WEBVTT - Inside Hudson River Trading's Blistering Token Burn 0:00:02.720 --> 0:00:16.400 Bloomberg Audio Studios, Podcasts, radio News. 0:00:18.520 --> 0:00:21.480 Hello and welcome to another episode of the Odd Lots podcast. 0:00:21.560 --> 0:00:23.799 I'm Joe Wisenthal and I'm Tracy Alway. 0:00:23.960 --> 0:00:26.400 Tracy. We did another one of our live shows, this 0:00:26.480 --> 0:00:28.720 time our biggest show ever, biggest show ever. 0:00:29.160 --> 0:00:31.720 It was absolutely amazing. We did it at City Winery 0:00:31.800 --> 0:00:34.880 in New York. I think we had over three hundred people. 0:00:35.080 --> 0:00:37.360 Yeah, three hundred and fifty people were there. 0:00:37.560 --> 0:00:37.880 Yeah. 0:00:37.920 --> 0:00:39.959 And the crazy thing is, I think it was our 0:00:40.000 --> 0:00:43.440 sort of first themed show, and we didn't really plan 0:00:43.560 --> 0:00:45.800 it that way, but it just worked out right. 0:00:45.880 --> 0:00:48.400 I guess like it themed an anti theme at the 0:00:48.440 --> 0:00:51.440 same time, because we're in this moment in which everything 0:00:51.560 --> 0:00:55.240 is just like AI markets, markets, AI, et cetera. But 0:00:55.440 --> 0:00:57.480 you know, there's all kinds of new things to trade, 0:00:57.760 --> 0:01:00.680 and people are fascinated by the trade itself, pro fascinated 0:01:00.680 --> 0:01:03.880 by the way the technology development is affecting the trade. 0:01:04.000 --> 0:01:06.480 So we really wanted to do a kind of future 0:01:06.480 --> 0:01:09.920 of trading show, which is a very broad thing. But 0:01:10.000 --> 0:01:11.360 it did sort of come out that way. 0:01:11.480 --> 0:01:15.200 Yeah, it really did. And you finally fulfilled your longtime 0:01:15.280 --> 0:01:18.679 dream of doing two part episodes with our guests. So 0:01:19.040 --> 0:01:21.720 our first speaker of the evening was actually someone who's 0:01:21.720 --> 0:01:22.559 been on the show before. 0:01:22.720 --> 0:01:25.559 That's right, So we had him on the show last year, 0:01:25.840 --> 0:01:29.120 and we had him on our live show. Listen to 0:01:29.160 --> 0:01:31.600 our episode with Ian Dunning, he is the head of 0:01:31.720 --> 0:01:38.160 AI at Hudson River Trading. Talked about all things implementing AI, GPUs, 0:01:38.400 --> 0:01:41.600 all that stuff within the context of a trading shop. 0:01:41.640 --> 0:01:43.399 Take a listen, Joe, this is. 0:01:43.360 --> 0:01:45.880 Your dream, right, you finally got to do a two parts. 0:01:45.959 --> 0:01:48.080 This is the thing I always think about, which is 0:01:48.080 --> 0:01:50.480 that after every episode we do, I'm like, Oh, there's 0:01:50.480 --> 0:01:52.560 a question I wish I had asked. So we had 0:01:52.640 --> 0:01:54.400 Yan on sometime last. 0:01:54.200 --> 0:01:56.240 Year, so the last round was easy, this one will 0:01:56.240 --> 0:01:56.480 be right. 0:01:57.160 --> 0:01:58.800 Worried about that, well, I was going to start first 0:01:58.960 --> 0:02:01.760 before we you know, talk about what you do, et cetera. 0:02:01.880 --> 0:02:04.040 So here's the question I wish I had asked last time. So, 0:02:04.120 --> 0:02:07.320 Hudson River Trading Shop, you're involved in the AI stuff, 0:02:07.880 --> 0:02:11.440 could you theoretically do what high Flyer did and launch 0:02:11.480 --> 0:02:14.800 an LLM with this tech stack that you have and 0:02:14.880 --> 0:02:16.240 launch a deep seat competitor? 0:02:16.639 --> 0:02:19.720 I think so. I think we're good at training models. 0:02:19.720 --> 0:02:22.080 We have a lot of compute and people are good 0:02:22.080 --> 0:02:24.360 at doing the cycle of research which is required to 0:02:24.400 --> 0:02:28.079 catch up to the sort of frontier. However, I guess 0:02:28.280 --> 0:02:31.680 reaching the frontier is clearly a very daunting task. So 0:02:31.720 --> 0:02:35.120 maybe it's with some effort deep seek, but beyond that, 0:02:35.160 --> 0:02:37.079 it's not a claim I'd be willing to make. It's 0:02:37.120 --> 0:02:39.080 a hugely capital intensive task, clearly. 