Latest / Deep Values: Value Investing Deep Dives / A16Z: Not A Bubble
Transcript
- 0:00$400 billion. It's a number that it's just
- 0:06hard to wrap your head around, isn't it?
- 0:07It's not just, I think it's impossible.
- 0:09We say billions so much now in tech news that the word has like
- 0:13lost all meaning. But let's actually try for a
- 0:16second. Let's really feel the weight of
- 0:17that $400 billion annualized in capital expenditures, CapEx and
- 0:22almost all of that, that whole mountain of money is being
- 0:26funneled into one single thing, AI infrastructure.
- 0:30Just to put some clothes on that number.
- 0:31Yeah, because it's so abstract. Yeah, that's about the GDP of
- 0:35Denmark or the Philippines. The whole country an.
- 0:37Entire mid sized nations economic output spent in one
- 0:40year mostly on chips, data centers and crucially the power
- 0:44to make them run. It's just it, it's staggering
- 0:46and you know, my brain immediately goes to one place
- 0:48when I hear a number that big pointed at 1 sector.
- 0:50I know exactly where you're. Going Is this a bubble?
- 0:52I mean, is this 1999 all over again?
- 0:54Are we just building the world's most expensive bridge to
- 0:56absolutely nowhere? That is the $400 billion
- 0:59question, isn't it? It's what every CIO, every fund
- 1:02manager is losing sleep over right now.
- 1:04And it should be keeping you, our listener, awake too.
- 1:07Because if this thing pops, it's not just some VC's in Silicon
- 1:10Valley who get hurt. It's it's our 401 KS, it's the
- 1:14whole economy. It is.
- 1:15So, you know, before we even dive in, if you actually want a
- 1:19tool to track this yourself, to get a feel for the market
- 1:22temperature so you're not just, you know, reacting to headlines,
- 1:25you might want to check out the Buffett Indicator app.
- 1:27It's on the iOS App Store. That's a good idea.
- 1:29It's just a great little sanity check to see where total market
- 1:32valuations are, historically speaking.
- 1:33Having your own dashboard is always smart, Yeah, but for this
- 1:36deep dive, we're going to try and answer that question by
- 1:38looking at a very specific set of documents.
- 1:40We're not just, you know, guessing.
- 1:42We're digging into a whole stack of internal reports,
- 1:44transcripts, presentation decks, all from the team at a 16 S
- 1:49Andreessen Horowitz. Yeah.
- 1:51And their growth fund team specifically?
- 1:53Right. And let's be super clear about
- 1:54who this is. These aren't just pundits on TV.
- 1:56These are the people writing the multibillion dollar checks to
- 2:00open AI, to data bricks, to XAI. They are on the inside.
- 2:04They see the server logs, they see the wire transfers.
- 2:06They have the ground truth. And their view, their house
- 2:09view, as the documents call it, is, well, it's pretty
- 2:12surprising. It kind of runs completely
- 2:14counter to that doom and gloom bubble story, doesn't it?
- 2:16It does. They're basically arguing that
- 2:18the whole bubble analogy is flawed, that the math we're
- 2:21using to measure it is just wrong because the underlying
- 2:25variables have fundamentally changed.
- 2:27OK, hold on, wait. This time is different.
- 2:28That's literally the most dangerous phrase in finance.
- 2:31Every single time someone says that six months later the whole
- 2:34market is on fire. Why should we listen to a
- 2:36venture capital firm, a firm whose entire job is to generate
- 2:40hype, when they say don't worry, it's different?
- 2:43And that is absolutely the right skepticism to have.
- 2:46If this was just some blog post, I'd be rolling my eyes right
- 2:49there with you. But what we're looking at is
- 2:51their internal CapEx models, their analysis of market
- 2:54structure. They're not saying it's not a
- 2:56bubble out of some like blind Optimism, OK?
- 3:00They're saying the mechanics of how this money is being spent
- 3:03are physically different for the mechanics of the 2000 crash.
- 3:06The plumbing is different. The plumbing, OK, I'm listening.
- 3:10My guard is a but I'm listening. Fair enough.
- 3:13So here's the plan. We're going to breakdown why
- 3:15they think that bubble math is wrong.
- 3:17Then we're going to talk about this huge structural shift in
- 3:20the stock market that is frankly locking most people out of the
- 3:24biggest gains. And finally, we'll get to the
- 3:27one bottleneck that actually does scare them.
- 3:29And it here's a spoiler. It's not a shortage of NVIDIA
- 3:32chips. It's something more fundamental.
- 3:34Much more And just real quick, a bit of housekeeping before we
- 3:37jump in. You can find full deep value
- 3:39research reports in the description links.
- 3:41Everything we talked about today is based on publicly available
- 3:44sources. It's all for educational and
- 3:46research purposes. Right.
- 3:47This is not investment advice. Please, please do your own due
- 3:50diligence. And if you do want to go deeper,
- 3:51you can subscribe to Deep Values dot AI.
- 3:54We share our latest research there and some pretty advanced
- 3:56AI agent tools for our fellow contrarians.
