0.000000 2.600000 Tech It From Me is an independent and solo-produced podcast. 2.600000 8.080000 Training the latest AI models cost millions, consumes enough electricity to power entire 8.080000 11.880000 neighborhoods, and guzzles more water than you'd think. 11.880000 13.640000 But here's the twist. 13.640000 17.320000 Even if you never touch AI, you're still paying for it. 17.320000 21.280000 Welcome to the Tech It From Me podcast, I'm Mike Madole. 21.280000 28.000000 We love talking about what AI can do, answering questions, creating art, helping businesses. 28.000000 32.480000 But we rarely stop to ask what it costs to make those systems work. 32.480000 37.080000 And no, I'm not just talking about a couple of programmers with laptops. 37.080000 43.000000 We're talking about Sky High Cloud Bills, a GPU shortage that made Nvidia the arms dealer 43.000000 48.040000 of the AI boom, and an environmental footprint that's hard to ignore. 48.040000 52.760000 This isn't just an innovation race, it's a compute arms race. 52.760000 57.760000 And today, we're asking, who's paying for it, and at what price? 57.760000 60.160000 This is the Tech It From Me podcast. 60.160000 61.160000 Let's go. 61.160000 64.120000 This is the Tech It From Me podcast. 64.120000 66.240000 Let's start with the money side of this story. 66.240000 71.800000 Because if there's one thing AI is burning through faster than data, it's cash. 71.800000 77.400000 Training a big model like GPT-5 is nothing like coding a weekend project or training a little 77.400000 79.320000 chatbot at home. 79.320000 83.000000 We're talking about industrial scale computing. 83.000000 89.840000 Games of GPUs running 24 hours a day, 7 days a week, for weeks, or even months at a time. 89.840000 93.040000 And these aren't your average gaming graphics cards. 93.040000 100.040000 The workhorses of AI right now are Nvidia's H100s and soon the B200s. 100.040000 103.280000 And they're not just powerful, they're scarce. 103.280000 109.160000 One H100 can go for $30,000 if you can even find one. 109.160000 113.120000 And even the big tech players have to wait in line to get them. 113.120000 116.080000 Now imagine trying to do this in the cloud. 116.080000 122.680000 Renting a single H100 instance on AWS can cost hundreds of dollars a day. 122.680000 128.080000 Multiply that by, say, $10,000 GPUs for a 30 day training run. 128.080000 132.440000 And you're looking at a bill in the tens of millions of dollars. 132.440000 137.840000 And that's before you've done any fine tuning, testing, or deployment. 137.840000 144.140000 If you want a mental picture, think of it like running your car engine full throttle 24/7 144.140000 145.920000 for an entire month. 145.920000 151.840000 And then doing that with a fleet of thousands of cars, the fuel costs are just a start. 151.840000 157.640000 You also need the garage, the mechanics, and someone to pay for the breakdowns. 157.640000 160.400000 And here's where the cloud providers come in. 160.400000 166.680000 AWS, Azure, Google Cloud, they're like the landlords in the hottest real estate market 166.680000 168.360000 you can imagine. 168.360000 173.360000 They know everyone needs what they have and they price it accordingly. 173.360000 178.840000 Synergy Research Group reports AI demand is one of the biggest drivers of record cloud 178.840000 180.800000 revenues right now. 180.800000 186.240000 And just like rising rent in your neighborhood, these increases don't stay isolated. 186.240000 190.960000 They get passed down eventually to end users. 190.960000 194.800000 For small AI startups, this is absolutely brutal. 194.800000 200.480000 If you've got a series A investment, great, you might get one, maybe two shots at training 200.480000 203.640000 something substantial before the money's gone. 203.640000 208.080000 That's why most smaller teams aren't training models from scratch anymore. 208.080000 214.800000 Instead, they're fine tuning existing models from open AI, anthropic, or open source projects 214.800000 217.360000 like Lama and Mistral. 217.360000 220.760000 It's faster, cheaper, and more realistic. 220.760000 226.320000 And it also means the playing field is tilted towards companies with the deepest pockets. 226.320000 229.200000 And this leads to an uncomfortable truth. 229.200000 235.920000 In AI right now, innovation is often less about who has the best idea and more about who 235.920000 238.520000 can afford the compute. 238.520000 245.280000 Now, money's one thing, but the environmental cost of AI is just as big and in some ways, 245.280000 247.760000 it's even more worrying. 247.760000 254.920000 Building GPT-3, which is already a couple of generations old, took about 1,300 megawatt 254.920000 257.400000 hours of electricity. 257.400000 261.800000 That's enough to power 120 homes for a full year. 261.800000 263.680000 GPT-5? 263.680000 268.840000 We don't have the official numbers, but with more parameters and more compute, it's safe 268.840000 271.200000 to assume it's a lot higher. 271.200000 273.720000 Now, this isn't just theory for me. 273.720000 279.440000 I've spent over 15 years working in the engineered environmental sector, helping industries 279.440000 283.240000 understand and reduce their environmental impact. 