# OpenAI COO on AI Infrastructure and Enterprise Adoption

**Podcast:** Handelsblatt Today - Der Finanzpodcast mit News zu Börse, Aktien und Geldanlage
**Published:** 2026-02-11

## Transcript

Hello to some and welcome and zwar to an ausgave from Handelsblad Today.
That is the COO this ChatGPT OpenAI.
What's the biggest misunderstanding you keep hearing about AI in your conversations?
Yeah, you know, I think the biggest thing is people tend to pe we we we're so used to technology being one done phenomenons.
Meaning there's something that gets invented, and once it's invented, it barely changes and it's constantly progressing.
And it's we get on day one is the beginning of uh I think what will be the world's journey uh to live with and incorporate into our lives of systems.
And so um for us, and you know, and being so close to our research, uh it always feels very close to being in motion.
Um and uh one of the things that we spent a lot of time doing is trying to translate that rate of progress uh to people here uh who have to think about how it applies to individual people, to businesses, to governments, to education, health, things like that.
So OpenAI has announced a device, and you've talked about this in Davos.
It uh might come out uh later this year.
What does the device look like?
Well, I can't I can't say much about it, but um one of the interesting things is every time you get uh a platform shift in technology, you tend to see entirely new uh ways of engaging with computing systems.
And so um when you had the microprocessor invented, for example, um, that led to uh the the personal computing revolution.
Um obviously mobile phones uh and applications on mobile phones completely changed the way people engage uh with technology day to day.
We think AI is gonna be no different.
And so um it's an entirely new interaction paradigm for the relationship that people have with computers.
Uh and we think that there's an interesting set of devices to be built that can accentuate and accelerate uh the impact that the technology can have.
We still use smartphones by then?
You know, I I I suspect we we will.
We'll have some form of um we'll have some form of uh of of a smartphone-like device because the compression of information on a smartphone is is so uh is so critical and and I think is still essential to be able to navigate uh the complex parts of of work.
But so you will be you'll believe there will be still apps around.
You know, it's it's it's hard to say.
Um uh I I it may be that the the way that we use smartphones and the way that we use computers changes.
So uh I'll give you an example.
Um I uh uh my wife and I um actually just had a baby.
Um we have congratulations.
Thank you.
Um and we have many friends who uh who are have uh small children and um and growing families.
And um one of the things that we've uh we've seen is ChatGPT has actually become uh an entirely new way for kids to think about using computers.
So for them it's much more natural to talk to a computing system, right?
Uh they see a a smartphone or a tablet um and they start talking to it, right?
Because the expectation is that it'll talk back to them because they're used to talking to something like ChatGPT in our voice mode.
And it's the same thing that you saw when uh smartphones and tablets were originally invented, um, when people just it became natural and expected to be able to touch a computer and point at the thing you want, um, which was a progression from the keyboard and mouse-based systems that dominated most of the 1990s and the early 2000s.
And so this is just the progression of technology is our interaction models change.
Uh devices get built that I think uh are built to accompany and accelerate the the what the technology is capable of ask one question because all what you just said would also be possible to do with an app on a smartphone so why do you need a new device and i i would have to carry another device with me in future then.
Well, I you know, I think I think that AI as a technology is one of the cool things about it is it's it's capable of taking a lot of context.
So uh you think about, you know, today we we actually are minimally using um the the context windows of AI systems so the surrounding the people I meet exactly all the things that um that really kind of comprise you know everything about your life your work um you think about the amount of data you come into contact with every day is many multiples the amount uh that ever uh is really understood or captured by an AI model.
Um and so but one of the things that we've seen consistently is that the more context AI systems have the better they perform at any given task.
And so part of the design challenge we have is how do we think about to make AI systems truly useful to build a truly useful super assistant how do you help expand that context window not in terms of what the AI system uh can take but how much um it it actually is able to capture.
Will this actually be this one super assistant?
Will we only have one LLM for everything or different AIs for different use cases.
So I I think that everyone will will feel like it they'll have one one super assistant that really knows them.
