Google Cloud on Adaptive Applications: The Next Phase of Enterprise AI

Google's own Kubernetes Engine has gone from three releases a year to shipping daily for some customers.

Daniel Newman is joined by Drew Bradstock, Senior Director of Product for Orchestration and Kubernetes at Google Cloud, for an AI Platforms, Ops, and Models Spotlight conversation at The Six Five Summit: AI Unleashed 2026.

Bradstock draws a sharp line between enterprises running flashy AI demos and the ones getting real production value: the differentiator is whether a team pushes through one full production workload, since everything gets easier after that first deployment.

He defines adaptive applications through Google Kubernetes Engine's own release cadence, which moved from three releases a year to weekly, and now daily for some clients, a shift driven by how quickly underlying models themselves change.

Bradstock points to targeted infrastructure modernization as the more effective pattern: successful customers upgrade specific infrastructure for specific new AI workloads, pairing that with the same experimentation fluidity they apply to models and applications. He also describes a shift toward workload-level benchmarking, where financial services customers test AI against actual user outcomes, like steps cut from a process, generating a clearer picture of business impact than model or hardware benchmarks alone.

Key Insights:
🔹 Bradstock says the real gap between AI experimentation and AI value is getting through one complete production workload, since teams that clear that bar find every subsequent deployment easier.
🔹 Google Kubernetes Engine has moved from three releases a year to weekly, and now daily for some clients, reflecting how much faster enterprises now expect applications to adapt as underlying AI models change.
🔹 Successful customers target infrastructure upgrades to specific new AI workloads, matching investment directly to each app's performance needs.
🔹 Bradstock describes the strongest-performing financial services customers benchmarking AI against real workload outcomes, like steps cut from a process, building a clearer picture of business impact than raw hardware or model specs provide.
🔹 Bradstock identifies budget for failure as a defining trait of successful organizations, rewarding teams for revisiting approaches that didn't work and treating failed experiments as part of the process.

Bradstock frames the choice between managed and DIY infrastructure as a decision made workload by workload, matched to each app's specific regulatory, speed, and control requirements.

Watch the full video at sixfivemedia.com, and subscribe to our YouTube channel so you never miss an episode.

Explore more sessions from Six Five Summit: AI Unleashed 2026 at sixfivemedia.com/summit.

Disclaimer: Six Five Media is for information and entertainment purposes only. Over the course of this video, we may discuss companies that are publicly traded, and we may reference their equity share prices. Nothing discussed during this webcast should be considered investment advice or a recommendation to buy or sell any security. We are not investment advisors, and you should not rely on this content as financial advice. Six Five Media collaborates with technology companies and industry leaders to produce research-driven interviews and multimedia programming for enterprise technology audiences.

‍

‍

Drew Bradstock:
If you do all DIY, it's expensive to build a whole platform. And then if you do all managed, you may not have the flexibility you want. So it's really finding what works for the workload, not one blanket statement for your entire app.

Daniel Newman: 

Hey, everyone, welcome back to the Six Five Summit. We are here in the AI Models, Ops, and Platforms track. So excited everybody can be here with us today. We're setting up for a very exciting conversation. I have Drew Bradstock joining me. He's a Senior Director of Product Orchestration and Kubernetes at Google. Drew, welcome to the 6.5 Summit. So excited to have you here.

Drew Bradstock: 

Yeah, thank you very much for having me. I'm excited too. Great conversation and one that we're practically having daily. with customers these days. So thanks again for the invite.

