An Interview with OpenAI President Greg Brockman: ‘There’s not going to be enough compute’

Key Context spoke with OpenAI cofounder Greg Brockman to discuss the latest model, the company’s AI infrastructure commitments and how AI agents will change everything.

OpenAI is set to go on a generational run this year.

That is my main takeaway after listening to several executives Thursday discuss the company’s newest flagship AI model release, GPT-5.5. The AI company called the model “our smartest and most intuitive to use model” yet.

“You can give GPT 5.5 a messy, multi-part task and trust it to plan, use tools, check its work, navigate through ambiguity, and keep going,” OpenAI said in a blog post.

OpenAI said the model showed significant improvements for agentic coding, broader knowledge work, scientific research, and other types of work that require more intelligence.

In an email to Nvidia employees, Jensen Huang stated GPT-5.5 was a “huge achievement” and said it was a great example of the arrival of “agentic AI” that does actual work, not just answers questions. He noted the model was trained on the chipmaker’s Blackwell GPUs.

OpenAI also said the model was better than Anthropic’s Opus 4.7 for tasks that required “reasoning and autonomy,” citing tests with senior engineers from several customers. An executive added GPT-5.5 outperforms Opus 4.7 in Terminal-Bench 2.0 and excels at long-horizon tasks.

GPT-5.5 may be the beginning of a flurry of model releases. OpenAI Chief Scientist Jakub Pachocki said on a press call with reporters to expect a faster pace going forward. “We see pretty significant improvements in the short term, extremely significant improvements in the medium term,” he said. “I would definitely expect that we will continue to see the pace of AI capabilities improvement to keep increasing. I would say the last few years have been surprisingly slow.”

Key Context spoke with OpenAI cofounder and president Greg Brockman late Thursday to discuss the latest model advances, the company’s AI infrastructure commitments, future roadmap, and the massive impending compute shortage.

Here are lightly edited highlights from our conversation:

Key Context: Your main rival, Anthropic, insinuated that OpenAI was recklessly spending on compute as recently as two months ago. But here we are with overwhelming demand and compute shortages as agentic coding has taken off. What did OpenAI see last year and how do you think about how much compute infrastructure to buy going forward?

Brockman: We’ve been very focused on where we see this technology going and the demand that we see in practice right now. And then we look at the capability that we expect six months, 12 months, and 18 months down the road and think about how much demand there will be for that.

I think we were heading to this compute-powered economy. There’s not going to be enough compute in the world to meet the demand. Where we are with software [agentic coding] is night and day before December versus after December. People are really leaning into these tools and using them to move so much faster. People say they move two to three times faster now, but they’re also working harder than ever.

The other thing to note is that within OpenAI, we have experienced compute scarcity for many years. Where every team has people whose productivity is directly tied to how much compute they have. The most intense internal conversations we have are about that allocation.

That is the world that we’re going to head to. We’ll have to answer questions like, which problems are most worthy of pouring the compute of these very large, gigawatt-scale data centers toward solving?

You’ve talked about how that type of AI coding traction will spread to other areas of knowledge work. Why are you confident that will happen?

Within OpenAI, we aspire to be the most AI-forward organization so we can live in the future and see what’s possible when you really embrace these tools. I’ve seen the adoption just over the past couple of months.

We started with the developers and then started with the people who were spending a lot of time with developers. For example, Lindsay from our comms team who was on the call earlier, she has been an early adopter of Codex for all sorts of things, using it to organize a press event, send emails to the participants and ask for people’s dietary restrictions. It just manages all of that. It took a little bit of effort to show her how to get set up. But once she was set up, she was flying.

These agents are just so capable. And we know how to make them very useful in every single domain. So we’re spending a lot of time with enterprises, a lot of time with different verticals in order to make these tools valuable. We’re seeing it in the results.

The big takeaway I took from the earlier press meeting was when Jakub said he sees “extremely significant” OpenAI model improvements in the medium term. Can you expand on that? What is he seeing in the medium term and where does he see significant improvements for AI models?

The way to look at our history is that I remember in the early days we set a goal of each year doing one thing that was previously totally impossible, totally unthinkable. Figuring out how to actually do it. That’s extremely hard. That execution, that vision, all these things have to come together. But we’ve succeeded at that. Every single year having something that was just like a new step function change. And that’s not stopping.

