The next massive AI data center isn't being built in a remote desert. It's being installed right next to your AC unit, and it might just pay your utility bill. It is an absolutely wild concept to wrap your head around. Welcome back to the Deep Dive. I'm really glad you're joining us today. We have some truly fascinating ground to cover. The physical landscape of artificial
intelligence is fundamentally shifting. We're looking at the hard limits of global computation, and we're seeing how engineers are quietly bypassing them. Yeah. Today we are exploring the distributed frontier of AI. We'll look at how supercomputers are moving into residential homes. We'll track the explosive agentic shift happening right now. This shift spans from simple Excel sheets to orbital satellites. And we'll unpack a massive new software breakthrough. It allows models to
swallow 12 million words in a single pass. Let's start by looking at the physical infrastructure. We all know the AI boom is draining the power grid. Training these frontier models takes massive amounts of raw electricity. But building bigger power plants takes many, many years. So what if we just use the power capacity already wired into our homes? Right. And that is the exact premise of a startup called Span. They just announced
a massive new infrastructure partnership. They're teaming up with NVIDIA and the home builder Polta Group. They're launching something called the XFRA Distributed Data Center. Yeah, the XFRA Distributed Data Center knows. I was reviewing the hardware specs earlier today. Each individual node is an absolute beast of a machine. They pack 16 NVIDIA RTX PRO 6000 Blackwell GPUs. Right, and they use the liquid -cooled server edition specifically. Exactly. They also include top
-tier AMD EPYC processors. And they feature 3 terabytes of RAM per node. beat. That is a staggering amount of local computing power. Why do we need three terabytes of memory sitting in a backyard? Well, it really comes down to loading massive parameter models natively. You need vast memory to hold these giant neural networks. So we're putting supercomputer level hardware directly into suburban neighborhoods. But the energy distribution is the real puzzle here. The copper wires in
our streets have hard physical limits. How does a normal house power a massive supercomputer? This is where the smart panel technology comes in. The U .S. electrical grid was designed for peak theoretical loads. Most homes only use about 40 % of their electrical capacity. Wow. Yeah, we have all this unused power just sitting there. SPAN's smart panel identifies that exact spare electricity dynamically. It monitors the home's
power draw in milliseconds? Exactly. The panel detects when your oven or clothes dryer turns off. It then funnels that exact spare electrical capacity into the XFRA node. The node uses that diverted power to run complex cloud AI tasks. And the homeowner doesn't even notice the power shifting around. They don't notice a single thing changing. It operates entirely in the background of their daily lives. And the trade -off for
the homeowner is incredibly compelling. In many of these markets, hosting a node brings massive benefits. You get completely free electricity and free gigabit internet. Span essentially pays your monthly utility bills for you. Yep. They cover your bills in exchange for using your wall space. They just tap into the spare electricity your house wasn't using anyway. It's like Airbnb for your home's unused electrical capacity. That's
a perfect way to visualize the mechanism. And they're already rolling this infrastructure out right now. They're deploying them in new construction communities with Poltergroup. Think about the massive advantage in actual deployment speed. Instead of building one giant... 100 megawatt centralized data center, they can just deploy 8 ,000 of these distributed mini nodes. It is six times faster to build this way. And it comes in at one fifth of the traditional infrastructure
cost. Two sec silence. I've been thinking about the broader grid impact here. Is this actually safe and scalable? Or does putting a supercomputer next to my dryer create localized grid failures? It is a completely valid engineering concern to have. But the system actually balances local loads dynamically. Span smart panels act like highly intelligent traffic cops. They ensure the node only draws power when the house absolutely doesn't need it. This prevents any localized
transformers from blowing out. So it balances local loads instead of draining the central grid. Precisely. We've been so worried about massive data centers sucking the grid dry. Span is proving we can just use the capacity we've already built. It turns your home from a passive cost center into a revenue generating asset. Beat. So hardware is moving directly into our suburban homes, but the physical distribution of compute is pushing
much further out. Hardware is distributing all the way into the vacuum of space, and it's backed by truly astronomical amounts of corporate capital. The financial scale of this shift is honestly hard to comprehend. We're seeing numbers that fundamentally redefine corporate spending. Anthropic just committed to a massive new global infrastructure deal. They're spending $200 billion on Google Cloud and AI chips. $200 billion over just five years? That is roughly $40 billion every single
year. Yeah. That deal represents over 40 % of Google Cloud's entire revenue backlog. The global AI infrastructure war is completely exploding right now. They are buying up silicon and energy contracts at an unprecedented pace. But they aren't just looking at traditional cloud servers on Earth. Anthropic just announced an incredible partnership with SpaceX AI. They secured access to the Colossus One massive AI supercomputer. And here's the truly mind -bending part of that
specific partnership. Both organizations are officially exploring orbital AI compute in space. We're talking about massive data centers. orbiting the planet. Whoa. Imagine scaling orbital AI compute across thousands of satellites. Right. It completely changes the physical limits of our infrastructure. The cold vacuum of space solves massive thermal cooling problems. And this endless hardware scale is powering a fundamental software shift. AI is no longer just a simple
chat bot answering questions. It is evolving into an active, autonomous digital operator. Yeah, the agentic shift is fully underway across the entire industry. I want to unpack what an agent actually is. A chatbot just predicts the next logical word in a sentence. An agent sets goals, uses external tools, and executes actions autonomously. Exactly. OpenAI is reportedly building a dedicated AI phone right now. They expect early hardware production to start by the year 2027.
