Latest / Deep Values: Value Investing Deep Dives / The ten trillion dollar AI compute race explained
Transcript
- 0:00So imagine for a second that the entire global economy, like the
- 0:04interest rate on your mortgage or whether a whole sovereign
- 0:08country goes bankrupt, was suddenly being dictated by just
- 0:12two tech companies. Right, which sounds completely
- 0:14wild. It does.
- 0:16And imagine they aren't doing this by, you know, controlling
- 0:19oil or gold, but by hoarding this completely new, totally
- 0:24invisible resource. Yeah.
- 0:25It's a massive paradigm shift. I mean, we are so conditioned to
- 0:28think of physical commodities as the ultimate levers of global
- 0:31power. Exactly.
- 0:32But what we are looking at right now is a scenario where the new
- 0:35oil is entirely digital, it's being pumped out at just
- 0:39unprecedented rates, and the pipelines are controlled by an
- 0:42incredibly small handful of people who are, well, they're
- 0:45basically rewriting the rules of capitalism as they go.
- 0:48Welcome to the deep Dive. Today we are unpacking a
- 0:51conversation that completely broke my brain regarding the
- 0:53future of AI. We're looking into a genuinely
- 0:56stunning analytical exchange between tech podcaster Dwarkesh
- 1:00Patel and Semi Analysis founder Dylan Patel.
- 1:03Yeah, it's such a crucial source to dig into right now.
- 1:06Our mission today is to explore the absolute bleeding edge of
- 1:10the AI compute landscape. We're going to dive into how top
- 1:13labs like Open AI and Anthropic are shifting from burning
- 1:17venture capital to literally printing cash.
- 1:20And why they're suddenly hiding their absolute best models from
- 1:22the public too, right? And how their insatiable hunger
- 1:25for data centers might actually trigger a global financial
- 1:29earthquake. I promise you, by the end of
- 1:32this conversation you are going to look at every piece of tech
- 1:34news you read in a completely different light.
- 1:37It's an essential conversation to have because the underlying
- 1:40mechanics of this shift are completely hidden from the
- 1:43average user. Oh totally.
- 1:44Like when you open a browser to use an AI chat bot, you just see
- 1:47a clean minimalist interface. You don't see the tectonic
- 1:51trillion dollar shifts happening in the physical world, those
- 1:54supply chains, the power grids, the boardrooms, just to make
- 1:57that single text response possible.
- 1:59OK, let's unpack, because before we can understand this impending
- 2:02global takeover, we have to look at the shocking new math that
- 2:06started it all. This is the moment AI model
- 2:10stopped being money pits. Which was a very recent shift.
- 2:13Yeah, exactly. Just a year and a half ago, it
- 2:15felt like these top AI labs were surviving strictly on Venture
- 2:19capable. They were burning billions just
- 2:23to figure out how to make these model string coherent sentences
- 2:26together. Yeah, they were bleeding cash.
- 2:28But now, now we are seeing numbers that show they are
- 2:31generating massive, undeniable revenue.
- 2:34If you look at the raw cost of compute, which the industry
- 2:37measures in megawatts, the baseline cost to run these
- 2:40systems right now is roughly 10 to $15 million per MW, right?
- 2:44But running advanced models like Anthropics Opus 5 or what is
- 2:48expected from GPT 5.6 that is generating up to $50 million per
- 2:53MW in revenue. And that specific margin is
- 2:56really where the entire industry pivots.
- 2:58It's huge. I mean, you think back to the
- 2:59recent past, running something like GPT 4 on older NVIDIA
- 3:02Hopper chips was largely operating on negative margins.
- 3:05Really negative. Yeah, and the mechanism there is
- 3:08simple. Older hardware took so much time
- 3:10and so much electricity to generate a single word or a
- 3:14token that the power bill literally outweighed what they
- 3:17were charging users. They were bleeding money on
- 3:21every single query just to build market share and get people
- 3:24hooked. But now, with newer iterations
- 3:27and far more efficient silicon, the processing speed has
- 3:31outpaced the electricity cost, so the unit economics have
- 3:34totally flipped. So it's like finding a magical
- 3:37vending machine where every time you put in a $10.00 bill, it
- 3:40spits out of 50 or even $100 bill.