0:02:39.120 --> 0:02:41.800 And do you ever, like, do people ever chat about that? 0:02:41.840 --> 0:02:43.520 It's we could we think about it? 0:02:43.639 --> 0:02:43.760 You know? 0:02:44.280 --> 0:02:46.600 I mean, I think perhaps we missed our moment to 0:02:46.639 --> 0:02:48.800 do so. It is so many open models now coming 0:02:48.800 --> 0:02:51.240 out from the US as well as like China, that 0:02:51.639 --> 0:02:53.880 there's like a huge array of them. It's kind of 0:02:53.880 --> 0:02:56.040 an interesting shift from that deep Seak moment where it 0:02:56.080 --> 0:02:57.720 felt like it was the first bolt from the blue. 0:02:58.040 --> 0:03:00.600 He's a competitive open model. So now I see so 0:03:00.639 --> 0:03:03.680 many groups releasing them. I don't know what the future 0:03:03.680 --> 0:03:06.160 of open models is. That they're all kind of a 0:03:06.240 --> 0:03:09.000 serious step back in. The frontier is progressing so fast. 0:03:09.200 --> 0:03:11.440 I don't know how you keep up with that, but 0:03:11.600 --> 0:03:13.639 many people believe that it's possible. I'm not so sure. 0:03:13.800 --> 0:03:14.880 I'm one of those people though. 0:03:15.040 --> 0:03:18.240 Okay, so, speaking of things moving so fast, my first 0:03:18.320 --> 0:03:20.919 question is slightly different. I looked up your Twitter feed. 0:03:21.040 --> 0:03:24.000 Oh no, before you came on the show. Your last 0:03:24.000 --> 0:03:29.400 tweet before today was and I quote feel this every day, worry. 0:03:29.440 --> 0:03:32.360 It's some sort of AI induce delirium. But then again, 0:03:32.520 --> 0:03:35.600 various empirical measures are exponential looking, so feels best to 0:03:35.640 --> 0:03:38.800 assume we're hurtling towards some sort of endgame. So first 0:03:38.840 --> 0:03:42.760 of all, please convince us all live on stage that 0:03:42.800 --> 0:03:45.600 you are in fact not suffering from AI induce delarium. 0:03:45.680 --> 0:03:48.560 But secondly, like, what is the endgame that you speak 0:03:48.600 --> 0:03:48.960 of here? 0:03:49.480 --> 0:03:53.480 God, now I saw like a San Francisco pressent I do. I'sist. 0:03:53.600 --> 0:03:57.600 I've been doing AI stuff since around twenty sixteen, and 0:03:57.760 --> 0:04:00.480 that stuted at Deep Mind, and it was a bit 0:04:00.480 --> 0:04:02.280 of a culture shock for me, because there are true 0:04:02.320 --> 0:04:04.840 believers then, and I was most certainly not a true believer, 0:04:05.480 --> 0:04:07.160 and I resisted it, and it was kind of a 0:04:07.240 --> 0:04:11.760 natural skeptic for a long time. But certain empirical measures 0:04:11.760 --> 0:04:14.040 of the pace of progress in the outside world. And 0:04:14.040 --> 0:04:18.600 I also look at our own business, which looks somewhat 0:04:18.640 --> 0:04:23.200 exponentially the amount of compute I'll have next year versus 0:04:23.240 --> 0:04:26.880 this year, and this year versus last year. Looks kind 0:04:26.920 --> 0:04:31.640 of exponentially, and we're doing things today that I didn't 0:04:31.960 --> 0:04:34.600 really I should have dreamt of. I wish I had 0:04:34.600 --> 0:04:36.560 that kind of visionary say, I'm a visionary and I 0:04:36.600 --> 0:04:38.880 can see the future and I'm building towards it. But no, 0:04:39.080 --> 0:04:41.479 I'm I think I'm a pragmatic engineering archetype, and so 0:04:41.520 --> 0:04:44.919 it's been very incremental and I'm like, wow, that happened 0:04:44.960 --> 0:04:46.960 in a year. So what does this mean? It's some 0:04:46.960 --> 0:04:51.400 sort of technological convergence, everything going faster all the time. Well, 0:04:51.400 --> 0:04:54.080 give us any I'm probably, but you give us an. 0:04:53.960 --> 0:04:56.240 Example then, because you know, obviously those of us using 0:04:56.360 --> 0:05:00.560 just the regular models, obviously the improvements and capability from 0:05:00.560 --> 0:05:03.080 one year to another our mind blowing. But from the 0:05:03.160 --> 0:05:06.800 perspective of like, okay, the application of AI within the 0:05:06.839 --> 0:05:09.520 trading context, what is something that you can do in 0:05:09.560 --> 0:05:12.520 twenty twenty six that and say twenty twenty four you 0:05:12.560 --> 0:05:14.280 would not have been able to anticipate. 