AWS Summit DC 2026 | Felipe Millon, OpenAI & Garrett Berntsen, Accenture Federal

Clips
More from AWS Summit DC 2026

Felipe Millon

Head of Government & Go to Market

OpenAI

Garrett Berntsen

Chief AI Officer

Accenture Federal

play_circle_outline ChatGPT Hits 900 Million Weekly Users as GPT‑5.5/5.6 Launch on AWS GovCloud with DISA Provisional Authorization

play_circle_outline Bringing OpenAI to Government: GPT-5.5/5.6 on AWS Bedrock with DISA Authorization for Mission‑Sensitive Data

play_circle_outline Code-Writing Models for Faster SDLC and Modernization: Daybreak-Enabled Vulnerability Detection, Patching, and Red/Purple Teaming

play_circle_outline Daybreak program: trusted access to GPT-5.5-Cyber for government red teaming and penetration testing

play_circle_outline Securing Mission Data Through Cultural AI Change: Leadership, Upskilling, Hands-On Use and FedRAMP/GovCloud Access

play_circle_outline From Leadership to Agents: Driving Adoption Through Hands‑On Experimentation and Codex Orchestration

play_circle_outline Tri-party collaboration—OpenAI, AWS, Accenture—building mission-focused AI solutions and integrations

play_circle_outline Knowledge Graphs Powering Mission-Critical Use Cases: Cancer Research, Fraud Detection, Retiree Benefits via Accenture

Info
Transcript

Felipe Millon, OpenAI & Garrett Berntsen, Accenture Federal

Felipe Millon

Head of Government & Go to Market OpenAI

Garrett Berntsen

Chief AI Officer Accenture Federal

In this interview from AWS Summit Washington, D.C. 2026, Felipe Millon, head of government go-to-market at OpenAI, joins Garrett Berntsen, chief AI officer of Accenture Federal Services, to talk with theCUBE + NYSE Wired's Gemma Allen about how the public sector is moving from AI experimentation to mission-ready execution on AWS. Millon details how OpenAI's models — including GPT-5.5, now available on Amazon Bedrock in government regions — are gaining the accreditations needed to unlock their use in mission-sensitive applications, highlighted by a recent prov... Read more

explore Keep Exploring

What is OpenAI's current user adoption, and how is it making its models available to government customers (including availability on AWS Bedrock/GovCloud and any relevant authorizations)? add

How is OpenAI enabling government customers to access and use its models (for example via AWS Bedrock and government cloud regions)? add

What are the most important use cases for these AI models—particularly in software development and cybersecurity—and how do you respond to concerns about their security, access, and the relative safety of different frontier models? add

How is OpenAI addressing the cybersecurity risks posed by increasingly capable AI models, including who gets access to cyber-capable models and how organizations are supported to harden their systems? add

How can government workers gain practical experience with AI tools like ChatGPT while addressing concerns about data use, authorization, and security? add

How can government organizations and their employees effectively learn to adopt and use AI tools? add

How can AI be applied within government agencies to solve mission problems, and how should technology providers collaborate with partners (like AFS) and agency mission owners to implement those solutions? add

How can AI providers partner with government agencies and partners (like AFS) to apply AI to mission problems efficiently—what approaches, acquisitions/contracting models, and near-term (next 6–12 months) actions are needed to deliver outcomes? add

bolt Powered by CUBE AI

Felipe Millon, OpenAI & Garrett Berntsen, Accenture Federal

search

Gemma Allen

>> Welcome back to The Cube's live coverage here on the ground at AWS Summit in Washington, D.C., 2026. We are bringing you all the live action and one thing is explicitly clear and that is that the conversation around AI has moved from what can AI do for us to what AI is doing for us and how fast across government and public sector organizations supported by AWS. Joining me now for a conversation on exactly that are two folks from some big branded names in this space, Felipe Millon from OpenAI and Garrett Berntsen from Accenture. Welcome folks.

Garrett Berntsen

>> Thank you.

Felipe Millon

>> Thank you for having us.

Garrett Berntsen

>> Thanks for having us.