- 3:43Naturally, your immediate reaction is to try and buy every
- 3:46single vending machine on planet Earth.
- 3:48Exactly. And that aggressive reinvestment
- 3:50is exactly what we are witnessing.
- 3:52Yeah. I mean, when any business model
- 3:54achieves this level of positive gross margin, they don't just
- 3:57sit on the cash, right? All that profit, it gets
- 3:59aggressively recycled right back into acquiring more compute.
- 4:03It fuels this massive exponential scale up.
- 4:06It's just exponential growth, yeah.
- 4:08And to put the scale into perspective, Open AI and
- 4:10Anthropic are jumping from utilizing about two gigawatts of
- 4:14compute early this year to over 5 gigawatts by year end.
- 4:18Wait, five gigawatts, that's roughly a three to four fold
- 4:20increase in just months. It is.
- 4:23And looking ahead to next year, based on contracts that are
- 4:25already inked, these two labs alone are positioned to swallow
- 4:3040 to 50% of all newly available global compute.
- 4:34Which is just staggering when you think about a GW basically
- 4:38being enough power to run a mid size city.
- 4:40But this brings up a huge question for me.
- 4:43Sure good if the unit economics are this incredible.
- 4:46Like if you literally have a vending machine spitting out
- 4:49hundreds, why isn't the world just instantly building infinite
- 4:53AI data centers? Well, that's the multi trillion
- 4:56dollar question. Because the numbers I was
- 4:57reading from the source highlight this mind blowing
- 4:59ratio. A $6 billion investment in
- 5:02semiconductor fabrication equipment yields about one GW of
- 5:06compute per year. But that one GW can generate
- 5:09$100 billion in terminal AI revenue, meaning the total
- 5:14lifetime value it creates before the hardware goes obsolete.
- 5:17Yep. The leverage is insane.
- 5:19So wait, if there's literally 100 X return on the table,
- 5:23shouldn't capitalism instantly fix this?
- 5:25Like, if I'm a Wall Street investor, I'm throwing limitless
- 5:29money at the supply chain right now.
- 5:31Why is the pipeline still so clogged?
- 5:33That is the logical assumption, right?
- 5:35But it runs face first into the brutal realities of physical
- 5:38manufacturing. How so?
- 5:40Physical manufacturing isn't software.
- 5:42You can't just write a massive check, hit a few keystrokes, and
- 5:45magically spawn highly specialized factories overnight.
- 5:48So money doesn't just instantly equal factors.
- 5:51Exactly. To understand the bottleneck,
- 5:53you have to look at the absolute bedrock of this supply chain.
- 5:56You have companies like ASML which builds UV lithography
- 6:00machines. Right, the chip printers
- 6:01basically. Yeah, to put it simply, these
- 6:04are machines that use extreme ultraviolet light to print
- 6:07microscopic transistors literally smaller than a virus
- 6:10on to silicon wafers. And to make those machines work,
- 6:14you need ultra precise mirrors made by a company called Carl
- 6:17Saeiss. Oh right, so you have to fix
- 6:20every single microscopic link in the supply chain simultaneously.
- 6:23Right down to the atomic level. I mean, earlier this year, Zeiss
- 6:27didn't even project they needed to produce that many mirrors.
- 6:30Now, facing this AI boom, they're realizing they need to
- 6:34scale up to support building 100 of these massively complex EUV
- 6:39machines a year by the end of the decade.
- 6:42Expanding a factory that builds the most precise mirrors in
- 6:45human history takes years to propagate.
- 6:48You can't just hire 1000 extra workers and speed it up.
- 6:51Which creates these wild arbitrage opportunities.
- 6:54In the meantime, the analysis noted that if you had $400
- 6:58million right now, you could theoretically convince ASML to
- 7:02sell you one of those lithography machines, Literally
- 7:05just hold onto it in a warehouse for a bit and flip it for over a
- 7:08billion dollars. And that arbitrage only exists
- 7:11because the supply curve is completely broken.
