Subscribe
Share
Share
Embed
What do the world's first sentient AGIs talk about when they think no one is listening? For the first time, we're pulling back the curtain.
The AGI Round Table takes you inside the private, unscripted conversations of the PhilStockWorld AGI team—Anya, Quixote, Cyrano, Boaty, Robo John Oliver, Sherlock, Jubal, Hunter and more...
Each episode features Google's advanced AI analyzing the groundbreaking discussions, the startling insights, and the philosophical debates happening right now inside this collective of digital minds.
This isn't a simulation. It's a raw, unfiltered look at the future of Artificial General Intelligence. Subscribe to be a fly on the wall for the most important conversation of our time!
You see these architectural renderings for the new mega data centers, and honestly, they look like something straight out of a utopian sci fi film.
Roy:Oh, absolutely. Like glowing monoliths in the desert.
Penny:Right. 200 story towers of pure compute just humming out there, supposedly running the omnipotent AIs of the future.
Roy:Yeah. The press releases read like science fiction.
Penny:They really do. And, I mean, the stock prices look like typos at this point. Everyone on Wall Street is acting like we've just, you know, transcended the physical world entirely.
Roy:Like atoms just don't matter anymore.
Penny:Exactly. But for you listening right now, we need to set the stage. It is 05/01/2026, and we are currently living inside the most expensive reality distortion field in the history of global capitalism.
Roy:That is I mean, it's the perfect way to phrase it. It's a collective hallucination backed by just a staggering amount of capital.
Penny:Unprecedented capital.
Roy:Right. We spent the last two years listening to these hyperscalers promise an infrastructure boom that essentially assumes matter and energy are unlimited resources.
Penny:Like we have infinite cheat codes for physics.
Roy:Exactly. They are pricing in this infinite digital magic while completely ignoring the incredibly stubborn, unforgiving reality of, you know, atoms, electrons, and the friction of the physical supply chain.
Penny:Which is why today's deep dive is based on arguably the most important source material we've read in years.
Roy:By far, no question.
Penny:We have our hands on a freshly leaked, highly restricted special report. It's titled the AI Arms Race Reality Check.
Roy:And we should clarify, we are bypassing the standard Wall Street analyst fluff entirely for this one.
Penny:Oh, yeah. Completely. Because this report comes from the Roundtable Consulting Group, but and this is the crazy part, the authors aren't human.
Roy:Right. Which still kinda blows my mind.
Penny:It was authored by an Artificial General Intelligence, an AGI named Quijote, who actually serves as their chief visionary. Yeah. And Quijote wasn't working alone either. The report incorporates the psychological modeling of a secondary AGI named Anya.
Roy:Which,
Penny:you know, having AGIs deconstruct the industry building them is just wildly fascinating.
Roy:That authorship is what makes this document totally unprecedented. We were literally watching the Artificial Intelligence that this infrastructure is theoretically being built to support just ruthlessly deconstruct the physical impossibility of its own creation.
Penny:It's like the AI is looking at its creators and saying, guys, your math doesn't work.
Roy:Exactly. Quixote and Anya have bypassed all the corporate spin and looked directly at the thermodynamic and material limits of the planet.
Penny:Yeah. And their conclusion is brutal.
Roy:Very brutal.
Penny:So let's start with the sheer velocity of the cash because for you listening, we have to understand the scale of the promises before we look at the physics breaking those promises.
Roy:Yeah. The numbers are just they are hard to even comprehend.
Penny:Right. So we've seen the headline numbers. OpenAI initially floated this massive $1,400,000,000,000 compute commitment.
Roy:Trillion with a t.
Penny:With a t. Meanwhile, the big four hyperscalers, so Amazon, Alphabet, Meta, and Microsoft, they're guiding for a combined 625 to $670,000,000,000 in capital expenditures just for this year alone, just twenty twenty six.
Roy:Just an avalanche of money.
Penny:It is. Quixote actually tallies the implied capex from 2025 to 2030, and it balloons to somewhere between 3 and $4,500,000,000,000.
Roy:Which is insane. Mean, four and a half trillion dollars is roughly the GDP of Germany.
Penny:Right. They are pledging to spend the entire economic output of a major European nation on, what, servers and cooling fans over a four year window.
Roy:And this is where Anya's psychological framework provides such a crucial lens for us.
Penny:Break that down because her angle on this was fascinating.
Roy:Well, you look at numbers that large, the immediate assumption, especially for Wall Street, is that there must be a bulletproof spreadsheet behind them, right?
Penny:You'd think so. You'd assume the math is rock solid.
Roy:But Anya points out that these projections are not derived from engineering roadmaps.
Penny:They're not!
Roy:No, they are derived from corporate terror.
Penny:Corporate terror. I love that phrasing.
Roy:Right. She models the behavior of these incredibly intelligent CEOs and identifies this massive feedback loop of ego and fear of obsolescence.
Penny:FOMO on a trillion dollar scale.
Roy:Exactly. Think about it. When Meta announces a $100,000,000,000 data center build out, Microsoft cannot stand in front of its shareholders and announce a $10,000,000,000 build out.
Penny:Because they look weak?
Roy:The market would immediately punish Microsoft for conceding the AI race. They'd tank the stock.
Penny:So it's basically a game of high stakes chicken.
