Latest / Deep Values: Value Investing Deep Dives / The Valuation Inversion: Andrew Ng's AI Investment Stack Analysis
Transcript
- 0:00Welcome back to the Deep Dive, where we plunge into the most
- 0:02consequential capital intensive trends shaping our future and
- 0:08give you the essential knowledge you need without all the hype.
- 0:11Today we are looking at something.
- 0:14Well, it's staggering. It's the sheer unprecedented
- 0:18flood of money into artificial intelligence.
- 0:21I mean, we're talking about global capital flows that really
- 0:25rivaled the late 90s, but they're moving so much faster.
- 0:28And it's all driven by what feels like, you know, genuine,
- 0:31verifiable breakthroughs. It truly is a moment of just
- 0:35dizzying scale when you see proposals for over a trillion
- 0:38dollars in infrastructure spending.
- 0:40And that's just for the foundational elements.
- 0:42And you watch single companies, you know, briefly hitting market
- 0:45caps in the multiple trillions, the classic unavoidable car and
- 0:49just pops up. Are we right now in the middle
- 0:51of an AI bubble? Exactly.
- 0:53But that binary question, you know, is it a boom or is it a
- 0:55bust? It's often all you hear.
- 0:57And it's it's just an overly simplistic framing.
- 0:59It really obscures the true nature of the risk and, frankly,
- 1:03the opportunity. It does.
- 1:04So our mission for this deep dive is to move entirely beyond
- 1:07that. We're going to be using the
- 1:10remarkably rigorous framework from one of the true fathers of
- 1:14modern deep learning, Doctor Andrew Eng.
- 1:16We're diving into his Seminole thesis, which he really
- 1:18articulated in the batch issue 329, and then refined it
- 1:22throughout his analysis in 2024 and 2025.
- 1:25And ENG's work is really the key here.
- 1:27He argues that the AI market, it can't be treated as a single
- 1:32thing. It's not a monolith.
- 1:34Not one giant speculative asset. No, instead it's a sophisticated
- 1:38composite structure. It's made-up of three distinct
- 1:42economic layers. And each layer has its own
- 1:44rules. Its own economic physics, as he
- 1:46puts it. It has its own unique risk
- 1:48profile and it's attracting capital in wildly different ways
- 1:51and that. Structure is what we're calling
- 1:53the tripartite stack. The conclusion Eng reaches,
- 1:56which is honestly it's shocking, is that we're currently
- 1:59experiencing A structural valuation inversion.
- 2:02Which sounds complicated, but the idea is simple.
- 2:06It means the foundational infrastructure, the stuff you
- 2:08need to create the AI, is possibly overheating.
- 2:11It's being overcapitalized. While the layer you need to
- 2:14actually extract value and generate revenue from it all.
- 2:17The application layer. That's being structurally
- 2:20undercapitalized. Yeah, Starved even.
- 2:22That inversion is the very definition of market
- 2:25instability. Right it is.
- 2:26It's the source of all the current anxiety, but it's also
- 2:30the blueprint for the massive value creation that could still
- 2:33be to come. So the question isn't whether
- 2:35the tech is real. No, the tech is real.
- 2:38The question is whether the capital has been routed
- 2:40correctly to actually realize all that potential.
- 2:43OK, so before we jump into the structure of this inversion,
- 2:45just a crucial word on context. As always, the analysis we're
- 2:49providing is based purely on publicly available sources.
- 2:52It is strictly for educational purposes.
- 2:55We're synthesizing information to help you get better informed.
- 2:57Exactly. This does not, and it must not
- 3:01constitute investment advice. We really urge you to conduct
- 3:05your own thorough due diligence before making any investment
- 3:08decisions whatsoever. OK, let's get into it.
- 3:10Let's establish the ground level, Let's start with the
- 3:13scale of the investment because you really need to feel the
- 3:15weight of this capital to understand Ang's analysis,
- 3:18right? Analysts are tracking over $1.4
- 3:21trillion that's specifically proposed for AI infrastructure
- 3:25projects. Global, and this isn't just
- 3:27theoretical money, you know, like a projection, this is
- 3:30actual CapEx planning. If this gets executed, it
- 3:34fundamentally changes global economic dynamics.
- 3:37And that scale translates directly into market cap figures
- 3:40that are, like you said, just hard to process.
- 3:42We've seen periods where single entities, you know, the dominant
- 3:45ones and heart are like NVIDIA. Briefly held market caps that
- 3:48exceeded the combined GDP of major established nations.
- 3:52It's unbelievable. It reflects this fundamental
- 3:54rapid reallocation of capital. It's moving away from
- 3:57traditional industries, oil, real estate, manufacturing.
- 4:01And going straight into specialized data centers and
- 4:03silicon fabrication, but. If everyone is pouring money
- 4:06into this one sector and we have multiple companies valued this
- 4:10highly, well, this is exactly what raises that prevailing
- 4:13anxiety skeptics out there. Like folks at Janus Henderson
- 4:16and other big institutional funds.
- 4:18Right, they point to the revenue gap.
- 4:20And the revenue gap is the core structural fear.
- 4:23It's simple really. It highlights the mismatch
- 4:26between the inputs and the outputs.
- 4:29The money in versus the money out.
- 4:30Precisely the inputs are the colossal amount of money, the
- 4:34CapEx being spent right now on chips, on power infrastructure,
- 4:38on cooling systems. And the output is the money
- 4:42that's currently being earned from AI software and services.