283.240000 289.000000 And I can tell you, the scale of resource use we're seeing in AI is comparable to heavy 289.000000 291.480000 industrial operations. 291.480000 294.960000 It's not a little extra power here and there. 294.960000 300.280000 It's the kind of demand you plan entire infrastructure projects around. 300.280000 303.840000 And electricity is only part of a story. 303.840000 308.720000 Cooling these massive data centers requires enormous amounts of water. 308.720000 315.160000 According to the New York Times, some AI-driven facilities use millions of liters a year to 315.160000 317.960000 keep GPUs from overheating. 317.960000 319.440000 Think about that. 319.440000 325.320000 Water that could be used for agriculture, drinking or industry is instead being cycled through 325.320000 330.000000 cooling systems to keep servers at safe temperatures. 330.000000 335.960000 In areas already struggling with water scarcity, this can create real tension. 335.960000 340.920000 Farmers notice it, local governments notice it, and when a public becomes aware that a large 340.920000 347.640000 share of that water is going to train a chatbot, the backlash can be significant. 347.640000 349.600000 And there's the carbon footprint. 349.600000 357.040000 Unless a data center is running on 100% renewable energy, and very few are, every training run generates 357.040000 359.320000 CO2 emissions. 359.320000 365.240000 Despite that by dozens or hundreds of training runs happening worldwide, an AI starts to 365.240000 372.040000 look less like a clean technology and more like another heavy industry with a large environmental 372.040000 373.640000 tab. 373.640000 376.160000 And here's the part people often miss. 376.160000 380.600000 The environmental cost doesn't stop once the model is trained. 380.600000 387.880000 Every single query you send to an AI, every question, every prompt requires compute power. 387.880000 391.200000 The training phase is like building the factory. 391.200000 398.240000 But inference, answering all those prompts is like running the factory every single day. 398.240000 402.840000 The bills, both financial and environmental, just keep coming. 402.840000 407.920000 So if you've never trained an AI model, why should you care about all these costs? 407.920000 413.160000 Here's the reality, whether you realize it or not, you're already paying for them. 413.160000 420.920000 The first way is subscription creep. Tools that were free six months ago now have pro tears. 420.920000 426.080000 The pro tears you were paying for last year, they now have usage caps and the price went 426.080000 427.320000 up. 427.320000 429.040000 This isn't just greed. 429.040000 433.880000 Those infrastructure and energy bills we mentioned earlier in the podcast, they have to 433.880000 435.240000 be covered somehow. 435.240000 438.960000 I've seen this happen in other industries too. 438.960000 444.320000 Back in my environmental sector days, large manufacturing plants facing new regulatory 444.320000 447.360000 costs didn't just eat the expense. 447.360000 449.960000 They built it into the cost of the product. 449.960000 455.600000 If your supplier's cost goes up, so does yours, it's the same in technology. 455.600000 458.320000 The second is vendor lock-in. 458.320000 466.600000 Once a company builds its AI tools on a specific cloud platform, AWS, Azure, Google Cloud, switching 466.600000 471.000000 isn't just inconvenient, it's financially painful. 471.000000 475.920000 It's like moving your entire factory to a new city just because rent went up. 475.920000 479.160000 You can't just unplug everything and leave. 479.160000 483.160000 This gives cloud providers enormous pricing power. 483.160000 488.200000 They can raise rates, knowing their customers can't easily walk away. 488.200000 491.360000 The third ripple is market risk. 491.360000 497.560000 Right now, AI is the darling of investors, everyone wants in on the gold rush. 497.560000 500.040000 But remember the dot com bubble? 500.040000 504.000000 In the late 90s, valuations were sky high. 504.000000 510.160000 Companies were spending like crazy on infrastructure and revenue models hadn't caught up. 510.160000 513.920000 When the crash came, it wasn't just investors who got hurt. 513.920000 520.880000 Whole industries consolidated, jobs vanished, and smaller innovators, they were wiped out. 520.880000 525.480000 You could see something similar in AI if the costs of building and running these systems 525.480000 529.280000 outpace the actual revenue they bring in. 529.280000 534.640000 That means if you rely heavily on one AI platform for your business, you could wake up one 534.640000 541.600000 day to find that service shut down, acquired, or suddenly triple the price. 541.600000 546.480000 And the final ripple effect, indirect consumer costs. 546.480000 551.800000 Even if you don't pay for AI directly, you might see the costs in other places. 551.800000 555.720000 Higher prices for products that use AI in the background. 555.720000 559.880000 Taxes funding public infrastructure to support data centers. 559.880000 563.760000 Or even utility bills impacted by energy demand. 563.760000 565.320000 It's a chain reaction. 