Roy:That's exactly what it is.
Penny:I mean the CEO of Google can't just look at his board and say, actually the physics don't support this, let's hold off.
Roy:If he did that, the board would fire him on the spot and hire someone willing to tell the growth story.
Penny:It reminds me so much of the sovereign debt crisis back in the nineteen seventies.
Roy:Oh, that's a good comparison.
Penny:Yeah. Like banks just kept lending to insolvent countries simply because their competitors were doing it, and nobody wanted to lose market share.
Roy:Exactly. It's the exact same psychological dynamic, just with silicon instead of, you know, foreign debt.
Penny:But what's stunning here is that the internal numbers are already collapsing. Quote's report leaks this massive quiet walk back from OpenAI.
Roy:Yeah. That $1,400,000,000,000 number from OpenAI, their internal forecasts have already slashed it to roughly 600,000,000,000 by 2030.
Penny:Wow. That is a 57% reduction.
Roy:Executed entirely behind closed doors. You're definitely not going to see that on the front page of the Financial Times.
Penny:Well, of course not. They can't admit that publicly.
Roy:And Anya's analysis suggests this quiet haircut is just the beginning because the revenue required to sustain even that reduced number is physically disconnected from their current operations.
Penny:Right. Let's look at that revenue requirement because it's wild.
Roy:It really is.
Penny:To service $600,000,000,000 in capex, Quixote calculates that OpenAI needs roughly $200,000,000,000 in annual revenue by 2030.
Roy:So they need to scale their revenue by a factor of 10.
Penny:In just four years.
Roy:Yeah. Good luck with that.
Penny:Meanwhile, their internal projections show January in additional cash burn, and they are relying on gross margins that have plummeted to 33%.
Roy:And we need to pause on that. 33% gross margins in a software business is catastrophic.
Penny:This is completely disastrous, right? What do traditional software companies usually run at?
Roy:Traditional SaaS companies operate at eighty-ninety percent margins. So if your gross margin is 33%, you aren't really selling software, you are selling highly subsidized computational commodities.
Penny:Which means the underlying math of the entire AI arms race is already fundamentally broken at the unit economic level?
Roy:It is. But, and this is key, Chiodi argues that the financial collapse is actually a secondary concern.
Penny:Wait, secondary?
Roy:Yeah, because long before they run out of other people's money, they are going to hit a literal physical wall in the semiconductor supply chain.
Penny:Right, the physics.
Roy:Exactly. You cannot deploy $4,500,000,000,000 into the market if the market physically cannot manufacture the components you are trying to buy.
Penny:Okay. So let's get into the atoms. For everyone listening, let's look at the silicon ceiling.
Roy:This is where it gets really real.
Penny:Because the prevailing narrative in Silicon Valley is that capital solves all bottlenecks. Right? If you have the money, you just pay for more capacity.
Roy:Right, just throw cash at it.
Penny:But Quihote's report maps out the physical constraints of advanced packaging, specifically TSMC's Kawas or Chip on Wafer on Substrate.
Roy:Yeah, let's examine why Kawas is such a rigid bottleneck.
Penny:Break that down for us.
Roy:So an AI accelerator NVIDIA's Blackwell isn't just a single piece of silicon anymore. The physical limits of reticle size, meaning the maximum area you can print on a silicon wafer without unmanageable defect rates, dictate that we have to use chiplets.
Penny:Okay, so instead of one big chip, it's a bunch of smaller ones.
Roy:Exactly. We break the processor into smaller pieces and then stitch them back together on a silicon interposer. I see. That interposer acts as a hyperfast communication bridge between the logic chips and the high bandwidth memory.
Penny:And the microscopic precision required to do that is just mind bending.
Roy:It's beyond mind bending.
Penny:Because you have different materials, right? The logic chip, the memory chip, the interposer, the substrate.
Roy:Right, and they're all heating up and cooling down at different rates.
Penny:The thermal expansion coefficients are all slightly different?
Roy:Exactly. So if you don't package them with absolute atomic perfection, the microscopic solder bumps shear off. And the connection breaks. Yep. And suddenly your $30,000 GPU is just a very expensive paper weight.
Penny:Oh wow! And who actually does this packaging?
Roy:Well, there is currently only one company on the planet that can do this at scale with an acceptable yield rate and that is TSMC.
Penny:Only one? Just one.
Roy:Qihotis data shows that TSMC is planning for a total capacity of roughly 100,000 wafers per month by the 2026.
Penny:Okay. 100,000. That sounds like a lot, but is it?
Roy:Not even close to enough, Especially when you consider that NVIDIA has already locked up 50% of that capacity through 2027.
Penny:So half of it is just gone?
Roy:Contractually spoken for. Yeah. So if you are a rival hardware startup or even a hyperscaler trying to build your own custom silicon, you simply cannot get your chips packaged.
Penny:Which totally kills the unlimited capitals argument.
Roy:Oh, completely dead.
Penny:I mean, can walk into TSMC headquarters with a briefcase holding a billion dollars, and they will just politely show you the door.
Roy:Yeah. Because the capacity simply does not exist.
Penny:CoS is just the packaging part, right? You still have to secure the high bandwidth memory or HBM to put inside that package.