Daniel Newman: 

Yeah, there's a lot going on. Of course, you know, I think we've gone through these different iterations over the past few years. It was very intensive. The infrastructure built out, compute built out. You know, we've seen different constraints coming along the way. Then, of course, we've had the model wars, frontier models, open models. Do we go open source? Do we go open weight? Do we just go open weight? Do we go open source? Do we do all the above? What happens if there are no frontiers? I don't know, a lot cooking here. But what we can agree on is that AI adoption, it's moving quickly, it's accelerating. We know that companies are shifting their focus from proof of concepts and trying to deploy AI in small pockets to really wanting to take it to production scale. I mean, at this point, I think we've all seen with the most advanced models, some of those coming from Google, some of those coming from companies that Google offers in their cloud. But with the most advanced models, with the open source models available, companies are trying to orchestrate, they're trying to build platforms, they're trying to operationalize and really do things at scale that make their businesses grow.

Drew Bradstock: 

I think, honestly, it's been amazing on two parts. One, just how much better all the models are getting. Like credit to all the Frontier builders, just what you can do with them now versus six months ago is incredible. And the amount of choice. And also the constraints of different choices also adds complexity for production usage. But we really are seeing people going from playing to using production to helping their businesses grow.

Daniel Newman: 

Absolutely. And a lot of them are using lots of different models, right? And so we know, I always like the joke, and you guys are a sponsor of F1, right? But you don't necessarily need tickets.

Drew Bradstock: 

I still don't have tickets. I've asked. I've yet to get F1 tickets, despite that.

Daniel Newman: 

I asked your team, too. But you know, what's the line? You don't necessarily need to drive, you know, the F1 car to get to Starbucks to grab a coffee, right? Like there's different models for different things, right? So, but let's start with kind of how I set this thing up, right? You know, enterprise conversations that have gone from experimentation to building AI powered applications. What are you seeing on the Google side that's driving the shift and why is the application layer specifically becoming the next big focus?

Drew Bradstock: 

Yeah, it's really interesting. Obviously, you've got the large digital natives and a lot of the whale producers or the AI producers just going for it. They're deploying a product and they're running all their operations with models at this point in production. But enterprise, we're seeing a great deal of difference. A lot of them are either, hey, we can't touch this, it's too risky, we can't do anything, but we'll let one little tiger team play with it. The challenge of that is you never get in production. And what we're seeing from a lot of customers right now is we're saying, hey, look, we've got teams playing with it, but how do we make that leap into production and get business value from it? So the ones who are getting something small and going from experimentation to actually making the task of getting a real production workload are the ones way further along, because once they get through that first workload, everything's way easier versus the ones who do a whole bunch of different experimentation with lots of cool stuff, but it never goes beyond the trial. I find those are the ones who are actually then being like, well, I don't see the value of AI because the business isn't actually going to return of it. They're just getting really cool demos.

Daniel Newman: 

Are there still people that actually say that? I can't imagine playing with the newest Gemini, Anthropic, any of these models and walking away and going, yeah, I can't get any value out of that.

Drew Bradstock: 

Well, okay. So they say, yeah, this is amazing. Like even personally, my entire product org has been using all the Gemini models, do parts for a job. And initially I was like, Oh, and now it's incredible. It can pull off. The difference is they go from saying, well, I see the value, but not my environment because of X, Y security compliance. I don't have the infrastructure data is not there. And it's a lot of excuses of why they can't take that value and actually achieve it. That's what I'm still seeing almost every single day. I'm

Daniel Newman: 

That I get, you know, I went through a phase right around the beginning of the year when like the first sort of what I would call really useful models hit. And you could argue which ones, and then there were some cool toys that came out.

Drew Bradstock: 

I love all the models, right?

Daniel Newman: 

Yeah, we love them all, right? Because whether you guys are building them, you know, you're the full stack at Google. You know I remember when perplexity computer first came I'm not a developer but it was like oh gosh a mixture of experts it can build an app for you it can it can actually handle the basic improv stand up a link and build a site and I'm just saying like I built but the one thing I was that really did quickly become evident to me is like. this is the world where everyone will build an app and they will all have one user. And if you actually want to deploy something in scale, right, if you actually want to deploy something in scale, there's a lot of these concerns. And this is why I thought some of the kind of the SaaS apocalypse commentary and stuff was a little overblown, because I have a phrase, I call it the rules and rails of business, but the security, the compliance, governance, sovereignty, the actual, oh, by the way, why does SaaS even exist just to actually own, manage and maintain the software all the time? Because like, yeah, you could stand something up. but you got to keep it running. Let's talk about adaptive applications. First of all, what defines that? What is an adaptive versus maybe another app that builds? How is that differing from what I just talked about, the enterprise, where the rubber meets the road?

