WEBVTT - AI’s Physical Bottlenecks Beyond Chips 0:00:00.320 --> 0:00:04.479 Financial markets can no longer ignore the impacts of climate change, 0:00:04.620 --> 0:00:10.610 shifting consumer expectations and increasing regulatory scrutiny. These forces are 0:00:10.690 --> 0:00:16.730 reshaping industries, redefining risk and creating new opportunities for long-term 0:00:16.810 --> 0:00:22.810 value creation. Through conversations with industry leaders, investors and subject 0:00:22.850 --> 0:00:28.430 matter experts, we explore where sustainability is translating into measurable 0:00:28.690 --> 0:00:34.740 financial outcomes. Welcome to Sustainability Currents, brought to you by 0:00:34.780 --> 0:00:36.360 Bloomberg Intelligence. 0:00:40.560 --> 0:00:45.260 Welcome, everyone. I'm Shaheen Contractor, Senior Sustainable Finance Research Analyst, 0:00:45.680 --> 0:00:49.000 and your host for today's episode. The AI boom has 0:00:49.040 --> 0:00:52.500 been largely framed as this race for chips and compute, 0:00:53.020 --> 0:00:56.720 but as AI scales, constraints are becoming increasingly physical. From 0:00:56.800 --> 0:01:00.270 access to electricity and grid, to the infrastructure needed to 0:01:00.310 --> 0:01:03.590 support it. So to help us unpack this, I'm joined 0:01:03.630 --> 0:01:08.150 by Stephen Bird, Global Head of Thematic and Sustainability Research 0:01:08.270 --> 0:01:11.990 at Morgan Stanley. Today, we'll explore where the biggest physical 0:01:12.050 --> 0:01:16.620 bottlenecks are emerging and where investors may still be underestimating 0:01:16.880 --> 0:01:19.480 both the opportunities and the risks. Stephen, welcome. 0:01:19.700 --> 0:01:20.720 Thank you. Thank you for having me. 0:01:21.560 --> 0:01:25.160 Stephen, you've been one to argue that as AI scales, 0:01:25.260 --> 0:01:29.500 these bottlenecks are increasingly physical. I think the one thing 0:01:29.540 --> 0:01:32.690 that comes to my mind immediately is electricity. So can 0:01:32.730 --> 0:01:37.630 you tell me about how significant is this gap or 0:01:37.670 --> 0:01:40.630 problem when it comes to electricity and this AI boom? 0:01:40.970 --> 0:01:44.360 It really is a significant issue. So in our model, 0:01:44.540 --> 0:01:50.100 we see over a 30% shortfall of power for U.S. 0:01:50.140 --> 0:01:53.460 data center developers through 2028. That is a fairly massive number. 0:01:54.080 --> 0:01:57.340 We've been working on this for about four years, and 0:01:57.780 --> 0:02:01.640 over the last six months, this shortfall that we're projecting 0:02:01.660 --> 0:02:04.520 has gotten a lot worse, mostly due to physical issues, though. 0:02:04.560 --> 0:02:09.109 We think of three issues, sort of power, people, and politics. Now, 0:02:09.290 --> 0:02:11.770 all three are related to each other very, very closely, 0:02:12.230 --> 0:02:14.829 but we're really running into limits. One way to think 0:02:14.870 --> 0:02:17.320 about it simply on the power side is for years 0:02:17.400 --> 0:02:20.200 and years, the U.S. really didn't see much incremental power 0:02:20.220 --> 0:02:22.890 demand growth. So we had a little bit of slack 0:02:22.930 --> 0:02:25.589 in the system, a little bit of access available. The 0:02:25.710 --> 0:02:28.549 AI players, data center players have taken that slack. And 0:02:28.590 --> 0:02:31.230 so now we have to actually add a lot to 0:02:31.270 --> 0:02:34.130 the system to permit more data centers to be connected. 0:02:34.169 --> 0:02:35.669 And that's where we're running into problems. 0:02:36.030 --> 0:02:41.030 Okay. So when it comes to this 30% shortfall, where 0:02:41.090 --> 0:02:43.720 do we lie on sort of where the gap is? 0:02:43.790 --> 0:02:46.010 Is it we're not able to generate enough? Is it 0:02:46.060 --> 0:02:48.239 we're not able to get access to the grid? Once 0:02:48.280 --> 0:02:52.730 you get access, the transmission distribution is Or sort of where, 0:02:52.770 --> 0:02:53.430 what is the problem? 