🤖 Moore's law never ended - it moved to GPUs and expert systems go next level

“the list price of an H100 is about 30,000, but the street price, so it's very hard to get it. So if you go to eBay and try to buy an H100, it's like 40 or 45,000.”

“We have a GPU shortage, and it's probably not going to get better for a year or two, if that.”

“My God, these people that put the money in the seed round should have just bought NVIDIA. Buy some call options so you can make more money.”

“Certainly the most obvious is demand for chips because all this infrastructure is being built out.”

“I don't think there's any problems with NVIDIA. I think they're going to continue to perform. I think they'll beat their numbers for the balance of the year.”

“I've told this story before, but I pinged those guys and I was basically like, I just wanna meet the team that built the TPU. Long story short, a year later, I put them in business. And we've been building silicon for this moment for years.”

“Look, I think as performance improves, as cost declines, like any economic model, there's a pretty nonlinear relationship with demand. So we'll find new ways to apply this technology. I think the demand is only going to go nonlinear.”

“So that could flip where the demand could actually decrease because of software efficiencies or hitting some benchmark that's reasonable enough.”

“So I think what I'm waiting for, Friedberg, is like in the next two quarters, if AMD, Facebook, Google, Microsoft and Amazon don't announce something substantive, there's a very good chance that Nvidia runs away with this.”

“But when you look at the valuation of the business, they are currently trading at 70 times the next 12 months EBITDA. So how do you think about valuation, even at a trillion dollar market cap, at a trillion dollar market cap, they're trading at 70 times next 12 months EBITDA.”

“Yeah, Nvidia has basically joined the trillion-dollar club now in terms of market cap companies. It's really amazing. I mean, I'm kind of kicking myself because this was the easiest buy ever.”

“And I think this is the great renewal for America, another amazing American company. It's only three decades old. It's best years, it's best decades are in front of it, obviously.”

“Well, I mean, I was talking with the CEO and CFO of a major data center REIT, and they shared with me that they're seeing more demand in the last couple of months than they've seen in the prior 10 years.”

“And six months from now when you're compounding at 48 or 72 hours is like 10 to 12 years in other technology solutions.”

“I think it's at the silicon layer, because you need to really re-architect how compute will actually work in a world of all of these models. Those folks will get paid. If you look at AMD and Nvidia, they've been getting paid for years.”

“The ability to run this at scale is going to happen because we're getting better and better at creating silicon that specializes in doing things in a massively parallelized way and the cost of energy at the same time is getting cheaper and cheaper along with it.”

“Well, we were building silicon for machine learning. That's different.”

“We sell a lot of machine learning hardware into this market. The biggest buyers are the US government and these ultra high frequency trading organizations.”

“One is that the marginal cost of energy goes to zero, and the second is that the marginal cost of compute goes to zero. And the second one is really about shifting compute to more parallelism on GPUs and ASICs and FPGAs. But that's why all of this stuff is possible.”

“So this is just an observation that I think that we are continuing to compound knowledge and intelligence effectively at the same rate as Moore's Law and we will continue to be able to do that because this makes it a problem of power and a problem of money. So as long as you can buy enough GPUs from Nvidia or build your own, and as long as you can get access to enough power to run those computers, there really isn't many problems you can't solve.”

“You look at the data sets that are available to train. So you have the GPUs still escalating massively the amount of computer power can do. You have cloud computing at the same time, and then you have these data sets.”

“the way that he described it to me, which is so true the minute he said it, I was like, my gosh, it's like Moore's law never ended, it just shifted to GPUs. Because the inherent lack of parallelization that CPUs have, you solve the GPUs. And so that's why the surface area of compute of innovation has actually shifted, Jason, to what you're saying, which is all of these new kinds of machine-learned models, because you can just now brute force and create such a tonnage of compute capabilities and resources to solve these problems that weren't possible before.”