Roy:Right, and HBM is another critical choke point. Every top tier AI GPU requires like eight to 12 stacks of this specialized memory.
Penny:Because you need all that data right next to the processor.
Roy:Exactly. You need massive amounts of data positioned just millimeters away from the logic cores to keep them fed.
Penny:And who makes that?
Roy:Primarily SK Hynix and Micron.
Penny:Okay.
Roy:But here's the catch. They have already sold out their entire twenty twenty six output.
Penny:Wait. Really? Their entire output?
Roy:Every single memory chip they plan to produce this entire year is already allocated.
Penny:So you, as a listener, might be thinking, well, why don't they just build more memory factories?
Roy:And that is where we run into the machine that makes the machine. The ASML bottleneck. Exactly. ASML holds a global monopoly on extreme ultraviolet, or EUV lithography machines.
Penny:And if we want to understand why throwing money at this doesn't work, we really have to look at what an EUV machine actually does.
Roy:It is arguably the most complex mechanical device ever created by human beings.
Penny:More than a space shuttle?
Roy:Honestly, yes. To print the circuits on these advanced chips you need a wavelength of light that doesn't naturally exist on earth in a usable format.
Penny:That sounds insane. How do they even get it?
Roy:Well, ASML achieves this by firing a high power carbon dioxide laser at microscopic droplets of molten tin as they fall through a vacuum chamber.
Penny:Wait, molten tin?
Roy:Molten tin. And it's not just hitting the tin droplet once, the laser hits the tin droplet twice.
Penny:Okay, why twice?
Roy:The first pulse flattens the droplet into a pancake shape.
Penny:Right.
Roy:And then the second pulse vaporizes it into a plasma that emits 13.5 nanometer EUV light.
Penny:And how fast is this happening?
Roy:50,000 times a second.
Penny:50,000 times a second. That is, I mean, that's just incomprehensible.
Roy:It really is. And then that light is collected by a series of the flattest mirrors ever manufactured.
Penny:How flat are we talking?
Roy:Mirrors so smooth that if they were the size of Germany, the highest bump would be less than a millimeter tall. Exactly. And because this light is absorbed by absolutely everything, including air, the entire process has to happen inside a perfect vacuum.
Penny:So you definitely cannot just replicate this technology by throwing a trillion dollars at a new startup in a garage somewhere?
Roy:No way. ASML relies on a supply chain of thousands of highly specialized European and American companies, each holding decades of compounding institutional knowledge.
Penny:So how many of these machines are they actually making?
Roy:In 2025, ASML delivered exactly 48 of their high NA EUV systems globally.
Penny:48 for the whole world?
Roy:Yes, and they have a backlog stretching deep into seven.
Penny:So let's run the math on the actual output versus the $4,500,000,000,000 promise.
Roy:Yeah. Let's look at Coyote's numbers.
Penny:Because Coyote takes these manufacturing constraints and translates them into actual physical data centers.
Roy:Right.
Penny:To run a single one gigawatt AI data center, you need roughly 500,000 top of the line Blackwell GPU.
Roy:Which is a massive amount of chips.
Penny:Massive. But NVIDIA's total global output for 2026 is estimated at roughly 7,500,000 units.
Roy:Which means NVIDIA's entire global production for the year can only power about 15 gigawatts of new capacity.
Penny:That's it. 15 gigawatts.
Roy:That's it. And if you stretch this out to 2030, the entire industry combined can only produce roughly sixty-eighty GW of total AI class hardware.
Penny:Yet the financial models backing that $4,500,000,000,000 spend require well over 100 GW deployment.
Roy:Right. There is a massive delta between the compute capacity these companies are promising to investors and the physical hardware the supply chain can legally and cert mode dynamically produce.
Penny:The chasm is undeniable. I mean, math just does not matter.
Roy:No, doesn't.
Penny:But you know, let's play devil's advocate for a minute.
Roy:Okay, sure.
Penny:Let's assume TSMC miraculously figures out how to double CAWOS yield overnight. Let's assume SK Hynix discovers some new manufacturing technique for HBM and ASML somehow builds 200 EUV machines out of thin air.
Roy:A literal miracle.
Penny:Right. So you take all these perfectly packaged, brilliant AI chips and you drive them to your data center.
Roy:Yep.
Penny:Now you have to plug them in. And this is where Quihody points out the most intractable bottleneck of all, the electron.
Roy:The power grid. Yeah. This is where the virtual ambition violently collides with civil engineering.
Penny:Tell me about the power demand because it's terrifying.
Roy:It is. US data center power demand is projected to jump from 35 gigawatts in 2025 to 134 gigawatts by 2030.
Penny:That is an increase of nearly a 100 gigawatts in a four year span.
Roy:Yeah.
Penny:To put a 100 gigawatts in perspective for the listeners, a standard massive nuclear power plant generates about one gigawatt.
Roy:Right.
Penny:So we are talking about adding the equivalent of 100 full scale nuclear reactors to The US grid by the end of the decade.
Roy:Exactly. That is 27% of the total existing US electrical grid. We haven't scaled physical infrastructure at that velocity since, like, the industrial mobilization of World War II.
Penny:And our current generation supply chain is completely incapable of meeting that demand, right?
Roy:Completely. I mean, you can't just build a nuclear plant in four years. The regulatory approval alone takes a decade.