I think um how that that system engages uh on any given set of tasks uh is still to be figured out.
You can imagine worlds where um they have you have LM systems that are delegating to other models um you've got uh multi-agent systems and uh protocols that are established between agent systems, for example, that can let them transact and exchange and um and and uh and engage.
Um in many ways, I think you'll have kind of an entire AI universe that in some ways mimics a lot of the structures we've built for people to interact.
Um and so what and you've seen glimpses of this already.
Um we introduced something uh called our agentic commerce protocol, um, let's uh agent systems execute commercial transactions.
Um you've seen now a lot of research.
Uh it's a good question.
We can get back to you on that.
But um, but it's the the future uh is that is the direction of travel, I think, is um AI systems that really start to integrate into the world, which as you know the world is is uh is complex, uh it's very systems-based, and um it's it's uh uh engaging uh between people and between systems is critical.
We're living in a rapidly changing world, and now more than ever, Europe needs to define its future.
Global markets have become battlegrounds, and fields like AI, biotech, digital sovereignty, security, and space exploration will mean the difference between success and failure.
These themes stand at the core of tech.
Europe's hub for innovation and collaboration.
From the 31st of May to the 2nd of June 2026, Handelsblatt, the Handelsblatt Media Group, and Schwarzdigits will welcome leading minds to Heilbronn.
Find out more at tech minus Europe.org.
Now that topic I keep hearing here in Davos, uh AI has gone from hype to infrastructure, and in some places also disappointment.
A lot of executives tell me, yeah, well, um, the pilots and themos are amazing, but uh productivity gains are not as good as we expected.
So what's the bottleneck right now?
Is it models?
Is it data, workflows, or the leadership inside companies?
Yeah, so we actually think that um right now the we're in what we call a capability overhang.
And what we mean by that is that the models are actually way more capable uh of doing productive work than uh than the environments in which they're put in.
Um so really we we actually see this partly as a systems problem and partly as a context problem.
So, what are the mistakes being made within companies?
Well, enterprises are complicated organizations.
And so I think a big part of the trend line here is going to be um how do enterprises set up their organizations and their systems to be able to um to ingest and inherit uh AI systems that are performing fundamentally new work.
Um, I think you know that's going to be really the next uh the next wave.
And I think what we're excited about into 2026 in the enterprise is you're going to start to see the beginnings now of um systems that look like uh more like teammates, um where they can make use of enterprise data in more interesting ways.
Uh they can make use of tools to solve problems, reasoning models being at the core of that.
Uh obviously, reasoning models are an entirely new paradigm of AI model.
Uh that these are systems that can think for a long amount of time.
Uh, they have the ability to experiment, to come up with ideas, try stuff, learn from feedback.
And so this is an entirely new way uh really that that software is even built to work.
Um, but that means that organizations and enterprises are gonna have to learn to adapt uh both their organization and their IT stacks and their systems to be able to get use of these uh of these types of AI.
Can you give an example of a company that has found amazing use cases that you could talk about here?
Well, there are many.
One that I love as an example is uh we work very closely with T Mobile in the United States.
Um the way that they have thought, for example, about uh customer interaction, and they're a very customer centric company.
They really do think about the user experience at every step of the company and every step of the journey.
And one of the things that we worked really closely with them on is how do we actually improve the way that customers engage with uh with T Mobile in a support capacity, in a contact center capacity.
And so we've really rethought that from first principles, meaning uh how do you actually bring language models into that workflow, make that experience much more seamless, bring more information to the customer, leveraging the data that T Mobile has in their systems.
Eventually we can start to transition that into other modalities, things like voice.
And so we're really just at the beginning of that.
And you can imagine that applying anywhere.
It can be in life sciences, it can be in retail, it can be in industrials, financial services.
Um, there's really no limit, and that's one of the cool things about this.
And you have massive amount of data, how people worldwide are using AI.
Do you see any difference between Europe and United States and Asia?
Well, I think you know, we we've seen as much enthusiasm in Europe as we've seen anywhere in the world.
