Picture a rusted out industrial site in western Pennsylvania, a relic, a defunct coal plant that's been sitting silent for years. But now you see construction crews all over it and they're not tearing it down. They're bringing it back. They're bringing it back to life, but not for coal. They're turning it into this massive gas plant. And here's the thing that just stops you in your tracks. All that power. Every single megawatt isn't for homes. It's not for a town. It's being resurrected
to feed a single data center campus. Yeah. It is such a stark imagery. We think of this AI revolution as this clean digital thing that lives up in the cloud. Right. Ethereal. Exactly. But then you see that plant or you see the pipelines going down in Texas and you realize the cloud is actually heavy. It's hot and it's hungry. And right now it's demanding we dig up the industrial pass to fuel our future. Welcome back to the
Deep Dive. I'm really glad you're here. Today we're going to step back and just look at that friction. The friction between our digital progress and, well, physical reality. We've got a stack of sources here that paint a very complex picture. It's not just about the code anymore. It's about physics. Right. And we're going to navigate this in four parts. First, that gas boom. There's a massive environmental cost to all this compute
and the numbers are, frankly, staggering. Then we'll shift to the business landscape, which is just chaotic right now. We're talking corporate espionage, a literal space race. Actual servers in orbit. Servers in orbit. We'll get there. Then we'll look at the toolkit new software that's changing how we work, plus a pretty serious security warning. And finally, a big technical breakthrough in how robots actually see the world. OK, let's unpack this. We started with that plant in Pennsylvania.
That's just one data point from a much bigger report we have here from Global Energy Monitor. Yeah, but the core tension here is just simple physics. Yeah. AI models are getting exponentially larger. And as they get bigger, their appetite for power just grows with them. You're talking about data centers with thousands of GPUs. These aren't like your laptop that goes to sleep. No. These things need always -on power. They cannot
afford any downtime. But hold on a second. I look at the sustainability reports from Microsoft, Google, Amazon. They all have these really aggressive net zero pledges for 2030. If they're pivoting to gas, doesn't that just kill those goals? It certainly puts them in serious jeopardy. But here's the engineering reality they're up against. Renewables are intermittent. Right. The sun doesn't always shine. The wind doesn't always blow. To run a GPU cluster at 100%, you need what's called
baseload power, consistent flatline energy. And batteries. I mean, we always hear that battery storage is the answer to that. Batteries are great for smoothing out a few hours, but right now, utility -scale battery storage, it just hasn't scaled enough or gotten cheap enough to bridge those long gaps for something this big. We're talking gigawatts. So the industry is pivoting. They're pivoting hard back to natural gas because it provides that instant, reliable baseload.
The data from Global Energy Monitor shows the U .S. is leading what they call a massive gas boom, specifically for AI's energy problem. I was looking at the emission projections in the report. They are difficult to wrap your head around. They're huge. If all the proposed U .S. projects go forward, we're looking at potential emissions of 12 .1 billion tons of CO2. Can we
put that in context? 12 .1 billion tons. To put that in context, that is roughly double the current annual emissions of the entire United States. Double, just from the infrastructure to support this one tech sector. That's the projection if the full build -out happens. And if you look globally, it could hit 53 .2 billion tons. It's a complete reshaping of the energy landscape. In the U .S., electricity demand could surge 60 % by 2050, and the main driver they cite is
AI. And this isn't happening evenly across the country, is it? It seems really concentrated. Texas is the epicenter. As of 2025, they had 57 .9 gigawatts of gas projects underway just in Texas. But it's elsewhere, too. Meta is building a $1 .5 billion data center in El Paso. And that facility, gas -powered. It feels like we're training the grid to depend on fossil fuels again at the exact moment we were supposed to be weaning ourselves off them. It's a pragmatic choice, I guess, but
a controversial one. They need the juice. There's a political angle in here, too. The report mentions comments from Donald Trump about this. That's right. Trump has promised that companies like Microsoft will, and I'm quoting, pay their own way for this stuff. So no public funding. Right. The idea is that tech giants shouldn't rely on rate payers to build out these huge power draws. But the report does note that the actual details on how you'd enforce that are pretty scarce right
now. So we're trying to invent the future AI autonomous agents. But to get there, we are doubling down on energy sources from the 20th century. We're burning the past to build the future. Yeah. So I have to ask, does the utility of AI justify doubling our carbon output? Immediate gains versus long -term climate debt. Long -term climate debt. That is a heavy bill. It really complicates the story of AI as this clean future. But speaking of complications, the business side is just as
messy. Let's look at the corporate battlefield. It is absolutely chaotic. You're seeing this mix of incredible profit, fierce rivalry, and some like straight up spy novel stuff. Okay, let's start there. The spy novel stuff. This story about the Google engineer. Yeah, this is just wild. A former engineer at Google was convicted for stealing over 2 ,000 files, AI trade secrets. But it wasn't just, you know, code snippets. He was stealing model weights, architecture diagrams.
