Google Cloud AI Lead Maps Out 3 Frontiers for AI Models

As a Vice President of Product at Google Cloud, Michael Gerstenhaber's work predominantly centers on Vertex, the company’s integrated platform for deploying enterprise AI solutions. This role provides him with a unique vantage point, offering profound insights into how organizations currently leverage AI models and what advancements are still necessary to fully unlock the potential of agentic AI.

During our conversation, one particular concept articulated by Michael resonated strongly, offering a fresh perspective. He explained that AI models are concurrently pressing against three distinct frontiers: their inherent intelligence, their response time, and a third crucial attribute that relates less to raw capability and more to economic feasibility—specifically, whether a model can be deployed affordably enough to operate at massive, highly unpredictable scales. This framework presents a novel and particularly valuable way of assessing model capabilities for anyone striving to advance frontier models in new directions.

This interview has been adapted for conciseness and clarity.

When asked to outline his journey in AI and his responsibilities at Google, Michael shared, "I’ve been in AI for about two years now. I was at Anthropic for a year and a half, I’ve been at Google almost half a year now. I run Vertex, Google’s developer platform." He elaborated that most of their clientele consists of engineers developing their own applications, who seek access to agentic patterns, an agentic platform, and the inference capabilities of the world’s most sophisticated models. "I provide them that, but I don’t provide the applications themselves. That’s for Shopify, Thomson Reuters, and our various customers to provide in their own domains."

Explaining his decision to join Google, Gerstenhaber emphasized the company's distinctive vertical integration. "Google is I think unique in the world in that we have everything from the interface to the infrastructure layer. We can build data centers. We can buy electricity and build power plants. We have our own chips. We have our own model. We have the inference layer that we control. We have the agentic layer we control. We have APIs for memory, for interleaved code writing. We have agent engine on top of that that ensures compliance and governance. And then we even have the chat interface with Gemini enterprise and Gemini chat for consumers, right?" He concluded, "So part of the reason I came here is because I saw Google as uniquely vertically integrated, and that being a strength for us."

It seems peculiar that, despite the differences between companies, the major AI labs appear to be remarkably close in their capabilities. This raises the question: is the competition solely a race for greater intelligence, or is the landscape more intricate?

Michael identifies three critical boundaries that define model performance. The first relates to raw intelligence, exemplified by models like Gemini Pro, which are optimized for tasks such as coding. In such scenarios, the paramount concern is generating the absolute best code possible, regardless of whether it takes 45 minutes, given the long-term requirements for maintenance and production. The objective is simply the highest quality output.

The second boundary revolves around latency. Consider applications in customer support, where an AI needs to apply a policy—for instance, determining if a return is permissible or if a seat upgrade is available. While intelligence is crucial for accurate policy application, the answer's correctness becomes irrelevant if it takes 45 minutes to deliver. Therefore, for these use cases, the goal is the most intelligent product within a defined latency budget, as additional intelligence offers no benefit once the user disengages due to delay.

The final category addresses scenarios where cost and scalability are paramount. Platforms like Reddit or Meta, tasked with moderating the entire internet, operate with substantial budgets but cannot assume enterprise risk without certainty about how solutions will scale. The unpredictable volume of "poisonous posts" or other content means they must align their budget with the highest intelligence model they can afford, ensuring it scales infinitely across countless subjects. In these instances, cost becomes an exceptionally critical factor.

One recurring question has been why agentic systems have been slower to gain widespread traction. Despite the existence of capable models and impressive demonstrations, the major shifts anticipated a year ago have not yet materialized. What, then, is impeding their adoption?

Michael attributes this to the nascent stage of the technology, which is only about two years old, and the substantial lack of foundational infrastructure. "We don’t have patterns for auditing what the agents are doing. We don’t have patterns for authorization of data to an agent." He explains that these essential patterns require significant development work to be ready for production. Production, by its nature, is always a trailing indicator of technological capability. Thus, two years is simply not enough time to observe the full impact of this intelligence in production environments, which is where many organizations are currently encountering difficulties.

He notes that the adoption has progressed uniquely quickly within software engineering because it integrates seamlessly into the existing software development lifecycle. "We have a dev environment in which it’s safe to break things, and then we promote from the dev environment to the test environment." At Google, for example, the process of writing code mandates two individuals to audit it and jointly affirm its quality before it can be released under Google’s brand to customers. "So we have a lot of those human-in-the-loop processes that make the implementation exceptionally low-risk." The challenge now lies in replicating these successful patterns in other domains and professions.