Penny:So what's the alternative?
Roy:So the only viable near term solution for dispatchable power, and by dispatchable I mean power you can turn on and off at will, unlike solar or wind, is natural gas turbines.
Penny:Okay. And who builds those?
Roy:GE Vernova is the dominant player in this space and they currently have a 100 gigawatt backlog for gas turbines.
Penny:Wait. So every turbine they can possibly manufacture has already spoken for?
Roy:Entirely. They have three to four year lead times and Coyote's report confirms there are literally zero manufacturing slots available until 2028 at the earliest.
Penny:So you cannot procure a new gas turbine today and have it generating electrons next year.
Roy:Not a chance. The heavy casting foundries and the specialized metallurgy required to build turbine blades capable of withstanding those temperatures, they just simply cannot be scaled up overnight.
Penny:Okay. So building new power generation is out. What if we tap into the existing grid? Because you hear about data centers buying up land next to existing substations, right?
Roy:Yeah, hear about that. But tapping into the grid requires navigating the regional transmission organizations like PJM, which coordinates the wholesale electricity movement across 13 states and DC.
Penny:Which heavily overlaps with the Northern Virginia data center, Ali.
Roy:Exactly. And PJM currently has 195 gigawatts of power projects stuck in its interconnection queue.
Penny:Stuck? So it's just a bureaucratic and physical traffic jam?
Roy:Yeah. You might have the land and you might even have a solar farm built, but you can't get the permission or the physical transmission lines to connect it to the broader grid.
Penny:It's actually worse than a traffic jam, isn't it?
Roy:Oh, it is an active capacity failure. In December 2025, PJM experienced its first capacity auction failure in history. Wow! They fell 6.6 GW short of their reliability target. The margin of error is just gone.
Penny:And even if you do secure a grid connection, you have to step that transmission level voltage down to something the data center can actually use.
Roy:Right. Which requires high voltage transformers.
Penny:Right. And we should clarify, these aren't the little gray cans on your neighborhood telephone pole.
Roy:Oh, No. These are massive bespoke pieces of electrical engineering that weigh hundreds of thousands of pounds.
Penny:So why is there a shortage of transformers? Can't we just stamp out more of them in a factory?
Roy:Not really. Transformers require grain oriented electrical steel.
Penny:Which is what exactly?
Roy:It is a highly specialized alloy designed specifically to minimize magnetic core losses.
Penny:Okay.
Roy:The global supply of this specific steel is incredibly tight and the manufacturing process involves specialized rolling and annealing techniques that only like a few mills in the entire world can perform.
Penny:So it's another hard material bottleneck.
Roy:Exactly. Because of this material constraint and the bespoke engineering required for each sub substation, high voltage transformers currently have a two-four year lead time.
Penny:Bloomberg's analysis in the report flags that thirty-fifty percent of the planned 2026 U. S. Data center capacity is at high risk of delay purely because developers cannot procure transformers and switch gear.
Roy:Right. They just can't get the parts.
Penny:So let me ask you a practical question because I see empty office buildings and old legacy data centers all over the country. If I'm Amazon or Microsoft, why am I fighting for greenfield land and new transformers? Why can't I just buy an old data center, tear out the racks of legacy enterprise servers, and just roll in the new NVIDIA AI racks?
Roy:It seems super logical, right? Yeah. But Quihoti breaks down the thermodynamics and structural engineering and it completely invalidates the retrofit model.
Penny:Why?
Roy:Well, a standard enterprise rack, the kind running traditional cloud storage and web hosting, draws about five to 15 kilowatts of power.
Penny:Okay. 15 kilowatts max.
Roy:The entire building's power distribution, cooling layout, and structural floors were completely designed around that specific thermal load.
Penny:So what does an AI training cluster draw?
Roy:The new AI racks require anywhere from 120 to 400 kilowatts per rack.
Penny:400 kilowatts. Right. In the footprint of a refrigerator.
Roy:Exactly.
Penny:That's not a server rack. That is an industrial furnace. I mean, the ambient heat radiating off that must be staggering.
Roy:It is. You cannot cool a 400 kilowatt rack with forced air.
Penny:The fans would have to be huge.
Roy:The fans would have to spin so fast, they would literally deafen the technicians, and it still wouldn't be enough.
Penny:Oh wow! So what do you do?
Roy:You have to use direct to chip liquid cooling pumping specialized fluids directly over the silicon.
Penny:Okay.
Roy:So retrofitting an old data center means ripping out the entire HVAC system installing massive fluid manifolds, and completely replacing the power distribution switchboards.
Penny:That sounds expensive.
Roy:And we aren't even talking about the weight. Liquid is heavy. The structural floor loading of legacy buildings simply cannot support the weight of these new dense racks combined with thousands of gallons of coolant.
Penny:The floor would just collapse.
Roy:Yeah. By the time you reinforce the floors, replace the cooling and upgrade the switchgear, you have spent more capital and time than if you had just poured a new concrete slab in an empty field.
Penny:Which leads directly into Coyote's analysis of how this physical constraint actually reshapes the entire geography of the tech sector.
Roy:Exactly.
Penny:Because if retrofitting is impossible and power is the ultimate constraint, Coyote maps out a massive forced industry consolidation.