0:02:53.570 --> 0:02:55.310 Yeah, great question. It is all of the above and 0:02:55.330 --> 0:02:59.459 it does depend on the location. So for example, often though, 0:02:59.500 --> 0:03:02.420 I'd say the single most often issue is lack of transmission. 0:03:03.320 --> 0:03:06.320 You know, we often will have enough power generation, though 0:03:06.360 --> 0:03:09.100 in places like Texas, we're also running out of power generation. 0:03:09.240 --> 0:03:12.500 Often the problem is, can we get that massive amount 0:03:12.520 --> 0:03:15.180 of electricity to this new data center? And this is 0:03:15.200 --> 0:03:18.230 something new to the power world in the sense that 0:03:18.250 --> 0:03:20.919 the usage of a single data center think of it 0:03:20.980 --> 0:03:26.440 as just this concentrated demand that's unbelievably high. And utility 0:03:26.460 --> 0:03:29.839 grids are generally not really well suited for this incredibly 0:03:29.930 --> 0:03:32.690 high level of demand, where also you could see fluctuations 0:03:32.730 --> 0:03:36.930 that can be very sudden. So typically it's transmission. But 0:03:37.010 --> 0:03:40.050 in some jurisdictions like Texas, we need to build more 0:03:40.070 --> 0:03:42.070 power plants. So it is a bit of both. 0:03:42.530 --> 0:03:46.850 Okay. I guess, you know, If electricity gets scarce, whether 0:03:46.890 --> 0:03:49.150 we can produce it or not, whether it can get 0:03:49.210 --> 0:03:52.850 transmitted or not, our data center is going to be 0:03:53.070 --> 0:03:57.510 increasingly competing with households for power. I mean, we have 0:03:57.530 --> 0:04:01.690 this age-old debate on affordability, which I'm guessing feeds into 0:04:01.710 --> 0:04:05.490 at least two of your triple Ps, right? It feeds 0:04:05.550 --> 0:04:08.090 into the politics and it feeds into the people. 0:04:08.340 --> 0:04:08.860 Absolutely. 0:04:08.880 --> 0:04:13.370 So how does energy affordability... How is that going to 0:04:13.410 --> 0:04:14.560 change the equation here? 0:04:15.320 --> 0:04:19.320 Here we have seen a dramatic improvement in how the 0:04:19.380 --> 0:04:22.640 tech world is thinking about impacts to other customers. So 0:04:23.100 --> 0:04:26.880 what I really like is that developers of data centers 0:04:26.990 --> 0:04:29.609 are extremely aware of the issue you just raised, and 0:04:29.650 --> 0:04:33.670 they do not want to have any adverse impacts, especially 0:04:33.910 --> 0:04:36.010 on customer power bills. So there are a few ways 0:04:36.050 --> 0:04:37.950 to deal with that. One way is you could just 0:04:38.050 --> 0:04:39.560 not go on the grid at all. You could build 0:04:39.600 --> 0:04:42.400 your own with your own power plant, and then you 0:04:42.440 --> 0:04:46.840 can irrefutably argue you're not having an impact. Another way 0:04:46.880 --> 0:04:49.700 to do that is a fully regulated utility could set 0:04:49.740 --> 0:04:52.340 a special rate, a special what we would call a 0:04:52.380 --> 0:04:56.700 tariff rate, that's fairly high for just the data center 0:04:56.740 --> 0:04:59.669 folks so that we can show, look, all upgrades required 0:04:59.850 --> 0:05:03.450 and then some will be paid for by these data centers. 0:05:03.529 --> 0:05:05.470 But you have to make sure you do that. There 0:05:05.589 --> 0:05:07.710 are a few situations where it's a little more gray 0:05:07.750 --> 0:05:10.849 than I just laid out, And those situations often will 0:05:10.890 --> 0:05:12.469 receive a lot of political pushback. 0:05:13.310 --> 0:05:16.110 And do you feel like one is being implemented more 0:05:16.190 --> 0:05:19.070 than the other? Or are these still work in progress, 0:05:19.150 --> 0:05:20.480 things that have to be figured out? 0:05:20.500 --> 0:05:23.560 That is so fascinating. So we are essentially running out 0:05:23.600 --> 0:05:26.660 of grid access. And so by necessity, we're starting to 0:05:26.720 --> 0:05:29.599 see a huge growth in demand for just off grid 0:05:29.640 --> 0:05:30.640 because we're out of options. 