WEBVTT 00:00:00.000 --> 00:00:03.419 artificial intelligence. It's really not just 00:00:03.419 --> 00:00:05.700 science fiction anymore, is it? It's kind of 00:00:05.700 --> 00:00:08.080 woven into our daily lives now. Oh, absolutely. 00:00:08.240 --> 00:00:10.580 It's everywhere. Your digital assistant, how 00:00:10.580 --> 00:00:13.900 you shop online, even those generative AI tools 00:00:13.900 --> 00:00:16.239 like ChatGPT that so many people are using. Right, 00:00:16.379 --> 00:00:19.280 exactly. It feels incredibly fast, almost seamless. 00:00:19.539 --> 00:00:22.679 How integral it's become. Genuinely revolutionary. 00:00:22.899 --> 00:00:25.300 It is revolutionary, definitely. But that speed, 00:00:25.559 --> 00:00:27.839 that innovation. Well, it comes with a hidden 00:00:27.839 --> 00:00:30.480 environmental cost, a pretty heavy one. OK. And 00:00:30.480 --> 00:00:32.619 our sources today really dig into that. They 00:00:32.619 --> 00:00:37.280 explore this massive growing footprint, not just 00:00:37.280 --> 00:00:40.380 the software, but the real world demands on energy, 00:00:40.719 --> 00:00:43.049 water. hardware. Yeah. And that's our mission 00:00:43.049 --> 00:00:45.229 today, right, to try and quantify that demand 00:00:45.229 --> 00:00:47.170 because the scale we're talking about here, it's 00:00:47.170 --> 00:00:49.329 pretty staggering. We're not talking small changes. 00:00:49.509 --> 00:00:51.109 Not at all. Yeah. Let's start with electricity. 00:00:51.310 --> 00:00:53.869 Just think about this. In 2023, data centers 00:00:53.869 --> 00:00:56.670 were already pulling 4 .4 percent of all electricity 00:00:56.670 --> 00:00:59.630 in the U .S. 4 .4 percent. That's huge already. 00:01:00.009 --> 00:01:03.009 It is. Enough to power whole states. But the 00:01:03.009 --> 00:01:05.989 projections, they suggest that number could triple 00:01:05.989 --> 00:01:11.409 by 2028. Triple. And AI is the main driver behind 00:01:11.409 --> 00:01:14.010 that surge. Triple is, yeah. Hard to even picture. 00:01:14.129 --> 00:01:17.430 And if you look globally, the strain on power 00:01:17.430 --> 00:01:20.930 grids sounds almost, well, unimaginable. Exactly. 00:01:21.430 --> 00:01:23.409 Some estimates are suggesting data centers could 00:01:23.409 --> 00:01:25.810 account for maybe 20 % of the entire world's 00:01:25.810 --> 00:01:29.430 electricity use by 2030, maybe 2035. 20%. That's 00:01:29.430 --> 00:01:31.620 not just. Like needing a few more power plants. 00:01:31.939 --> 00:01:33.840 That fundamentally changes global energy demand. 00:01:33.939 --> 00:01:36.420 It really does. And you can see this growth directly 00:01:36.420 --> 00:01:39.200 in the AI specific hardware. The energy used 00:01:39.200 --> 00:01:41.120 just for those servers. It jumped from under 00:01:41.120 --> 00:01:44.920 two terawatt hours back in 2017 to over 40 terawatt 00:01:44.920 --> 00:01:48.019 hours in 2023. 40. That's a massive leap. It's 00:01:48.019 --> 00:01:50.670 an explosion. And it's tied directly to how these 00:01:50.670 --> 00:01:52.810 big AI models, these large language models, or 00:01:52.810 --> 00:01:55.409 LLMs, are actually built and then how we use 00:01:55.409 --> 00:01:57.290 them every day. OK, let's get into that, the 00:01:57.290 --> 00:01:59.430 mechanics. Researchers often talk about two main 00:01:59.430 --> 00:02:01.829 phases, training the model and then inference. 00:02:02.370 --> 00:02:04.230 Why is the training part so incredibly power 00:02:04.230 --> 00:02:07.329 hungry? Well, training involves this incredibly 00:02:07.329 --> 00:02:11.409 complex process of adjusting billions, sometimes 00:02:11.409 --> 00:02:14.870 trillions, of parameters in the model. It's like 00:02:14.870 --> 00:02:16.689 trying to teach it everything all at once by 00:02:16.689 --> 00:02:20.169 crunching massive data sets repeatedly. This 00:02:20.169 --> 00:02:23.270 needs serious computing power. High -performance 00:02:23.270 --> 00:02:27.289 computing, HPC, we're talking thousands of specialized 00:02:27.289 --> 00:02:31.169 chips, GPUs and TPUs running flat out nonstop 00:02:31.169 --> 00:02:33.729 for weeks or even months. That's constantly running. 