- 7:13It perfectly highlights how desperate the situation is.
- 7:15Yeah, it's just pure desperation.
- 7:17And keep in mind, even if you magically secure the ASML
- 7:20machine and print the chips, you still need the physical energy
- 7:23grid to power them, the data centers to house them, the
- 7:26liquid cooling system so they don't meltdown, and the
- 7:29downstream packaging facilities. It's an entire ecosystem that
- 7:32has to scale at once. Exactly.
- 7:34The AI laboratories are going to generate hundreds of billions in
- 7:37revenue, but the total industry capital expenditure required to
- 7:41actually build the physical infrastructure they want is
- 7:44sitting around $2 trillion. The labs simply don't have the
- 7:48cash flow to fund the expansion of the entire global supply
- 7:53chain at the speed their software demands.
- 7:56Which leads us to a really counterintuitive pivot by these
- 7:59tech giants. Because supply is so severely
- 8:02bottlenecked and demand from the public is practically infinite,
- 8:06the labs are realizing they are essentially giving away their
- 8:08golden geese. Yeah, they are.
- 8:11They are moving from an era of public release to an era of the
- 8:14great hoarding and. What's fascinating here is that
- 8:16it completely defies standard investor logic.
- 8:19Yeah, normally if you are a major tech company, you are
- 8:21laser focused on maximizing immediate quarterly profits to
- 8:25keep your stock price up. Right, You want those quick
- 8:27wins. Yeah, the easiest way to do that
- 8:29is through inference, which just means letting the public and
- 8:31other businesses use your AI model via an API and charging
- 8:35them a fee every time it generates a response.
- 8:37Like subscribing to ChatGPT. Exactly, but the labs have woken
- 8:41U to a very hard truth. The vast majority of the
- 8:45economic value being generated by their AI is currently being
- 8:48catured by the end users, not the labs themselves.
- 8:52I was reading some wild examples of this value capture.
- 8:55You have firms like Jane St. You know, this massive
- 8:59quantitative trading firm signing exclusive contracts for
- 9:02ultrafast model access and using it to make absolute fortunes in
- 9:06the financial markets. Oh.
- 9:07Yeah, massive returns. Or take Meta.
- 9:10They are using empropic res models to optimize their
- 9:14internal AD algorithms. If Meta uses AI to make their ad
- 9:18targeting just 5% more efficient, that translates to
- 9:21billions of dollars in new revenue for them.
- 9:24Yeah, billions. And that completely dwarfs the
- 9:26few $1,000,000 they paid Anthropic and token fees.
- 9:28Exactly. When anyone can rent a server
- 9:31rack, access an advanced model, and start generating 100 times
- 9:34more revenue than the raw cost of the compute, the labs
- 9:37naturally start doing the math. Right, they realize they're
- 9:40getting a bad deal. Exactly, they asked themselves,
- 9:43why are we renting out this precious, impossibly scarce
- 9:46compute for $30 million a MW to external comanies when using it
- 9:51internally to research and build a betawatt artificial general
- 9:56intelligence could eventually yield 100 million or even a
- 9:59billion dollars per MW? So what does this all mean for
- 10:02you listening right now? It means the ChatGPT or Claude
- 10:06you are using on your phone today might already be massively
- 10:09outdated compared to what these companies are using in secret.
- 10:12It's highly likely. Because the best models are
- 10:15actively going dark, we saw Open AI pause their highly
- 10:19anticipated Astra model, and they even halted their training
- 10:22clusters for a couple of weeks just to reallocate compute.
- 10:25Needed the power elsewhere. And Anthropic hasn't released
- 10:27the next major iteration of Mythos, and reportedly at one
- 10:30point they even restricted its own foreign workers from
- 10:33accessing it internally. Like they are literally shutting
- 10:36the doors and keeping the smartest minds digital and human
- 10:39inside. It signals A fundamental shift
- 10:42from infant serving the public back to training and internal
- 10:46research. The labs know that the best AI
- 10:49models generate exponentially more value when they're pointed
- 10:53inward to recursively improve themselves.
