Super glad to be again on stage. It's a bit emotional because a year ago we were all here together with Eric Schmidt, with Andrew. 12 months have passed. A lot have happened. So, maybe today Andrew is the most let's say bankable company in the streets. You just did your IPO and it's fantastic to see two builders working together and walking us through today how do we go from energy infrastructure to tokens and intelligence at scale? So, maybe you want to quickly run us through your collaboration and how Cerebras is working with OpenAI to power that next phase of intelligence.
I got to start. I think maybe the way we see it is that starting in about 2025 models got smart enough to be widely used. And as these models became sort of smart enough to make a meaningful impact on jobs, it became more and more important that they be fast. And OpenAI was one of the first to recognize that and Sam called me in the summer of 25 and said we're really interested in fast inference. And we see the time is right. And so we began a collaboration that resulted in signing one of the largest deals in Silicon Valley history together. A deal closed in December 24th, north of 20 billion dollars over the next several years of compute to deliver to exactly what Sachin was talking about which was sort of a portion of the workload that is made vastly better by being faster.
Yeah, I know. I think Andrew summed it up well, which is we started seeing last fall that as users started, as customers started using AI for more and more complex tasks, as the models became capable, just like we went through what happened with search in the 2010s, right? So initially people paid for quality. Google search was just a better answer than anything else. And then Google had to optimize for latency, right? So the amount of investment Google made in optimizing for latency was what led to a lot of growth for Google itself. Because lower latency is directly correlated to search revenue for Google, for example. And we believe that AI is going to go through the same transformation. As it becomes part of our day-to-day work, first and foremost we'll always want quality and accuracy. But immediately right after comes how interactive does it feel? What is the user experience? And so latency is a very critical product ingredient for us going forward. And so that's where we started thinking about what are the options out there on the hard compute side to actually build such kind of capability. And Cerebras has solved some really hard problems getting us there. And so super thrilled with this collaboration. I think 56 that we announced is going to be available on Cerebras. It's probably the only frontier model, not probably, it is the only frontier model that is going to run at 750 tokens per second. That's an unheard of speed. That's probably an order of magnitude faster than anything else that's out there. I think it'll feel magical when you use it.
I mean that's the coolest thing on earth. Frontier intelligence at blisteringly fast speeds. I mean that's what we do this for.
I think like the core DNA of your collaboration is adoption, adoption for developers but also enterprise. Before we had a conversation on why do you see Codex as a broader knowledge work empowerment than more like a software engineering tool. Can you walk us through this session?
Yeah, I think we're seeing this internally, right? As I was showing in the data from our internal usage. Codex is the default user interface for the company now. Right? So, people don't even interact with browsers. I know many engineers who interact with the browser through Codex because Codex has computer ears. So, anything we do, we basically start with Codex. It's a new user interface. And I think that's a pretty fundamental statement to make. And so, as we looked at usage, as I said, it is going much more broadly beyond engineering, legal, go-to-market, finance, you name it. Everyone's using Codex for more and more complex tasks. And any task that has output that can be measured, and you can actually figure out how good or bad a particular output is, Codex is really good at just iterating. The AI just continuing to work through it and solving more and more complex problems. And as the models become more and more powerful, they can run longer and solve more and more complex tasks. The joke I had, a joke, HR department in OpenAI built an agent to do human reorgs. So, agents are doing reorgs of human teams now inside the company. But, reorgs, as many of you probably appreciate, it's very complex topics, right? These are not easy things to do. And we are increasingly getting AI to basically tell us kind of what's the best way to organize.
But, I think it says a lot that when you're owning like the compute topic within OpenAI that usage is being the one metric that you're looking at on a daily basis. Okay, every decision that I'm taking at the infrastructure level is going to have obviously huge implications for the end users who are the again the developers and enterprise. Maybe Andrew, can you walk us through why latency is such, why latency is basically the backbone of whatever decision for enterprise and developers and intelligence at scale.
I think we can begin with sort of the reverse. We could ask ourselves why there's no market for slow search. Right? Why is there no market for dial-up internet? You know, if your teenage daughter is being naughty and you shouldn't take away her phone, you should set it to dial-up rates. Right? That will change behavior. I think that what Sachin is describing is a world in which AI is integrated into the way we work. It is part of the fabric of how you do your job every day. And from big things to little things in corporate life. And in that world, waiting is horrible. I mean, how long will you wait for a website to resolve before you click away? 2 seconds? 3 seconds? And what if you used it constantly? Right? That part of your job. This was the tool you used to do your work. And in enterprises, whether it's a CRM, whether it's coding, whether it's in the HR organization, that we're using tools constantly. And if you ask people to wait, the experience is horrible. And you know, what has been shown again and again in every study, both academic study and business study, is that if you give people fast tools, they use them more often. They enjoy using them and they use them on harder and more interesting problems. And so as we take sort of these extraordinary models that OpenAI is making and we deliver them to a whole range of different users at real time. So they're interactive. All right. There is nothing standing between the user and the benefit from the AI. The value. The productivity. And I think that's why we're seeing this sort of across the board.
Okay, so we're living in a new era where token maxing is basically being the very popular word we see on X. So basically the more token you use, the more AI native you are. But so how do you, let's say you're a large enterprise. How can you go above the POCs, the pilots? How can you be sure that okay, Codex is going to be distributed at scale within those enterprise? What is the productivity metric that you are benchmarking? Sachin.
I think we are seeing it in company visible metrics. I mean, that's why I was showing the model releases stat. The pace of us releasing new models has now gone on to like a month. We are releasing a new model every month. And I do mean it sincerely that Codex is the primary reason that pace has quickened because we are just able to do a lot of the work a lot more faster than we used to do before because we were previously AI research was human limited fundamentally. And we are increasingly getting to the point where recursion begins to become real where AI is going to if not do the AI research itself, right? And so that's the most important productivity metric for OpenAI. How quickly do we release a frontier model, right? It's showing up. I do believe that every obviously part of the enterprise will have different metrics, but we are beginning to see it in the output of the company itself on what agents are capable of accelerating.
Yeah, I don't think you should count your tokens as a measure of how AI forward you are. I think we're building AIs to do work. You should count the productivity of the work. And if it's an extremely sort of rough measure of productivity the number of tokens you use. And I think if you remember back when the cloud emerged and we gave every engineer, we gave them access to AWS to bypass our own IT organizations because our IT organizations were so slow. We discovered that was expensive. But we also discovered that our IT organizations were in the way of productivity. Both. And that what we needed to do is we needed not to just sort of give it to everybody, but to think and measure productivity along the way. And I think each time Sachin talks about this, what they're seeing and they measured unbelievably carefully and they show it to us is that sort of these are spikes first in productivity. The token use is how you get there, but what you want to measure is sort of how fast their models come out, how much smarter each generation is. And if you're shortening the time and increasing the intelligence of the models, you're being vastly more productive. And as you guys think about deploying these agents, that's sort of the way I would recommend. I don't know if you agree, Sachin, but pick your business metric, and I think you'll find that the agents can vastly improve your business metrics.
I think like the conclusion that we got from every session over the last two days is every part of the supply chain, everything is exponential. Everything is also bottleneck. And so, how do you see the, I mean, Agent TKI Codex is going to boom in term of usage. So, what are the implication at the infrastructure level? And how do you, yeah, what is the strategy on a daily basis to work with that exponential?