00:02:34.150 --> 00:02:36.650 Constantly. So that initial training phase, it's 00:02:36.650 --> 00:02:39.189 one of the most resource -intensive computing 00:02:39.189 --> 00:02:41.870 tasks humans have ever really undertaken. And 00:02:41.870 --> 00:02:44.009 we have a number for that, right, for GPT -A4. 00:02:44.110 --> 00:02:46.389 That really puts it in perspective. Yeah. The 00:02:46.389 --> 00:02:49.379 estimates for training GPT -4 alone... are around 00:02:49.379 --> 00:02:52.060 50 gigawatt hours of energy. 50 gigawatt hours. 00:02:52.180 --> 00:02:55.680 And costing over $100 million just for the training 00:02:55.680 --> 00:02:58.199 run. And to make that concrete for you listening, 00:02:58.439 --> 00:03:00.599 that amount of energy could power, what, the 00:03:00.599 --> 00:03:02.659 whole San Francisco metro area for about three 00:03:02.659 --> 00:03:04.719 days? That's the estimate, yeah. Just for that 00:03:04.719 --> 00:03:07.300 one initial training session. OK, so that's the 00:03:07.300 --> 00:03:09.419 huge one -time upfront cost. But you said the 00:03:09.419 --> 00:03:12.240 energy demand shifts after that. It shifts dramatically. 00:03:12.319 --> 00:03:15.219 And this is where it really impacts daily use 00:03:15.219 --> 00:03:18.530 and the overall footprint. After training, the 00:03:18.530 --> 00:03:21.189 vast majority of the energy, like 80 to 90 percent, 00:03:21.389 --> 00:03:23.949 according to research, goes into inference. Inference. 00:03:24.129 --> 00:03:26.030 That's when we actually use it, right? When I 00:03:26.030 --> 00:03:28.930 type a question into ChatGPT or ask an AI to 00:03:28.930 --> 00:03:31.449 generate an image? Exactly. It's you, the user, 00:03:32.030 --> 00:03:35.189 making a request. Yeah. Because millions, maybe 00:03:35.189 --> 00:03:37.210 billions, of these requests are happening all 00:03:37.210 --> 00:03:40.770 the time now, that's where the bulk of the continuous 00:03:40.770 --> 00:03:42.919 energy drain comes from. So all those seemingly 00:03:42.919 --> 00:03:45.580 small interactions summarize this, write that, 00:03:45.659 --> 00:03:48.599 create this picture, they add up massively. They 00:03:48.599 --> 00:03:52.300 really do. It's pretty striking that a single 00:03:52.300 --> 00:03:55.240 query to something like ChatGPT can use about 00:03:55.240 --> 00:03:57.080 five times more electricity than just doing as 00:03:57.080 --> 00:03:59.159 a full Google search. Five times more, just for 00:03:59.159 --> 00:04:01.500 one question. And what's really interesting and 00:04:01.500 --> 00:04:04.319 maybe a bit concerning is how much that energy 00:04:04.319 --> 00:04:06.699 use can swing depending on the model. How so? 00:04:07.000 --> 00:04:08.800 Well, it depends heavily on the size of the model 00:04:08.800 --> 00:04:11.379 you're querying and how complex the task is. 00:04:11.819 --> 00:04:14.219 Our sources looked at open source models, like 00:04:14.219 --> 00:04:16.800 Meta's Lama series, to try and get a handle on 00:04:16.800 --> 00:04:19.100 this variance. OK, yeah. Let's look at that data. 00:04:19.180 --> 00:04:22.379 That sounds like a key point. Right. So take 00:04:22.379 --> 00:04:26.740 a smaller text model, like Lama 3 .1 8b. The 00:04:26.740 --> 00:04:30.279 8b means 8 billion parameters. A single response 00:04:30.279 --> 00:04:33.220 from that might use around 114 joules of energy. 00:04:33.439 --> 00:04:36.560 Joules are tiny, right? Yeah. 114. How can we 00:04:36.560 --> 00:04:39.259 picture that? Researchers use comparisons. So 00:04:39.259 --> 00:04:42.060 114 joules is like riding an e -bike for maybe 00:04:42.060 --> 00:04:44.699 six feet, or running your microwave for just 00:04:44.699 --> 00:04:46.500 a fraction of a second, like a tenth of a second. 00:04:46.579 --> 00:04:49.519 Sounds tiny. OK, almost nothing. Seems efficient. 