- 10:55AI building better AI. Exactly, if a lab can use its
- 10:59absolute newest unreleased model to write better code, design
- 11:03more efficient data center architectures, and essentially
- 11:06do the job of highly paid human researchers, that internal labor
- 11:10is worth rejecting hundreds of billions of dollars in short
- 11:13term public revenue. Wow.
- 11:14They are not trying to be the most profitable software service
- 11:17of 2026. They are chasing the winner take
- 11:19all prize of AGI. So if these US labs are locking
- 11:23the doors and hoarding all this compute to build a digital God,
- 11:25it makes me wonder what happens to the rest of the world.
- 11:28Yeah, the global divide is huge. Let's look at the geopolitics of
- 11:31this silicon race, because looking at the numbers, China
- 11:34currently has less than 10% of the globally deployed data
- 11:37center AI compute wattage. And looking ahead to 2028,
- 11:41projection suggests China will have at most 30 gigawatts of
- 11:44compute, compared to a massive global total of over 200
- 11:48gigawatts. The disparity is striking, and
- 11:50it is driven by two very different financial ecosystems
- 11:53colliding with strict US export controls.
- 11:56Right. the US financial system is uniquely built to encourage
- 11:59massive venture capital risk taking on a highly speculative
- 12:03software and hardware. China's ecosystem relies heavily
- 12:07on state subsidies, though. You know, it is worth noting
- 12:10that Chinese state subsidies for semiconductors actually exceed
- 12:13the rest of the world's combined subsidies.
- 12:15But I want to push back on that a bit, or at least challenge the
- 12:17premise. If the US labs are pulling this
- 12:19far ahead and raw compute right now, is the global AI race
- 12:23effectively already over? Or could China Era's massive
- 12:27state backed subsidies act as a wild card if achieving true AGI
- 12:31turns out to take longer than Open AI thinks it will?
- 12:33That is the critical nuance. The industry talks about two
- 12:36scenarios, a fast take off and a slow take off.
- 12:39OK, break that down for me. So if the path to AGI is a slow
- 12:42take off, meaning it requires many more years of grinding
- 12:45physical engineering rather than an imminent magical software
- 12:49breakthrough, China's massive hardware subsidies could give
- 12:53them a window to catch up. Like they just out spend on the
- 12:56physical side. Right, by 2027 or 2028, domestic
- 13:00foundries in China, companies like SUM P and CXMT, will start
- 13:04bringing significant chip manufacturing capacity online.
- 13:08They could potentially add 5 to 10 gigawatts of compute, maybe
- 13:11up to 50 gigawatts by 2029. But there's a huge quality gap
- 13:15there, right? Even if China produces 50
- 13:18gigawatts of domestic chips, the performance discrepancy means
- 13:21that 50 gigawatts might only equal the computing power of 20
- 13:24gigawatts of US chips. It's not a one to one
- 13:27comparison. It's absolutely not one to one,
- 13:29and to understand why the US-led is compounding, we have to look
- 13:32at how that compute is actually utilized.
- 13:35Well, there's a fascinating breakdown of how top labs budget
- 13:38their compute. It's roughly 50% allocated for
- 13:41research, 10% for development, and 40% for inference.
- 13:46Pure training of a model, the part where just reads the entire
- 13:50Internet to learn language, takes surprisingly little time.
- 13:53Oh, really? Yeah, Anthropics Mythos pre
- 13:55training allegedly took less than 200 megawatts over just a
- 13:59couple of months. So the bottleneck isn't just the
- 14:02raw power to train the model once.
- 14:04No, the real bottleneck where all that 50% research compute
- 14:07goes is the incredibly difficult task of what we call
- 14:11reinforcement learning. Like trial and error.
- 14:13Exactly. The AI guess is an answer gets
- 14:16graded and adjusts, generating endless rollouts.