- 3:59All right, let's do it. So the core tension here, as I
- 4:02see it, is between what feels like pure market mania, the
- 4:06hype, the stock charts, you know, and what the A Sixteens
- 4:10team is calling a fundamental shift in production.
- 4:12That's it exactly. And to get that shift you have
- 4:15to zoom out first. You have to look at the macro
- 4:17picture. The 1st and maybe biggest point
- 4:20in these reports is that tech isn't really a sector anymore.
- 4:23What do you mean by that? I mean, I open my brokerage app,
- 4:25I see a pie chart. It's got a slice for energy, a
- 4:28slice for industrials, a slice for technology.
- 4:30It's right there. You see the slice?
- 4:32Yes, but that's an illusion functionally, in terms of what
- 4:35actually drives the market, what moves the S&P 500, tech has
- 4:39basically eaten the whole pie, right?
- 4:41I mean, look at the S&P 500 today.
- 4:43Who are the top five most valuable companies?
- 4:45Companies, all of them. All of them.
- 4:46Microsoft, Apple, NVIDIA, Amazon, Google.
- 4:50Sometimes it's six of the top 6. If you zoom out to the top ten,
- 4:53it's usually 7 or 8 of them. So the stock market, for all
- 4:56intents and purposes, is big tech.
- 4:57Precisely. The diversification you think
- 5:00you're getting in an index fund is well, it's largely A myth.
- 5:03You're mostly just making a concentrated bet on a handful of
- 5:06mega cap tech stocks. OK, that's one thing.
- 5:09But here's the part of the report that really, it really
- 5:12made me sit up. It explains this feeling I think
- 5:15a lot of us have that the public markets feel harder or just more
- 5:19boring. It's this private for longer
- 5:21trend? Yeah, this is a structural
- 5:23earthquake that I don't think most people have fully grasped.
- 5:25Yet the style was just crazy. I had to read it a couple of
- 5:28times. It changes the entire game for
- 5:29the retail investor. OK, so lay it out.
- 5:31It used to be a company would take what, 5 to 10 years to go
- 5:35public, Something like that. Right.
- 5:36Think about Amazon or Google. They IP owed when they were
- 5:39still, you know, relatively young companies.
- 5:41They went public to get the capital they needed to fund
- 5:43their growth. And if you were a regular
- 5:45person, if you were paying attention, you could buy Amazon
- 5:48stock in the late 90s and ride that wave.
- 5:51You could, you could participate in that hyper growth phase, but
- 5:55now the average time to IPO according to these reports is
- 5:59pushing 14 years, sometimes longer. 14 years in the tech
- 6:03world, That's like 3 different geological eras.
- 6:05It's an eternity, and the consequence of that is just
- 6:08massive for you, for the listener.
- 6:10Think about a company like Databricks or Open AI, which the
- 6:13reports mention a lot. When Amazon went public, it was
- 6:16worth maybe a few $100 million. So if you bought in, you
- 6:19captured the growth all the way up to a trillion dollars.
- 6:22You got the best part of the ride.
- 6:24But a company like Databricks, by the time they even think
- 6:27about an IPO, they're already a giant.
- 6:28They're already a DECA corn worth 10s of billions of
- 6:31dollars, sometimes much, much more.
- 6:33So that whole explosive growth phase, the part where a company
- 6:35goes from a billion dollars to save $50 billion, where is that
- 6:40value creation happening? It's happening entirely in the
- 6:43private markets. It's locked away inside funds
- 6:45like 16, inside university endowments, inside sovereign
- 6:48wealth funds. The public doesn't get a piece
- 6:49of it. We're basically locked out of
- 6:51the best part of the growth curve.
- 6:53That's the reality and the numbers are staggering.
- 6:56The report points out that the total private market cap for
- 6:58these unicorns, these private billion dollar companies, has
- 7:02gone up 7X in the last decade. 7X.
- 7:04From about 500 billion to 3 1/2 trillion dollars.
- 7:07Yeah, that's a chunk of value equivalent to like 10 or 12% of
- 7:12the entire NASDAQ just sitting there, invisible to the public
- 7:15exchanges. So why?
- 7:16Why is this happening? Is it just that they don't need
- 7:19the money from public markets anymore?
- 7:21That's a huge part of it, yeah. There's just so much private
- 7:24capital available from these giant funds, funds that they
- 7:26don't need to IPO for cash. But it's also about the
- 7:29headaches of being public. The regulation, the quarterly
- 7:32earnings calls. All of it.
- 7:33The Sarbanes oscillate compliance the pressure from
- 7:36Wall Street analyst to hit a quarterly number.
- 7:39If you're Sam Altman at Open AI and you have a 10 year plan to
- 7:42build AGI, do you really want to spend your time explaining your
- 7:45quarterly GPU burn rate to some analyst who's worried about the
- 7:48next three months? No, you want to be left alone to
- 7:50burn cash and peace until you've basically taken over the world.
- 7:53Exactly. You want to execute your grand
- 7:55vision without the constant scrutiny.
