Bill Gurley

Then they got to buy the servers. So the way this works, I think, is they sell a four-year deal to Microsoft or a four or five-year deal to OpenAI or four-year deal to Google or whatever. They expect to pay back all the CapEx, OpEx, and GPUs in three years. And so the fourth year, which is a four-year guaranteed contract, the fourth year is your profit margin, right?

And then anything you earn past the four years, that's all gravy on top. And the consensus earnings are not giving them any credit for anything after the end of those contract periods. Now, what I'll tell you is, and we've done a lot of research on this, There's still a lot of A100s in use in the world today. In fact, Jensen has talked at length about that. That's a 2020 product.

And then anything you earn past the four years, that's all gravy on top. And the consensus earnings are not giving them any credit for anything after the end of those contract periods. Now, what I'll tell you is, and we've done a lot of research on this, There's still a lot of A100s in use in the world today. In fact, Jensen has talked at length about that. That's a 2020 product.

And then anything you earn past the four years, that's all gravy on top. And the consensus earnings are not giving them any credit for anything after the end of those contract periods. Now, what I'll tell you is, and we've done a lot of research on this, There's still a lot of A100s in use in the world today. In fact, Jensen has talked at length about that. That's a 2020 product.

So we're in the fifth year and A100s are still out there being used by almost everyone that bought A100s. And then if you look at it, I think Jensen at GTC said last year that that OpenAI had just retired the V100s. That was a 2017 GPU. So that's like a seven-year lifecycle that they were using those for.

So we're in the fifth year and A100s are still out there being used by almost everyone that bought A100s. And then if you look at it, I think Jensen at GTC said last year that that OpenAI had just retired the V100s. That was a 2017 GPU. So that's like a seven-year lifecycle that they were using those for.

So we're in the fifth year and A100s are still out there being used by almost everyone that bought A100s. And then if you look at it, I think Jensen at GTC said last year that that OpenAI had just retired the V100s. That was a 2017 GPU. So that's like a seven-year lifecycle that they were using those for.

And so I think that we have a lot of comfort that at a minimum, people are going to be using these things for four years, a couple of years for training, a couple of years for inference. I've yet to hear of anybody throwing away any GPU because it doesn't have value. Remember the way CUDA works. The software that runs these GPUs, it constantly gets upgraded. It's like my Tesla, right?

And so I think that we have a lot of comfort that at a minimum, people are going to be using these things for four years, a couple of years for training, a couple of years for inference. I've yet to hear of anybody throwing away any GPU because it doesn't have value. Remember the way CUDA works. The software that runs these GPUs, it constantly gets upgraded. It's like my Tesla, right?

And so I think that we have a lot of comfort that at a minimum, people are going to be using these things for four years, a couple of years for training, a couple of years for inference. I've yet to hear of anybody throwing away any GPU because it doesn't have value. Remember the way CUDA works. The software that runs these GPUs, it constantly gets upgraded. It's like my Tesla, right?

I had an old Tesla Model S, like seven years old, but it felt like a new car because my software got updated all the time. And frankly, it still got me to the places I needed to get to. It wasn't as good as the new model I bought in December with full FSD and everything else, but it didn't feel like a really old car because the software was constantly updating.

I had an old Tesla Model S, like seven years old, but it felt like a new car because my software got updated all the time. And frankly, it still got me to the places I needed to get to. It wasn't as good as the new model I bought in December with full FSD and everything else, but it didn't feel like a really old car because the software was constantly updating.

I had an old Tesla Model S, like seven years old, but it felt like a new car because my software got updated all the time. And frankly, it still got me to the places I needed to get to. It wasn't as good as the new model I bought in December with full FSD and everything else, but it didn't feel like a really old car because the software was constantly updating.

I kind of think of that the same way for these GPUs. The GPUs are getting better every year, even though the hardware remains the same. So I'm not nearly as worried about that depreciation schedule. It seems to be a, you know, a big hit on the company and lots of people are talking about it, but they're out the door and, you know, you know, kudos to them on this big new deal today.

I kind of think of that the same way for these GPUs. The GPUs are getting better every year, even though the hardware remains the same. So I'm not nearly as worried about that depreciation schedule. It seems to be a, you know, a big hit on the company and lots of people are talking about it, but they're out the door and, you know, you know, kudos to them on this big new deal today.

I kind of think of that the same way for these GPUs. The GPUs are getting better every year, even though the hardware remains the same. So I'm not nearly as worried about that depreciation schedule. It seems to be a, you know, a big hit on the company and lots of people are talking about it, but they're out the door and, you know, you know, kudos to them on this big new deal today.

And the answer may lie somewhere in between. And like I said, I don't think they need it to be more than four in order to achieve the margins that they have. But they're also, to your point, it's a highly levered business. They got to de-lever the business. So there's a lot of things in play here with CoreWeave. That's why, again, if you look at the multiples it's trading at,

And the answer may lie somewhere in between. And like I said, I don't think they need it to be more than four in order to achieve the margins that they have. But they're also, to your point, it's a highly levered business. They got to de-lever the business. So there's a lot of things in play here with CoreWeave. That's why, again, if you look at the multiples it's trading at,

And the answer may lie somewhere in between. And like I said, I don't think they need it to be more than four in order to achieve the margins that they have. But they're also, to your point, it's a highly levered business. They got to de-lever the business. So there's a lot of things in play here with CoreWeave. That's why, again, if you look at the multiples it's trading at,

Well, I don't know what they are today, but the multiples that came public were not overly taxing from our perspective. But there's a lot of headwind for all these AI companies. I mean, you have NVIDIA trading at 19 times fully taxed earnings.