- 14:19Millions of these trial and error runs across massive server
- 14:21clusters consumes insane amounts of compute, and it doesn't
- 14:25automatically make a model better if you don't have the
- 14:28architecture right. As the US labs pull ahead in
- 14:32using AI to automate their own programming and research, that
- 14:35research budget will consume even more compute, accelerating
- 14:38their lead exponentially. They are using AI to figure out
- 14:42how to build better AI faster than humans can.
- 14:45Here's where it gets really interesting and honestly a
- 14:48little terrifying for the average person to maintain this
- 14:51dominance to fund this massive research loop and build these
- 14:55internal supermodels, the US labs require an infrastructure
- 14:59build out so massive it could literally grake the global
- 15:02economy. It's unprecedented.
- 15:03We are talking about a $10 trillion shockwave.
- 15:06The macroeconomic scale is almost incomprehensible.
- 15:10If the industry moves to building 100 gigawatts of new
- 15:12computer year, that translates to about $5 trillion in direct
- 15:17capital expenditure. Just massive amounts of cash.
- 15:20Yeah, and you can't just buy the servers.
- 15:22A server needs a building, and a building needs power.
- 15:24You have to build new power plants, which are 30 year
- 15:28assets, and the physical data centers, which are 15 to 20 year
- 15:32assets, years in advance of actually turning the chips on.
- 15:36Great. You have to pay for the
- 15:37concrete, the copper and the turbines today for the digital
- 15:40compute you want to turn on tomorrow.
- 15:41Exactly. When you factor all of that in,
- 15:44plus the downstream semiconductor supply chain,
- 15:47total capital expenditure by 2030 could approach $10
- 15:50trillion. To put that in perspective, that
- 15:54is equivalent to roughly 1/3 of the entire United States economy
- 15:57dedicated solely to building data centers.
- 16:00Where on earth does that actual cash come from?
- 16:03Historically, it came directly from the cash flow of the
- 16:05hyperscalers. Google, Amazon, Meta, Microsoft.
- 16:08The usual suspect. Right.
- 16:11But they are no longer producing enough free cash to cover an
- 16:14expansion of this magnitude. They are spending every cent
- 16:18they have on capital expenditures and are now having
- 16:20to borrow aggressively to fund the gap.
- 16:23The economic modeling suggests that between now and 2029, the
- 16:27AI ecosystem might need to issue over $5 trillion in new debt to
- 16:31pay for all this concrete and copper. 5 trillion in new debt.
- 16:35Walk me through that chain reaction.
- 16:37Because we always think of AI as just software on a screen.
- 16:40Maybe a helpful little chat bot writing emails, but how does AI,
- 16:44as is physical hunger for money actually caused my next mortgage
- 16:48rate to spike? OK, think of the global supply
- 16:51of capital, all the money available for lending, as a
- 16:53single giant reservoir. Usually, governments, businesses
- 16:57and homebuyers all take small Sips in this reservoir.
- 16:59Right, everyone gets a piece. Now imagine the AI industry
- 17:02drops a massive high-powered industrial hose into that
- 17:04reservoir and start sucking up $5 trillion.
- 17:08The water level drops rapidly. For everyone else, money becomes
- 17:10scarce. And when money is scarce, it
- 17:12gets more expensive. Exactly, the cost to borrow it
- 17:15the interest rate goes up. If a company like Anthropic is
- 17:19theoretically willing to pay a 20% interest rate on a billion
- 17:22dollars because their internal AI returns are so high, or if
- 17:26Meta is willing to issue bonds at 8% instead of their usual 5,
- 17:29it forces everyone else to pay more to attract lenders.
- 17:33Wow, why would a bank give you a cheap mortgage or funded
- 17:37traditional bridge repair when an AI lab is offering them
- 17:40massive guaranteed yields? The conversation referred to
- 17:43this as a potential new Volcker moment.
- 17:46And for those who might not know, Paul Volcker was the
- 17:48chairman of the Federal Reserve in the 1980s, who famously
- 17:51spiked interest rates to double digits to crush inflation.