00:04:49.639 --> 00:04:52.100 Seems efficient. Now, compare that to a much 00:04:52.100 --> 00:04:56.360 bigger model, Lama 3 .1405B. That's over 50 times 00:04:56.360 --> 00:04:59.100 larger in terms of parameters. OK. A single response 00:04:59.100 --> 00:05:02.060 from that model jumps to around 6 ,706 joules. 00:05:02.220 --> 00:05:06.930 Whoa, OK. From 114 to over 6 ,700, that's a huge 00:05:06.930 --> 00:05:09.189 leap. What's the microwave comparison for that? 00:05:09.370 --> 00:05:11.149 That's like running your standard microwave for 00:05:11.149 --> 00:05:13.170 about eight full seconds. Eight seconds versus 00:05:13.170 --> 00:05:15.509 a tenth of a second, just for a similar text 00:05:15.509 --> 00:05:17.529 output. Potentially, yeah. The researchers use 00:05:17.529 --> 00:05:19.430 microwaves because they're a familiar high -power 00:05:19.430 --> 00:05:21.089 appliance. It gives you a more visceral feel 00:05:21.089 --> 00:05:23.589 for the difference in energy consumption. So 00:05:23.589 --> 00:05:27.019 why? Why use the big, energy -hungry model if 00:05:27.019 --> 00:05:29.620 a smaller one could do the job for many simple 00:05:29.620 --> 00:05:32.720 queries? Why the massive jump in energy use? 00:05:33.019 --> 00:05:36.120 Well, the thinking, which has some backing, is 00:05:36.120 --> 00:05:39.639 that bigger models are generally smarter, more 00:05:39.639 --> 00:05:43.139 versatile, less likely to make errors or hallucinate. 00:05:43.459 --> 00:05:46.879 They can handle a wider range of tasks. So capability 00:05:46.879 --> 00:05:50.709 trumps efficiency? Often, yes. A platform might 00:05:50.709 --> 00:05:53.209 default to the bigger, more capable model, even 00:05:53.209 --> 00:05:55.610 if your specific query didn't really need it. 00:05:55.889 --> 00:05:57.870 So you end up paying the higher energy cost for 00:05:57.870 --> 00:06:01.170 every single request, basically. Scale is kind 00:06:01.170 --> 00:06:02.930 of the enemy of efficiency here. And we've only 00:06:02.930 --> 00:06:05.050 talked about text so far. What about generating 00:06:05.050 --> 00:06:08.040 images or even video? Oh, yeah. That's where 00:06:08.040 --> 00:06:09.819 the numbers just go into another league entirely. 00:06:10.500 --> 00:06:12.439 Generating, say, a five second video with an 00:06:12.439 --> 00:06:15.040 open source model like Cog Video X. That single 00:06:15.040 --> 00:06:18.180 task takes about 3 .4 million joules. Million? 00:06:18.339 --> 00:06:20.959 Wow. That's over 700 times the energy needed 00:06:20.959 --> 00:06:23.079 to generate a pretty high quality static image, 00:06:23.540 --> 00:06:25.759 or sticking with a microwave comparison. Let 00:06:25.759 --> 00:06:27.600 me guess, it's longer than eight seconds. It's 00:06:27.600 --> 00:06:30.180 like running your microwave continuously for 00:06:30.180 --> 00:06:32.720 over an hour, just for one five second video 00:06:32.720 --> 00:06:35.100 clip. OK, that's... That's genuinely shocking. 00:06:35.399 --> 00:06:38.060 An hour of microwave power for five seconds of 00:06:38.060 --> 00:06:40.360 video. Yeah. It really highlights how the type 00:06:40.360 --> 00:06:43.959 of AI task drastically changes the impact. Absolutely. 00:06:44.279 --> 00:06:46.819 It's not just if you use AI, but what you're 00:06:46.819 --> 00:06:50.879 asking it to do. And this enormous energy use 00:06:50.879 --> 00:06:55.259 leads directly to another major, often overlooked 00:06:55.259 --> 00:06:58.699 issue. Which is? Water. All those chips working 00:06:58.699 --> 00:07:01.100 so hard generate a phenomenal amount of heat. 00:07:01.360 --> 00:07:03.420 Right, they need cooling, massive cooling systems. 00:07:03.620 --> 00:07:05.740 Exactly, and modern data centers rely heavily 00:07:05.740 --> 00:07:07.920 on water for that cooling. We're talking vast 00:07:07.920 --> 00:07:10.379 amounts of fresh water. The estimate is around 00:07:10.379 --> 00:07:13.160 two liters of water for every single kilowatt 00:07:13.160 --> 00:07:15.220 hour of energy the data center consumes. Two 00:07:15.220 --> 00:07:17.399 liters per kilowatt hour, and is that water recycled? 