- 7:58But the side effect for the rest of us is that the public market
- 8:01is slowly becoming a collection of mature, slower growth
- 8:05companies. The S&P 500 is turning into a
- 8:08retirement home for companies whose best growth days are
- 8:11behind them. That is a deeply depressing
- 8:13thought for my retirement account.
- 8:14But OK, let's connect this back to the $400 billion, because
- 8:18even if the spenders are private or huge public companies, that
- 8:21cash is real. And it brings up all the scar
- 8:24tissue from the year 2000 the.com crash.
- 8:26Great trauma. Totally.
- 8:28I remember that. The whole story then was we
- 8:30overbuilt, we dug too many trenches for fiber optic cables,
- 8:33we laid all this dark fiber demand didn't show up and the
- 8:36whole thing imploded. So why isn't this $400 billion
- 8:39AI build out just fiber bubble 2 point O?
- 8:42This is where we get to what the report calls the strong balance
- 8:45sheet thesis, and honestly, I think this is the single most
- 8:48compelling argument against the bubble narrative right now.
- 8:52It all comes down to who is spending the money and,
- 8:55critically, how they're paying for it.
- 8:57OK, so in 2000, who was it? It was the telecom companies,
- 9:01Worldcom, Global Crossing, and the key detail is that they were
- 9:06debt heavy companies. They were borrowing insane
- 9:08amounts of money to fund all that construction.
- 9:11Their balance sheets were incredibly fragile.
- 9:13They were leveraged up to their eyeballs.
- 9:14So they were basically betting with the bank's money. 100% So
- 9:17when demand for broadband lagged even by a little bit, they
- 9:20couldn't make their interest payments, they couldn't service
- 9:22the debt, they went bankrupt. And then the contagion starts.
- 9:25Exactly because they owed all that money to the banks, their
- 9:28failure created a systemic risk. It threatened to take down the
- 9:32whole financial system with them.
- 9:33OK, so that's the gambler who borrows from a loan shark to bet
- 9:36big at the casino. He loses the loan shark, breaks
- 9:39his legs, and the whole system feels the pain.
- 9:42Right. A very vivid analogy, but yeah,
- 9:43that's basically it. Now contrast that with today.
- 9:46Who's building the data centers? Who's buying all the NVIDIA
- 9:49GPU's? It's the hyperscalers.
- 9:51It's Microsoft, Google, Meta, Amazon.
- 9:55Right. And these companies are, without
- 9:56any exaggeration, the most profitable cash rich entities in
- 10:01the history of capitalism. They are not borrowing money to
- 10:03do this. This is coming out of their own
- 10:05free cash flow. They have literal mountains of
- 10:07cash just sitting there. So if Google spends $50 billion
- 10:12on chips this year and it turns out to be a total waste, what
- 10:15actually happens? They have a really bad quarterly
- 10:17earnings report. Their stock probably drops 20%.
- 10:19Shareholders get mad. Sundar Pichai has a very bad
- 10:22day, but Google does not go bankrupt.
- 10:25There's no creditor knocking on the door to repossess the
- 10:27servers. None.
- 10:28There is no systemic contagion risk.
- 10:30The whole thing is self-contained within their own
- 10:33fortress like balance sheet. So it's not the gambler
- 10:35borrowing money, it's the billionaire using his pocket
- 10:38change to build a new wing on his mansion.
- 10:41That's the perfect analogy. If the billionaire decides he
- 10:44doesn't like the new wing, or he never uses it, it's a waste.
- 10:48It's an embarrassing, expensive mistake, but he doesn't lose the
- 10:51mansion. He doesn't default on a loan.
- 10:54So the risk of a systemic meltdown, a 2008 style crisis,
- 10:57it just isn't there in the same way.
- 10:58Right. What we have is a valuation
- 11:00risk, not a solvency risk. The stocks could go down, but
- 11:03the companies won't go under. But the risk of waste is still
- 11:05there, isn't it? I mean the overbuild fear.
- 11:08Just because Meta can afford to light $50 billion on fire
- 11:11doesn't mean it's a good idea. Oh, absolutely not.
- 11:13And the report is very upfront about this.
- 11:16They believe we are almost certainly in an overbuild phase
- 11:18right now. They are building more capacity
- 11:20than is being used today, but this is where their field of
- 11:23dream strategy comes into play. If you build it, they will come
- 11:26sort. Of but it's a more defensive
- 11:28version of that. It's more like if I don't build
- 11:30it, I will die. For a company like Google or
- 11:33Microsoft. The risk of overspending on AI
- 11:36infrastructure is a financial risk.
- 11:38It hurts your earnings. But the risk of underspending?
- 11:41The risk of missing the AI wave entirely.
- 11:44That is an existential risk. If you miss it, you're not just
- 11:47less profitable, you're irrelevant.
- 11:49You're Yahoo. That is their nightmare, Yeah.
- 11:51They're terrified of being the Yahoo of the AI era, and that
- 11:55fear is a very powerful motivator.
