Founder & CIO @SargassoCap [email protected] {Prev: biotech, Chief of Staff to US Presidential Candidate, & @Yale alumnus}
Greenwich, CT
Joined March 2024
- Tweets541
- Following1.7K
- Followers2.4K
- Likes2K
Pinned Tweet
NVIDIA CEO @JensenHuang tells @altcap on @BG2Pod that people misunderstand Nvidia's demand. About a third is classical computing growing 30 to 40% a year with no AI labs. A quarter to a third is five or six frontier labs, not two. And a third is startups renting Nvidia like next-gen AWS. In my view, treating Nvidia as a two-lab capex story is a real mistake.
"People misunderstand what's going on. There are basically three large segments.
The first is baseline computing. Data processing, molecular dynamics, physics, simulation, computer graphics, classical machine learning, fraud detection, recommender systems. Classical computing, not AI, not deep learning, not agentic systems. Call it a third of our business. And that business is growing call it 30%, 40% per year. In the absence of any AI labs, Nvidia is not zero. We're quite a large company.
The second part is the AI labs. People see two, but we work with every one of them. Meta Muse is doing fantastic. It's vastly Nvidia. Grok didn't exist three years ago. We didn't run Gemini before, we run Gemini now. Starting this year, we started working with Anthropic. People see a couple of labs, but I see more than that. I see five, six labs. And if you told me that represents not quite one-third of our business, let's say a quarter of our compute, fantastic.
And another third, this is the part that's really great. Because we're at the beginning of a new computing platform shift, a vast number of new startups are coming out. Thousands of them. And what do they do? They rent NVIDIA compute. So a third of our business is kind of cloud computing for the next generation. It's kind of like AWS for today. And that represents about a third of our business altogether.
My sense is that we're in the early innings of the buildout of this new computing architecture. And it's a real big mistake to recognize just a couple of companies and think that they were funding the infrastructure.”
A few months ago, I launched a hedge fund. The firm runs on @grok @Bot, set up with an org chart: COO, analysts, assistants, marketing, IR, & legal. Each bot owns a function, and they hand work off to each other. We even have meetings together.
Tonight's task was a website redesign. Total time was 2 hours. Old vs. new below.
Patrick F. Feeley retweeted
.@Citadel CEO Ken Griffin asked what the useful life of a GPU really is. Cutting it by just one or two years raises the AI revenue Big Tech needs from about $500 billion a year to about $700 billion.
What is the useful life of a GPU? Ken Griffin asked the right question this week. He said people are earning extraordinary pricing on five and six year old chips, and that probably doesn't last another five or six years.
So my team went into the filings, and the big builders don't agree with each other. In 2025, Amazon cut the life of some servers from six years to five, citing the pace of AI. The same year, Meta went the other way and lengthened its servers to five and a half. Microsoft, Google and Oracle use six. And that choice is worth a lot of money.
By our rough estimate, if the servers Amazon, Microsoft, Google, Meta and Oracle bought in 2025 and 2026 last four years instead of five or six, their depreciation goes up by roughly $50 billion a year. For depreciation and interest to stay at 30% of AI revenue, that revenue has to go from about $500 billion a year to about $700 billion.
.@Citadel CEO Ken Griffin asked what the useful life of a GPU really is. Cutting it by just one or two years raises the AI revenue Big Tech needs from about $500 billion a year to about $700 billion.
What is the useful life of a GPU? Ken Griffin asked the right question this week. He said people are earning extraordinary pricing on five and six year old chips, and that probably doesn't last another five or six years.
So my team went into the filings, and the big builders don't agree with each other. In 2025, Amazon cut the life of some servers from six years to five, citing the pace of AI. The same year, Meta went the other way and lengthened its servers to five and a half. Microsoft, Google and Oracle use six. And that choice is worth a lot of money.
By our rough estimate, if the servers Amazon, Microsoft, Google, Meta and Oracle bought in 2025 and 2026 last four years instead of five or six, their depreciation goes up by roughly $50 billion a year. For depreciation and interest to stay at 30% of AI revenue, that revenue has to go from about $500 billion a year to about $700 billion.
Patrick F. Feeley retweeted
Asked on @SquawkCNBC whether this is 1999 again, Altimeter's Brad Gerstner told @andrewrsorkin the constraint has flipped.
"In 2000, we had constrained demand because we only had 30 million people on the internet. Today, we have unconstrained demand for superintelligence."
A glut needs spare supply, and wafers and power take years to build. The question was never whether there is too much compute. It is who knows what to do with it.
Yep.
Starboard's Jeff Smith told @davidfaber on @SquawkStreet that AI has changed the first conversation in the boardroom.