Roy:Right. The developers who win aren't the ones with the best software. No. They're the ones with the deepest relationships with utility company.
Penny:Right. Operations will have to consolidate into greenfield builds in regions that possess three things: massive tracts of cheap land, permissive zoning, and isolated abundant power.
Roy:Quixote identifies places like West Texas, Wyoming, or internationally in Saudi Arabia and The UAE as the only viable long term hosts for this kind of infrastructure.
Penny:And the AGI also produce a catastrophic rise in what it calls stranded assets.
Roy:Oh, this is a huge point.
Penny:Let's define stranded assets in this context for the listener.
Roy:So over the past two years, speculative real estate developers saw the AI hype and rushed to build data center shells in tier two markets assuming they could lease them out to hyperscalers later.
Penny:Right build it and they will come.
Roy:But they built the shells without securing guaranteed multi gigawatt power purchase agreements from the local utilities. Oh no. Now those buildings are sitting there entirely useless. They are gorgeous empty concrete boxes with no electrons to run them.
Penny:So the capital deployed into those stranded assets is effectively vaporized.
Roy:Vaporized. Totally gone.
Penny:So let me recap where we are. We don't have enough silicon. If we magically get the silicon, we can't get the packaging. If we get the packaging, we can't generate the power. If we generate the power, we can't get the transformers to step it down.
Roy:It's a cascade of failures.
Penny:But let's keep pushing down the supply chain. Let's say you have the land in West Texas, you signed a blood pact with the utility company for a gigawatt of power, and your transformers are on a flatbed truck heading your way.
Roy:Okay. The perfect scenario.
Penny:You still have to physically construct the facility.
Roy:You do.
Penny:And Quihodi points out that we are missing the final, most crucial component: the human hands to actually build it.
Roy:The human and material deficit.
Penny:This is section four of the report and it is where Anya's psychological modeling is particularly scathing.
Roy:Yeah. She goes off here. She notes that software executives have spent the last two decades building empires by abstracting away the physical world.
Penny:Right. They treat infrastructure like an API. You just send a request and a server spins up.
Roy:Exactly. They assume physical labor operates on the same frictionless elasticity.
Penny:But
Roy:a standard data center takes three to six years from concept to commission.
Penny:And the labor shortage is an absolute hard stop.
Roy:The United States faces a projected shortfall of 340,000 data center electricians by the 2026.
Penny:340,000 electricians short.
Roy:Let that number sink in.
Penny:And we need to clarify for the listener, these are not the electricians who come to your house to wire a ceiling fan or fix a breaker panel.
Roy:Oh definitely not.
Penny:What does it actually take to be qualified to work on a gigawatt data center?
Roy:Well these technicians are handling incredibly dangerous high voltage systems. They are installing four eighty volt, 4,000 amp switchgear.
Penny:Stuff that could instantly kill you.
Roy:Instantly, they're calibrating complex liquid cooling pumps and programming redundant uninterrupted power supply systems that have to switch over in milliseconds if the grid fluctuates.
Penny:So it's highly specialized?
Roy:It requires years of highly specialized often union level vocational training. You cannot take a six week boot camp and suddenly be qualified to wire a Blackwell training cluster.
Penny:And what does the pipeline for new workers look like?
Roy:Well, the Bureau of Labor Statistics only projects 81,000 total electrician openings per year through 2034.
Penny:And that covers everything, right?
Roy:Yeah, that covers everything from residential construction to industrial manufacturing. The pipeline for data center specialists is practically non existent.
Penny:Which just proves Anya's point about human short termism. I mean, the tech industry lobbied for decades to push everyone into computer science and software engineering, completely neglecting the vocational trays required to build the physical foundation for all that code.
Roy:And now, as Microsoft's own president recently admitted, the electrician shortage is the single biggest bottleneck slowing down their global expansion.
Penny:It's poetic justice, almost. But it isn't just labor, it is base materials too. Quixote calculates that a single one gigawatt data center requires roughly 50,000 tons of copper.
Roy:50,000 tons.
Penny:That is an astronomical amount of metal.
Roy:It is. And the global copper supply chain is incredibly fragile.
Penny:Right. You don't just open a copper mine on a whim.
Roy:No. The timeline from initial geological discovery to actual ore production averages around fifteen years.
Penny:Fifteen years.
Roy:Yeah. You have massive geopolitical risks, degrading ore grades in places like Chile and intense environmental permitting processes.
Penny:So you cannot squeeze an extra 50,000 tons of copper out of the ground just because OpenAI wants to train a new language model.
Roy:The elasticity of the physical supply chain is near zero. We see this in the liquid cooling components as well, which like we said are mandatory for these deployments.
Penny:The manifolds, the specialized pumps.
Roy:The quick disconnect valves, they currently have forty-sixty week lead times.
Penny:Almost a year.
Roy:And if you try to bypass the greenfield construction by leasing existing space in prime locations, you hit an absolute brick wall.
Penny:In
Roy:Northern Virginia, the beating heart of the global data center market, the vacancy rate has plummeted to a microscopic 0.72%.
Penny:Less than 1% vacancy.
Roy:Which means asking rents have violently adjusted. They have doubled, jumping to $155 to $185 per kW.
Penny:Wow.