- 11:57They see this spending not as an offensive play for profit, but
- 12:00as a defensive insurance policy against extinction.
- 12:03They'd rather waste $10 billion on unused chips and wake up in
- 12:06five years and realize the world has passed them by.
- 12:08That's the calculation they're making.
- 12:10It's survival maximization, not profit maximization.
- 12:13OK, that makes a ton of sense. On the supply side, the buyers
- 12:16are rich and they're scared. But what about the other side of
- 12:18the equation? Demand.
- 12:20Because you can have the world's most robust cash funded
- 12:22infrastructure, but if nobody actually uses the AI, it's still
- 12:26a dud, right? And the report had some just
- 12:28insane numbers on adoption velocity here.
- 12:31This is where any comparison to previous tech cycles like the
- 12:34Internet just completely falls apart.
- 12:36The speed is on a different planet.
- 12:38The report sites this one stat that's just mind blowing.
- 12:40It took ChatGPT just two years to hit 365 billion searches.
- 12:46Two years. OK, that sounds fast, but I need
- 12:49a comparison. Well, any guess how long it took
- 12:52Google search to reach that same milestone?
- 12:54That same volume I mean. Google was the fastest growing
- 12:57thing ever back then. I don't know, maybe five years,
- 12:596. 11 years. Wow, 11 versus 2.
- 13:04More than five times faster. And the reason for that isn't
- 13:06just that it's a great product, it's this concept the report
- 13:08calls infrastructure inheritance.
- 13:10I love that phrase infrastructure inheritance.
- 13:13It sounds like they were born rich.
- 13:14What does it mean? Think about the friction
- 13:16involved in getting online in 1999.
- 13:18To use Google, you had to go to a store and buy a computer.
- 13:21You had to buy a modem. You had to sign up for an ISP
- 13:23like AOL. I remember the dial up screech
- 13:26the you've got mail CD's they sent you every week.
- 13:28Exactly. There was immense physical and
- 13:30financial friction. Same for the iPhone.
- 13:33Apple had to convince millions of people to go out and spend
- 13:35$600.00 on a totally new kind of device.
- 13:38Right, big hurdle. But AI, it just showed up on the
- 13:40rails that were already built. There are 5 billion people who
- 13:43already have a supercomputer in their pocket.
- 13:45The cloud infrastructure was already in place.
- 13:47AI didn't have to sell a single piece of hardware.
- 13:50All I had to do was ship a URL the.
- 13:51Distribution was instant and global from day one.
- 13:56It's like building a house from scratch versus moving into a
- 13:58fully furnished luxury penthouse.
- 14:00You just show up and turn the lights on.
- 14:01And that lack of friction, that speed, leads to this idea of
- 14:05consumer surplus that they talk about.
- 14:07The argument being that we're all getting a ridiculously good
- 14:10deal on AI right now. A massively good deal.
- 14:12We're all underpaying. I mean I pay 20 bucks a month
- 14:15for ChatGPT pro, sometimes I wonder if it's worth it.
- 14:18But think about the value exchange.
- 14:20For that $20, you're getting access to an intelligence that
- 14:22can write code, analyze data, draft legal documents,
- 14:27brainstorm ideas, things that would previously required hiring
- 14:30a human assistant for, what, $60,000 a year?
- 14:33That huge gap between the value you get in the price you pay,
- 14:37That's the consumer surplus. It's like that iPhone example
- 14:40from the report. Yes, perfect example, you pay
- 14:43what, $1000 for an iPhone? But if someone you know put a
- 14:46gun to your head and said, what's the absolute maximum
- 14:49you'd pay to keep this device that runs your entire social
- 14:51life? Your banking, your navigation,
- 14:53your memories. Oh God, I mean easily 5 grand,
- 14:55maybe more. Don't tell Tim Cook I said that.
- 14:59He knows, but that's the point. Apple captures $1000 of value,
- 15:03but you capture the other 4000 in surplus.
- 15:06And the argument here is that the surplus with AI is even
- 15:09bigger right now. The pricing is just totally out
- 15:12of whack with utility. I'm thinking of that anecdote
- 15:14about buying a baseball bat, one of the A 16's partners, David
- 15:18George, told it. Yes, that one really landed with
- 15:20me. He talks about trying to buy a
- 15:22baseball bat for his son. In the old world, the Google
- 15:25world, that's a 30 minute chore, right?
- 15:27Oh, it's the worst. Best youth baseball bat.
- 15:30You click through 7 sponsored links.
- 15:32You try to read some blog post. It's just SEO garbage written by
- 15:36a bot to get affiliate clicks. You've got three tabs open
- 15:39trying to compare prices. It's madding.
- 15:41I literally went through this last week trying to buy a new
- 15:43coffee maker. 45 minutes of my life I'll never get back.
- 15:46Just trying to figure out which top ten list wasn't a complete
- 15:48scam. I was ready to just give up and
- 15:50drink tea. Exactly that search fatigue is a
- 15:52real tax on our time. But in the AI world, he just
- 15:56asked the model. My son is 10.