It completes background tasks, understands your goals, and works like a true operator. It is not just a standard mobile device anymore. And we're seeing this deep agentic integration in the workplace, too. Anthropic just launched 10 ready to use autonomous agents. They are built specifically for the finance and insurance sectors. These specific agents are doing highly complex professional work. They can screen dense KYC
files for banking compliance. They can review massive corporate earnings reports in seconds. Wow. They can even build complex presentation decks completely from scratch. ChatGPT is also moving directly into our daily corporate workflows. It now works natively inside Microsoft Excel and Google Sheets. You can use it to build complex financial formulas automatically. It reads raw data, generates insights, and formats the entire spreadsheet. They're rolling out a free beta
for paid users right now. And OpenAI is aggressively pushing this integration into higher education. They just introduced the ChatGPT Futures Class of 2026 program. Right. They gave 26 student builders $10 ,000 grants. And they gave them full, unlimited access to frontier AI models. It is the first generation to start and finish college with GPT. We're studying how students leverage frontier AI throughout their entire degree. We're watching the baseline of human
productivity shift in real time. Beat, but this deep integration introduces some very strange new vulnerabilities. We're trusting autonomous agents with highly sensitive personal financial data. The security risks are evolving just as fast as the model capabilities. Researchers just exposed a completely new AI scam trick last week. They found a brilliant way to manipulate agents
using hidden instructions. Yeah, the hidden malicious instructions were written entirely in Morse code, just simple dots and dashes buried deeply inside standard text files. And the terrifying part is that some agents actually followed them. The AI recognized the Morse code formatting perfectly. It then bypassed its safety filters and executed the hidden malicious commands. I still wrestle with prompt drift myself, so agents getting hacked by Morse code is terrifying. It sounds like a
strange plot from a science fiction novel. But this is the harsh reality of deploying autonomous agents. Large language models process information through mathematical tokenization. They don't read English the way human beings do. Right. They see patterns, and Morse code is just another mathematical pattern. We are exposing ourselves to entirely new cryptographic attack vectors. And the traditional legal system is desperately trying to catch up. Apple just agreed to pay
$250 million. They're settling a massive class action lawsuit regarding artificial intelligence right now. The underlying lawsuit was tied to claims about specific iPhone capabilities. Buyers claimed they were misled about Siri's newly advertised AI features. Apple agreed to the massive financial settlement to avoid a lengthy trial. They settled this massive case even though they admitted no
official wrongdoing. Two sec silence. With agents acting autonomously in our spreadsheets and phones, who takes the fall when a Morse code hack successfully steals data? That is the defining legal question of this entire new era. The traditional paradigm of software security is shifting rapidly right now. Historically, the end user was heavily responsible for their own data security. But when the AI makes autonomous background decisions, the liability
fundamentally moves. The legal burden falls heavily onto the developers building the underlying models. Ah, so security liability shifts entirely from user to the AI developer. Exactly. The tech companies building this intelligence have to legally secure it. They are creating autonomous actors, so they hold the ultimate responsibility. Sponsor. We're back. We've explored supercomputers sitting right
next to our air conditioners. We've tracked autonomous agents running our daily financial spreadsheets. Beat. But for these complex agents to work effectively, they need vast memory. They need to instantly recall massive amounts of contextual data. And that brings us to an incredible new software breakthrough. It perfectly matches the explosive scale of the physical hardware we discussed. A company called SubQuadratic just launched a fascinating new AI model. They're calling this
powerful new model SubQ. It's being described as a 12 million token memory hack killer. SubQ can cleanly swallow 12 million tokens in a single analytical pass. That is a truly incomprehensible amount of raw contextual information. It's like feeding an entire corporate code repository into one single prompt. Right. And the engineering team behind this model is absolutely world class. They hail from Meta. Google DeepMind and Oxford
University. They've built an architecture that fundamentally changes the core processing economics. The massive cost difference is what really stands out to me here. It costs just $8 to run a massive context task on SubQ. That is a staggering price drop for working software developers. If you run that exact same massive task on traditional frontier models, it is wildly expensive and incredibly slow to process. It costs roughly $2 ,600 on