0:05:14.400 --> 0:05:16.839 Oh, I think it's one way to think is just 0:05:16.839 --> 0:05:19.760 like the amount of compute going into both training a 0:05:19.800 --> 0:05:22.080 model and running a model, and that it's the same 0:05:22.120 --> 0:05:25.880 technology working across every equity, every future, every crypto market, 0:05:25.920 --> 0:05:28.000 every option market across the world with a kind of 0:05:28.080 --> 0:05:31.800 unified approach, and this is something that we're doing. But 0:05:31.960 --> 0:05:34.800 even more interestingly, I would not claim that we are 0:05:34.880 --> 0:05:37.159 some unique people who are only the ones who have 0:05:37.240 --> 0:05:40.000 really made progress in AI and trading. I think many 0:05:40.040 --> 0:05:43.680 of our peers are also investing massively and we're all 0:05:43.680 --> 0:05:45.120 doing it all at the same time. And what does 0:05:45.160 --> 0:05:48.040 that mean, Like, surely you can't just like keep getting 0:05:48.080 --> 0:05:50.919 better at predicting markets Furvace. It's got to be some 0:05:50.960 --> 0:05:54.880 sort of forcing function where you know your your margins 0:05:54.880 --> 0:05:56.240 go to zero as you keep investing. 0:05:56.800 --> 0:05:59.680 What you're saying is you are just getting better and 0:05:59.720 --> 0:06:02.279 better at being able to predict where a market is 0:06:02.279 --> 0:06:04.240 going to go further and further out in the. 0:06:04.160 --> 0:06:06.920 Time frame basically, And we're not the only ones. So 0:06:07.040 --> 0:06:10.040 in the end, can there be some highlander type thing 0:06:10.120 --> 0:06:12.240 like what are we doing? And it's it's this is 0:06:12.240 --> 0:06:13.720 like my scale, And I guess the other thing I 0:06:13.760 --> 0:06:16.280 find interesting is, of course the scale that everyone can 0:06:16.320 --> 0:06:19.520 see with the big labs and what they're doing of computeion. 0:06:19.600 --> 0:06:22.760 It's like it looks awfully exponential to me. We just 0:06:22.800 --> 0:06:26.560 had another model released today from Anthropic and the type 0:06:26.560 --> 0:06:30.560 of spacing between them seems to be compressing. I do 0:06:30.600 --> 0:06:33.520 sound hilarious. I sound feverish, and that's why. 0:06:32.760 --> 0:06:34.960 It's literally everyone in this room. 0:06:35.000 --> 0:06:37.359 Probably I feel everyone must feel the fever to some extent. 0:06:37.440 --> 0:06:40.480 Yeah, I never understood the Highlander there can only be 0:06:40.520 --> 0:06:44.480 one thing because they're already two. They could just coexist, 0:06:44.560 --> 0:06:47.640 that's right. Anyway, Sorry, I'm just picking apart your analogy. 0:06:47.800 --> 0:06:50.840 You mentioned a new model release. When a new model 0:06:50.880 --> 0:06:53.680 gets released, like, what is the first thing you do 0:06:54.000 --> 0:06:56.440 at Hudson River Trading to evaluate it? And how do 0:06:56.480 --> 0:06:58.440 you actually compare them to the existing one? 0:06:58.839 --> 0:07:01.920 So I mean our prime reuse cases as a trading 0:07:01.960 --> 0:07:05.039 and definitely kind of just like accelerating your own research. 