- 15:58He's a power hitter. My budget is under $100.
- 16:01What's the best bat? And in 30 seconds, it
- 16:03synthesizes all the reviews, the specs, the prices, and gives him
- 16:07the answer that 29 minutes of his life he got back.
- 16:11That's pure value. So the idea is that we're in
- 16:13this kind of golden age of cheap AI before the companies figure
- 16:17out how to properly price it and capture more of that surplus.
- 16:20It could be the report hints at what's coming next.
- 16:22Maybe it's not a flat subscription, maybe it's usage
- 16:25based pricing for power users, or maybe it's a whole new kind
- 16:29of affiliate model where the AI doesn't just recommend the BAT,
- 16:33it buys it for you and takes the cut of the transaction.
- 16:36The models are still evolving, but the core point is the demand
- 16:39is real because the utility is off the chart.
- 16:40OK. So the demand is there, but this
- 16:43brings up the really big question that the skeptics
- 16:45always come back to How big can this market actually get?
- 16:48They say, sure, Salesforce is big, but is AI really going to
- 16:52be 100 times bigger than all of existing software combined?
- 16:55Is there enough money in the world's IT budgets to justify a
- 16:58$400 billion annual spend? And this, for me, was the most
- 17:02profound insight in the entire deck of documents.
- 17:05It's a total perspective shift. If you take away only one thing
- 17:09from this entire deep dive today should be this.
- 17:12Stop looking at the software market, OK, Look at the labor
- 17:14market. The labor market?
- 17:16Unpack that for me. That sounds like something from
- 17:17an economics textbook. What does it mean for this AI
- 17:20build out? Look at the split of the US
- 17:22economy. US GDP.
- 17:24If you add up all the money companies spend on software,
- 17:26it's about 1% of GDP. It's a huge number in absolute
- 17:30terms, hundreds of billions, but it's only 1% of the.
- 17:33Pie, right? It's a budget line item, the CI
- 17:34OS budget. Now look at a different number.
- 17:36The total US white collar payroll, The money we spend
- 17:39paying people to sit at desks and think and write and code and
- 17:43analyze spreadsheets and answer emails.
- 17:45That is roughly 20% of GDP. Whoa.
- 17:48OK, 20% versus 1%, the labor pie is 20 times bigger than the
- 17:52software pie? Exactly.
- 17:53And the entire thesis of this AI investment wave is that AI is
- 17:57not just better software. It is not just a tool to make a
- 18:00human 10% more efficient. In many cases, it is service as
- 18:03software. It is a replacement for labor
- 18:05itself. So AI isn't attacking that 1%
- 18:07slice of the economy, it is going after the 20% slice.
- 18:10That completely changes the total addressable market, the
- 18:12Tam. We're not talking about the tool
- 18:14budget anymore, we're talking about the salary budget.
- 18:16The entire salary budget, and this is amplified by the insane
- 18:20economics of intelligence itself.
- 18:23The report highlights this deflationary force that is just
- 18:27it's hard to believe. They point out that the input
- 18:29cost, the cost to generate intelligence measured in tokens,
- 18:32has dropped by more than 99% in just two years.
- 18:35Hang on, let's define tokens for anyone who isn't living and
- 18:38breathing this stuff. It's not a crypto thing.
- 18:40No, no, definitely not. A token is basically a piece of
- 18:42a word. Think of it as about 3/4 of a
- 18:44word. On average, when you pay to use
- 18:46an API from Open AI, you're paying per million tokens
- 18:50processed. It's the base unit of AI work.
- 18:53Got it. So the price of that unit of
- 18:54work has fallen by 99%. In about 20-4 months, yeah.
- 18:57At the same time, the capability of the models, how good they
- 19:00are, is doubling roughly every seven months.
- 19:02So you have this insane dynamic where the cost is racing towards
- 19:050 while the quality is racing towards Infinity.
- 19:07It's like Moore's Law on hyperdrive.
- 19:09So what's the end game? What happens when the cost of
- 19:11intelligence actually hits 0? It just becomes a utility like
- 19:14electricity or Wi-Fi. You don't think about the cost
- 19:17of an electron when you flip on a light switch, you just expect
- 19:19the light to be there. And the a 16's view is that
- 19:22intelligence will become the same.
- 19:23It'll just be an ambient utility that's always on.
- 19:25But wait, if it gets that cheap, how does anyone make money?
- 19:30This gets to that whole debate around gross margins, right?
- 19:32I hear investors complain all the time that AI apps are bad
- 19:35businesses because they have to pay Open AI or NVIDIA for every
- 19:39query. That's the classic sauce
- 19:41investor critique. Yeah, a traditional software as
- 19:43a service company like Adobe has beautiful 80 plus percent gross
- 19:48margins. Once the software is built, it
- 19:50cost them almost nothing to serve another customer.
- 19:52Right. But an AI company has a high
- 19:54cost of goods sold or COD GS because every single answer a
- 19:58user gets requires an expensive GPU to spin up and burn
- 20:01electricity. So their margins might be more
- 20:03like 50 or 60%. And a traditional investor looks
- 20:05at that and says, yuck, I want my 80% margins.