Gemma Allen

>> So let's get straight into it. We've had a lot of folks on the show talking about Bedrock, talking about orchestration, bringing workflows to life using AI. OpenAI is kind of synonymous with AI, right? A lot of us just think of ChatGPT. It's like almost we think about Coca-Cola or something. Talk about what's happening here from your perspective at a government level, Felipe. Bring it to life for us. I mean, huge time for you guys, huge opportunity.

Felipe Millon

>> Absolutely. So what we're seeing is there's clearly a massive adoption of AI. People are starting to use it. We have shared publicly, we have well over 900 million weekly active users of ChatGPT. That means 10% of the world's population used our AI tools in the last seven days. But people are able to use it themselves personally, able to use all sorts of different things. But in the government side, what we've struggled with for some time is how do we make that technology available to them where their mission is, where their data sits, where it resides. And a lot of that time that means here where we are with AWS is how do we bring it to the AWS cloud and to the AWS government regions such that government customers have that technology, the same technology they have at home, that they can really use in their mission. And so that's what we've been really focused on and super excited that just a couple months ago we announced that OpenAI models, including our GPT-5.5 model and coming soon our GPT-5.6 models, are coming to Bedrock and 5.5 available today on Bedrock, including in the government region. We have our GPT-4 model, 5.4. And the model recently, actually just last week, got a provisional authorization from DISA. So it's a very bureaucratic thing, but really important because that means the government can actually now start using this technology with their most mission-sensitive applications.

Gemma Allen

>> So let's talk about bureaucracy on technology for a second, Garrett.

Garrett Berntsen

>> Sure. Yeah.

Gemma Allen

>> You're at Accenture.

Garrett Berntsen

>> Yeah.

Gemma Allen

>> Accenture has been in government helping government work for a long time, right. Globally too. I actually did a lot of work with Accenture in Dublin 20 years ago.

Garrett Berntsen

>> Global company, for sure.

Gemma Allen

>> Absolutely. I mean, things are changing very, very quickly. Expectations are changing. Every organization now, whether it's a small cap company, a government organization, or an enterprise, under this pressure to get ahead of AI, to use it, or to fall behind, right?

Garrett Berntsen

>> Yep.

Gemma Allen

>> Talk about what you're seeing from the perspective of the workflows and the sows you have at a government level. I'm sure huge.

Garrett Berntsen

>> Yeah. So I talk with clients all the time and all of them are saying a lot of the same stuff you just said right now. It's like, we want to move faster. How do we move faster? And the answer is we need to use AI to do a lot of the work that historically was done with large teams. And so we are doing things internally to upskill everyone so that we're using the right tools to do things faster and we're partnerships like with OpenAI to get early access to training to go-to market collaboration so that we can meet our clients where they are. They say, "Hey, we want to move faster. We don't want to be on our back foot." We want to say, "No, here is how you go faster. These tools are ready. They're accessible. Let's turn them on." Because I've spent time in government actually. We were talking before we got kicked off here. Even some of the things you've done, which will get us there to access, there still will be folks in government who say, "Do I really have access?" And we have to say, "Yes, you do. Here's how," and actually guide a lot of our clients through the process to turn it on. Someone's got to turn it on in the system to plug in these models to do the work faster. So that's where we're here to help.

Gemma Allen

>> Let's talk for a second about a typical government environment because it is uniquely different, right.

Garrett Berntsen

>> Yeah.

Gemma Allen

>> It is intentionally secure. It is not the same as a commercial environment per se. It has a whole load of nuances.

Garrett Berntsen

>> Totally.

Gemma Allen

>> So when we think about what you described, Felipe, which is bringing AI to the data, making sure that OpenAI can visit it where it lives to optimize it, it's a different workflow. It's a different challenge, right?

Garrett Berntsen

>> Totally.

Gemma Allen

>> Talk about what you're seeing. Give me some use cases at a government level.