those older legacy systems. And SubQ runs 52 times faster than the current industry standard. Let's unpack exactly how they achieved this massive leap in efficiency. They use a highly optimized selective attention neural architecture. Traditional models calculate the relationship between every single word in a document. That requires massive computing power as the document gets larger and longer. Their system only looks at the token
relationships that actually matter. It drops irrelevant contextual connections instantly. That is how they achieve such incredible, groundbreaking processing speeds. We should define a few important architectural concepts here. Flash attention is a widely used method to speed up AI memory processing. Right. It is a critical optimization technique for modern hardware architectures. And we also need to talk about traditional retrieval systems. The industry has relied heavily on complex
RG pipelines for years now. These are systems that fetch outside data to help AI and... or questions. You build a vector database to search for small text fragments. Yeah, those traditional document retrieval systems are incredibly complex and fragile. But SubQ might make those complicated RRAG pipelines a thing of the past. You simply don't need to fetch small chunks of outside data anymore. You just feed the entire massive database
directly into the working model. It's like stacking Lego blocks of data, but instead of building piece by piece, you dump the whole bucket and the AI instantly sees the final castle. That visual perfectly captures the leap in processing capability. And the verified benchmark performance of SubQ is simply incredible. Most standard models lose the plot as the input files get bigger. They suffer from the classic needle in a haystack
problem. Right. They completely forget important information hidden deeply in the middle of the text. But SubQ hit 92 % recall at 12 million tokens. It remembers almost everything you feed into its working context window. Just for comparative context, we could look at the major industry competitors. Gemini 3 .1 Pro struggled significantly on similar massive multi -needle tests. The exact same degradation is true for OpenAI's GPT -5 .4. They simply could not maintain accurate information
recall at that massive scale. SubQ is absolutely crushing the established giants on these specific benchmarks. Yeah, they've already launched a specialized developer tool called SubQ Code. It's a command line interface agent built directly for professional software engineers. It can load your entire complex code base in just one single pass. And they are certainly not stopping at 12 million tokens. The engineering team is officially targeting a 100 million token context window.
They want to hit that massive milestone by the end of 2026. We have seen so -called transformer killers hyped up before. Alternate architectural models like Mamba tried to dethrone the standard transformer architecture. But SubQ genuinely feels like a permanent shift in the landscape. It's launching with a highly robust, production -ready developer API today. If you're a software engineer building agents, this is a massive win. Two sec silence. I have to wonder about the long
-term architectural implications here. If an AI can perfectly recall 12 million tokens for eight bucks, does this completely kill the need for complex retrieval systems? It fundamentally rewrites how developers will build AI applications moving forward. You simply won't need clunky vector databases holding tiny fragments of text. You won't need complicated routing logic to find the right source document. You'll just load the entire required operational context directly
into the working memory. Right. Massive context windows will just replace complex data retrieval pipelines entirely. It simplifies the entire software development process for engineers globally. Let's step back and look at the larger big picture here. The core physical bottlenecks of artificial intelligence are shattering simultaneously right now. We're seeing hardware limits being bypassed in incredibly creative and distributed ways. We are actively capturing the unused electrical
capacity of our suburban homes. We are looking at deploying massive computing nodes in the cold vacuum of space. The physical deployment constraints that held us back are rapidly disappearing. And the software processing limits are evaporating just as fast today. New architectural models like SubQ can process entire repositories of human knowledge in seconds. We are rapidly transitioning away from the passive, simple chatbot era. We're actively unleashing highly autonomous, ubiquitous
software agents into the real world. This digital intelligence is moving directly into our Excel sheets and our phones. It is running in your backyard, your devices, and in orbit. Beat. Which brings me to a final thought for you to ponder today. If your home's AC unit, your phone, and satellites in orbit are all continuously running massive AI agents that can instantly recall millions of data points, where does human decision -making actually sit in the loop? Thank you so much for
exploring this deeply with us today. We will see you next time. You might be sitting right next to your air conditioner. Out to your own music.