We have this continuous curve of improvements and that we see exactly in front of us. The highest-level answer is every single piece of the pipeline, we’re improving, we’re making more focused, we’re trying to bring it to real world utility.

One thing that we really have shifted on is not just thinking about, can we pass the benchmarks? Because at this point all the benchmarks are saturated, right? If you even want to know if you’re making progress, you need to see this stuff helping people. That’s the real metric. That’s the thing that we want.

And so I think we just continue to see step function improvements that we know how to make with all the research that’s in the pipeline that we’ve been investing in for years. They’re kind of overlapping bets that are coming to fruition. And then again, it all feeds back into the, we’re going to need a lot more compute.

Talk about your chip strategy. OpenAI uses a lot of Nvidia, but also has deals with AMD and Cerebras. You are also working on internal custom chips. How are you thinking through that, and what’s the roadmap?

Nvidia remains our closest partner. We work closely with them. I’m a huge fan of Jensen and the team. They’re incredible and do very innovative work. Nvidia is our primary compute partner.

We also use other chips, depending on the workload and the comparative advantage of those. And we have our own chip that we’ve been working on. That’s a team that is also doing incredible work. They are punching above their weight compared to other programs out there.

So the short answer is I think that we’re in a world that’s compute scarce. We need more compute. We need more chips. This is a field where it’s just beneficial for everyone to be making all these chips.

Do you still believe 8 billion GPUs is the total addressable market?

I think that every person having a dedicated GPU is a good baseline. It’s a very evocative thing to think about because we’re not on trajectory to build anything close to that. Not anytime soon. We’re talking 10 million GPUs, 50 million GPUs, maybe you could squint and get a little bit more than that. But getting to billions of GPUs, getting to one GPU per person, that is not in anyone’s cards.

There’s this concept where knowledge work will shift away from the repetitive tedious tasks we’ve have to do manually to employees all becoming orchestrators of AI agents. How do you see knowledge worker life changing as we start managing these fleets of agents? Because I imagine you already have that vision of the future playing out to some degree.

What we’re seeing is that the barrier to doing things goes down. And you start to really think about how much of your time is spent on things that are not really things you enjoy, the things that are important, the things that uniquely require your skills.

We’re seeing already with these agents that you can solve these problems or just make them 10 times easier. The tasks that people do with Codex, you can go through your email and organize your email. It can find action items that are outstanding. If you have a bunch of unread Slack DMs, it can go through those. It can reply to people.

Now, I think that interesting ways of working are going to evolve and emerge. For example, using workspace agents. If you have a feedback channel internally, where people are posting bug reports or here’s something broken in the product, you have an agent that’s in there that’s able to actually triage those. That’s able to find and bundle together duplicate reports. That is able to talk to the team that needs to fix it. And also reply to the person who posted it. All those things are happening now.

But there’s something even deeper happening, right? If you look at Codex, which is becoming the super app. And you hook it up to computer use. If you try something like Chronicle, it’s able to see what you’re doing on your screen and turn those into memories. And then you just realize that for anything that you’re thinking of that you can just ask your computer and it’ll be able to do it for you.

And then I think that also we’re going to be in a world where these agents are interacting with each other and that the humans are providing oversight and guidance. But you can’t be totally checked out of the details. You need to be accountable and be looking at the end result. Just like if you hire a contractor to build your house or work on your house. If there’s a mistake, it’s going to be on you.

Thanks for your time, Greg.

Tae’s Take: I believe GPT-5.5 shows the scaling laws are alive and well. It seems the major frontier model makers like OpenAI, Anthropic, and Google are benefiting from a new AI self-improvement feedback loop. There are going to be advances in novel synthetic data techniques, algorithms, and more compute power.

Brockman sees overwhelming demand for current and future models, meaning there won’t be enough AI compute to go around. That’s bullish for AI chipmakers like Nvidia, HBM memory suppliers like Samsung Electronics/SK Hynix/Micron, and CPU suppliers Intel and AMD. All names we have repeatedly written about.

We’re still in the early innings of an AI exponential liftoff that will accelerate for the rest of the year. Hold on to your hats.