00:07:17.579 --> 00:07:19.839 Often, no. It's what's called consumptive use. 00:07:20.199 --> 00:07:22.620 A lot of it evaporates in cooling towers or becomes 00:07:22.620 --> 00:07:24.500 contaminated with chemicals and can't just be 00:07:24.500 --> 00:07:26.439 put back into the local water system. So it's 00:07:26.439 --> 00:07:28.740 competing directly with local water needs. Drinking 00:07:28.740 --> 00:07:31.860 water, agriculture. Directly. And we're seeing 00:07:31.860 --> 00:07:34.279 the impact in corporate water usage figures. 00:07:35.040 --> 00:07:37.939 Google's data center water use jumped 20 percent 00:07:37.939 --> 00:07:41.639 between 2021 and 2022. They use five billion 00:07:41.639 --> 00:07:45.339 gallons in 2022 alone. 5 billion gallons. And 00:07:45.339 --> 00:07:48.079 Microsoft's use went up 34 % in the same period. 00:07:48.379 --> 00:07:50.879 These are huge numbers, but they become really 00:07:50.879 --> 00:07:52.899 concrete when you look at local impacts. Like 00:07:52.899 --> 00:07:54.819 in Newton County, Georgia? That was a case study 00:07:54.819 --> 00:07:57.300 in the source material, wasn't it? It was. A 00:07:57.300 --> 00:07:59.899 really stark example. Metta built a huge data 00:07:59.899 --> 00:08:02.600 center there. And almost immediately, nearby 00:08:02.600 --> 00:08:04.959 residents who relied on private wells started 00:08:04.959 --> 00:08:07.319 having serious problems. Yeah, the local reporting 00:08:07.319 --> 00:08:09.779 described people's taps running dry or the water 00:08:09.779 --> 00:08:12.259 coming out full of sediment. Wells they relied 00:08:12.259 --> 00:08:14.980 on for years suddenly became unusable. And replacing 00:08:14.980 --> 00:08:17.259 a well is incredibly expensive for a homeowner. 00:08:17.389 --> 00:08:19.850 Meanwhile, the Meta facility is estimated to 00:08:19.850 --> 00:08:21.910 be using around half a million gallons of water 00:08:21.910 --> 00:08:24.410 per day. Half a million gallons a day. But it's 00:08:24.410 --> 00:08:27.389 not just that one facility. The systemic issue 00:08:27.389 --> 00:08:29.310 is that Newton County is already projected to 00:08:29.310 --> 00:08:32.490 face a water deficit by 2030. Residents are looking 00:08:32.490 --> 00:08:34.789 at potential rationing and water rates are expected 00:08:34.789 --> 00:08:38.009 to jump by about 33%. That's terrifying for the 00:08:38.009 --> 00:08:40.350 people living there. And yet, new data centers 00:08:40.350 --> 00:08:43.620 are still being proposed. Oh yeah. The reporting 00:08:43.620 --> 00:08:45.899 mentioned new applicants requesting permits for 00:08:45.899 --> 00:08:48.980 facilities needing up to six million or even 00:08:48.980 --> 00:08:52.100 nine million gallons of water a day. Nine million? 00:08:52.159 --> 00:08:54.659 That's more than the entire county uses currently, 00:08:54.740 --> 00:08:56.700 isn't it? That's what the sources suggest, yeah. 00:08:57.019 --> 00:08:59.320 It points to a fundamental disconnect in planning. 00:08:59.899 --> 00:09:03.519 How do the tech companies justify this? Prioritizing 00:09:03.519 --> 00:09:06.620 cheap electricity over local water security seems... 00:09:06.639 --> 00:09:09.399 Hard to defend. Is it just about the jobs and 00:09:09.399 --> 00:09:11.419 tax revenue they bring? That's the argument, 00:09:11.600 --> 00:09:14.740 jobs and taxes. But critics argue those benefits 00:09:14.740 --> 00:09:17.039 often don't cover the immense strain on local 00:09:17.039 --> 00:09:19.639 infrastructure, like water and the power grid. 00:09:20.100 --> 00:09:22.759 The core problem seems to be site selection criteria. 