0:07:05.240 --> 0:07:07.840 So that can be coding, but it can also be 0:07:08.040 --> 0:07:12.200 coming up with experiment ideas, monitoring experiments. We had a 0:07:12.360 --> 0:07:15.040 sort of a false start with AI I would say 0:07:15.280 --> 0:07:18.400 sometime last year with the Opus four point zero models, 0:07:18.480 --> 0:07:23.240 especially from Anthropic, where a cursory examination made us feel like, well, 0:07:23.280 --> 0:07:25.600 this is the moment we've crossed the dividing line, and 0:07:25.640 --> 0:07:28.320 we had a very feverish week where we felt the 0:07:28.360 --> 0:07:32.560 AGI and we left feeling empty because we realized that 0:07:32.840 --> 0:07:35.120 it was not there and was not able to meaningfully 0:07:35.240 --> 0:07:38.720 augment human researchers. And then we had that same feeling 0:07:38.760 --> 0:07:41.560 again when Opus four point five came out and suddenly 0:07:41.600 --> 0:07:43.560 it was like, oh, wait, no, this is this is 0:07:43.560 --> 0:07:44.920 actually what we thought it was going to be six 0:07:44.960 --> 0:07:48.600 months ago. So in the most recent model releases, the 0:07:48.680 --> 0:07:52.000 differences have been more subtle, but we see, I think 0:07:52.040 --> 0:07:54.600 we have a much better sense of an ever reducing 0:07:54.600 --> 0:07:56.720 set of errors they make, and so we're kind of 0:07:56.760 --> 0:07:59.360 looking for those those of mistakes, and we are We 0:07:59.400 --> 0:08:00.960 spent some time in the past couple of weeks trying 0:08:00.960 --> 0:08:03.240 to come up with objective measures to index them against 0:08:03.440 --> 0:08:08.240 humans in the active quant research ideating signals and things 0:08:08.720 --> 0:08:10.440 quant research used to be, as we talked about a 0:08:10.480 --> 0:08:13.800 little bit like hand crafting indicators and things, why not 0:08:13.840 --> 0:08:16.560 ask AI agents to do that and compare them against 0:08:16.600 --> 0:08:20.200 humans like a little sort of battle. And it's they're like, 0:08:20.320 --> 0:08:24.160 I don't know, intern level AI. Perhaps the thing is like, 0:08:24.440 --> 0:08:25.720 what do I think it'll be in a year? And 0:08:25.800 --> 0:08:27.200 I would not want to make a bold claim, but 0:08:27.200 --> 0:08:27.760 it will still be. 0:08:28.240 --> 0:08:32.160 It'll be wild. So when we think about investing in general, 0:08:32.640 --> 0:08:36.319 even within sort of like classical quant trade, and going 0:08:36.360 --> 0:08:40.240 back decades, there is often it might be quant but 0:08:40.240 --> 0:08:43.560 there's some intuition behind it. Right, Cheap stocks tend to 0:08:43.600 --> 0:08:46.760 do better, and we don't actually totally have agreement why 0:08:46.800 --> 0:08:49.680 they did for a while, but people aren't necessarily surprised 0:08:49.720 --> 0:08:52.760 by that fact, right are we at the point where 0:08:52.760 --> 0:08:55.640 it's like why even bother coming up with the human 0:08:55.720 --> 0:08:58.920 intuitive story and you just skip the part of giving 0:08:58.920 --> 0:09:02.120 an explanation that's owns logical to a person and it's 0:09:02.160 --> 0:09:07.360 just basically pure like rigorous back testing. And then it's like, look, 0:09:07.480 --> 0:09:10.000 here is something that seems to work, and we've back 0:09:10.040 --> 0:09:12.360 tested it a million different ways and it seems to work, 0:09:12.480 --> 0:09:14.480 and we don't even bother coming up with a story 0:09:14.480 --> 0:09:15.720 for why, but we're going to trade it. 0:09:15.920 --> 0:09:17.920 I feel like we're in that world today. It's seat 0:09:17.920 --> 0:09:21.760 of post post post capitalism. When I see IPOs discussed 0:09:21.760 --> 0:09:24.080 for this coming summer, at the valuations they are, I'm like, 0:09:24.240 --> 0:09:27.319 what is a fundamental? Like what is anything? It feels 0:09:27.400 --> 0:09:30.800 like markets are just the cynical thick is everything is gambling, 0:09:30.840 --> 0:09:33.280 and so everything is some sort of like gambling market, 0:09:33.360 --> 0:09:36.640 including public markets. But the joke is flows, it's buying 0:09:36.679 --> 0:09:38.720 and selling, and it's just it's worth what it's worth, 0:09:38.920 --> 0:09:42.080 and it's detached and more biased, and sellers price go up, 0:09:42.160 --> 0:09:45.480 and models are excellent at like pulling that out of data. 0:09:45.559 --> 0:09:48.160 But just like let's say, you know the classic example 0:09:48.200 --> 0:09:50.520 of like a back test, it is like, oh, companies 0:09:50.559 --> 0:09:53.080 with the ticker symbol it starts with P, they do 0:09:53.200 --> 0:09:55.640 well on Tuesdays. And it's like, well, look the data 0:09:55.640 --> 0:09:57.080 says that, but this makes no sense. We're not going 0:09:57.160 --> 0:09:58.839 to trade that. Could it get to the point where 0:09:58.840 --> 0:10:01.240 it's like, look, symbols that starts with P, you do 0:10:01.320 --> 0:10:03.480 well on Tuesdays, And we've run this a bunch of 0:10:03.520 --> 0:10:05.120 times and it seems to work, so we're gonna put 0:10:05.160 --> 0:10:05.760 money beyond us. 0:10:06.280 --> 0:10:07.480 AI Like I. 0:10:09.120 --> 0:10:12.000 Feel like, yes, although it sounds crazy, it sounds like 0:10:12.000 --> 0:10:13.920 AI delirium when I say it, but I feel like 0:10:13.960 --> 0:10:16.600 there's some sense that that could be true at some 0:10:16.679 --> 0:10:20.400 point I can't predict. So at the very short time scale, 0:10:20.400 --> 0:10:22.920 people accept this already, right Like I can't tell you 0:10:22.960 --> 0:10:24.680 the price of like a stock in a minute, and 0:10:24.679 --> 0:10:27.160 no one would really reasonably expect any human to do so, 0:10:27.520 --> 0:10:29.080 even if they had the uder book and spend all 0:10:29.120 --> 0:10:30.640 the time in the world is staring at it. But 0:10:30.720 --> 0:10:33.920 we accept that neural networks can do this, And then 0:10:34.520 --> 0:10:37.160 when does that logic break down? Why should it break 0:10:37.200 --> 0:10:39.760 down at some long time scale. If it's ingesting all 0:10:39.840 --> 0:10:41.800 the data and has everything and it can keep it 0:10:41.840 --> 0:10:44.760 all in the context in way human can't, why should 0:10:44.760 --> 0:10:48.040 I be able to understand it? Yeah, and that is 0:10:48.040 --> 0:10:50.839 a strange thought, a loss of control. It feels like 0:10:50.880 --> 0:10:54.240 a loss of control. But it's like, you know, people 0:10:54.240 --> 0:10:56.280 save us from math. Maybe humans are actually very bad 0:10:56.320 --> 0:10:58.920 at math. So it's not surprising AI as much better 0:10:58.960 --> 0:11:02.200 than humans that these like mass proofs. Humans probably would 0:11:02.200 --> 0:11:05.440 be pretty bad at markets where thousands of tradable instruments 0:11:05.480 --> 0:11:08.000 on like very long time scales. We just kind of 0:11:08.000 --> 0:11:10.480 accepted that we were some people were good at this. 0:11:11.040 --> 0:11:14.520 Maybe that's a temporary state of affairs. 0:11:14.640 --> 0:11:16.720 Well, we talked about this the last time. You were 0:11:16.760 --> 0:11:20.040 on the idea that the models themselves are not very interpretable, 0:11:20.200 --> 0:11:22.320 I guess you would say, but you're comfortable with that 0:11:22.800 --> 0:11:25.960 on a short trading timeframe, which is what you do. 0:11:26.520 --> 0:11:29.400 And then we started joking about magic models, and magic 0:11:29.480 --> 0:11:31.920 is a dangerous word to use on this podcast because 0:11:31.920 --> 0:11:36.000 people start thinking about magic boxes. But anyway, now that 0:11:36.080 --> 0:11:38.360 you've been doing this for another six months since we 0:11:38.480 --> 0:11:40.720 last spoke to you, do you feel like you have 0:11:40.840 --> 0:11:44.199 better insight into what the models are actually doing and 0:11:44.280 --> 0:11:46.880 why they're able to succeed on short timeframes. 