And he was pitching investors while he was still there. While he was still employed. He was downloading files, pitching investors in China to build a rival startup, and then he tried to flee. It just highlights how valuable this IP is right now. Governments are treating these model weights like state secrets. Because if you have the weights, you don't need to spend the $100 million to train the model. You can just run it. Exactly. You skip the whole R &D cost and go right to deployment.
And speaking of money, we saw Microsoft's earnings. The numbers are just eye -watering. Microsoft pulled in $7 .6 billion in profit from OpenAI in a single quarter. A single quarter. Yeah. But here's the stat that really got me. Forty five percent of their six hundred twenty five billion in total deals came just from that partnership. OK, that explains why they're so intertwined. This isn't a side project for them. It's a massive pillar of their whole valuation. Right. And the
market is shifting all over. We see QAI raising one point five billion and now they're teaming up with Apple. That's a huge strategic move. Yeah. But while some are rising, others are pivoting. OpenAI is actually retiring models next month, including GPT -4 .0. I saw that, and the reason was that barely anyone uses it. Which is funny because I know for a fact some power users absolutely
loved that specific iteration. But in this field, if a model isn't hitting critical mass, the maintenance cost, the inference cost, it's just too high. They cut it loose. Okay, so we talked about the energy limits on Earth. Maybe that explains this next story, which feels like it's pulled from science fiction. China announced plans to launch AI data centers into space. This is where it gets really interesting. China's laid out a roadmap
to have this going by 2030. It's being framed as a direct competition, Beijing versus Elon Musk's SpaceX. The goal is to process AI data off Earth. I'm assuming the logic is cooling, right? I mean, space is cold. You solve the heat problem instantly. Well, that's actually a common misconception. Is it? Yeah. I mean, space is cold, sure, but it's a vacuum. On Earth, we cool things with air or water. That's convection. The air carries heat away. In a vacuum, there's
no air. So there's nowhere for the heat to go. Thermodynamically, getting rid of heat in space is an absolute nightmare. So you can't just open a window. Exactly. You would literally roast your servers. You have to use these massive radiators to radiate the heat away as infrared light. So the engineering challenge isn't just launching them. It's keeping them from melting down without an atmosphere. Whoa. Imagine scaling that. managing a billion queries in orbit while trying to radiate
heat into the void. It's the ultimate high ground. But the physics are brutal. Still, they pull it off. They bypass all the land and energy constraints we just talked about on Earth. So looking at this whole dynamic, is the rush for dominance compromising security? Speed and scale are outpacing safety protocols. All right, let's bring it back down to Earth, to our laptops. The newsletter highlighted a toolkit section, and I see two categories here. tools for bureaucracy, and tools
that manipulate perception. But first, that security warning. Claudebot. Right, Claudebot. So this is an open source AI agent. It looks super powerful because it can execute code on your local machine to solve tasks for you. I have to admit, I still wrestle with that urge to try new code like this on my main machine. You see the demo, you get excited, you just want to install it. I totally get it. The curiosity takes over. But the source material is very, very clear here. This is experimental
code. Do not install this on your main laptop. If there's a vulnerability or if the agent just hallucinates a command, it has access to your entire file system. So use a burner laptop, basically. Or use the alternative they mentioned, Cloudflare's Moldworker. It's pretty much the same capability, but it's all cloud -hosted. It's safer because it's sandboxed away from your personal data. Okay, let's look at what I'm calling the bureaucracy
killers, Pandata and StoryCV. These are fascinating because they signal the end of the junior analyst grind. Pandata takes messy input like CSVs, PDFs, even photos of whiteboards, and it turns them into these McKinsey -level reports. It structures the data, writes the narrative for you. And StoryCV does the same thing for resumes. Right. It just interviews you. You talk to it about your work history, and it writes the resume. It removes that writer's block. But think about the implication.