Garrett Berntsen

>> Do you want me to start? Yeah. So I mean, this is the challenge is it's such a wide-ranging technology. I could literally give you a thousand different use cases, but I'll hit on a couple. The first is just the most obvious, which is software development lifecycle. I mean, these models are exceptional at writing code. They are just exceptional at it. And so a thing that you'd take a massive code base of a million lines of code, that would take a team a year. I mean, it could take longer than that to try to think through all the dependencies and how do we modernize it? How do we transform it? And we're really seeing an incredible acceleration of the ability to take legacy code and actually make it modern and do it faster. But that also ties to a second really important use case, which is interrelated, which is cyber. And there's been a lot of news around cyber and how these models are so good at that, identifying vulnerabilities. What we're really trying to do is say to our clients, "Yes, you can identify the vulnerabilities, but these models can also close these vulnerabilities. They can help you get ahead of finding them, repairing them, and actually having more secure systems." So those are just two really big ones. Again, there's a million interesting use cases around business systems, and productivity, and all of that stuff, but we are a large integrator who does a lot of code development and a lot of cyber, so those are really important to us.

Gemma Allen

>> Felipe, let's talk about security for a second because there's narratives out there around which frontier model is more secure, which isn't, and I feel like it's like a ping pong machine of enthusiasm. People's enthusiasm wane and grow in different directions, but really security is extremely important. Mythos obviously created a huge narrative, a little bit of fear, inertia around what's possible. What is your response to that when you're talking to some prospective customers or government agencies? How do you respond to the security question from an OpenAI perspective?

Felipe Millon

>> Yeah. I think we strongly believe in the cyber capabilities of these models and we actually have a model that we have made available with a trusted access program we call Daybreak and that is our GPT-5.5 cyber model. So we actually have a cyber version of a model. We do not make it generally available, the average user cannot access, but we are making it available to trusted red teamers, specifically within government and at large institutions like financial institutions, so that they can use this technology to do pentesting of red teaming on their own systems. And partially this goes because we think the U.S. has a lead right now in these models. So the U.S. AI labs, we've been able to develop these models. There is an open way ecosystem, mostly out of China, they are nine months to a year behind us. So this technology will become available. And so I think the biggest thing we can do right now as a country and as leaders of enterprise organizations is ensure that your systems are hardened to potentially cyber capable models that are more broadly available. And so what we are doing under our Daybreak program is providing that trusted access to organizations, to government agencies, both to the federal level, state and local level, folks that are operating critical infrastructure within the United States to ensure that their systems are hardened before more cyber capable models are more broadly available.

Gemma Allen

>> Let's talk about something that I think is very important when we think about public sector and government and that is that culture is as important as technology, right. That adoption curve that we all know and love, it's extremely important. And when you mention red teaming or purple teaming, there's been process in place for a long time, similar with patching, right. Some of the narrative around the Mythos revelation was it could take a hundred days to build a patch that's needed for some of these operating systems. People have certain inertias around this. How do you think about that side of it, about the culture, the workflow, the ownership, and how quickly that needs to change to keep up?

Garrett Berntsen

>> Yeah, I can start. It's certainly a hard problem, but I think what we have to remember is a lot of these people work in government, they are passionate about the mission. They are. They care about the mission. And so when you frame it that way, when you frame it as, actually, honestly how you just framed it, we don't have a lot of time and we want to run secure systems because we care about national security, we care about people's privacy, all of that, when you frame it around that, from the interactions I'm having, people are saying, "Okay, I'm willing to like give this a try." You have to try to get out the noise of, "You don't need your people anymore," and just say, "No, this is going to make your people a lot better and your system's a lot more efficient and secure. And you're going to be able to have better impact for the public, or for national security, or all of the missions of government." When we frame it that way, I mean the people I talk to are like, "I get it. Let's do it."

Felipe Millon

>> I think what we also need to see in government orgs especially is the people who just try it and use the technology can get it. There's an expression I use a lot, which is you unfortunately do not get fit by reading about working out. You have to go to the gym.

Garrett Berntsen

>> I wish.

Gemma Allen

>> I tried.

Felipe Millon

>> Yeah.

Garrett Berntsen

>> That'd be great.