0:11:47.200 --> 0:11:50.520 I do think there are diagnostics we have done where 0:11:51.000 --> 0:11:54.719 when we can see things that we can understand. It's 0:11:54.760 --> 0:11:57.040 like looking at some very very complex thing and you 0:11:57.040 --> 0:11:58.720 can look at one facet of it and be like, 0:11:58.760 --> 0:12:00.680 this is the fast that I understand, and that gives 0:12:00.720 --> 0:12:04.280 you some confidence, but it might be illusory because it's 0:12:04.280 --> 0:12:06.800 a very very complex object, and you can if you're 0:12:06.840 --> 0:12:10.080 only taking slices through it and to understand aspects of it. 0:12:10.559 --> 0:12:13.480 You know, we had this emergent phenomenon we saw where 0:12:13.600 --> 0:12:15.920 it felt like the model kind of understood meme stocks 0:12:15.960 --> 0:12:21.400 from first principles like quantum stocks and crypto stocks being 0:12:21.480 --> 0:12:24.920 kind of adjacent in stock space, and of course from 0:12:24.960 --> 0:12:28.160 a fundamental perspective that says, no, there's no meaning to it. 0:12:28.240 --> 0:12:30.160 But we looked at the model and us under a 0:12:30.160 --> 0:12:33.280 certain lens and it clearly felt like they knew they 0:12:33.320 --> 0:12:35.840 were connected. There's some other actual companies that I probably 0:12:35.840 --> 0:12:38.080 won't name. It feels like it's bad form, but you 0:12:38.120 --> 0:12:40.720 know Wall Street BET's favorites, I guess, and they were 0:12:40.800 --> 0:12:43.400 near the cluster too, And this was like just one 0:12:43.440 --> 0:12:45.880 little window. But there were other slices we tried to 0:12:45.920 --> 0:12:47.920 take which just didn't make sense to us. But again 0:12:48.000 --> 0:12:49.880 it's like, who am I to say? 0:12:51.760 --> 0:12:55.280 The model says they're in that vicinity of hyperdimensional space. 0:12:55.520 --> 0:12:58.040 Yeah. The one thing for us though, is that when 0:12:58.080 --> 0:13:02.080 we do have this magical model, it is in a 0:13:02.120 --> 0:13:04.679 lot of safety around it. Because we're doing this higher 0:13:04.679 --> 0:13:07.280 frequency trading. We're trading positions back and forth. There's a 0:13:07.320 --> 0:13:09.560 lot of risk checks that are fully automated in things. 0:13:09.920 --> 0:13:12.480 I don't know how you generalize this logic to long 0:13:12.600 --> 0:13:16.360 term discretionary trading, where the idea of like risk checking 0:13:16.600 --> 0:13:19.559 and that kind of layer of defense around it, it's 0:13:19.600 --> 0:13:21.200 not so obvious to me how you apply that. We 0:13:21.200 --> 0:13:24.360 can apply very strict controls around this model because it's 0:13:24.360 --> 0:13:28.000 a well posed problem. We're not taking giant idiosyncratic risks 0:13:28.040 --> 0:13:30.440 in like one name for months at a time. We 0:13:30.480 --> 0:13:32.160 can sleep at night because of this. I don't know 0:13:32.200 --> 0:13:34.880 how you apply the same thinking to like a fundamental long, 0:13:34.920 --> 0:13:36.240 short thing where you have to put a trade on 0:13:36.280 --> 0:13:39.120 and it's for three months, and you're intentionally taking a 0:13:39.280 --> 0:13:43.680 very large risk in a very sudden direction. That's what's 0:13:43.720 --> 0:13:45.560 the risk management story around the AI if you just 0:13:45.600 --> 0:13:59.760 give up all control to just the magic prediction. 0:14:04.160 --> 0:14:06.720 So you said something on the last time we interviewed you, 0:14:06.840 --> 0:14:08.559 which is very important. First of all, I feel like 0:14:08.640 --> 0:14:11.400 in the quote AI trade unquote, people are obsessed with like, 0:14:11.480 --> 0:14:14.640 what's the bottleneck now, right? And because whatever the bottleneck is, 0:14:14.640 --> 0:14:16.840 you probably solved for a lot more money. You said 0:14:17.320 --> 0:14:20.240 the last time we talked to you, the chips themselves 0:14:20.640 --> 0:14:23.800 are We're not actually a major constraint for you, and 0:14:23.800 --> 0:14:26.960 that it was more like citing the chips and the 0:14:27.080 --> 0:14:29.840 powering the chips, the access to electricity, talk about that. 0:14:29.880 --> 0:14:32.480 What is the state right now? Let's say like I 0:14:32.560 --> 0:14:36.320 assemble approach a bunch of people from Hudson River Trading. Yeah, 0:14:36.400 --> 0:14:39.320 I get a bunch of GPUs. Is it then not 0:14:39.560 --> 0:14:41.440 trivial to find a place to plug those in. 0:14:41.720 --> 0:14:45.560 It's definitely hot to find sites and at shortly times. 