- 20:08But the report makes a very sharp counter argument.
- 20:11They say that's looking at it through the old lens first.
- 20:14As we just said, those input costs are collapsing.
- 20:16So the margins are improving naturally.
- 20:18But the more important point is that if you're selling labor
- 20:21replacement, the margin math is totally different.
- 20:23How so? If I sell you a software tool
- 20:25for $100 a month, that's coming out of your tool budget.
- 20:29But if I sell you an AI agent that completely replaces a
- 20:32junior engineer who is costing you $100,000 a year and I charge
- 20:36you $50,000 a year for. It I just saved $50,000.
- 20:39I don't care what your margin is.
- 20:40I could care less if your code GS is 10% or 90%.
- 20:44I just cut my biggest line item in half.
- 20:45Exactly. The customer isn't buying a
- 20:47software license, they're buying an outcome.
- 20:49They're buying labor cost reduction and the total dollar
- 20:52value is so much larger that the margin percentage becomes way
- 20:55less important than the value delivered.
- 20:56That's a huge mental shift, but it creates another problem.
- 20:59If everyone can tap into these super powerful, super cheap
- 21:02models from Open AI or Anthropic, where's the mote?
- 21:06What stops me from just building a slightly different version of
- 21:09your app tomorrow using the same underlying AI?
- 21:12This is the question that is killing a lot of AI startups
- 21:14right now and the A 16's documents are crystal clear on
- 21:18what is not a defensible mode and that is wrapper apps.
- 21:22Wrappers. So basically apps that are just
- 21:24a pretty user interface on top of a call to the GPT 4 API.
- 21:28Exactly. I built an app that helps you
- 21:29write marketing emails. Well, guess what, ChatGPT can do
- 21:32that natively now, or a competitor can pop up tomorrow
- 21:35using a different model. There's no customer lock in.
- 21:38There's no stickiness. So what is sticky?
- 21:40What's the real Moat in this new world?
- 21:42It's deep, painful, messy workflow integration.
- 21:45The report uses the example of medical scribing.
- 21:48There's a company called a bridge.
- 21:49I've heard of them, So what are they doing that's so different
- 21:51from just transcribing a conversation?
- 21:53Think about the day-to-day reality of a doctor.
- 21:56They spend hours every night on paperwork.
- 21:58It's a #1 cause of physician burnout.
- 22:00It's awful. A Bridges AI sits in the room,
- 22:03listens to the doctor patient conversation, and automatically
- 22:06writes the clinical note. OK, that's cool.
- 22:08But that's just the start. The real magic, the Moat, is
- 22:11that it's deeply integrated into the hospital's electronic
- 22:14medical record system, the EMR. It knows which fields to
- 22:17populate in Epic or Cerner. It suggests the right billing
- 22:21codes. It understands the compliance
- 22:23rules. Once a doctor gets that two
- 22:25hours of their life back every single day, are they ever going
- 22:28to switch to a different tool? Never.
- 22:29Not a chance. The switching cost is
- 22:31astronomical. Not in money, but in pain.
- 22:33Exactly that is the Moat. It's not the AI model itself,
- 22:36it's the integration into the complex, regulated real world
- 22:40workflow. The report calls it a shift from
- 22:43AUI of forms and databases, which is what Salesforce has
- 22:46been for 20 years, to 1 of proactive agency.
- 22:48Proactive agency. I like that.
- 22:50It means the software stops waiting for you to input data.
- 22:53It starts telling you what to do next, or better yet, it just
- 22:55does the task for you. That's the new frontier of user
- 22:58experience. OK, this is a pretty compelling
- 23:00picture. We have the money from cash rich
- 23:02giants. We have massive demand driven by
- 23:06labor replacement. We have a strategy for building
- 23:08moats through workflow. It feels like this thing is
- 23:11unstoppable. But the report does identify a
- 23:15major bottleneck coming. And kind of shockingly, they say
- 23:18it's not chips anymore. No, it's not.
- 23:20The chip shortage narrative is kind of last year's story.
- 23:22NVIDIA is ramping production like crazy.
- 23:24The supply chain is catching up. The new wall we're about to hit
- 23:28is a wall of pure physics, energy and heat.
- 23:31This is where it gets a little bit sci-fi.
- 23:32We're literally running out of power to run the AI.
- 23:35In the specific places you need it, yes.
- 23:37You have to appreciate the scale.
- 23:38These new AI data centers are the most power hungry structures
- 23:42humanity has ever built and they need a specific kind of power.
- 23:45Massive, stable, 204 seven base load power.
- 23:48You can't train a foundational model on intermittent wind or
- 23:50solar. You need the lights to stay on
- 23:52always. And This is why we're suddenly
- 23:54hearing about a nuclear renaissance from all the big
- 23:58tech companies. It's the only logical
- 23:59conclusion. The report explicitly mentions
- 24:02Microsoft getting involved in restarting the Three Mile Island
- 24:06nuclear power plant. Which is just wild to think
- 24:08about about. 3 Mile Island is like the Chernobyl of America.