0:05:31.240 --> 0:05:31.460 Okay. 0:05:31.660 --> 0:05:31.900 Okay. 0:05:31.920 --> 0:05:36.600 That's interesting. I think, you know, given how important energy 0:05:36.720 --> 0:05:42.200 obviously is to AI, Indeed. How does that change potentially 0:05:42.290 --> 0:05:45.500 which countries are best positioned to win the AI race? 0:05:45.550 --> 0:05:50.070 Because right now it's very clearly squarely divided, or not divided, 0:05:50.110 --> 0:05:51.430 but in certain countries. 0:05:51.529 --> 0:05:53.610 Yeah, absolutely. And this is a prediction that we call 0:05:53.650 --> 0:05:56.229 the two worlds of AI, where you have different spheres 0:05:56.270 --> 0:06:02.789 of influence. To compete in AI, meaning to develop frontier-level capabilities, 0:06:03.170 --> 0:06:07.250 is very, very challenging for a nation. you would need 0:06:07.270 --> 0:06:11.210 to have tremendous amount of technical talent, tremendous amount of energy, 0:06:11.830 --> 0:06:15.360 and a tremendous amount of compute and government support more broadly. 0:06:16.180 --> 0:06:19.940 Very few nations will possess all of those at a 0:06:19.980 --> 0:06:23.000 level sufficient to compete at the frontier. Though many nations 0:06:23.040 --> 0:06:25.700 will say, look, I don't need to be at the frontier. 0:06:25.860 --> 0:06:29.370 I can develop very practical solutions that will work really 0:06:29.410 --> 0:06:32.909 well for my people. An example of that is India. 0:06:33.250 --> 0:06:34.990 I was in India for the big AI summit they 0:06:35.029 --> 0:06:37.990 had several months ago. I think that strategy is really 0:06:38.170 --> 0:06:43.440 interesting and thoughtful. The government there wants open-weight models, so 0:06:43.450 --> 0:06:46.260 you understand what the models are doing, freely available to 0:06:46.279 --> 0:06:50.260 their people that function extremely well. Now, do you need 0:06:50.279 --> 0:06:52.240 to be at the frontier? Meaning, do you need to 0:06:52.300 --> 0:06:55.500 be at the level of some of these American developers 0:06:55.560 --> 0:07:00.210 where they're really going to the frontier, cutting-edge capabilities? Not necessarily. 0:07:00.550 --> 0:07:03.710 So we will see different strategies globally, though certainly the U.S. 0:07:03.730 --> 0:07:07.480 and China both have ambitions. to have the best capabilities. 0:07:07.580 --> 0:07:09.140 And there is a lot at stake in the sense 0:07:09.180 --> 0:07:12.700 that as the models become more capable, they are becoming 0:07:12.810 --> 0:07:18.870 very clearly geopolitically critical. An example, cyber. Now that these 0:07:18.930 --> 0:07:24.710 models are able to compromise many systems, these become geopolitically important. 0:07:24.730 --> 0:07:30.120 It's very important for not just countries, but businesses, individuals, 0:07:30.560 --> 0:07:32.700 to have the best protections against cyber attack. I think 0:07:32.740 --> 0:07:36.450 what we're learning now is with these AI capabilities proliferating, 0:07:36.870 --> 0:07:42.530 it's increasingly cheap for parties to attack websites and other... Cheap. 0:07:42.930 --> 0:07:47.140 Very cheap. Very. In the sense that these often open-weight 0:07:47.180 --> 0:07:50.500 models can be replicated and then you can set them 0:07:50.540 --> 0:07:53.680 forth to attack websites fairly broadly. So I say all 0:07:53.700 --> 0:07:57.580 this to say that essentially really strong capabilities on defense 0:07:58.480 --> 0:08:00.630 against these cyber attacks will increasingly be important. 0:08:01.210 --> 0:08:04.670 Okay. That makes sense. I guess, you know, given what 0:08:04.710 --> 0:08:09.150 you've said, I'm wondering the costs and benefits of AI 0:08:09.270 --> 0:08:12.510 and the energy transition. Yes. So on one hand, you 0:08:12.560 --> 0:08:16.890 have all this need for power, which basically can create 0:08:16.920 --> 0:08:18.960 economies for scale for clean energy. 0:08:19.040 --> 0:08:19.260 True. 0:08:19.860 --> 0:08:22.760 At the same time, it's also just, you know, it's 0:08:22.840 --> 0:08:25.560 anything that creates power, it's also fossil fuel. So how 0:08:25.580 --> 0:08:29.580 do you, do you see this AI boom as accelerating 0:08:29.620 --> 0:08:33.250 the energy transition or or does it have more cost 0:08:33.270 --> 0:08:33.740 than benefit? 