00:00Silicon Data, which provides pricing and performance benchmarks for AI compute, has raised a $30.5
00:06million Series A to help grow its tools and capabilities. Joining us is Silicon Data CEO
00:11Carmen Lee. She's also the former head of Strategic Alliances for Enterprise Data at Bloomberg.
00:17Congrats on the round. Really interesting group of investors that have backed you.
00:22You've grown very fast. And I think that the best place to start is how are you going to use
00:26the
00:26funds? What is it you're trying to get going on? Thank you. Thank you, Ed. Thanks for having me on
00:32the show. So it's a quite exciting week for Silicon Data. As the video story rolled out, as CME
00:40announcement came out yesterday, our goal will really become the independent referee for the
00:48whole compute stack. So as independent referee, we need to keep developing indices as there are more
00:55chips. There's more tokens. There's more LIM models coming online. We need to design next
01:01generation of benchmarking services. This is all going to be used to design products and software
01:06applications. Carmen, very quickly, the valuation wasn't disclosed. Can I just ask you what the
01:12approximate valuation of Silicon Data was in this round? It is a very healthy range.
01:19Are we at unicorn status yet?
01:22Not quite yet, but I will definitely pin you. We'll get that out of the way. Really interesting.
01:29Gavin Baker with Valor through the Valor Atreides AI Fund led this round. What should we take from
01:36that? Beyond just the check and the capital, what are you able to get from their expertise and
01:42background here? So Gavin and the team, they are the pioneer in the whole compute industry. They're
01:52the one backed a lot of new clouds from a very early stage. And we're also backed by CME, F
02:00-Prime,
02:01which is part of the Fidelity ecosystem, VanEck, obviously DRW, geometry trading. So we,
02:09as independent referee, we want to make sure our product is very well aligned and is useful for the
02:17market participants who are the natural long and natural shorts. So that's why we're excited about the
02:23whole suite of investors that's going to be heavily our clients at the same time.
02:29Right. Right. The big thing that's to come is CME planning GPU futures that are settled against your
02:36benchmark. What's the progress there? And what is it that actually will be traded? You know,
02:42just explain to the Bloomberg Tech audience that might not be familiar with this story, the basics.
02:48So it's quite exciting. So the news came out, CME, the plan date is October 5th. The first suite of
02:57indices is going to be cash settled against our H100 and H100 new cloud on-demand indices.
03:04So what's interesting, there's two actually volatility the market want to hedge. Number one is if you are
03:13natural loans of use, meaning you have your own tons of servers, obviously your revenue is tied with
03:19the rental rates. And for you to hedge out the future volatility, you really want to short futures.
03:25You do not care about physical delivery because you already have your servers.
03:30All you're looking to do is hedge out the price volatility in the future years. If you are
03:36a consumer of compute, consumer of GPU, obviously you're paying the rental prices.
03:43Ideally, you will lock prices in for a duration of time so you can loan the futures and lock in
03:50a
03:50certain prices. So that's really a use case for natural markets.
03:54That takes us to the bigger picture, and you referenced it. The NVIDIA $500 billion with all
03:59of those investment firms raised a lot of questions. For example, well, are these not assets that
04:05depreciate? How do you, in your benchmark, reflect computers collateral? Just go through some of those
04:13ideas. Yeah, so it's actually the same story. The NVIDIA $500 billion story is about a financing layer.
04:24CME and Silicon Data story is about risk management layer. So you need both. You can't have a market
04:30about half a trillion size without a way for the investor, the people with exposures, having a
04:39place for them to hedge and have prices coming. Let me ask you this as a quick follow-up,
04:47and I'm sorry to interrupt. What does your data tell you about how quickly a GPU generation can
04:53depreciate? What I see in the news right now is A100's very high utilization, very high pricing,
05:00but they're generations old.
05:04So great, great observation. So what's interesting is A100 and H100 prices have been pretty stable for
05:13the past 20 days. The price went up about 20% for both chips since January this year. The last
05:2020 days
05:20has been pretty stable. The way I will read is, so there's a few points you raised before. Just because
05:27something depreciates, server is a machine, right? The machine depreciates just like aircraft carriers,
05:33just like ships, right? Everything depreciates in terms of its lifespan. Doesn't translate to just
05:40because something depreciates, meaning they have no economic value, right? So you look at residual
05:44value population for any asset class is looking at future cash flow of that particular assets. And for
05:52the A100, to your point, right, it's an older chip, but people can still use that for different
05:56workflows and for smaller model use cases, then they can keep charging at whatever price that, you know,
06:02That's the utility.
06:04Exactly. And even L40s, right? Forget about A100. And L40s has machine learning use cases, which is
06:12a complete old chip. Right, that takes us back many generations. And to end, I've got a question from
06:16the audience quickly, very interesting. Jucan on X. What are three signals in silicon data's data over the next
06:22two to three years that would tell you AI infrastructure over supply has begun?
06:28So, great question. So there are three data points I can actually really get to take a look. Number
06:34one is spot prices, right? Spot prices is where supply and demand meet. If spot prices keeps coming
06:40down for a period of time, you can argue either there's less demand, or there's just outpacing supply,
06:48which is not really the case right now. Number two is forward curve, right? If the forward curve
06:54is contangled, which is kind that we have now, that translates to people are paying even higher price
07:00to lock in a longer term contracts versus short term rental. So that doesn't say it's oversupply,
07:06right? However, if the forward curve became very steep downward sloping curve,
07:10then the implied rate for the future, meaning it well comes down quite a bit. The third indicator is
07:18residual value, right? If the secondary market transactions of any servers, those prices came
07:24down, that represents people's expectation of those future revenue of those servers will be coming down,
07:31right? So residual value from the secondary transactions is also a good point.
07:35So-called-to-dead option is to increase depending on your server. So thank you.
Comments