Drew Bradstock: 

Rather than pick on my customers or many of my SaaS partners, I'll actually look at Kubernetes. If I look at, part of my remit is only Google Kubernetes Engine, We used to release on the monthly, even before it was three times a year at the Kubernetes release, do a big push and some sub-releases. Then that became a month, then it became bi-weekly, then it became weekly for clients. Now it's daily, not for all clients. But the expectation of speed and responding to change a lot quicker has fundamentally altered our own release velocity. Now, on the outside, what we're seeing is that two things, right? One, that the models change so frequently and get better, that the old way of I release an app, I test it, I make sure it's secure, CISO signs off, deploy some hardwares there, I'm good. And that thing runs banking systems, whatever, for years and years and years, and you leave it alone other than some occasional patches. Now with the adaptive apps, because the models change, the capabilities change, people are actually updating their apps much more frequently. Because even if a model doesn't update and you don't check your app or don't look at new capabilities, it can behave quite differently, for better or for worse, especially if you're looking for something that's designed for your business. So I think the expectation from the business owners is that you've got to constantly updating. And this old static version we deploy at the bank website, or the retailer page that's there, It's not there. The business is expecting a lot more of these changes and that's pushing app teams to fundamentally alter their own views of how they do things. And that's a greater burden on these poor teams too, right? Because they just don't want the teams to be that adaptive with their old ways of doing things.

Daniel Newman: 

So that kind of leads me to the next question though, because you point out the obvious is like, I mean, we can argue the infrastructure is a multi-time, we're updating infra now at the same pace we used to update software, right? And so with more and better infra, what we're doing is building more powerful software models and applications. And you kind of just, you ended the last thought with organizations are running into the challenges of basically like, hey, a new model drop, like, you know, we just deployed software Friday and Monday, a better model has come out. Yeah, but by the way, better doesn't inherently just mean plug it in and go because, like, you know, you get some benefits and you'll miss some things. And so sometimes it's like, hey, we want to deploy the new model for this part of the application, but not like, how does that work? Like, how do our organizations overcoming this sort of barrage of possibilities that this model.

Drew Bradstock: 

Pace is creating. I think it's one of the things we're seeing work quite well is when customers look at using new infra or their new apps rather than trying to modernize everything. You know the classic migrate and modernize there instead saying hey if I'm going to get this new AI app that's going to require better performance. better responsivity for our customers, be it inference or training, they're buying new hardware for that specific bit, or looking at extending their databases to run better for this as well, because people's expectations for AI responsiveness are really high. Like, think how long you'll wait for a query that just dies out. You move on, right? And they're also expecting just the throughput, but the business doesn't want to spend a lot of money. So people are actually modernizing their infrastructure for that area of the app and looking at being as fluid with infra and experimenting as they are with the models and the apps versus I've got a whole bunch of racks, I deploy them, I'm done. That's changing right along with the way they view app development.

Daniel Newman: 

As a industry analyst that tries to track everything that's going on in tech, completely overwhelmed on a daily basis as to how, and again, it's even just trying to look at bench, as we do benchmarking too, we benchmark hardware and we benchmark models here. And increasingly, we're actually trying to get to the point where we can benchmark workloads, meaning the model itself, it's like which workload on which model on which hardware provides the highest Because in the future, again, we used to, you know, the industry, like it used to be all about like, you know, who had the most memory, who had the fastest processor. But in the end, it's like, I am a bank and I need to know for solving customer service problems within FINRA and other regulatory that I have, which model on which hardware gives me the highest outcome all the way down while protecting the data and da, da, da, da, da. And it's like, and, you know, trying to benchmark that with the rate of change is just, It's just really, really hard.