0:08:34.280 --> 0:08:35.500 Unfortunately, I see a mix. 0:08:35.940 --> 0:08:36.220 Okay. 0:08:36.679 --> 0:08:40.020 I am an optimist. Very fortunate. It's not all bad. 0:08:40.200 --> 0:08:42.959 Let's start with the concerns, and then let's talk about 0:08:42.980 --> 0:08:46.480 the mitigants. So we've tallied up the likely carbon emissions 0:08:46.559 --> 0:08:50.239 profile of the data center sector through 2030, all the 0:08:50.280 --> 0:08:52.160 way through scope three, and we think that will be 0:08:52.220 --> 0:08:56.559 about 4.5 billion tons of incremental CO2. Say that again, sorry. 0:08:56.559 --> 0:09:01.520 4.5 billion tons cumulative through 2030, not a small amount. However, 0:09:01.960 --> 0:09:05.110 on the positive side, we are seeing a wide variety 0:09:05.130 --> 0:09:11.189 of benefits from an environmental perspective. One is much better 0:09:11.730 --> 0:09:16.300 efficiency of grid management, much better usage of renewables. So, 0:09:16.340 --> 0:09:20.500 for example, if AI can more accurately predict wind power generation, 0:09:21.040 --> 0:09:23.460 then we can better incorporate that and draw that power 0:09:23.500 --> 0:09:26.500 into the grid and use it more effectively. So I'm 0:09:26.520 --> 0:09:29.880 really excited about that. On the material sciences side of things, 0:09:30.020 --> 0:09:33.440 I think we should expect really strong breakthroughs in battery chemistry, 0:09:33.960 --> 0:09:39.260 for example, and storage is just absolutely critical to transitioning 0:09:39.300 --> 0:09:43.160 to cleaner energy. We're learning that. Even more broadly, when 0:09:43.179 --> 0:09:46.330 I think about climate risk and extreme weather risk, I 0:09:46.400 --> 0:09:48.589 am really excited, especially in the last month or two, 0:09:49.030 --> 0:09:51.140 with what I'm seeing from some of these models in 0:09:51.140 --> 0:09:55.210 terms of predictive weather capabilities that are so important to 0:09:55.210 --> 0:09:57.980 minimize damage and loss of life. So I think the 0:09:57.980 --> 0:10:00.300 benefits are going to start to tally up quite a bit, 0:10:00.340 --> 0:10:03.340 especially as we watch this non-linear rate of AI improvement. 0:10:04.040 --> 0:10:06.360 That's very exciting. And then if we think away from 0:10:06.420 --> 0:10:09.470 just environmental benefits, I think there are clearly a long 0:10:09.510 --> 0:10:12.510 list of potential benefits to humanity. On the top of 0:10:12.550 --> 0:10:15.410 that list would be life sciences for me. I think 0:10:15.550 --> 0:10:18.830 a very exciting to see what might be possible. And 0:10:19.110 --> 0:10:22.050 this may sound a bit nerdy, but in the last 0:10:22.090 --> 0:10:24.610 two weeks, the world of mathematics has been shaken to 0:10:24.670 --> 0:10:28.250 its core by AI in terms of the frontier capabilities 0:10:28.350 --> 0:10:32.450 and sort of uncovering solutions to problems mathematicians have never 0:10:32.470 --> 0:10:35.270 been able to uncover. I think we should expect similar 0:10:35.309 --> 0:10:38.809 breakthroughs in life sciences and material sciences. So those benefits 0:10:38.890 --> 0:10:41.610 over time, I think, will add up to be quite significant. 0:10:42.110 --> 0:10:44.550 I think what you said reminds me of what somebody 0:10:44.570 --> 0:10:48.230 has previously said on the podcast, which is that you know, 0:10:48.370 --> 0:10:50.929 AI can, the benefits of AI, we don't even know 0:10:50.960 --> 0:10:53.580 what they are because it evolves so fast. And he 0:10:53.600 --> 0:10:56.600 had this example of, you know, when it comes to 0:10:56.640 --> 0:11:00.200 critical materials, suppose we need a certain material for battery tech. 0:11:00.520 --> 0:11:02.900 Why can't AI just one day create it? It was as, 0:11:02.960 --> 0:11:06.120 you know, far out as that. So I think that 0:11:06.179 --> 0:11:09.060 makes a lot of sense. Maybe one follow-up. So you 0:11:09.080 --> 0:11:12.390 had this big number, 4.5 billion tons by 2030. That's 0:11:12.429 --> 0:11:17.309 a huge number. I guess my question is, Are companies 0:11:17.630 --> 0:11:22.239 or data centers doing anything to limit those emissions from 0:11:22.280 --> 0:11:26.720 an operational standpoint? The benefits are far beyond grid efficiency. 