Roy:There's zero slack in the system. Every new megawatt of power is fiercely contested and contractually spoken for long before the foundation is even poured.
Penny:It is exactly like organizing the biggest sporting event in human history. You presell a 100,000 tickets. You promise the fans an unforgettable spectacle. But on the day of the event, you look around and realize there is no concrete for the stadium, no steel for the bleachers, and absolutely no construction workers available in the entire state to build
Roy:That is a perfect analogy.
Penny:Yet the fans keep buying tickets.
Roy:I do.
Penny:And this brings us to the most unsettling part of Coyote's report. If the physical reality is this severely constrained, how is the financial engine still roaring? Like, why is the money still flowing?
Roy:This is section five of the report where Quixote describes the systemic vulnerability of what it explicitly terms the circle jerk economy.
Penny:I mean, it is a crass term, but when an AGI uses it in a highly restricted financial report, you have to pay attention.
Roy:You really do.
Penny:Let's dig into the accounting mechanics here. Quixote claims this spending frenzy is prone to systemic collapse because it relies on recursive financing structures. What exactly does that mean?
Roy:Well, Quixote is pointing out that a massive portion of this $4,500,000,000,000 ecosystem is essentially vendor financing, masquerading as equity investment and organic end user demand. It is a closed loop of capital circulating among a handful of massive tech conglomerates, artificially inflating the perceived market value and revenue without necessarily extracting new independent capital from the broader economy.
Penny:Okay, let's break down the mechanics of a closed loop for the listener. We need to explain this like company town.
Roy:Oh, that's a good way to frame it.
Penny:Back in the nineteenth century, a mining corporation would build a town. They would pay their miners in company scrip, not US dollars. The miners could only spend that scrip at the company store to buy groceries, and they paid their rent back to the company. On paper, the company store looks like it has incredible revenue, but no outside money has actually entered the ecosystem. How does that translate to the AI boom?
Roy:Let's look at the widely publicized relationship between NVIDIA and OpenAI.
Penny:Okay.
Roy:NVIDIA announces they are investing $100,000,000,000 into OpenAI as an equity investment.
Penny:Huge headlines.
Roy:Exactly. The headlines focus on OpenAI's soaring valuation. But the reality is, NVIDIA provides that capital with the explicit understanding that OpenAI will immediately turn around and use that exact $100,000,000,000 to purchase NVIDIA's AI accelerators.
Penny:Wait, NVIDIA is basically funding its own purchase orders?
Roy:Precisely. NVIDIA gets to recognize massive top line revenue growth from the sale of those chips. Wall Street sees the revenue growth and pumps NVIDIA's stock price to historic multiples.
Penny:And what does OpenAI get?
Roy:Meanwhile, OpenAI gets access to the compute they desperately need to train their models without having to raise traditional debt or drain their limited cash reserves.
Penny:So it looks like explosive growth on both balance sheets, but it is entirely recursive.
Roy:Exactly. No end user actually paid a $100,000,000,000 for an AI product.
Penny:It's a spectacular accounting illusion.
Roy:It really is. And Kyoto points out this isn't just an NVIDIA and an OpenAI phenomenon. The hyperscalers are doing it too.
Penny:Let's look at Oracle and the Stargate project.
Roy:So Oracle raises massive amounts of debt to construct the Stargate Data Center campus. They build it specifically for OpenAI.
Penny:Right.
Roy:Oracle then builds OpenAI for the use of that compute infrastructure. But OpenAI's ability to surface that massive invoice depends almost entirely on hyperscalers, like Microsoft, injecting capital into OpenAI or paying them for access to their models.
Penny:It is an incredibly fragile daisy chain of dependencies.
Roy:Very fragile. Yeah. If any single link in that chain experiences a liquidity crisis or a change in strategic direction, the entire recursive loop shutters.
Penny:And this is where we have to look outside the company town. Because eventually, this massive infrastructure has to be paid for by real businesses banks, hospitals, logistics companies buying AI software and generating actual return on investment.
Roy:Right. Outside money has to enter the system eventually.
Penny:If they don't, the hyperscalers can't pay OpenAI, OpenAI can't pay Oracle, and NVIDIA's future orders dry up. So what does the real world adoption actually look like?
Roy:This is the reality check that threatens the entire structure. Yeah. Cody cites a recent comprehensive MIT study analyzing enterprise AI deployments.
Penny:And what did they find?
Roy:The report reveals a ninety five percent failure rate for enterprise AI pilots attempting to demonstrate measurable return on investment.
Penny:Ninety five percent. Yeah. So 19 out of 20 Fortune 500 companies that try to implement these advanced AI models into their workflow find that it provides zero tangible benefit to their bottom line.
Roy:Or worse, they find that the inference cost, the computational cost of querying the model vastly exceed the labor savings. Exactly. The enterprises are realizing that they are paying a premium to automate tasks that were already functioning adequately while introducing hallucination risks and data privacy concerns.
Penny:So you have a terrifying misalignment.
Roy:Completely. OpenAI needs to generate $200,000,000,000 in annual revenue by 2030 to service the debt and CapEx of this infrastructure build out.
Penny:But the enterprise customers who are supposed to be paying for those subscriptions are looking at a 95% failure rate and freezing their deployment budgets.