Drew Bradstock: 

I think you hit on something really important. One, it's great that you're benchmarking. I think that's incredibly important and looking not just at traditional hardware software, but extending it out. The customers I've seen, especially in financial services who have done really well in this, actually benchmark the workload they're looking at and the user experience, either their own like analysts or quants or their customer experience. And we'll repeat the test from that angle, not necessarily hardware or software. And then they're able to see, well, how much did they improve the process? How many steps did it cut? That's been the big switch. And that's those customers have just ramped up AI usage greatly because then there's, we did this work and here's the impact of the business versus we deployed a model. dot, dot, dot, and there isn't that impact. So that's a great way of doing actually looking at the workload itself.

Daniel Newman: 

It's just layering like a CSAT or an NPS on top of hardware benchmarking, meaning like they're- Another dimension, like you said at the top of this, right?

Drew Bradstock: 

It's a challenge, but I think it leads to a lot quicker AI adoption and keeps the developers honest too.

Daniel Newman: 

So you started talking about with finance, but maybe more broadly, like what's, you know, separating, because you get to talk to tons of organizations across many industries, right? Uh, we know the regulated industries bring extra complexity, but the ones with less regulation, I would say, have more scale and opportunity because they can go faster. But none of us want to spill our data or get hacked in the process by a rogue agent. Um, by the way, when's Google? I'm joking, but like when Google gonna announce the ...

Drew Bradstock: 

That’s definitely a no comment, I actually know where that question is going. I’m not touching that one. 

Daniel Newman: 

Announce the next agent has run out and break into something. I'm kidding. I'm kidding. I just, it was. It’s been a, It was a fun month leading up to the summit here of different models

Drew Bradstock:

Unless you’re in PR then it’s not so much fun, right? So. 

Daniel Newman: 

So I don't, I don't even know. It feels very, it gets a lot of attention. They say no PR is bad PR, Drew. Anyways, plead the fifth. You don't have to answer that.

Drew Bradstock: 

I'm going to plead the fifth on this one. Despite being Canadian, I'm going to plead the fifth. I'm going to leverage you as law at this point, right?

Daniel Newman: 

Going back, my point is, across the industries, what are you seeing that separates the successful orgs versus the ones that are still struggling to be autopilots?

Drew Bradstock: 

I think two major things. One, the ones that are really successful are constantly experimenting. Like even within Google, we're constantly trying to use new things, look at different models. But the customers who are doing really well in any industry are the ones that will try, and even if they get something working, revisit what didn't work before. So it's having more of a budget for failure and also recognizing when things fail and rewarding employees for it. Because often people are reticent to admit this month's worth of effort went in the garbage. That's okay. It's more an approach of, we're just going to keep on developing, we're going to change, and some will actually be a step back. That's okay. That's one thing. Two is that one size does not fit all. Meaning that some customers, and FSI and retail are great examples, where they're going to use managed offerings, like Selfishly, Gemini Enterprise. We've got all the controls around it. And you can get through it easier such you can get production with the things in place, or other offerings from other model producers too. But they want that managed offering. The other side, though, and in Blockstep, customers are doing DIY. They may want to have it fully controlled such they can experiment a lot more, and they don't want all that manage offering because, if I look at finance, some of them may want regulatory, and others, they need specific controls on things, and they're going to do it themselves. Not having a pigeonhole to just manage or say deploying with Kubernetes or building your own, and being willing to try both to figure out what meets with you, is actually working out really, really well. Otherwise, you get delayed. If you do all DIY, it's expensive to build a whole platform. And then if you do all manage, you may not have the flexibility you want. So it's really finding what works for the workload, not one blanket statement for your entire enterprise.