"Today, if I walk into a company and say, you need to be more efficient, they all say, we understand."
Agreement was never the hard part. Every board now nods along. The real gap is between companies that say "we understand" and the ones that actually rebuild how the work gets done.
Amazon is sold out of AI/SI capacity past 2028. Quite a situation...
.@bhorowitz says @ajassy told him that directly, and he explains on @theallinpod why that's the opposite of the dot-com bubble. GPUs now last six to eight years instead of three to five, and they're getting more expensive. The internet bubble died because there weren't enough users to absorb what got built, and AI has more demand than supply.
"The other interesting thing is GPUs, not only have they been going up in duration, they've been going up in price. So that's been a big tailwind for him as the owner of such things. It was three, four, five [years], now it's six, seven, eight. And there's a shortage, a real shortage of GPUs with power and cooling, which is what he has.
And this shortage is so bad. I was speaking with Andy Jassy, who runs Amazon, yesterday, and I said, yeah, we're out of everything till 2028. He's like, oh, it's beyond 2028. End of 2028, maybe. So everything else is sold out. So he has the advantage that he's got supply in a great shortage, which is extremely different than the internet bubble.
The internet bubble, we didn't have enough demand because we had to build the internet. People had to get broadband. People had to get their browser, all this kind of thing. There weren't enough consumers on the internet to absorb all the bandwidth that had gotten built out, and they had a massive amount of debt, not unlike what we're building today, although not as big. But the lack of demand killed it.
We don't really have a lack of demand in AI. So I think the dynamics will be different.”
.@eglyman of Ramp tells @JTLonsdale on @AmOptimistShow that the world now spends about $150 to $170 billion a year on AI inference, up roughly 5x from $30 billion a year ago, and that's still only about 2% of the $8 trillion the US spends on white collar wages. I think the demand story for AI is much earlier than the capex debate assumes.
"With AI, it's quite different. One, it's growing very, very quickly. About a year ago this time, the entire world was spending on the order of $30 billion a year on inference across the major labs. Today, estimates are something closer to $150 to $170 billion per year, depending on how you measure it.
To put that in perspective, I think in the United States, there's something like $8 trillion a year spent on white collar wages, sort of back office wages. In other words, the amount spent on AI is equal to about 2% of all white collar wages. And it just went up 5x over the last year and could go up, who knows, some other multiple again.
We're a knowledge economy and information ecology. It doesn't strike me as so crazy if you were to say, hey, we spend 10% of what we bring in on thinking. You might even say, perhaps that's low. These machines are getting smarter."
Patrick F. Feeley retweeted
.@DavidSacks on @theallinpod makes an excellent point when he said the one number to watch in the AI trade is payback period.
CoreWeave earns its money back in 2.5 years on 5-year contracts, while SpaceX does it in about one, because it sells near spot at roughly $50 billion per gigawatt against CoreWeave's $20 billion, and scarcity has kept pushing spot prices higher. I think short paybacks are what make this cycle different, since debt gets repaid fast and everyone's risk drops. If payback starts stretching because of weaker demand, falling spot prices or delays bringing capacity online, that's the signal to get nervous.
"There's one metric that, the more I think about it, I would really keep an eye on, and that's payback period. To me, that's like the North Star metric here.
So CoreWeave says that its payback period is 2.5 years. Their typical deal is 5 years. They pay back within 2.5 years. SpaceX, they pay back in roughly one year. That makes sense because I think that SpaceX is monetizing at about $50 billion per gigawatt and CoreWeave at about $20 billion per gigawatt.
And the reason for that is because SpaceX is basically getting spot. They're doing these deals that can be canceled by either side within 90 days, so they're basically getting spot minus 90. Whereas CoreWeave is signing up for 5-year commitments, so it's contracted way in advance and they're getting this contracted price. And what's happened, I think, over the past year is that spot has moved way up, because again of the scarcity, and the demand is so off the charts.
So in any event, payback period is really critical, and as long as the payback periods are really short, I don't see how these companies can lose, or at least you'd have to be really incompetent to screw it up. You're making these big capital investments, but if you're getting your money back in one year or two and a half years, that means you're paying back the debt quickly and the debt's rolling over more quickly. It just reduces everyone's risk.
Now, if that payback period starts taking longer and longer, then I get nervous, because that magnifies the risk. It also means the economics aren't as good. And why would the payback period take longer? It could happen for a bunch of reasons. If the demand goes down and the offtake's not as good, the spot price going down, or there's just execution problems and it takes you longer to get the capacity online.
There's a lot of reasons why payback could take longer and longer. But as long as payback remains as short as it is today, to me that would be the one metric I would really keep an eye on."