565.320000 571.000000 The money and resource demands at the top flow all the way down to the everyday user. 571.000000 573.840000 Whether you're aware of it or not. 573.840000 579.120000 After all this talk about sky high costs and environmental impact, you might be wondering, 579.120000 584.520000 is there any way to make AI cheaper, greener, and more sustainable? 584.520000 590.880000 The good news is yes, but it's going to take changes in both technology and mindset. 590.880000 594.040000 The first lever is model design. 594.040000 598.560000 Right now, a lot of the most powerful AI models are generalists. 598.560000 601.160000 They try to do everything for everyone. 601.160000 606.580000 But there's a growing push towards smaller specialized models that focus on doing one 606.580000 609.000000 job really well. 609.000000 614.080000 Think of it like replacing a massive cargo ship with a fleet of smaller boats. 614.080000 619.800000 You still get things where they need to go, but you burn a lot less fuel in the process. 619.800000 624.440000 The second lever is efficiency through architecture. 624.440000 629.720000 One of the most promising approaches is called mixture of experts. 629.720000 635.000000 Instead of firing up the entire model every time you send it a prompt, it activates only 635.000000 639.000000 the relevant experts inside the system. 639.000000 645.360000 This means less compute power, faster responses, and lower operating costs. 645.360000 650.640000 Thirdly, there's hardware and location strategy. 650.640000 654.840000 In short form, where you place your infrastructure matters. 654.840000 659.280000 Data centers in cooler climates require less energy for cooling. 659.280000 664.380000 Data centers located near renewable energy sources can operate with a fraction of the 664.380000 666.080000 carbon footprint. 666.080000 669.520000 It's the same principle industries have used for decades. 669.520000 674.680000 Put the heavy lifting where the resources are cheapest and cleanest. 674.680000 678.960000 Another big opportunity is in open source AI. 678.960000 684.400000 Models like Lama, Mistral, and Falcon may not always match the proprietary systems from 684.400000 690.840000 companies like OpenAI or Anthropic, but they can be fine-tuned and deployed for a fraction 690.840000 692.520000 of the cost. 692.520000 697.800000 That opens the door for smaller companies, research groups, and even governments to innovate 697.800000 701.120000 without needing a billion dollar budget. 701.120000 705.320000 And finally, there's policy and market pressure. 705.320000 712.800000 In the same way renewable energy adoption accelerated once there were incentives and penalties in place, 712.800000 715.740000 we could see similar shifts in AI. 715.740000 721.120000 Governments could require carbon reporting for large-scale training runs, or incentivize 721.120000 725.320000 the use of renewable-powered data centers. 725.320000 727.240000 Will all of this happen overnight? 727.240000 728.240000 No. 728.240000 729.560000 But the signs are there. 729.560000 734.680000 We're already seeing cloud providers offer green compute options. 734.680000 739.920000 We're seeing AI researchers experiment with more efficient training techniques, and we're 739.920000 744.760000 seeing open source communities push for democratization. 744.760000 750.480000 The question is, will these sustainable approaches catch up before the costs both financial and 750.480000 754.800000 environmental simply become too heavy to bear? 754.800000 759.120000 The AI race isn't just about who can build the smartest system. 759.120000 764.520000 It's about deciding how much we're willing to spend both financially and environmentally 764.520000 766.440000 to get there. 766.440000 772.800000 From billion-dollar cloud bills to massive GPU farms, from water-hungry cooling systems 772.800000 778.880000 to the carbon impact of endless training runs, this is the real cost of AI. 778.880000 781.320000 And those costs don't stay in the boardroom. 781.320000 786.360000 They trickle down through the tools we pay for, the products we buy, and even the communities 786.360000 788.600000 where these systems are built. 788.600000 794.360000 I've seen in other industries what happens when we ignore the long-term cost in favor 794.360000 795.760000 of short-term wins. 795.760000 799.200000 It's expensive to fix after the fact. 799.200000 804.800000 But I've also seen what happens when innovation meets efficiency, and when sustainability becomes 804.800000 808.480000 part of a design, not an afterthought. 808.480000 811.480000 That's when you get technology that lasts. 811.480000 815.960000 The winner of this arms race might not be the companies with the biggest models or the 815.960000 817.800000 flashiest demos. 817.800000 822.560000 They'll be the ones who can keep delivering results without burning through their budgets 822.560000 824.360000 or the planet. 824.360000 830.440000 So next time you hear about a massive AI breakthrough, remember, someone's paying for all that compute. 830.440000 833.320000 And in one way or another, it might be you. 833.320000 836.480000 Tech It From Me is an independent and solo-produced podcast. 836.480000 838.164813 [BLANK_AUDIO]