- 24:11It's the most infamous nuclear site in the country, and now
- 24:15tech companies were bringing it back to life to power ChatGPT.
- 24:18They have no choice. They're pursuing what are called
- 24:21behind the meter deals. What does that mean behind the
- 24:23meter? It means you don't even use the
- 24:25public electrical grid. You literally build your data
- 24:28center on the same piece of land as the power plant.
- 24:32You run a giant private cable directly from the reactor to
- 24:35your servers. You bypass the congested public
- 24:38grid entirely. You get the raw, undiluted power
- 24:41fresh from the source. It guarantees your supply.
- 24:43That is intense. It's like something a James Bond
- 24:46villain would do. Build their lair next to a
- 24:48volcano. It's the only way they can get
- 24:49the gigawatts of power they're forecasting they'll need of
- 24:52getting the power in. It's about getting the heat out,
- 24:54the cooling problem, I don't think.
- 24:56People really get how big a deal this is.
- 24:57Not at all. Let's put it in numbers.
- 24:59A standard server rack in a normal cloud data center draws
- 25:03maybe 10 kilowatts of power. OK, 10 kilowatts A.
- 25:06Single rack of. New NVIDIA GPU's for AI training
- 25:10that draws 100 to 120 kilowatt, 10 times the power.
- 25:13Density 10 times you're. Concentrating an incredible
- 25:16amount of energy into a very small space, and all that energy
- 25:19becomes heat. When you put 100,000 of these
- 25:22GPUs together in a single building, you've essentially
- 25:25created a blast furnace. And if you can't dissipate that
- 25:28heat fast enough, the chips literally melt.
- 25:31Or they throttle themselves down to prevent melting and your
- 25:34multibillion dollar investment stops working.
- 25:36So just blowing cold air on them.
- 25:37Isn't going to cut it anymore, Not even close.
- 25:40The report says the next major frontier of innovation is in
- 25:43liquid cooling. Literally running pipes of fluid
- 25:46directly over the chips to carry the heat away, and just general
- 25:49thermal management there. Was that amazing?
- 25:51Story in the transcripts about X AI, Elon Musk's company building
- 25:56their Colossus supercomputer in Memphis.
- 25:58Yes, that was used as a case. Study and just raw brute force
- 26:02execution. They managed to stand up one of
- 26:04the world's largest AI training clusters in a fraction of the
- 26:08normal time. Usually takes years.
- 26:09They did it in a few months. How?
- 26:11What was their secret? They did what the report.
- 26:14Calls unnatural things. They hit the power wall.
- 26:17The local utility company in Memphis physically couldn't
- 26:20build out the substation fast enough to power the cluster, so
- 26:23they just had to wait. No.
- 26:24They sent. Teams out across multiple States
- 26:27and bought every single industrial backup generator they
- 26:29could find. Huge portable gas turbines.
- 26:33They line them up in a field next to the data center and
- 26:35created their own private temporary power plant to bridge
- 26:39the gap until the grid could catch up.
- 26:40They just brute forced. Physics they did.
- 26:43And it's a perfect illustration of the new reality.
- 26:46The primary bottleneck is no longer money or even silicon.
- 26:50It's the physical constraints of infrastructure.
- 26:52It's pouring concrete and stringing copper wire.
- 26:56And the teams that can bend that physical reality are the ones
- 26:58who win, which leads perfectly into the.
- 27:00Final section of their analysis, which they call the Tale of two
- 27:03markets, because it really does sound like there's one group of
- 27:06winners who can afford to buy their own nuclear plants and
- 27:08private power grids, and then there's everybody else.
- 27:11It's a bimodal. Distribution a barbell market is
- 27:14another way to put it. On one side of the Garbell you
- 27:16have this tiny handful of super winners open AIXAI anthropic
- 27:21data bricks. These elite teams are just
- 27:24they're on another planet in terms of growth and execution.
- 27:27The report had this great. Term for the founders, hard
- 27:29capitalist scientists. I love that phrase, it's
- 27:32perfect. Because that's what they are,
- 27:33These people. Like the founders of Open AI,
- 27:35they started as pure researchers, academics.
- 27:38Their goal was this lofty scientific pursuit of artificial
- 27:41general intelligence. But then the competitive reality
- 27:44kicked in. They realized that to build the
- 27:46thing they dreamed of, they needed access to billions and
- 27:49then 10s of billions of dollars of capital and compute.
- 27:52So they had to morph. They became these ruthless world
- 27:55class operators and fundraisers. They are scientists who were
- 27:58forced to learn the game of capitalism and ended up being
- 28:01better at it than most capitalists.
- 28:03And on the. Other end of that barbell.
- 28:06The other market? The struggling incumbents.
- 28:08The report has a sobering statistic. 95% of existing
- 28:11public software and Internet companies are now forecasting
- 28:14less than 25% annual growth. Less than 25% for a.