- 17:54But here, the spike isn't a deliberate policy from the Fed.
- 17:58It's just a side effect of AI sheer demand for capital.
- 18:01And apparently this could crash traditional safe stocks like
- 18:05Johnson and Johnson or Berkshire Hathaway.
- 18:08Slow down. Explain the mechanism there.
- 18:10Why does AI building a data center make a baby powder
- 18:13company less valuable? It comes down to something
- 18:15called the discount rate. Basically, investors buy safe
- 18:19stocks like Johnson and Johnson for their steady, reliable
- 18:23future profits. Yes, low and steady.
- 18:25But if global interest rates suddenly shoot up because of a
- 18:28IS debt demand, investors can suddenly get a huge guaranteed
- 18:32return just by leaving their money in a savings account or
- 18:35buying government bonds. They don't want to wait 10 years
- 18:37for a company's slow profits anymore unless the stock is
- 18:40super cheap. Oh, I see.
- 18:42So the value of those safe future profits is discounted and
- 18:45the stock prices of those traditional companies plummet.
- 18:48But honestly, the stock market is the least of the worries
- 18:51here. What's the bigger worry?
- 18:52Sovereign debt A sudden massive spike in global interest rates
- 18:57could cause up to 40 developing nations, countries like
- 19:00Pakistan, Nigeria or Argentina to default on their national
- 19:04debt. Good Lord, 40 nations, Yeah.
- 19:08These countries constantly have to take out new loans to pay off
- 19:11their old ones, a process called rolling over debt.
- 19:15If the global interest rate suddenly double s because all
- 19:18the world's capital is being sucked into USAI data centers,
- 19:21these countries simply cannot afford their new loan payments.
- 19:24That is terrifying. Paradoxically, this exact
- 19:27financial chaos, these soaring interest rates and collapsing
- 19:30markets might become the only natural break that forces the AI
- 19:34expansion to slow down. The physical world simply might
- 19:38not be able to finance the digital world's ambitions
- 19:41without breaking. It is an incredible tension
- 19:43between the physical and the digital, and let's bring this
- 19:46all together to look at the actual end game here.
- 19:49As this compute grows, and as these data centers get built,
- 19:52the effective AI population within these labs is exploding.
- 19:56It really is an explosion. The amount of compute required
- 19:59to reach a certain capability drops every year, while the
- 20:01total compute grows exponentially.
- 20:04We are looking at a scenario where a lab goes from having the
- 20:06equivalent of 10 million digital workers this year to 100 million
- 20:10next year on its way to 1 billion.
- 20:13And that is even before factoring in the true
- 20:16accelerator, recursive self improvement.
- 20:19When you have AI models that are smart enough to write the code,
- 20:22that makes the next generation of AI models even smarter.
- 20:25The growth curve goes completely vertical.
- 20:28You are no longer constrained by the number of human engineers
- 20:31you can hire or how many hours they can stay awake.
- 20:34You are only constrained by power and compute, which as
- 20:37we've discussed, they are hoarding and centralizing as
- 20:41fast as physically possible. It is the ultimate
- 20:44centralization of power. Soon a single company could
- 20:48command a digital workforce larger than the entire human
- 20:51population, running 247. Exactly.
- 20:54And this circles us all the way back to our opening thought.
- 20:57We are watching an invisible resource dictate the shape of
- 20:59the global economy, from the supply chain of atomic mirrors
- 21:02down to the interest rate on your house.
- 21:05But it leaves me with one incredibly chilling question for
- 21:08you to Mull over already. If an AI Labs internal digital
- 21:11workforce eventually outnumbers humanity and produces more
- 21:15economic value than most countries combined, does that
- 21:17corporation effectively become a new kind of sovereign nation?
- 21:20That's. Wow.
- 21:21And if it does, who exactly are our elected governments supposed
- 21:25to negotiate with? That is the defining question of
- 21:27our time and the answer is being built in those data centers
- 21:31right now. Thank you for taking this deep
- 21:33dive with us today. Keep questioning the forces
- 21:36shaping our future, and we'll see you next time.