Um, in Germany in particular, uh, some of the numbers are really staggering.
I think Germany is is top three uh among users, both of all uh of of many of our products uh in in uh in the world.
When it comes to number of users, number of users, number of developers, and number of businesses.
I think uh Germany comes in top three on all.
Um and we've seen just tremendous growth uh uh in in Europe and a lot of demand uh from European users and enterprises alike, uh wanting to bring AI into into their uh into their lives.
What do you uh do people use ChatGPT for the most?
What is the top three?
Uh well there it's it's very you know it's very individual.
So um every person's got a little bit of their own uh their own pattern of use.
One of the things we're particularly excited about, though, uh is health.
So we recently released ChatGPT Health.
Um you're going to talk about it definitely.
Yeah, and health is is one of the areas that we saw as um something that was really from the beginning of ChatGPT a consistent area of interest from users, um wanting to take more control of their health care.
Um, and that can mean um understanding uh things like chronic conditions that they're managing, can mean understand uh understanding the healthcare environment that they live in, so um whether that's uh things like insurance and benefits, um, care pro you know, providing care for others.
Uh there are many people who are caretakers for other people.
So is health one of the top three things people use ChatGPT for?
It's definitely up there.
Yeah.
Okay.
What else?
Um, well, there's there's a bunch of other things we see.
Learning is a is a is a critical use case.
That was really one of the original use cases, I think, for ChatGPT.
We saw adoption among students um very early on, and we've worked really closely with uh educational institutions.
Education is now one of our largest segments, actually.
Um, and we've worked really closely with uh both in in the United States and elsewhere in the world, um, organizations that are um focused on how AI adoption in schools and in learning uh can really impact outcomes.
Talk a little bit about how the health um focused ChatGPT will look like in future.
Will you help users directly?
Will you help health organizations?
Yeah, yeah.
So right now we're really focused on individual use within ChatGPT.
So my personal doctor, exactly.
So you'd be able to link your medical records.
Um we have models that are specialized uh for uh uh giving um information about health.
Um we outlink to resources that are related to health, whether those are medical or otherwise.
Um the models are really good at doing things like ingesting lab results and helping you interpret and understand lab results.
So we're just starting to scratch the surface of what this experience looks like.
Obviously, critical that we do this in a privacy-preserving way.
So that's been a big focus of how we've architected the health uh feature in ChatGPT.
Um, and there's a lot more to come on this.
And um, when will this start in Europe?
Uh so we I think we we just we're just in our in our alpha now.
Um I I think anyone that that's that signs up can get on a wait list, and our hope is to start letting people in fairly soon.
Okay, I tried, but it wasn't possible last weekend, so maybe it's we can maybe I can try and try it.
Try and help you out.
Very interesting point is where do you draw the line between helpful information and actual medical advice?
Well, I think you know, one of the ways that we uh we really see language models evolving is that the in the language models themselves are not necessarily the the sources of information, um, especially as you enter uh this new age of reasoning models that are inherently more agentic.
One of the things that they're able to do is use tools to help you solve problems.
And so the same way that a person would go look up sources of information from databases or from uh you know uh sources of truth, systems of record, um, you know, qualified sources, um, AI models are capable now of doing the same exact thing.
And so one of the things we really want to do and is emphasized in our health work and elsewhere is being able to help direct people toward the places that uh that are you know those systems of of record and sources of truth and making sure people have access to resources.
Okay, but it's not medical advice, it's only it's not medical advice, yeah.
Exactly.
Okay, will this become a business model at some point?
Well, you know, I think the way we look at it is uh our uh it comes back to our mission, right?
Um we really see broad benefit and access as critical to the work we do.
And so our view is if we build products that people want to use and AI systems that are useful to people, we'll figure out the business model later.
But um the core of it is really just A, understanding how people are using AI, and then B is uh making sure that it's meeting people where they're at.
Uh and that's gonna be different for everyone.
In another field, you have changed your business model, uh, let's say so.