- 28:17Tech company that's basically going backwards, that's a death
- 28:20sentence. It's stagnation.
- 28:22The traditional software as a service model, the whole pay us
- 28:24$50 per user per month thing is coming under immense pressure as
- 28:29AI starts automating jobs. Companies need fewer seats, so
- 28:32the business model that created all the wealth and software over
- 28:35the last 15 years is starting to stall out.
- 28:37So you have these AI natives. On a rocket ship and the legacy
- 28:40software guys hitting a wall. If you're an investor or just
- 28:43anyone trying to make sense of this, what's the take away?
- 28:46The A16Z. Strategy is what they call the
- 28:48ALSA strategy. It's basically a belief that in
- 28:51this new world, you have to make highly concentrated asymmetric
- 28:55bets on the top five, maybe top ten teams in the entire world.
- 28:59Just the top five that feels. Incredibly risky, very
- 29:02concentrated it is, but. Their logic is that the talent
- 29:05density at a place like Open AI or Anthropic is so extreme and
- 29:09so scarce that it's actually safer to pay a Nosebleed
- 29:12valuation for a piece of that team than it is to bet on the
- 29:1550th best AI company. Because that 50th best company
- 29:18is just a. Rapper app that's going to be
- 29:19obsolete in six months exactly, they argue.
- 29:22That the talent itself is the downside protection.
- 29:24Even if the company's first product fails, that collection
- 29:28of human capital is so valuable that it will find a way to win
- 29:30eventually. It's a consolidation of power
- 29:32and talent into a very, very small number of elite
- 29:35organizations. Wow it.
- 29:37Is a fascinating and honestly a kind of terrifying vision of the
- 29:40future. We've gone through a ton of
- 29:42ground here. Let's try and just crystallize
- 29:43the big takeaways before we wrap up.
- 29:45Let's do it. 4 main points. One, this is not a debt fueled
- 29:48bubblelikethe.com crash. It's a cash fueled
- 29:52infrastructure build out by companies with fortress balance
- 29:54sheets. The risk is valuation, not
- 29:57solvency. OK, 22.
- 29:59The. Product here is intelligence
- 30:01itself. The cost of that intelligence is
- 30:03collapsing towards 0. And the target market isn't the
- 30:051% of GDP that is IT budgets, it's the 20% of GDP that is
- 30:09white collar labor. The Tam is enormous.
- 30:12Got it. 33. The real bottlenecks are no
- 30:15longer digital, they're physical.
- 30:16We're hitting the limits of energy and thermal dynamics.
- 30:18The cloud has a very heavy, very real footprint on the planet.
- 30:21And four. And four.
- 30:23The market is splitting in two, a bimodal distribution of a few
- 30:26hyperscaling winners and a vast sea of stagnant incumbents.
- 30:30The alpha, the outsize returns will likely come from betting on
- 30:32that tiny handful of league teams.
- 30:34It really reframes everything. When I see a headline now that
- 30:37says Microsoft spends $10 million on data centers, my
- 30:41first thought used to be bubble. Now my first thought is
- 30:43insurance policy. Exactly.
- 30:45And when you? Hear nuclear power.
- 30:46You don't think disaster moving, you think compute.
- 30:48It totally changes the lens. But before we sign off, I want
- 30:51to end with one final thought for our listener.
- 30:53We've talked so much about the company's the capital, the
- 30:56technology, but let's bring it back to the human level, always
- 30:59the most important. Part we call you the.
- 31:01Listener, the learner, that's who this is for.
- 31:04But if this technology AI is the ultimate learner, if it can read
- 31:09every book ever written in a second and write perfect code
- 31:12and analyze any data set instantly, what happens to the
- 31:15human learner? What happens to us is the
- 31:17question that. Underlies all of this if this
- 31:19whole thesis is. Correct.
- 31:21And AI captures the value of Labor the way Google captured
- 31:24the value of information. What's our role?
- 31:27Are we all training to become the managers of AI systems, or
- 31:31are we just competing for the handful of jobs left that still
- 31:34require a physical human body? The way I think about it is
- 31:37this. If intelligence becomes as cheap
- 31:40and abundant as electricity, then all the value shifts.
- 31:44It's shifts to where you decide to plug it in.
- 31:46Electricity itself is a cheap commodity, but a hospital
- 31:50operating room lit up by that electricity is incredibly
- 31:54valuable. A Broadway stage is valuable.
- 31:56The electricity just enables the value.
- 31:58It isn't the value itself. So the skill is no longer.
- 32:01Being the light bulb, the skill is being the person who decides
- 32:04where to point the beam. That's it exactly.
- 32:06The question for all of us is no longer how smart are you?
- 32:09The question is, where are you plugging all this new
- 32:11intelligence in? What unique problem are you
- 32:14solving? Because that's the only thing
- 32:16that's going to matter. Something to think about if you.
- 32:18Want to keep plugging into these kinds of deep dives.
- 32:20All the links are in the description.
- 32:21Thanks for joining us. My pleasure.