00:09:22.940 --> 00:09:25.360 Meaning? Companies often prioritize places with 00:09:25.360 --> 00:09:28.560 the cheapest, most abundant electricity. Water 00:09:28.560 --> 00:09:31.100 availability often seems to be a secondary concern 00:09:31.100 --> 00:09:33.480 even if the location is already water -stressed 00:09:33.480 --> 00:09:36.159 like say Phoenix, Arizona or parts of Chile or 00:09:36.159 --> 00:09:38.440 Uruguay. So they build in deserts or drought 00:09:38.440 --> 00:09:41.039 -prone areas because the power is cheap. It appears 00:09:41.039 --> 00:09:43.820 that way sometimes, yeah. Water is treated almost 00:09:43.820 --> 00:09:47.080 like an externality, a cost someone else often 00:09:47.080 --> 00:09:49.480 the local community has to bear. And this lack 00:09:49.480 --> 00:09:52.000 of foresight, this lack of maybe community consultation 00:09:52.000 --> 00:09:54.879 or impact assessment links back to another big 00:09:54.879 --> 00:09:57.279 issue, doesn't it? The accountability gap. Exactly. 00:09:57.399 --> 00:10:00.179 The water issue is partly visible because, well, 00:10:00.200 --> 00:10:03.080 you can see the impact on wells. But the energy 00:10:03.080 --> 00:10:05.500 consumption and the resulting carbon emissions, 00:10:05.919 --> 00:10:08.000 that's often much harder to track. It's kind 00:10:08.000 --> 00:10:10.899 of a black box. Because the big AI companies, 00:10:11.159 --> 00:10:15.159 OpenAI, Google, Microsoft, they run... closed 00:10:15.159 --> 00:10:16.820 models, right? They don't share all the details. 00:10:17.080 --> 00:10:19.519 Precisely. Key operational data, like the exact 00:10:19.519 --> 00:10:21.679 size of the models, how many queries they handle, 00:10:22.100 --> 00:10:24.820 their precise energy use per query, the energy 00:10:24.820 --> 00:10:27.860 mix they use. A lot of that is guarded as proprietary, 00:10:28.059 --> 00:10:30.159 as a trade secret. So researchers trying to figure 00:10:30.159 --> 00:10:32.500 out the true environmental costs have to essentially 00:10:32.500 --> 00:10:35.159 guess or reverse engineer the numbers? Pretty 00:10:35.159 --> 00:10:37.179 much. They have to make estimates based on the 00:10:37.179 --> 00:10:39.240 open source models, like we discussed with Llama, 00:10:39.679 --> 00:10:41.940 or based on published papers, which might be 00:10:41.940 --> 00:10:44.779 out of date. It makes getting accurate real -time 00:10:44.779 --> 00:10:47.539 figures incredibly difficult. And this secrecy 00:10:47.539 --> 00:10:50.220 matters hugely because of the type of power these 00:10:50.220 --> 00:10:53.899 data centers need. Yes. They need power 247, 00:10:54.059 --> 00:10:58.080 365 days a year, non -stop. You can't run these 00:10:58.080 --> 00:11:01.320 massive AI operations solely on intermittent 00:11:01.320 --> 00:11:04.360 renewables like solar and wind, not with current 00:11:04.360 --> 00:11:07.139 grid technology and storage. So they need that 00:11:07.139 --> 00:11:10.940 constant, reliable baseload power. Which often 00:11:10.940 --> 00:11:13.940 means? Often means fossil fuels. Natural gas, 00:11:14.340 --> 00:11:15.980 sometimes coal, depending on the region's grid. 00:11:16.419 --> 00:11:18.460 One study found the electricity used by data 00:11:18.460 --> 00:11:21.259 centers had a carbon intensity that was, on average, 00:11:21.779 --> 00:11:24.879 48 % higher than the U .S. grid average. 48 % 00:11:24.879 --> 00:11:26.840 higher. So they're not just using a lot of power, 00:11:26.960 --> 00:11:29.220 they're often using dirtier power. That's the 00:11:29.220 --> 00:11:31.620 implication. And data centers tend to cluster 00:11:31.620 --> 00:11:33.820 geographically, like in Virginia, which has the 00:11:33.820 --> 00:11:35.940 highest concentration in the U .S. And Virginia's 00:11:35.940 --> 00:11:38.059 grid relies heavily on natural gas, doesn't it? 00:11:38.200 --> 00:11:40.960 It does. Plus, you have the issue of backup power. 