00:00As an owner of NVIDIA, what are you going to be scrutinizing from the earnings?
00:05Probably the most important forward-looking indicator for us are purchase commitments
00:11because this doubled quarter on quarter in the last print. We went from 50 billion in Q3 to 90
00:24billion last quarter. And this is essentially NVIDIA locking out the supply chain. It's no surprise
00:31for all of us that the supply chain is very tight. We all hear about there not being enough memory
00:36to
00:36go around. TSMC silicon is a meaningful bottleneck. Optics is the same. Well, NVIDIA have got between
00:44two-thirds and three-quarters of the supply, these key components through 2026 and into 2027. So we're
00:53expecting a significant step up again. So to be clear, this is NVIDIA's purchase commitments for its
00:58suppliers or purchase commitments from hyperscalers for NVIDIA chips? Purchase commitments to suppliers
01:04precisely. To suppliers. Okay. So that for you is an indication that they have continued control over
01:09that supply chain. Absolutely. Because the risk going into this print and really throughout this
01:13year we think is not about demand. We do not think that that is a risk anymore. This is solely
01:19a
01:20question about supply. Six months ago the market was focused on are we overbuilding. The biggest risk
01:25now is can we build fast enough to keep these earnings moving in line with the stock prices of
01:33the entire ecosystem. Is NVIDIA the way to play that though or are some of the Asian memory suppliers the
01:40way to play it? Are they a cleaner, more monetizable kind of short-term play on how this AI build
01:47-out story
01:48is going to develop? Have we moved on from NVIDIA to those kinds of companies? Well, that's exactly what
01:54the market has done over the past six months. You look at the stock prices of SK Hynix,
01:58Micron versus NVIDIA. Right now we see more. Is that the correct move? Possibly in the short term. I mean,
02:05what the market has done and we've all spent a lot of time doing is following the bottlenecks. And
02:09over the past six months those have been particularly in high bandwidth memory. Optics has been a key
02:14one emerging. But right now the first derivative of the AI trade is now offering the most compelling
02:20valuation. And what we saw last quarter with NVIDIA and we're expecting to see continue throughout
02:262026 is re-acceleration in their top line and on their bottom line. So NVIDIA grew earnings about 50%
02:33last year, year on year. A very admirable achievement for a company of that scale. That's set to grow
02:38almost 100% this year. And so right now we will be trimming some of those Asian component suppliers
02:46because, yes, they are in pure place on some of these bottlenecks. But their stock price has also
02:52moved an incredible amount in the past couple of months. And can I ask you about the life of some
02:57of
02:57the products that NVIDIA puts into these hyperscaler businesses and others? Because the life of them
03:04and the period over which you'd appreciate them was something of a topic of conversation for quite a
03:08while. And you've got some interesting stuff in your notes about how the installed base becomes
03:11more invaluable, more valuable, not less. Explain how that can be possible.
03:15Absolutely. We do think that there has been a misunderstanding about this because with NVIDIA
03:23architectures, they've now got a one-year release cadence. And you buy the latest and greatest for
03:31training your large language model at the frontier. So that's Blackwell, Ultra right now, moving to
03:36Rubin next year. What happens to Hopper, which was the prior architecture, even Ampere, which was six,
03:43seven years ago, is that those N-1, N-2 architectures moved to inference. And inference is
03:49when customers actually monetize these assets. So the ROI on these architectures, they actually go
03:56up, not down. And what we've seen right now, because we're in such an acutely demand supply
04:02constrained imbalance, is that the prices for Hopper, they've gone up 35% year to date in the cloud.
04:11Yes. And the reason these get so more valuable over time is because they're optimized by software.
04:17Yes. And that's that's the whole QDA ecosystem mode. And one of the reasons why the market's
04:21attention switched to Asia was because of the memory story. We've also seen a lot of businesses
04:26trying to follow NVIDIA and trying to grab some of that market share. And some of them arguably trying
04:31to, you know, managing to deliver higher profits as a result. But does NVIDIA need to worry about that
04:35competition or not at all? Competition from the... From others who are trying to do what NVIDIA does.
04:41Well, it comes down to locking out the supply. And so these... It's not about Ruben and what it can
04:48do
04:48versus the competition. Well, Ruben is definitely a massive leap forward versus competition. And what
04:55we saw from Google with the TPU last year, TPU version seven, this is very competitive for Blackwell.
05:02That does change with Ruben. Google moves to its next TPU architecture. And Ruben particularly moves
05:11to more memory bandwidth intensive architecture, which is really, really crucial for inference. It's got
05:17like 2.4 times more memory bandwidth. And that's really, really important as we move into the
05:25agential era, which is highly more compute intensive. It still does come down to capacity at TSMC and
05:33capacity with these memory players. And these ASICs players just do not have that same weight.
05:41Can I just... Everything costs a lot in this industry at the moment. And you're making that point very
05:45clearly. Are we as users, our corporates as users, paying the correct price for the AI that they are
05:52consuming at the moment? And if they did, would they use the same amount or would they use less?
05:58Are we being subsidised at the moment by the industry, and you look at the cash flow story,
06:02for what we're using? And does that need to change?
06:06So we expect the cost of AI to continue declining. And the cost of inference right now, it's about a
06:12thousand times cheaper than it was last year. And you achieve this through technological progress.
06:18Every time NVIDIA moves to a new architecture, this brings down cost per dollar, cost per...
06:24The companies are still burning money.
06:25The companies are burning money, but that's only because they're continuing to invest.
06:31In the hyperscale capex, if we look at this, it's slated to be over 700 billion this year.
06:36If there's a 12 to 24-month hiatus between standing up this infrastructure and then actually
06:47producing revenues from it. Andy Jesse painted this clearly in his shareholder letter last quarter.
06:53And because the hyperscalers are seeing ROI on the capacity laid two, three years ago,
07:01they're continuing to invest, but more than what the revenues are. So we think the right metric to be
07:05looking at is revenue inflections, cloud revenue inflection, not free cash flow. We're expecting
07:11free cash flow to actually be materially negative for all the hyperscalers next year. But the crucial
07:16point is that revenues are now inflecting. We saw this from Google, 63% year on year.
07:22and to have a benefit of seeing the power of manufacturing is available.
07:22Thanks!
07:22It's the aim of the game.
Comments