Felipe Millon

>> And the same is true in AI. The only way to learn AI is by using AI and using these frontier capabilities. So the number one thing is take them and apply them to your mission. And so that's why it's so critical that we make them available in the right accredited ways. So there's been a lot of fear, uncertainty, and doubt of am I allowed to use this data? Can I put this data into ChatGPT or Codex, or what does that actually look like? Can I take this document? Can I use it? And so by making it available, by getting the authorizations, we remove those barriers of trying it and just saying people, "Go try it." And they'll try it as a toy. Try it with a real problem that you really have that's something you're actually working on. And that is really what allows you to get that frontier experience. And that's really what I would push anyone listening to is these tools exist. A lot of times I work at OpenAI and people come up to me and say like, "Oh, could ChatGPT do this or can it do that?" And I normally tell people, "That is a very bad question to ask me. ChatGPT exists. There's a free version available."

Garrett Berntsen

>> Go to try.

Felipe Millon

>> "Try it and then tell me whether it worked or not." So how do you get in that mind shift of ... We at OpenAI and other labs are building these incredible capable models. I am not an expert in the National Cancer Institute within NIH's processes on how they are developing and processing grants or whatever that might be doing, but those individuals working there own that mission and that's what they know well. But what we need to do is then apply this technology to the mission, to the areas that they know well, and trying to do it in a way that can do even more. And so we're getting so much more advanced beyond just ChatGPT to what we're doing with Codex and the Jentic workflows that there's so much more that these models can now do things as well. We actually published a report last week at OpenAI about how we're using AI internally and 98% of tokens at OpenAI are generated via Codex today. So actually within OpenAI, we're not even using ChatGPT anymore. It's really the minority. We've almost moved entirely ... And all of our teams. So Codex has this idea of that it is a developer tool, but we're seeing our recruiting teams, our HR teams, our finance teams are the biggest adopters. And because what we're allowed by putting it into the Codex harness is these models can now start doing things. A very simple example I'll give is yesterday I was talking with a coworker and we were here at the AWS summit and there was a reception in the evening and I was like, "Oh, did you register for this accept reception?" And she's like, "Oh, I didn't get the link. Can you send it to me?" I did a quick look. I couldn't find it, but then I just opened up my phone, remoted into Codex and said, "Hey, Codex. I'm looking for the federal civilian happy hour this evening. Can you go ahead and find the link for me and send it to Lauren?" And that's how I did a quick voice dictation, walked away, Codex in the background ran, searched for it, found the Slack message, identified this was the right one, knew I meant Lauren Oliphant on my team, and then sent her the Slack message automatically. And so rather than the back and forth, now it's able to actually just do things. And that's a very, very simple example. There are so many more amazing examples that we're seeing of what these tools can do.

Garrett Berntsen

>> I think, if I can, one other thing that we talk a lot about internal to the company about the importance of leadership leaning in on this stuff because leadership engaging with these tools, it's worth more than any training just because people do what they see. So my CEO, Ron Ash, I'm his chief AI officer, I'm in a meeting and he starts talking about how he created an OpenClaw to like do his travel. And I'm sitting there, this is pretty recent in the OpenClaw, that month or whatever it was where it just blew up, I had not done that yet. I'd not played around with it and I'm sitting there, I'm embarrassed. He wasn't calling me out.

Gemma Allen

>> There was a long list for Mac Minis, right?

Garrett Berntsen

>> Yeah. But he didn't call me out or anything. But I sat there thinking my boss is doing this, he's playing around with it. I have to figure it out. I go home. I put time in the gym. I stand one up. That weekend my wife's like, "What are you doing?" I was like, "This is incredible. I'm sending things on Telegram." It's doing actions for me. So you do have to show. You can't just mandate, you have to show people, and then reward people, and give them praise, and all that important stuff, and show that it's useful to the company. And that's the mode we try to take.

Gemma Allen

>> Let's talk about the future. Let's talk specifically as it relates to AWS and Bedrock. So FedRAMP's a great win because it unlocks a huge amount of opportunities. Any restriction or concern, they have a throat to choke now, let's be honest. So what does it mean from the perspective of this partnership? Consultancy is an interesting space too right now. There's a lot of noise. Like I told you, I'm a consultant myself. I get it. There's a lot of opportunity. Sometimes it can be somewhat distracting because there's so many options out there. How do you guys think about the next even six to 12 months from the prospective of AWS, Accenture, and OpenAI?