Um I find that quite interesting because Sam Altman said in October 2024, I kind of think of ads as uh last resort for us as a business model.
Now you implemented ads in your business model.
So what changed since then?
Well, it's a few things.
I would say one is you know, we're really early in um in in ads as a uh as a format and a uh a business at all.
Um and so part of it is just gonna be experimentation for a long time to make sure we get it right.
Um that means respecting user privacy, it means uh making sure that we're really clear uh and upholding the principles that we communicate around ads, which we published.
Um so I expect we'll be in that stage here for for a while.
Um but I think it comes back to access.
So one of the things that we want to make sure is that anyone in the world, anywhere with an internet connection can have access to a really high quality and powerful AI system.
Um if you look at the direction of travel for AI systems, they are becoming much more compute intensive.
Um, in part that's because of what I mentioned is this new reasoning paradigm where you have models that now can go off and think, solve hard problems, use tools.
Um, and we've we're starting to see that appear in uh in in in uh in consumer use very consistently.
So um, you know, there's still a gap there.
We still see uh a fairly wide gap, and that's what we call this capability overhang of the power users uh using AI seven times more.
But subscribers don't see ads.
Then the average subscribers don't see ads.
And won't see ads.
Neither will businesses.
That's yeah, yeah, correct.
So um, you know, you can choose to opt out of those if you if you want.
But um, I think what we heard from users consistently is if I can get access to uh to more powerful systems that are capable of doing more for me um you know it's it's that's that's the priority will ads become a core part of ChatGPT uh again we're still experimenting with with how this will work how do you find out if people find ChatGPT less useful when they see ads how can you monitor that?
Uh well you know we we obviously can can see can see it in data um we we talk to users but um I think the the important thing here is that we we really want to make sure that we're thoughtful about the implementation of this and I think you know ultimately the uh the way that we ideally the way we can do this is is that they're helpful to people right um a lot of people use ChatGPT in search of things it was one of the inspirations for why we launched uh commerce in ChatGPT in shopping um is because people use ChatGPT as an assistant to help them find things that they need in their life.
And so in some sense, you know uh this is actually a way for us to to possibly add to the experience if we do it thoughtfully okay.
Not so many look for ads I guess but let's see do you think we are in an AI bubble already?
Uh no I mean, I think, you know, if you look consistently at uh at how uh we've been able to um invest in compute, um a we've we've never been on the right side of the demand curve there.
We've always felt uh that we could use many multiples more compute than we have.
Um and then two is the the the kind of geometric increase in the rate of of growth of users, and then also the rate of consumption for each individual user.
Sure, but also you need an ARR to finance all of that, right?
Yes, and we've seen consistent ARR growth too.
So what we've seen pattern-wise is every dollar that we've been able to invest in compute has generally produced about three dollars of revenue for us.
So we think that's a pretty good return.
Um we expect that could increase over time.
And again, we're gonna invest aggressively ahead of what we see as the trend line.
So no bubble.
Uh I I don't right now see any any reason to think that.
But so what would be a clear sign um that the bubble narrative is right?
A key sign that the bubble narrative is right.
Um, you know, I I think uh like look, there's gonna be aggressive investment, and I think in some sense that's that's a feature, not a bug.
Um what it means is investors recognize the opportunity um and capital floods into into the space.
Not not every company is gonna work out.
And so I think we just, you know, there's there's we have to be prepared for the fact that this is a dynamic environment uh and there's a lot of people trying a lot of things.
Some of those things are highly capital intensive, and so they're attracting investment, but um I think we can focus on what we can focus on, and that's building great models and great products for users.
Let's go to the big promise.
OpenAI started as a research lab with a goal of AGI, and Sam Altman has said in the past that AGI could be closer than many expect.
However, critics and that is a discussion here in Davos as well, um, argue the opposite, that progress is slowing, jumps are getting smaller, scaling is delivering less.
Who's right?
Well, we we see the the the trend line as um as more exponential than not.
I would say we still feel um that the mod the rate of model progress is uh is extraordinarily high.