0:11:26.760 --> 0:11:29.160 We can get economies of scale in getting renewables on 0:11:29.200 --> 0:11:34.250 the grid, but are they doing anything innovative to reduce 0:11:34.330 --> 0:11:36.370 emissions on their side or right now? Is it just 0:11:36.550 --> 0:11:39.220 a rush to produce power and then deal with it? 0:11:39.530 --> 0:11:45.000 There is a wide range of technologies that are improving efficiency, but... 0:11:45.420 --> 0:11:48.040 A lot of what we're seeing is natural gas fire generation. 0:11:48.440 --> 0:11:52.700 Which is most of the... Exactly. That's the primary issue. 0:11:53.320 --> 0:11:55.970 That said, there are tremendous economic incentives to make those 0:11:55.990 --> 0:12:01.050 data centers as efficient as possible. So cooling advances, building efficiency, 0:12:01.510 --> 0:12:05.290 even just model usage, finding ways to maximize the usage 0:12:05.330 --> 0:12:08.250 of the chips in a way that is energy efficient. 0:12:08.970 --> 0:12:12.220 All of that is underway. Even chip design is increasingly 0:12:12.820 --> 0:12:17.150 with the thought of power efficiency. All that's great. But 0:12:17.390 --> 0:12:19.729 the bottom line is we still see the amount of 0:12:19.790 --> 0:12:22.230 carbon emitted to grow quite quickly. 0:12:22.270 --> 0:12:23.370 Because of the natural gas. 0:12:23.429 --> 0:12:23.849 Exactly. 0:12:24.230 --> 0:12:25.130 Generation behind. 0:12:25.340 --> 0:12:25.550 Okay. 0:12:25.880 --> 0:12:29.020 Everything you just said reminds me of one of the 0:12:29.120 --> 0:12:34.839 takeaways we saw from our survey that we released in October. 0:12:35.220 --> 0:12:36.859 So let me tell you the question and the answer. 0:12:36.920 --> 0:12:39.569 So the question was, and this was a survey of 0:12:39.590 --> 0:12:44.490 about 250 investors, about 30% owners, about 70% asset managers. 0:12:44.929 --> 0:12:48.840 So it was which which priority do you think will 0:12:48.860 --> 0:12:52.720 define sustainable investing in the years to come? Now, we 0:12:52.760 --> 0:12:57.660 had obviously AI, tariffs, energy independence, transition, etc. What I 0:12:57.700 --> 0:13:01.880 thought was interesting is that energy independence and security was 0:13:02.140 --> 0:13:04.180 far above the energy transition. 0:13:04.240 --> 0:13:04.440 Yes. 0:13:04.920 --> 0:13:06.540 And I don't think that would have been the case 0:13:06.600 --> 0:13:08.460 a few years back when we didn't have wars. 0:13:08.480 --> 0:13:08.620 Yes. 0:13:09.559 --> 0:13:12.970 What is your, maybe just a reaction to that, given 0:13:13.090 --> 0:13:16.410 a lot of this is not only driven by war, 0:13:16.460 --> 0:13:18.560 but also driven by the amount of energy that we 0:13:18.600 --> 0:13:23.060 need today. Is the transition sort of a second thought now? 0:13:23.840 --> 0:13:26.439 It is lower on the priority list for many policymakers 0:13:26.500 --> 0:13:28.689 around the world. I'd say that this speaks to a 0:13:28.870 --> 0:13:32.450 theme we think of as multipolar world, where increasingly nations 0:13:32.510 --> 0:13:37.190 are looking for higher levels of self-sufficiency. And that shows 0:13:37.290 --> 0:13:40.900 up in energy and defense in many ways, AI capabilities. 0:13:41.429 --> 0:13:43.790 And that is at the forefront of what we see. 0:13:43.809 --> 0:13:45.650 And the stakes are very high there. So I'm honestly 0:13:45.690 --> 0:13:47.449 not surprised at all by the survey results. 0:13:47.830 --> 0:13:50.190 I wasn't surprised. I just thought it was interesting. I 0:13:50.230 --> 0:13:53.449 thought it's something we would not have seen in probably 0:13:53.450 --> 0:13:56.990 2023 when everyone was just saying, oh, she transitioned. Yay. 0:13:57.010 --> 0:13:59.690 The world is very different now. You know, in many ways, 0:13:59.770 --> 0:14:03.050 the stakes around AI are so high to the positive 0:14:03.070 --> 0:14:07.010 and negative, right, that they tend to dwarf other considerations. 