why agentic AI systems haven't been adopted more rapidly. The underlying models seem advanced, and the demos are impressive, yet we haven't witnessed the transformative shift many anticipated a year ago. What are the primary barriers?**

This technology is essentially only two years old, and significant infrastructure gaps remain. We lack established patterns for auditing agent actions or for authorizing data access for agents. These are critical frameworks that require development before widespread production deployment is feasible. Production readiness always lags behind technological capability. Two years simply isn't enough time to fully translate the intelligence of these systems into reliable, deployed applications, which is where organizations are currently encountering challenges.

The adoption has been notably faster in software engineering because it aligns well with existing development lifecycles. Developers work in safe, isolated environments where experimentation is low-risk, and code progresses through staged reviews before promotion. At Google, for instance, the process mandates that two engineers audit and approve any code before it carries the company's brand to customers. These human-in-the-loop processes make implementation exceptionally low-risk. The key now is to develop and establish similar safe, structured patterns for other professions and use cases.