As a product VP at Google Cloud, Michael Gerstenhaber works principally on Vertex, the firm’s unified platform for deploying enterprise AI. It provides him a high-level view of how corporations are literally utilizing AI fashions, and what nonetheless must be achieved to unleash the potential of agentic AI.
When I spoke with Michael, I used to be notably struck by one thought I hadn’t heard earlier than. As he put it, AI fashions are pushing in opposition to three frontiers without delay: uncooked intelligence, response time, and a 3rd high quality that has much less to do with uncooked capability than with value — whether or not a model may be deployed cheaply sufficient to run at large, unpredictable scale. It’s a brand new means of occupied with model capabilities, and a very useful one for anybody attempting to push frontier fashions in a brand new route.
This interview has been edited for size and readability.
Why don’t you begin by strolling us by means of your expertise in AI thus far, and what you do at Google?
I’ve been in AI for about two years now. I used to be at Anthropic for a yr and a half, I’ve been at Google nearly half a yr now. I run Vertex, Google’s developer platform. Most of our clients are engineers constructing their very own functions. They need entry to agentic patterns. They need entry to an agentic platform. They need entry to the inference of the smartest fashions in the world. I present them that, however I don’t present the functions themselves. That’s for Shopify, Thomson Reuters, and our varied clients to supply in their very own domains.
What drew you to Google?
Google is I feel distinctive in the world in that now we have every part from the interface to the infrastructure layer. We can construct knowledge facilities. We can purchase electrical energy and construct energy crops. We have our personal chips. We have our personal model. We have the inference layer that we management. We have the agentic layer we management. We have APIs for reminiscence, for interleaved code writing. We have agent engine on prime of that that ensures compliance and governance. And then we even have the chat interface with Gemini enterprise and Gemini chat for shoppers, proper? So half of the motive I got here right here is as a result of I noticed Google as uniquely vertically built-in, and that being a power for us.
Techcrunch occasion
Boston, MA
|
June 9, 2026
It’s odd as a result of, even with all the variations between corporations, it seems like all three of the large labs are actually shut in capabilities. Is it only a race for extra intelligence, or is it extra difficult than that?
I see three boundaries. Models like Gemini Pro are tuned for uncooked intelligence. Think about writing code. You simply need the finest code you may get, doesn’t matter if it takes 45 minutes, as a result of I’ve to take care of it, I’ve to place it in manufacturing. I simply need the finest.
Then there’s this different boundary with latency. If I’m doing buyer assist and I must know the best way to apply a coverage, you want intelligence to use that coverage. Are you allowed to transact a return? Can I improve my seat on an airplane? But it doesn’t matter how proper you’re if it took 45 minutes to get the reply. So for these instances, you need the most clever product inside that latency price range, as a result of extra intelligence now not issues as soon as that particular person will get bored and hangs up the cellphone.
And then there’s this final bucket, the place someone like Reddit or Meta needs to average the total web. They have massive budgets, however they’ll’t take an enterprise danger on one thing in the event that they don’t know the way it scales. They don’t know what number of toxic posts there will probably be right now or tomorrow. So they’ve to limit their price range to a model at the highest intelligence they’ll afford, however in a scalable strategy to an infinite quantity of topics. And for that, value turns into very, crucial.
One of the issues I’ve been puzzling about is why agentic techniques are taking so lengthy to catch on. It seems like the fashions are there and I’ve seen unbelievable demos, however we’re not seeing the form of main adjustments I might have anticipated a yr in the past. What do you suppose is holding it again?
This know-how is principally two years outdated, and there’s nonetheless lots of lacking infrastructure. We don’t have patterns for auditing what the brokers are doing. We don’t have patterns for authorization of knowledge to an agent. There are these patterns which are going to require work to place into manufacturing. And manufacturing is all the time a trailing indicator of what the know-how is succesful of. So two years isn’t lengthy sufficient to see what the intelligence helps in manufacturing, and that’s the place individuals are struggling.
I feel it’s moved uniquely shortly in software program engineering as a result of it matches properly in the software program improvement lifecycle. We have a dev atmosphere during which it’s protected to interrupt issues, after which we promote from the dev atmosphere to the check atmosphere. The course of of writing code at Google requires two folks to audit that code and each affirm that it’s adequate to place Google’s model behind and provides to our clients. So now we have lots of these human-in-the-loop processes that make the implementation exceptionally low-risk. But we have to produce these patterns in different places and for different professions.