Um we still see uh leaps in capabilities that we measure as consistent with the trend line that we would expect uh in a world where we had a uh accelerating capabilities.
Can you give me uh one concrete capability that felt impossible, let's say a year ago and is now uh reliably working?
Of course.
So software engineering is a really interesting category for this.
That entire field, in my opinion, is uh is about to change.
Um if you spend time in San Francisco and kind of the epicenter of technology, the way that uh individual software engineers are using AI uh is fundamentally different than anywhere else in the world.
Um AI systems like Codex, for example, where you can now delegate fairly complex software engineering tasks to an agent that will go off and spend as much as a few hours thinking about how to solve that task, writing code, uh trying that code, um, writing unit tests, debugging, um, doing everything that a software engineer would normally go through and do, and can now do it fully autonomously only by enabling access to a code base.
That's a transformative change in how software engineering works as a practice.
And the engineers that I work with, who are uh especially at OpenAI, um, who work with these systems every day are now just fundamentally thinking differently about uh where they spend their time and and and the entire discipline of software engineering.
But this has changed for quite some time now.
Is that really a difference to one year ago?
Oh, definitely.
Yeah.
We, I mean, even a year ago, that that was it was it wasn't conceivable that you could have an AI system that could think and reason about a complex software engineering task for five, six hours, something like that.
And I expect we'll be back here in a year and talking about AI systems that are reasoning for days.
So, so um what is your definition of AGI as a benchmark?
Yeah, so we think about it as systems that can do most economically valuable work.
Um I think maybe one way to think about that is an AI system that can effectively use a computer to do anything that a human could do at a high level of proficiency.
Um I think we still have a ways to go to get there.
I think there's still some pieces that are missing, but um, I think we're making really good progress in that direction.
And how long will it take that we see that moment?
Very hard to say.
I I've sworn off making predictions, but uh, you know, this field, I've I've been doing this now for for a while.
I've been at OpenAI almost eight years.
Um, and one of the things that's been consistent is progress is surprising.
I think things that feel like they should be two, three, four years off, it's not uncommon that they're more like six, twelve, eighteen months off.
And so we don't see any reason why that that will be different uh as as we look out on the horizon, but again, it's very hard to predict.
How how will the world change the day after AGI came to our everyday life?
I you know, I think I think people will wake up.
Um, I think it will feel it will feel very much the same.
Um, one of the interesting things that we see is um, you know, despite the there's a great saying that I that I invoke now more and more frequently which is the the future's here it's just not evenly distributed yet.
And um I don't know if that was Arthur Clark or who who said that but um uh it really is apropos for for AI and how we see the trend line here and is is why we've been here at Davos talking a lot about this capability overhang where you now have a a very small group of people whose whose lives have been fundamentally upended by these types of systems.
Their workflows have changed they manage their life differently they set and track goals differently and then you have a lot of people who are still you know 99% of people who are still using one percent of of the capability of the AI.
But over time I think you're gonna see that start that gap start to close.
So nothing will change overnight I I don't know that anything will change overnight.
I think I think there's an ingestion process uh that that the world has to go through to absorb the technology one thing you uh most looking forward to the day after AJI happened what is changing well if it can solve email that would be nice um or Slack that would be great but uh but I would I don't believe this will ever happen.
Um, I would say that the the biggest thing is um anyone with really high agency, I think is going to be in a great position.
I think if you have ideas, if you're creative, if there are things that you want to see exist in the world, whether that's your personal life or in your community, um, you you now really have no reason why you can't go solve that problem.
If you want to start a business, um, if you uh if you want to uh accelerate uh health outcomes, there's there's so much opportunity.
And in some sense, part of the challenge we have at OpenAI is trying the work that we do uh into that opportunity, the realization of that opportunity.
And so it's why we lean in heavily to the community.
Um, that can be both from a political perspective and you know, with the work we do with countries, can be all the way down to the level of individuals, um, students, uh, creatives.
Um, all of these are really really important parts of the ecosystem, and we can't do it alone.
So yeah.
Brad, thank you so much.
Thank you.