Google Cloud AI chief outlines key frontiers for advancing model capabilities

As the Vice President of Product for Google Cloud, Michael Gerstenhaber's primary focus is on Vertex, the company's comprehensive enterprise AI deployment platform. This role affords him a broad perspective on how businesses are practically implementing AI models and the key hurdles that remain in unlocking the potential of agentic AI.

My conversation with Michael featured a particularly original insight. He frames AI model advancement as a push against three distinct frontiers simultaneously: raw intelligence, response latency, and a crucial third factor—cost efficiency. This third frontier is less about pure capability and more about affordability, determining whether a model can be deployed cheaply enough to operate at massive, unpredictable scale. It's a valuable framework for anyone looking to guide the development of cutting-edge models in a new direction.

This interview has been condensed and edited for clarity.

To begin, could you outline your background in AI and your current role at Google?

I've been working in AI for roughly two years. After a year and a half at Anthropic, I joined Google about six months ago. I lead the Vertex developer platform. Our primary users are engineers building custom applications. They seek access to agentic workflows, a robust agentic platform, and the inference power of the world's most advanced models. My team provides that foundational platform, while the end-user applications are developed by our customers—companies like Shopify and Thomson Reuters—within their specific industries.

What motivated your move to Google?

Google's unique position lies in its vertical integration, spanning from the user interface down to the infrastructure layer. We design our own data centers, manage power procurement, develop custom silicon (TPUs), train our own foundational models, and control the inference and agentic orchestration layers. We offer APIs for capabilities like memory and code generation. On top of that, we've built an agent engine for compliance and governance, plus consumer and enterprise interfaces like Gemini. This end-to-end control is a significant strength and a key reason I joined.

Save Up to $300 or 30% on TechCrunch Founder Summit

Join 1,000+ founders and investors at TechCrunch Founder Summit 2026 for a full day dedicated to growth, execution, and scaling. Gain insights from industry-shaping founders and investors. Network with peers who are at similar growth stages. Leave with actionable strategies you can implement right away.

Offer ends March 13.

Save Up to $300 or 30% on TechCrunch Founder Summit

Join 1,000+ founders and investors at TechCrunch Founder Summit 2026 for a full day dedicated to growth, execution, and scaling. Gain insights from industry-shaping founders and investors. Network with peers who are at similar growth stages. Leave with actionable strategies you can implement right away.

Offer ends March 13.

It's interesting—despite their differences, the three major AI labs appear closely matched in core capabilities. Is this purely a race for greater intelligence, or is the landscape more nuanced?

I see three distinct performance boundaries. First, models like Gemini Pro are optimized for raw intelligence. Tasks like complex code generation are a good example; you want the highest quality output possible, even if it takes longer, because you'll be maintaining and deploying that code.

The second boundary is latency. In a live customer support scenario, an AI needs both the intelligence to apply policy (e.g., processing a return or a seat upgrade) and the speed to deliver an answer before the user disengages. Here, you need the smartest model that can operate within a strict time budget.

The third boundary is cost at scale. A company like Reddit or Meta, aiming to moderate vast volumes of content, faces unpredictable demand. They can't take on unlimited enterprise risk without knowing how costs will scale. They need the most intelligent model they can afford that is also cost-efficient enough to handle a potentially infinite number of tasks.

A persistent question is why agentic AI systems haven't seen wider adoption. The model capabilities and impressive demos exist, yet the transformational change many anticipated a year ago hasn't fully materialized. What are the main bottlenecks?

The underlying technology is only about two years old, and critical infrastructure is still missing. We lack established patterns for auditing agent actions or authorizing data access for agents. Implementing these production-ready patterns requires significant work. Production adoption always lags behind technological capability. Two years simply isn't enough time for the full scope of the intelligence to be reflected in mature, deployed systems, and that's where the current challenge lies.

Adoption has been uniquely rapid in software engineering because it integrates well with existing development cycles. Engineers have safe dev and test environments. At Google, code requires review and approval from two engineers before it bears the company's brand for customers. These human-in-the-loop processes make implementation very low-risk. We need to develop equivalent, safe patterns for other domains and professions.