00:11:41.320 --> 00:11:43.220 When there are grid fluctuations or failures, 00:11:43.460 --> 00:11:45.980 data centers often rely on massive diesel generators 00:11:45.980 --> 00:11:48.039 to keep things running. That adds another layer 00:11:48.039 --> 00:11:50.879 of emissions. It seems like a cycle of high demand 00:11:50.879 --> 00:11:54.779 met by often carbon -intensive energy. Are governments 00:11:54.779 --> 00:11:56.960 or regulators starting to push back on this lack 00:11:56.960 --> 00:11:59.759 of transparency? There are signs of movement, 00:12:00.000 --> 00:12:02.519 yes. In the U .S. there's a proposed bill that 00:12:02.519 --> 00:12:04.879 would require federal agencies to assess the 00:12:04.879 --> 00:12:08.960 environmental impact of their own AI use. And 00:12:08.960 --> 00:12:12.360 the EU's AI Act includes requirements for developers 00:12:12.360 --> 00:12:15.340 of high -risk AI systems to report on energy 00:12:15.340 --> 00:12:18.460 and resource consumption. So some initial steps 00:12:18.460 --> 00:12:21.019 towards accountability. Small steps, but important 00:12:21.019 --> 00:12:24.090 ones. Opening up that black box is crucial. But 00:12:24.090 --> 00:12:26.250 in the meantime, who's paying for all the new 00:12:26.250 --> 00:12:28.269 infrastructure needed to meet this demand? The 00:12:28.269 --> 00:12:30.850 grid upgrades, the new power lines? Well, that's 00:12:30.850 --> 00:12:33.110 another really contentious point. Research looking 00:12:33.110 --> 00:12:35.029 at utility agreements suggests that, in some 00:12:35.029 --> 00:12:37.129 cases, the costs associated with building up 00:12:37.129 --> 00:12:39.190 the grid specifically for these massive data 00:12:39.190 --> 00:12:42.009 centers might actually be passed on to regular 00:12:42.009 --> 00:12:44.429 consumers. Wait, so ordinary households could 00:12:44.429 --> 00:12:47.090 end up paying higher electricity bills to subsidize 00:12:47.090 --> 00:12:49.110 the power needs of these giant tech companies? 00:12:49.490 --> 00:12:52.120 That's the concern. One analysis for Virginia 00:12:52.120 --> 00:12:54.279 suggested residents could potentially face an 00:12:54.279 --> 00:12:58.720 extra $37 .50 on their monthly bills just to 00:12:58.720 --> 00:13:01.080 cover these data center -related grid costs. 00:13:01.320 --> 00:13:05.379 $37 extra per month to power facilities for some 00:13:05.379 --> 00:13:07.919 of the richest corporations on the planet. How 00:13:07.919 --> 00:13:10.360 is that justified? That's the question utility 00:13:10.360 --> 00:13:13.019 regulators are grappling with right now. It really 00:13:13.019 --> 00:13:15.559 challenges the narrative that these centers are 00:13:15.559 --> 00:13:19.059 purely economic windfalls for a region if the 00:13:19.059 --> 00:13:21.700 public has to subsidize the essential infrastructure 00:13:21.700 --> 00:13:24.539 they require. Okay it feels like we've painted 00:13:24.539 --> 00:13:27.279 a pretty concerning picture so far. Massive energy 00:13:27.279 --> 00:13:31.120 use, water scarcity, lack of transparency, costs 00:13:31.120 --> 00:13:34.259 potentially pushed on to consumers. It is concerning 00:13:34.259 --> 00:13:36.419 no doubt. But the source material also pointed 00:13:36.419 --> 00:13:38.740 out that AI isn't just the problem, right? It 00:13:38.740 --> 00:13:41.120 can also be part of the solution. Yes, absolutely. 00:13:41.259 --> 00:13:43.840 That's the paradox. AI is incredibly powerful 00:13:43.840 --> 00:13:47.039 at optimization, finding efficiencies. And sometimes 00:13:47.039 --> 00:13:49.220 those efficiencies can lead to huge environmental 00:13:49.220 --> 00:13:52.340 benefits that might even outweigh AI's own footprint. 00:13:52.659 --> 00:13:55.440 Like the example with airplane contrails. Exactly. 00:13:55.759 --> 00:13:58.220 Contrails those white lines. Planes leave in 00:13:58.220 --> 00:14:01.299 the sky. They actually trap heat and contribute 00:14:01.299 --> 00:14:03.639 significantly to warming. Surprisingly so. I 00:14:03.639 --> 00:14:05.299 didn't realize they were that significant. They 00:14:05.299 --> 00:14:07.740 are. And AI algorithms have been developed that 00:14:07.740 --> 00:14:10.179 can guide pilots to make very slight adjustments 00:14:10.179 --> 00:14:12.820 to their flight paths to avoid the atmospheric 00:14:12.820 --> 00:14:15.830 conditions where contrails form. OK. One estimate 00:14:15.830 --> 00:14:18.190 