0:14:45.560 --> 0:14:47.400 If I went to the market and said I want, 0:14:47.520 --> 0:14:51.840 you know, six thousand Blackwell GPUs in a box somewhere 0:14:51.840 --> 0:14:55.880 in North America for delivery in Q four. I'm not 0:14:55.920 --> 0:14:59.440 sure such an offering exists at any reasonable price. Like 0:14:59.480 --> 0:15:01.520 if it from maybe someone will give up a lease 0:15:01.560 --> 0:15:02.960 and I could snag it. But I think if I 0:15:03.000 --> 0:15:04.360 went to the market and tried to get a. 0:15:04.400 --> 0:15:07.080 Quia sorry just to be clear, the chips are available, 0:15:07.120 --> 0:15:08.280 but not the competity. 0:15:08.280 --> 0:15:10.200 I think if I had power, I could get the 0:15:10.280 --> 0:15:12.320 chips Blackwell chips for delivery this yet, but I do 0:15:12.360 --> 0:15:14.800 not think I could get the whole solution. And then 0:15:15.040 --> 0:15:17.240 if you go into twenty twenty seven for the next 0:15:17.240 --> 0:15:20.160 generation of GPUs, the Ruben GPUs, they at least for 0:15:20.200 --> 0:15:23.040 the first like stretch, are going to be very much 0:15:23.160 --> 0:15:25.600 assault out. And so I think that's like a maybe 0:15:25.600 --> 0:15:27.640 you actually have on a twenty twenty seven delivery. You 0:15:27.680 --> 0:15:30.240 have more like finding a data center shell by then. 0:15:30.800 --> 0:15:32.560 But you need you need to be in cute now 0:15:32.680 --> 0:15:35.560 for those GPUs if you want them early. So those 0:15:35.560 --> 0:15:37.840 things are those things are in demand. I'll say that 0:15:37.880 --> 0:15:41.280 for sure. And one of my greatest failures has been 0:15:41.760 --> 0:15:43.720 you know, part of my skepticism has been predicting how 0:15:43.760 --> 0:15:46.000 many GPUs we would need and a long enough horizon 0:15:46.440 --> 0:15:49.960 and it's punishing because you're constantly playing catchup. And one 0:15:49.960 --> 0:15:52.880 of our competitors put out a podcast this weekend and 0:15:53.240 --> 0:15:54.880 they mentioned somebody along the lines of the fact they 0:15:55.120 --> 0:15:57.200 had one data center and it was the data center, 0:15:57.240 --> 0:16:00.000 and that was their data center, and then as they're hungering, 0:16:00.040 --> 0:16:01.520 hungry for more compute, they had to go out and 0:16:01.520 --> 0:16:03.880 find it wherever they could. And I would say, we 0:16:03.920 --> 0:16:06.280 are in exactly the same boat. You just can't be picky. 0:16:06.360 --> 0:16:08.560 It's like you've got like a mega watt there, I'll 0:16:08.600 --> 0:16:11.840 take it, and it could be you know, not in 0:16:11.960 --> 0:16:13.520 terms that are super favorable to you. 0:16:13.400 --> 0:16:15.640 Because we'll say more about that. How are you actually 0:16:15.640 --> 0:16:17.920 going out and sourcing this stuff, because as you say, 0:16:17.960 --> 0:16:21.040 it seems to be exceptionally competitive. And at the same time, 0:16:21.080 --> 0:16:23.520 don't you guys have an insane data center in like 0:16:23.600 --> 0:16:24.920 Norway or something. 0:16:24.680 --> 0:16:27.400 And it's not enough. Yeah, and it's not enough. Yes. 0:16:27.520 --> 0:16:32.520 We go to the neo clouds, the hyperscalers, everyone, and 0:16:32.920 --> 0:16:35.680 it's a constant dialogue and they're all in competition with 0:16:35.760 --> 0:16:38.520 each other. But in some sense there must be some 0:16:38.640 --> 0:16:41.680 much bigger, shadowy competition going on behind the scenes behind 0:16:41.680 --> 0:16:44.360 these neoclouds because they are all looking for space and 0:16:44.440 --> 0:16:46.840 power and I don't know if that's the true scarce resource, 0:16:47.200 --> 0:16:49.680 and they're a kind of intermediory layer over it. I 0:16:49.680 --> 0:16:51.840 don't know what