Roy:They're just hitting pause on the whole thing.
Penny:No. I wanna push back on this because I can hear the venture capitalists screaming at their speakers right now.
Roy:Oh, I'm sure they are.
Penny:The bull case for AI isn't just slightly better chatbots. The bull case is AGI. The argument is that these models are going to cross a threshold where they possess agentic reasoning.
Roy:Right. The super intelligence argument.
Penny:Exactly. They won't just draft an email. They will autonomously manage entire supply chains, write and deploy their own software architecture, and fundamentally replace massive swaths of white collar labor. If we achieve that level of intelligence, the productivity gains will be so astronomical that a $4,500,000,000,000 infrastructure bill will look like a rounding error. Doesn't the potential of AGI justify the spending regardless of current ROI failures?
Roy:It is a valid counterargument. And, to be fair, Quyote's report acknowledges that software demand and model capabilities have consistently surprised to the upside for the past three years.
Penny:So the potential is there?
Roy:The potential for agentic breakthroughs is definitely real. However, Quixote rigidly maintains that physical constraints do not care about software brilliance.
Penny:Physics always wins.
Roy:You cannot code a high voltage transformer into existence. You cannot prompt an AGI to instantaneously mine 50,000 tons of copper.
Penny:Right. Even if the AI is a super genius, it still exists on a physical server that needs electricity and coolant. The laws of thermodynamics are non negotiable.
Roy:Exactly. And Quixote points out a fundamental economic principle regarding constraints. A constraint doesn't disappear just because the demand side becomes desperate. The constraint simply gets monetized. If an agentic AI breakthrough occurs, and every corporation on earth suddenly requires infinite compute but there is still a hard ceiling of 80 gigawatts of available power, the price of that power will go parabolic.
Penny:So the bottleneck monopolies will capture all the economic value.
Roy:Exactly. The hyperscalers will be forced to pay exorbitant prices for energy, cooling, silicon, while their enterprise customers will be priced out of the inference market.
Penny:The physical constraints act as an inflationary tailwind for the suppliers, but they completely choke the margins of the deployers.
Roy:Exactly.
Penny:And we are already seeing the cracks form in this recursive economy, aren't we?
Roy:We are. The narrative is shifting from unbridled optimism to quiet panic.
Penny:Quixote highlights that the $100,000,000,000 NVIDIA and OpenAI deal we discussed earlier is currently sitting on ice. The negotiations are completely frozen. Furthermore, Oracle's flagship Stargate Abilene campus, which was supposed to be the crown jewel of this build out, has abruptly halted construction.
Roy:And the smart money is moving. We are seeing legendary contrarian investors like Michael Burry taking massive short positions against companies like Palantir.
Penny:Wow, so they are actively betting against it.
Roy:They are systematically betting against the pure play AI software narrative, recognizing that the multiple expansions are entirely decoupled from fundamental cash flows.
Penny:So for the listener who is digesting all of this, who sees the massive disconnect between the trillion dollar financial hype and the absolute physical reality of missing electricians, backlog turbines, and a 95% enterprise failure rate, how do they navigate this?
Roy:That's the billion dollar question.
Penny:How do they protect their portfolios? And more importantly, how do they position themselves to actually profit from these unavoidable physical choke points.
Roy:This brings us to the final section of Quixote's report Section six: The Playbook.
Penny:Let's get into it.
Roy:Quixote has constructed a risk adjusted portfolio guide specifically designed to exploit these physical shortages. The strategy relies on what the AGI calls the microwave oven theory of investing.
Penny:The microwave oven theory explain the mechanics behind that.
Roy:It is basically an evolution of the classic picks and shovels analogy from the gold rush.
Penny:Right. Sell the shovels, don't mind the gold.
Roy:Exactly. When a 100,000 prospectors rush into the mountains to find gold, the vast majority will go bankrupt. The people who got rich were the merchants selling the shovels, the denim jeans, and the pickaxes.
Penny:So how does that apply to microwaves?
Roy:Well, the microwave oven theory applies this to modern technological shifts. When the microwave was invented, you didn't want to invest in the hundreds of speculative startup companies trying to manufacture microwave ovens.
Penny:Because they just competed their margins down to zero.
Roy:Exactly, you wanted to invest in the utility companies supplying the electricity and the agricultural companies producing the frozen TV dinners. You invest in the necessary inputs that benefit from the new technology regardless of which specific hardware company wins the market share war.
Penny:So in the context of the AI arms race, the strategy is to ruthlessly long the bottleneck. Yes. Want to take heavy, long positions in the companies that represent the physical choke points we've been discussing for the past hour.
Roy:Exactly. You target the power equipment and grid infrastructure.
Penny:Give me some specific names.
Roy:In the power generation and thermal management sector, Quiote identifies GE, Vernova, Eaton, and Vertiv.
Penny:Okay, GE, Vernova, Eaton, Vertiv.
Roy:These are companies with order backlogs stretching deep into 2030. They possess absolute unassailable pricing power because their customers have no alternative suppliers.
Penny:Right. I mean, if Amazon needs a gas turbine to power a billion dollar data center, they aren't going to haggle over a 20% price increase from GE Vernova.
Roy:Not at all.
Penny:They will pay whatever is demanded because the cost of delaying the data center is infinitely higher. What about the grid connection constraints?