suggested that if the entire aviation industry 00:14:18.190 --> 00:14:21.409 adopted this AI guidance, the reduction in warming 00:14:21.409 --> 00:14:24.370 effect the CO2 equivalent saved could be greater 00:14:24.370 --> 00:14:26.990 than all the emissions generated by global AI 00:14:26.990 --> 00:14:30.950 use back in 2020. Wow. So one specific optimized 00:14:30.950 --> 00:14:34.090 AI application could potentially offset the entire 00:14:34.090 --> 00:14:36.330 sector's footprint at that time. Potentially, 00:14:36.509 --> 00:14:39.950 yes. It shows the immense power of AI for positive 00:14:39.950 --> 00:14:42.129 environmental impact when it's applied strategically. 00:14:42.379 --> 00:14:45.159 So the path forward involves harnessing that 00:14:45.159 --> 00:14:47.759 optimization power while simultaneously tackling 00:14:47.759 --> 00:14:49.960 the hardware and energy issues we've discussed. 00:14:50.519 --> 00:14:52.960 What are the main solutions being proposed? Researchers 00:14:52.960 --> 00:14:56.039 point to a few key areas. First, making the models 00:14:56.039 --> 00:14:58.860 themselves more efficient by focusing on developing 00:14:58.860 --> 00:15:02.120 smaller, more specialized models designed for 00:15:02.120 --> 00:15:04.779 specific tasks rather than constantly pushing 00:15:04.779 --> 00:15:07.860 for ever larger general purpose models that are 00:15:07.860 --> 00:15:09.720 often overkill for what users actually need. 00:15:10.360 --> 00:15:12.860 tailoring the tool to the job essentially. Makes 00:15:12.860 --> 00:15:15.799 sense. Less wasted computation. What else? Second 00:15:15.799 --> 00:15:18.679 is hardware innovation. Moving beyond the current 00:15:18.679 --> 00:15:22.559 generation of GPUs. Things like neuromorphic 00:15:22.559 --> 00:15:24.799 chips. Neuromorphic? What are those? They're 00:15:24.799 --> 00:15:27.299 designed to mimic the way the human brain processes 00:15:27.299 --> 00:15:29.980 information very parallel. And crucially, they 00:15:29.980 --> 00:15:32.460 use energy mainly when they're actively processing 00:15:32.460 --> 00:15:34.360 data, not just sitting idle. That could offer 00:15:34.360 --> 00:15:36.820 huge energy savings. Interesting. And also things 00:15:36.820 --> 00:15:39.220 like optical processors, which use light instead 00:15:39.220 --> 00:15:41.879 of electricity to move data around. Less resistance, 00:15:42.279 --> 00:15:44.700 less heat, potentially much lower energy use. 00:15:44.740 --> 00:15:47.740 These are longer term hardware shifts. OK. So 00:15:47.740 --> 00:15:50.399 smarter models, better hardware. What about the 00:15:50.399 --> 00:15:52.620 grid problem, the reliance on dirty energy in 00:15:52.620 --> 00:15:54.539 places like Virginia? That needs systemic change. 00:15:54.639 --> 00:15:57.649 Yeah. Obviously, transitioning data centers fully 00:15:57.649 --> 00:16:00.429 to renewable energy sources is key. But they're 00:16:00.429 --> 00:16:02.929 more ambitious ideas, too. Like distributing 00:16:02.929 --> 00:16:06.009 the computing tasks globally based on time zones. 00:16:06.730 --> 00:16:09.649 Shifting demanding AI workloads to regions where 00:16:09.649 --> 00:16:12.649 renewable energy production, solar or wind, is 00:16:12.649 --> 00:16:15.490 peaking at that particular moment. Sort of like 00:16:15.490 --> 00:16:17.950 following the sun and wind around the planet. 00:16:18.090 --> 00:16:21.100 That sounds complex, but clever. using renewable 00:16:21.100 --> 00:16:23.360 energy where it's most abundant at any given 00:16:23.360 --> 00:16:26.039 time. It's a big logistical challenge, but it 00:16:26.039 --> 00:16:28.519 tackles the intermittency problem head on. Okay, 00:16:28.620 --> 00:16:31.000 so there are paths toward a more sustainable 00:16:31.000 --> 00:16:33.340 AI, but we can't wrap this up without touching 00:16:33.340 --> 00:16:36.220 on that classic economic concept, the Jevons 00:16:36.220 --> 00:16:38.940 Paradox. Ah yes, the rebound effect. Explain 00:16:38.940 --> 00:16:41.940 that. If we make AI way more energy efficient, 00:16:42.080 --> 00:16:44.519 doesn't that just solve the problem? Not necessarily. 