We're automating the entry -level tasks that used to teach people how to think strategically. That's a really deep point. If the AI does the grunt work, how do juniors ever learn the nuance? OK, then there's the second category, the prompting company. This sounds like reverse SEO. That is the perfect way to describe it. In the old days, you'd optimize your website so Google search would find you. The prompting company finds what questions people are asking chatbots, chat GPT,
Claude, and it optimizes your content. So those AIs will cite your product as the answer. That is incredibly clever. And also. A little dystopian. We're optimizing our content so machines will read it and then recommend us to humans. It's the new marketing frontier. If the training data doesn't know you exist, you're invisible. And then highlight GPT. Simple browser extension. You just highlight text, get an explanation.
It just reduces the friction of learning. So looking at all these tools together, are we optimizing for algorithms or humans? We're tailoring our reality for machine consumption. Okay. For our final segment, we're going to look at how robots see the world. Or maybe how they struggle to see it. Yeah, this is about a breakthrough from China's ant group called Lingbot Depth. The problem they identify is broken depth. Can you explain
what that is? Sure. When a robot looks at the world with, say, a LiDAR or a camera, the data isn't perfect. You have surfaces like glass or shiny metal or just bad lighting, and they create these holes in the data. So the robot literally thinks there's a hole in the table. Or it just can't tell how far away the cup is. Right. It makes them clumsy. Lingbot Depth is a model designed to fix this. It takes that broken depth map and the color image, and it basically reconstructs
the missing 3D information. It fills in the blanks. But the way they trained it is what's so brilliant. They call it masked training. How does that work? Okay, think of it like learning to read a sentence where someone has taken a black marker and crossed out every third word. Okay. At first, it's just nonsense. But if you practice enough... You start to understand the context. You know that after the cat sat on the word, the next word is probably Matt. You don't even need to see the word to
know it's there. So they blinded the AI on purpose during training. Exactly. They forced it to learn the context of the physical world. So now when it sees broken data in the real world, it has the intuition to fill it in accurately. And the result is a robot that can handle like transparent objects and messy environments. It creates a really tight 2D to 3D understanding. And this is crucial for what they call embodied AI, getting robots out of the lab and into the real world.
So the question here, does this bridge the gap between software and the physical world? Yes, it gives machines intuition for physical space. We're going to take a quick beat. We'll be right back. And we're back. So let's try to tie all this together. We started with a coal plant being resurrected as a gas plant. We went to space with servers that need massive radiators. We looked at tools that write our resumes and robots that learn to see by reading between the lines.
There is such a clear through line here. We are witnessing a massive physical build out just to support a digital revolution. We love to focus on the software, the magic of the chat bot, right? But the reality is steel and concrete and gas and thermodynamics. It feels like a moment of friction. We want the magic of Lingbot. through broken pixels. We want the convenience of Pandata. But the cost is being paid in old world fossil fuels and really intense geopolitical rivalry.
Exactly. We're building tools that are effectively giving machines intuition and creativity. But to power them? We're reverting to these very traditional, heavy industrial methods. The digital future has a very, very heavy physical footprint. Before we go, I want to leave you with a thought on that prompting company tool. It really stuck
with me. If we start creating content primarily to please AI chatbots, so they'll recommend us, are we entering an era where humans no longer write for humans, but for the training data of the future? Are we just feeding the machine that feeds us? That is a thought that will keep me up tonight. Something to mull over. If you enjoyed this deep dive, make sure you're subscribed. We've got some great topics coming up. Thanks for listening, everyone. See you in the next deep dive.