0:14:07.050 --> 0:14:09.319 And I actually, I agree with that in the sense 0:14:09.380 --> 0:14:15.199 that AI holds tremendous promise and risk. And the nonlinear 0:14:15.240 --> 0:14:19.080 rate of that technology improvement really means we have to 0:14:19.120 --> 0:14:22.130 take that extremely seriously. And we need all eyes focused 0:14:22.870 --> 0:14:24.430 on that, both the positives and the negatives. 0:14:24.590 --> 0:14:28.790 All eyes focused comes down to my next question on priorities. 0:14:28.830 --> 0:14:31.330 And I think I ran this by you, but another 0:14:31.370 --> 0:14:34.790 interesting finding from our survey. So when we split the 0:14:34.840 --> 0:14:37.180 data in terms of, you know, what is the biggest 0:14:37.240 --> 0:14:44.300 priority for sustainable investing? Asset owners seem to factor AI 0:14:44.420 --> 0:14:47.860 as less of a priority than managers. Asset managers had 0:14:47.960 --> 0:14:52.140 AI at the top of the list, 34% thought AI 0:14:52.260 --> 0:14:56.920 would be the biggest thing shaping sustainable investing for asset owners. 0:14:57.360 --> 0:15:00.120 It was in the middle and on top was human capital, 0:15:00.230 --> 0:15:06.850 things around energy, independence. That confounded me. I was wondering 0:15:06.890 --> 0:15:10.730 if you had any thoughts given, I think everybody, Everybody 0:15:10.870 --> 0:15:14.510 would care about AI, maybe not equally, but quite a 0:15:14.530 --> 0:15:16.090 bit more than what I saw. 0:15:16.230 --> 0:15:19.380 I think that is fascinating. What I observe is that 0:15:19.500 --> 0:15:22.660 asset managers have to be extremely focused on the second 0:15:22.680 --> 0:15:26.640 derivative and on rates of change very quickly because stocks 0:15:27.380 --> 0:15:31.910 move in valuation so quickly based on expectations changes and 0:15:31.930 --> 0:15:35.310 you have to be laser focused on that rate of change. 0:15:35.910 --> 0:15:38.390 For an asset owner, that of course does matter a 0:15:38.830 --> 0:15:41.210 but it's not as if their asset is real-time getting 0:15:41.530 --> 0:15:46.010 tremendously repriced by millions of investors real-time. So that's, I think, 0:15:46.410 --> 0:15:50.340 the difference. I noticed the velocity of that among asset 0:15:50.380 --> 0:15:54.100 managers has increased as investors are thinking about this technology. 0:15:54.920 --> 0:15:57.320 That said, to your point, if you're an asset owner, 0:15:57.380 --> 0:16:00.020 you have to be focused on this. And there are 0:16:00.080 --> 0:16:03.359 risks there and opportunities that need to be considered. I 0:16:03.380 --> 0:16:05.610 would just say the asset managers just have this heightened 0:16:07.040 --> 0:16:10.470 because of just how quickly things can get repriced in 0:16:10.490 --> 0:16:11.330 the markets. Okay. 0:16:11.390 --> 0:16:17.090 So maybe faster to not only just react, but to integrate. 0:16:17.130 --> 0:16:19.270 Really to price it in. Exactly. To integrate and price 0:16:19.290 --> 0:16:21.930 it in. That's right. That said, I would say just 0:16:22.270 --> 0:16:26.200 stepping way back, I observe of all kinds of asset owners, 0:16:27.080 --> 0:16:31.680 whether it's a stock or a physical asset, we humans 0:16:31.760 --> 0:16:36.740 don't do a great job in nonlinear thinking. And... You know, 0:16:37.220 --> 0:16:39.500 this is so important when it comes to AI. What 0:16:39.580 --> 0:16:42.900 AI can do today, a year ago, I would have doubted, 0:16:43.580 --> 0:16:46.890 or two years ago. Where are we headed? I suspect 0:16:47.170 --> 0:16:50.510 the degree of capability gain is being much, much, much greater, 0:16:50.530 --> 0:16:54.810 especially given recursive self-improvement of these models. We're really not 0:16:55.030 --> 0:16:57.210 pricing a lot of that in at this point. And 0:16:57.550 --> 0:16:59.110 I know that's challenging for all of us to think 0:16:59.150 --> 0:17:01.250 in those nonlinear terms, but that's really what we have 0:17:01.270 --> 0:17:01.790 to try to do. 0:17:01.820 --> 0:17:02.540 Okay. 0:17:02.920 --> 0:17:04.880 Okay, maybe coming back to power. 