Felipe Millon

>> Yeah. I mean, there is so much opportunity and so ways to do different. Actually, I'll give Garrett, Ron, and the AFS team a ton of credit where they're really revolutionizing themselves to say, "All right, well, how do we embrace it and how do we become more efficient?" And there's so much opportunity because in government there is a lot of contracting and supports that support these missions. There's one aspect of it that we can support with a first party product and we have a lot of deployments, whether it be ChatGPT enterprise or our Codex product, out there. But where I think a lot of the mission is going to happen is how do we now automate this entire process and bring agents to that process and do all that? And the reality is we need to work with the people who understand the mission the most. AFS is a great partner. They've been there. A lot of this runs on AWS, and there are tools, and that's where the data lives, and excited. So I think the tri-party collaboration is going to be incredible ways to really get to the mission and say, "How do we solve the important mission problems?" So again, the end goal is not applying AI for AI's sake. It is applying AI because it'll help the National Cancer Institute come up with better treatments for cancer and because it'll help the treasury figure out how to better find financial fraud and how do we quickly do that so that we as taxpayers benefit. It'll help the office of personnel management help bring retiree benefits to federal retirees faster in a more automated way. So how do we get to those missions? And a lot of it is what we have is we have models and tools and capability, but there is this bridge that needs to gap because what the mission owners need, what those folks at OPM or HHS Treasury need is they need their mission solved. They need that accomplished. They need partners, they'll bring it. And that's where I think-

Gemma Allen

>> And just to interrupt before you answer, Garrett.

Garrett Berntsen

>> No, please.

Gemma Allen

>> Because I want you to answer this in your financier. They also need context, right? They need the knowledge graph behind this and context is huge, especially in industries like government where people have held roles and responsibilities for a long time and a lot of data lives here in the head. I should say that because people could listen to this on audio. How do you think about that in terms of the next six to 12 months?

Garrett Berntsen

>> So I think, honestly, Felipe hit on something that we are all in on and you're seeing this already in government acquisitions, you're seeing executive orders and stuff like that, which is outcomes-based acquisitions and contracts. Without naming any competitors, we are the most, amongst our competitors, outcomes managed service and fixed price work.

Gemma Allen

>> Please name them.

Garrett Berntsen

>> That is our bread and butter. That is where we're good. We don't want an economic incentive to add 10 people to the contract. We don't want that. We want an economic incentive to deliver the most exceptional result as efficiently as humanly possible and as fast as possible. There's two things happening at the same time. People are getting a lot more efficient because of these models and more of the work needs to move to outcomes based. There's a lot, but increasingly we want to see more of that because we think it just aligns the incentives and it will allow us to move a lot faster. So it also allows us to take risk. We can say, "Hey, listen, we don't have perfect requirements on this issue. We're not going to go into a discovery sprint for nine months to figure out the requirements. We will just make a bet that we can get this done fast, we'll take some risk. And if we don't deliver, it's a fixed price. We're going to eat the cost." And so we're willing to do more of that and our clients love it. That's why they come back.

Gemma Allen

>> Best PowerPoint promises sold from Accenture. I love it. Well, listen, I think what you said there, Garrett, is important because the reality is conviction only comes with outcome. And we're in a space right now where everyone needs to see the proof.

Felipe Millon

>> The last thing I'll add is it is changing so fast. And again, you have to be able to use this technology to stay on the cutting edge. What someone experiences of what these models could do even six months ago is already out of date. And so the biggest thing is how do you stay on the frontier? And the number one thing is access and use. Provide access, provide the right authorizations and then start using it and start trying to apply it to the areas that you know the best.

Gemma Allen

>> Well, it's certainly a very interesting and optimistic time and I hope for the sake of all of our tax paying money that you guys could do some wonderful work in government. So Garrett, Felipe, thanks much for joining me on The Cube.

Felipe Millon

>> Thank you very much.

Garrett Berntsen

>> Thanks for having us.

Gemma Allen

>> I'm Gemma Allen here live at The Cube at AWS Summit 2026 in Washington, D.C. We are having some great conversations with some true innovators and builders building this next paradigm for AWS. Don't go anywhere.