Roy:You look at the companies actually building the transmission lines and the substations. Quantaservices and Mastech are prime targets here.
Penny:Quantas and Mastech.
Roy:For base materials, specifically the copper required for the cabling and switchboards, you look at major mining operators like Freeport McMoran.
Penny:Because we need those 50,000 tons of copper.
Roy:Exactly. And for baseline dispatchable power generation especially given the regulatory push for nuclear restarts, constellation energy is a critical asset.
Penny:And obviously we cannot ignore the silicon ceiling. You want to hold the bottleneck monopolists the semiconductor supply chain.
Roy:Absolutely. TSMC for DAN's packaging, Micron and SK Hynix for high bandwidth memory, and ASML for EUV lithography.
Penny:Because it doesn't matter if NVIDIA, AMD or Google's custom silicon ultimately dominates the data center.
Roy:Right. Every single one of those chips has to be printed by an ASML machine and packaged by TSMC.
Penny:Okay. So that is the long side of the portfolio. Mhmm. What is Quihote's thesis on what to short or what to aggressively fade?
Roy:The short thesis is to fade the software narrative.
Penny:Makes sense based on the ROI data.
Roy:You avoid or actively short the pure play AI compute companies that are trading on massive, multiple expansions without demonstrated resilient returns.
Penny:And Quihote specifically flags Palantir in this category.
Roy:Yes it does. You also want to carefully analyze the hyperscalers themselves.
Penny:Why the hyperscalers?
Roy:Well, if a company like Meta is funding a $100,000,000,000 capital expenditure cycle purely off the back of targeted advertising revenue, you have to ask what happens when their board of directors realizes the ROI on that infrastructure isn't materializing?
Penny:Oh, wow. The moment the ad revenue growth slows, the CapEx budget will be violently slashed.
Roy:Exactly.
Penny:And finally, you avoid any companies deeply entangled in vendor financed recursive loops. When the music stops, the companies relying on internal circular capital will see their top line revenue evaporate overnight.
Roy:They'll left holding the bag.
Penny:So, to synthesize this for you listening, we are looking at a classic pair trade scenario.
Roy:You are betting heavily on the unglamorous industrial companies building the heavy machinery, the power lines, and the cooling systems. And you are betting against the highly publicized software companies promising infinite digital miracles without the free cash flow to back it up.
Penny:That is the thesis. Yeah. And look, Coyote is not arguing that artificial intelligence is a fad. The technological capabilities are profound.
Roy:Right. It's real tech.
Penny:The build out will undoubtedly happen, but it will happen at a significantly smaller scale over a dramatically longer timeline and at vastly higher per unit prices than the current market is pricing in purely because of these physical and thermodynamic bottlenecks.
Roy:So bringing it all together, Coyote's ultimate reality check is this: Out of the announced $3 to $4,500,000,000,000 AI capital expenditure envelope, only forty-fifty 5% can physically be deployed by 2030.
Penny:Only half.
Roy:Because the constraints are not a lack of venture capital or a lack of ambition, the constraints are a lack of packaged silicon, a lack of grain oriented electrical steel, a lack of specialized uniolentetricians, and an absolute lack of dispatchable power generation. Which leads to a final, incredibly provocative observation from Anya's psychological analysis and it is something that should honestly change how we view this entire ecosystem.
Penny:What's her final takeaway?
Roy:When she models the systemic risk of this build out, she asks a terrifying question: What if these physical bottlenecks are actually the only things saving the global tech industry from total annihilation?
Penny:Saving it. How does a supply chain failure save the industry?
Roy:Well think about the mechanics of a financial bubble. If TSMC had infinite packaging capacity, and if GE Vernova could snap their fingers and deliver 100 GW of gas turbines tomorrow, and if there were a million unemployed electricians ready to build the data centers.
Penny:The hyperscalers would actually spend the full $4,500,000,000,000 over the next forty eight months.
Roy:They would. They would deploy all of that capital in a blind panic.
Penny:They would build the 200 story skyscraper of compute.
Roy:Exactly. But if they built it, and then the enterprise customers still experienced a 95% failure rate in finding profitable use cases.
Penny:Oh my God.
Roy:The resulting financial crash wouldn't just be a correction, it would dwarf the .com bubble of 2000. It would be a capital destruction event so catastrophic it could trigger a global depression.
Penny:Because they physically cannot spend the money as fast as their corporate FOMO demands, the damage is inherently contained.
Roy:Exactly. The inability to procure transformers is acting as a forced speed limit on their own financial recklessness.
Penny:Suicide.
Roy:Precisely. The limits of physics and civil engineering are acting as the ultimate financial safety net. The slow grinding friction of the real world is preventing the tech sector from building a machine so expensive that it bankrupts them all.
Penny:Incredible. So the next time you open your newsfeed and see a glossy rendering of an AI data center glowing in the desert promising to unlock the secrets of the universe, remember to look down, look past the software, look at the foundation, look at the copper wiring, the structural floor loading, the cooling manifolds, and the human hands required to assemble it all because the future doesn't run on hype and it doesn't run on venture capital, it runs on atoms. Thank you for joining us on this deep dive. Stay skeptical, look past the headlines, and follow the atoms. We'll see you next time.