0:17:05.619 --> 0:17:05.800 So. 0:17:06.400 --> 0:17:08.200 BYOP, bring your own power. 0:17:08.260 --> 0:17:08.480 Yes. 0:17:09.270 --> 0:17:13.790 So we're increasingly hearing about data centers securing or even 0:17:13.869 --> 0:17:17.610 developing their own power. My question is, how far can 0:17:17.650 --> 0:17:22.010 this go? Can the biggest technology companies become energy companies? 0:17:23.230 --> 0:17:28.100 Yes, is your answer. And by necessity in the sense that, well, 0:17:28.119 --> 0:17:30.300 let's put some numbers around it. Numbers are always fun 0:17:30.320 --> 0:17:33.780 in this. You know, to build a data center can 0:17:33.820 --> 0:17:38.520 cost $ 50 a watt. to build a power system could 0:17:38.580 --> 0:17:43.040 cost three, four, five, six dollars a watt. So yes, 0:17:43.100 --> 0:17:47.400 it's expensive, but if that spending three to four, five, six, 0:17:47.420 --> 0:17:51.050 whatever it is per watt on that energy system allows 0:17:51.090 --> 0:17:56.010 you to deploy intelligence, which has tremendous economic value that's much, 0:17:56.070 --> 0:17:57.649 much greater than the cost of that power, you're going 0:17:57.670 --> 0:18:00.690 to find a way to do that. However, what we're 0:18:00.710 --> 0:18:03.880 learning is The two worlds don't know each other very well. 0:18:03.940 --> 0:18:06.260 The world of power doesn't understand the tech world and 0:18:06.280 --> 0:18:08.939 vice versa. We're having a lot of painful learnings. Let 0:18:08.960 --> 0:18:11.560 me give you an example. The engineers for the AI 0:18:11.680 --> 0:18:14.280 community sat down with the power engineers to talk about 0:18:14.420 --> 0:18:19.139 off-grid needs. You can't rely on the grid. What the 0:18:19.180 --> 0:18:22.679 power engineers quickly realized was the fluctuation in power usage 0:18:22.800 --> 0:18:25.119 at a data center could be so rapid that a 0:18:25.220 --> 0:18:28.920 turbine alone would really be inadequate. It just wouldn't ramp 0:18:29.030 --> 0:18:31.230 up and down fast enough. So you have to add 0:18:31.250 --> 0:18:34.330 a bunch of other equipment, including batteries and capacitors and 0:18:34.390 --> 0:18:38.740 software to make that work. Actually, quite a complicated engineering challenge. 0:18:39.000 --> 0:18:41.580 It will get sorted out. But what concerns me is 0:18:41.619 --> 0:18:45.439 the pace of announcements of those deals has been slow. 0:18:45.900 --> 0:18:48.679 And so this bottleneck concern we have just appears to 0:18:48.720 --> 0:18:51.230 get worse month after month because we're just not seeing 0:18:51.250 --> 0:18:53.650 the kind of volume of off-grid deals that we need 0:18:53.690 --> 0:18:54.440 to see. Okay. 0:18:54.580 --> 0:18:57.040 And that takes a lot of shared learning, I'm guessing, 0:18:57.080 --> 0:18:59.340 because this is the first time that's probably happening. 0:18:59.440 --> 0:19:02.179 Exactly. First time. That's right. We just don't have that 0:19:02.220 --> 0:19:06.700 kind of experience in developing these huge microgrids. We just 0:19:06.770 --> 0:19:07.199 don't have that. 0:19:07.240 --> 0:19:13.230 Okay. That's interesting. I guess coming to the questions on 0:19:13.310 --> 0:19:16.570 the U.S. midterms. Now, I think there are going to 0:19:16.650 --> 0:19:19.250 be a number of things that are going to be 0:19:19.290 --> 0:19:22.470 interesting to watch. The main one, I think, is affordability. 0:19:23.150 --> 0:19:23.310 Yes. 0:19:23.530 --> 0:19:27.190 I guess, how do you think the midterm elections will 0:19:27.330 --> 0:19:30.770 change the energy game when it comes to AI or 0:19:30.810 --> 0:19:32.230 just the game when it comes to AI? 0:19:32.270 --> 0:19:36.169 Absolutely. We've done a lot of work looking at American 0:19:36.210 --> 0:19:39.990 voter priorities. And to your point, affordability is important. And 0:19:40.060 --> 0:19:44.740 also opposition to data centers, especially anything that might be 0:19:44.980 --> 0:19:49.000 located near where people live. There's a lot of opposition. 