Daniel Newman: 

Let's bring this all together, Drew. You kind of just said something that's kind of interesting to me, because you're talking about DIY. I mentioned this earlier. I still think DIY is very difficult end-to-end. I still do think a lot of enterprises will build a it's going to be sort of a hybridized version. Like, I don't think I do see some software being deprecated, like I think , was famous for saying deprecated CRUD databases with logic. I do think some software probably gets phased or kind of becomes a feature inside. I do think some software remains critical, and I don't think it entirely goes away. How are you reading this adaptive building going forward? Because in the end, what I just don't think is that all companies are going to return back to a full end-to-end DIY circa 1990, building their own ERPs and their own CRMs. It's like, where do we land here and how much better does this get? Because it's also- that's one of the things- like, I only can use what I see now. Like, can these agents get so good that it can solve all those other issues that, you know, the compliance and sovereign and things that I think are still hard for people that are building homegrown apps? The number one thing is just keep playing.

Drew Bradstock: 

Like, that's even the advice I give my son as he's learning and he's 12. And he's like, what he's produced using these app tools. I'm like, I couldn't do that at age. Need to see us to go and learn what the code produced, and I think that's part of the problem of by coding. But what we are seeing is people are choosing to modernize what they're running on for a small set for DIY. It's more what's specific to them. Do I think they're going to police replace Salesforce or systems or HR systems with? But good ones, no. There's a lot of burden beyond, hey, we've made something quick. So I think that will really remain well. But I think you'll see a lot of people choosing to update their info to look at DIY systems for their own tasks. The things, be it a retail system, like chatbots for retail, where it's very specific based on merchant data. And there's some great examples of that, where that's added so much value, trying to use an inside vendor doesn't make sense. That's a good DIY one. And when you want to constantly adapt to the new models versus if you've got something that's doing back-office work, you want some stability; you want the greatest of the models, but you can't afford to have the churn if you've got 500 people processing loan records. You don't want the randomness. So I think you'll find that blend where the stability and the constantness will, you know, overwrite. I want the latest and greatest model every two weeks because it's going to impact your business. And that's where you see that bifurcation. And customers have to find out by really seeing what the impact of their business is and which to choose.

Daniel Newman: 

Drew, I think that's fascinating by the way, I have a 10 year old son he came in yesterday he you know he he coded out by the way on Google on Gemini he was building Avenger. Plot movies for Doomsday. He's trying to predict the, you know, the new Avengers Doomsday plot. So he came out; he built a whole- should sell as Disney, right? He built a whole plot line with it. He used, you know, AI to, to, to kind of code out and create the, all the- you can't create the actual images, you know, because all the trademarks, but the story and stuff. And I was like, you know, I was still like thumbing keyboarding at that point. So, you know, we've come a long way, and it's so exciting where it's all going. And again, you know, the only thing I could be absolutely certain of is, like, the models in a couple of years- like, Fable Five- are going to be like a toy. Like people were, you know, talking about, like, a mythos, like, you know, in two years, like this is going to be like a toy. It's like going back and using GPT-3.1, like, which was amazing at the time.

Drew Bradstock: 

When you think about it, it was incredible. And now we look at it like, oh, that's cute.

Daniel Newman: 

Yeah. And I mean, so that's the exciting part. And that's the part, like, I think people often struggle. We can think very linearly, but we struggle to think exponential. And the exponential effect here is really, really great. And so what you're saying about, you know, the power and enterprise or a business or whether it's our kids solopreneuring can build something that can be governable and can run. It's just so exciting. I really appreciate you taking the time and joining us here at the Six Five Summit to talk about this.

Drew Bradstock: 

It's my pleasure. Thanks again and have a great time.

Daniel Newman: 

Yeah, thank you, everybody, for being part of this Six Five Summit spotlight session. So much going on across AI OPS platforms and models. Probably a new model dropped since this conversation started. Stick with us more coverage.