00:16:44.779 --> 00:16:47.039 The Jevons Paradox basically states that when 00:16:47.039 --> 00:16:49.440 you make the use of a resource cheaper or more 00:16:49.440 --> 00:16:51.840 efficient, demand for that resource tends to 00:16:51.840 --> 00:16:54.620 go up, often increasing overall consumption in 00:16:54.620 --> 00:16:57.299 the long run. So making AI cheaper and easier 00:16:57.299 --> 00:17:00.019 to run might just lead to us using it way more, 00:17:00.220 --> 00:17:02.600 potentially erasing the affilency gains. That's 00:17:02.600 --> 00:17:05.680 the risk. AI is often described as an accelerant 00:17:05.680 --> 00:17:08.759 for everything. Making it more efficient accelerates 00:17:08.759 --> 00:17:11.160 its adoption across the board for good applications 00:17:11.160 --> 00:17:13.900 like the contrail example, but potentially also 00:17:13.900 --> 00:17:16.160 for more energy intensive, maybe less critical 00:17:16.160 --> 00:17:20.039 uses, the total footprint could still grow. That's 00:17:20.039 --> 00:17:22.579 a sobering thought. So efficiency alone isn't 00:17:22.579 --> 00:17:25.980 a silver bullet. It's necessary, but probably 00:17:25.980 --> 00:17:28.220 not sufficient on its own. This deep dive has 00:17:28.220 --> 00:17:31.150 really hammered home one thing for me. This AI 00:17:31.150 --> 00:17:34.289 revolution is driving energy and water demand 00:17:34.289 --> 00:17:37.250 at a pace and scale we maybe haven't seen since, 00:17:37.250 --> 00:17:39.410 I don't know, the Industrial Revolution or mass 00:17:39.410 --> 00:17:42.130 electrification. It's a valid comparison, I think. 00:17:42.470 --> 00:17:44.869 The trajectory is incredibly steep. And tackling 00:17:44.869 --> 00:17:47.410 it needs more than just individual users thinking 00:17:47.410 --> 00:17:50.309 twice about a query. It needs real transparency 00:17:50.309 --> 00:17:52.289 from the companies building these systems, and 00:17:52.289 --> 00:17:55.250 it needs smart systemic regulation. Exactly. 00:17:55.509 --> 00:17:57.529 The focus really needs to be on the system level, 00:17:57.609 --> 00:17:59.880 demanding accountability on energy sources. on 00:17:59.880 --> 00:18:02.859 water use, on model efficiency. Pinning it on 00:18:02.859 --> 00:18:05.359 individual consumption kind of misses the forest 00:18:05.359 --> 00:18:07.559 for the trees, given the scale and the secrecy 00:18:07.559 --> 00:18:10.400 involved. Right. Which brings us to our final 00:18:10.400 --> 00:18:13.039 provocative thought for you, our listener, to 00:18:13.039 --> 00:18:16.000 chew on. We're moving pretty fast towards a future 00:18:16.000 --> 00:18:18.279 where AI isn't just answering questions, but 00:18:18.279 --> 00:18:21.420 acting as agents. Autonomous agents performing 00:18:21.420 --> 00:18:24.940 complex tasks. Exactly. Agents that might plan... 00:18:24.799 --> 00:18:27.779 research, generate multiple options, run simulations, 00:18:28.480 --> 00:18:30.799 tasks that are inherently sequential and much 00:18:30.799 --> 00:18:32.940 more computationally intensive than a single 00:18:32.940 --> 00:18:35.549 query is today. they'll require significantly 00:18:35.549 --> 00:18:38.529 more energy per job. Yeah, the energy profile 00:18:38.529 --> 00:18:40.589 could shift again, dramatically. So the question 00:18:40.589 --> 00:18:43.390 is, in that near future, with these powerful, 00:18:43.630 --> 00:18:46.150 energy -hungry AI agents becoming potentially 00:18:46.150 --> 00:18:49.289 ubiquitous, how do we policymakers, citizens, 00:18:49.509 --> 00:18:51.690 consumers, ensure that the environmental toll 00:18:51.690 --> 00:18:54.230 of all this amazing innovation doesn't just completely 00:18:54.230 --> 00:18:57.410 overwhelm our planet's finite resources, especially 00:18:57.410 --> 00:19:00.089 energy grids and fresh water supplies? How do 00:19:00.089 --> 00:19:01.529 we keep that acceleration in check?