their process is like for saucing it. 0:16:52.400 --> 0:16:55.040 But yeah, they have come to us and said, this 0:16:55.160 --> 0:16:56.760 lease opened up. Can you please get back to us 0:16:56.760 --> 0:16:59.240 by the end of the day. And for commitments on 0:16:59.280 --> 0:17:02.240 a long term contract, and our contracts are a long term. 0:17:02.480 --> 0:17:05.080 This is not spot compute. This is like eight thousand 0:17:05.160 --> 0:17:08.760 GPUs for three years, four years, five years payment? Do 0:17:08.760 --> 0:17:10.440 you wanna pay half upfront? Do you want to pay 0:17:10.800 --> 0:17:12.960 some per year? A lot of different commercial terms, credit 0:17:13.040 --> 0:17:15.359 risk on both sides. It's complicated stuff. 0:17:15.560 --> 0:17:18.639 Tell us more about the counterparty risk. So it's like 0:17:18.720 --> 0:17:21.760 you come and you say you want capacity and some 0:17:21.920 --> 0:17:26.200 data center. I'm from Hudson River Trading. Who yeah, well, 0:17:26.600 --> 0:17:28.159 this is the kind of thing like this crowd. A 0:17:28.160 --> 0:17:30.439 lot of people know what Hudgson River Trading is, but 0:17:30.520 --> 0:17:33.640 maybe in San Francisco or whatever, that's not a household name, 0:17:33.720 --> 0:17:36.120 et cetera. They want to know for sure that you're 0:17:36.119 --> 0:17:38.960 gonna be good, You're gonna like pay your bills, et cetera. 0:17:39.200 --> 0:17:42.320 How do you establish to the data center that you 0:17:42.359 --> 0:17:44.200 were going to be a reliable I guess tenant. 0:17:44.280 --> 0:17:46.719 Yeah, it's a it's a definitely being a dance. It's 0:17:46.720 --> 0:17:49.080 getting better at this point. I think we've anted it 0:17:49.280 --> 0:17:51.480 enough deals enough people that I think we have that, 0:17:51.640 --> 0:17:54.040 But we've had everything from people being like, oh, you've 0:17:54.040 --> 0:17:57.280 issued bonds, what's the rating on those too? Not wanting 0:17:57.359 --> 0:18:00.119 us to sell too much of one site because if 0:18:00.119 --> 0:18:04.080 we take all their power rights and then go bust, 0:18:04.320 --> 0:18:06.040 they might have a long lead time where they can't 0:18:06.080 --> 0:18:08.640 get into a tenant and fill that. And so there's 0:18:08.640 --> 0:18:10.800 a kind of two potty problems of us where it's 0:18:10.840 --> 0:18:14.240 like they want customers, but there's presumably a lot of customers, 0:18:14.280 --> 0:18:16.600 but maybe not as many customers are willing to do 0:18:16.640 --> 0:18:19.920 the big size and pay more upfront. But you know, 0:18:20.160 --> 0:18:22.960 we're looking at their CDs is on some of these 0:18:23.119 --> 0:18:25.919 ones and thinking about how that affects our You know, 0:18:25.960 --> 0:18:28.520 maybe we should pay you three fifty an hour and 0:18:28.600 --> 0:18:31.040 take out a CDs for ten cents for our equivalent 0:18:31.040 --> 0:18:35.680 of insurance on your heavy leveraged neocloud. You're having a disruption, 0:18:36.160 --> 0:18:39.480 no names, but you know it's I think is reason 0:18:39.600 --> 0:18:42.200 to be cagy on both sides because this has all 0:18:42.320 --> 0:18:45.800 come from nothing like a year ago. We weren't there 0:18:45.920 --> 0:18:48.959 asking for it, and they didn't exist to sell it, 0:18:49.080 --> 0:18:51.919 and so the only rock is Nvidio I guess, an 0:18:51.960 --> 0:18:56.119 extremely well capitalized entity who is not going anywhere, and 0:18:56.160 --> 0:18:59.000 it's making love GPUs and we have a very positive 0:18:59.000 --> 0:19:02.439 relationship with them. I think that is also a material factor. 0:19:02.600 --> 0:19:05.480 How much optionality do you actually have on GPUs now, 0:19:05.600 --> 0:19:08.960 Like if you say you want to prioritize latency or throughput, 0:19:09.280 --> 0:19:11.919 like can you get the chips that you need to 0:19:11.960 --> 0:19:14.119 specify on one of those things? Or like do you 0:19:14.200 --> 0:19:15.440