0:19:49.990 --> 0:19:53.440 Both parties are discussing ways to make sure that data 0:19:53.460 --> 0:19:56.300 centers are developed in a way that really benefits communities, 0:19:56.800 --> 0:20:01.560 addresses concerns like water usage, energy bills, etc. To your point, 0:20:01.580 --> 0:20:03.440 let's cut right to it in terms of the election. 0:20:04.040 --> 0:20:07.080 I'd say the level of the election that we're most 0:20:07.119 --> 0:20:11.480 focused on for AI growth is is at the governor level, 0:20:11.619 --> 0:20:16.680 and let me explain why. Governors do have significant authority 0:20:16.850 --> 0:20:21.609 in considering whether to approve data center projects. Recently, we 0:20:21.650 --> 0:20:26.030 saw Governor Abbott in Texas essentially order a pause while 0:20:26.070 --> 0:20:29.330 the state of Texas reviews community impacts and other impacts 0:20:29.770 --> 0:20:33.060 from data centers. I suspect we will see some governors 0:20:33.230 --> 0:20:36.199 after the election, newly elected governors, say, I want to 0:20:36.240 --> 0:20:39.160 go through a blank month review, maybe it's a six-month review, 0:20:39.980 --> 0:20:42.080 of all of the data center projects in the state, 0:20:42.119 --> 0:20:47.880 let's look at impacts to customer power bills, property taxes, jobs, 0:20:48.560 --> 0:20:51.619 a wide range of dynamics. And the new one or 0:20:51.680 --> 0:20:55.120 new were is what are the net benefits that these 0:20:55.220 --> 0:20:58.210 AI players are bringing to the state? And we've noticed 0:20:58.230 --> 0:21:01.190 that that's been somewhat unequal among states. So for example, 0:21:02.030 --> 0:21:05.450 in Louisiana, we're aware of some projects where because of 0:21:05.490 --> 0:21:09.590 these data center projects, teachers and people in police and 0:21:09.609 --> 0:21:13.149 fire services are receiving significant raises as a result of 0:21:13.190 --> 0:21:18.280 these data center projects. How's that actually? Well, it's interesting. 0:21:19.260 --> 0:21:21.399 In a parish in Louisiana where a data center is 0:21:21.420 --> 0:21:25.740 going to get built, the tech firm backing that project 0:21:26.480 --> 0:21:30.590 is willing to provide the money to the community to 0:21:30.690 --> 0:21:34.710 permit raises of all of these folks. And that's a 0:21:34.750 --> 0:21:37.340 big deal. But in other states, we're seeing very little 0:21:37.380 --> 0:21:40.000 of that. It would be tempting for a new governor 0:21:40.040 --> 0:21:42.399 to say, look, I want a level set. I want 0:21:42.420 --> 0:21:44.369 to look at what other states are receiving, and I 0:21:44.410 --> 0:21:47.210 want to see what our state should be receiving in 0:21:47.270 --> 0:21:50.070 terms of support. In the state of New York, Governor 0:21:50.109 --> 0:21:54.250 Hochul basically recently said that she believed the appropriate level 0:21:54.290 --> 0:21:59.040 of community support was about $ 1 million per megawatt of project, 0:21:59.180 --> 0:22:02.280 and that's significant. So I think we may see more 0:22:02.300 --> 0:22:04.600 of that. So it's really the governor level that I 0:22:04.640 --> 0:22:06.820 think about from an AI point of view. When we 0:22:06.840 --> 0:22:10.179 think about Congress, of course, the challenge would be anything 0:22:10.200 --> 0:22:12.970 that would inhibit AI growth may very well be vetoed 0:22:12.990 --> 0:22:16.410 by President Trump. And so it really comes back down 0:22:16.470 --> 0:22:18.750 to the state level and also at the local level. 0:22:18.790 --> 0:22:24.540 We could see politicians reject zoning requests for data centers, 0:22:24.640 --> 0:22:27.540 air permit requests, a variety of stumbling blocks at the 0:22:27.580 --> 0:22:28.960 local level as well. Okay. 0:22:29.440 --> 0:22:33.750 This sounds very disjointed, if I have to say. I 0:22:33.790 --> 0:22:39.290 guess what policy changes could have the biggest impact on 0:22:39.310 --> 0:22:41.429