- 4:45Yeah, the actual realized revenue.
- 4:47So if you commit 1.4 trillion to building these AI factories, but
- 4:52the software running inside those factories only generates,
- 4:54let's say 150 billion in annual net profit.
- 4:58That equation doesn't work. It's economically unstable.
- 5:01The money spent vastly, vastly exceeds the money being earned
- 5:04back to justify the expense. And the risk, they argue, is
- 5:07simple if the demand for that AI software, the application
- 5:11revenue, if it fails to accelerate fast enough to close
- 5:14that gap. A rapid corruption is
- 5:16inevitable. Which leads to a deflationary
- 5:18shock across the whole tech ecosystem.
- 5:21And that's where ANG steps in and says wait a minute, you are
- 5:24asking the wrong question. He thinks the critics are
- 5:27looking at it the wrong way. Yeah, they're treating the AI
- 5:30market as a single speculative asset, a monolith.
- 5:35Eng fundamentally rejects this framework.
- 5:38He argues that viewing hardware AP is in the end users solutions
- 5:43as just one big bucket of risk. It blinds us.
- 5:47It blinds us to the actual structural drivers of value and
- 5:51volatility. OK, that's a crucial point.
- 5:53But if the entire value chain ultimately depends on the
- 5:55underlying silicon data centers, why isn't the risk profile
- 5:59shared? I mean, if the foundation model
- 6:01collapses, doesn't the application layer collapse with
- 6:04it not? Necessarily.
- 6:05And this is where the analogy of the early 20th century
- 6:08electrification era becomes so powerful and uses it
- 6:11masterfully. Right.
- 6:12Think back to that time of immense transformational
- 6:15technology. Capital flowed rapidly, creating
- 6:18multiple distinct speculative waves.
- 6:20Exactly. You had bubbles and things.
- 6:22Like copper mining, right? The raw material infrastructure.
- 6:24Copper mining and specific local utilities building out the
- 6:27distribution networks and those investment cycles.
- 6:29They had their own busts and consolidations, but N points out
- 6:34that those infrastructure bubbles were structurally and
- 6:37financially separate from the eventual massive durable value
- 6:41created by the electrical appliances themselves.
- 6:43The. Refrigerators.
- 6:44The motors, the lighting systems.
- 6:46Exactly. The fact that some copper miners
- 6:49or specific utility companies failed did not negate the long
- 6:52term value of electricity itself and the applications it enabled.
- 6:56So applying that analogy, now the bubble risk in the copper
- 7:00mining, which is the manufacturing of stabilized
- 7:02training chips. It's fundamentally separate from
- 7:04the lasting value creation in the electrical appliance.
- 7:07Which is the high ROI AI application?
- 7:10We have to decouple them to assess risk intelligently.
- 7:13And that brings us to clearly defining the three layers of
- 7:17ENG's tripartite stack, because the rest of our discussion
- 7:20really hinges on this segmentation.
- 7:22It does. OK.
- 7:22So layer one training infrastructure.
- 7:25This is the factory environment. It represents that colossal
- 7:28burst oriented computation you need to create the intelligence
- 7:31to train the massive foundation models like Jeep, AT or Llama.
- 7:34It's the specialized high intensity environment of
- 7:38discovery. Right, then you have layer 2
- 7:41inference infrastructure. The runtime environment.
- 7:44Yes, these are the computational resources required to execute
- 7:47those trained models. For end users delivering the AI
- 7:50services in real time, day in and day out, it's the utility
- 7:55company delivering the electricity.
- 7:56And finally layer 3, the application layer.
- 7:59This is where the rubber meets the road.
- 8:01It's the user interface, the final product, the specialized
- 8:05service that solves a real world high value problem.
- 8:08The economic utility, the appliance itself, The medical
- 8:12diagnosis agent, The automated legal review system.
- 8:15Right. And we have to remember this
- 8:16structure, Eng asserts that capital is currently pooling
- 8:19aggressively and potentially dangerously in layer one.
- 8:23Well, layer. Three, the only layer that can
- 8:25provide the cash flow to justify layers 1 and 2 is being starved.
- 8:28That is the valuation inversion we need to understand.
- 8:31OK, so let's focus on the bottom of that stack.
- 8:32Training infrastructure. This is where Nang explicitly
- 8:36suggests that while he's, you know, still optimistic about the
- 8:38overall AI sector. He's cautious.
- 8:41He's very cautious. He says.
- 8:42This particular layer is facing intense pressure and could also
- 8:47be a bubble. It represents the highest
- 8:49financial risk. And the risk is just baked into
- 8:51the nature of the investment required.
- 8:53I mean, training state-of-the-art models involves
- 8:55assembling hundreds of thousands of the highest performance
- 8:58GPU's. We're talking clusters that
- 8:59require immense power draw specialized cooling.
- 9:03All into singular bespoke computational centers.
- 9:07It's just an inherently capital intensive process, a massive
- 9:11nondilutable CapEx commitment. And the primary structural
- 9:15mechanism that really amplifies this bubble risk is what Ang
- 9:18highlights as the depreciation trap.
- 9:20Which is made even worse by the physics of Moores law, right?
- 9:23Traditional assets like a steel mill or commercial real estate,
- 9:27they might hold their value for decades.
- 9:30But AI training hardware? It's operating on this
- 9:32accelerated timeline where technological obsolescence can
- 9:35happen in just months. Just consider the financial
- 9:38absurdity of the speed. Let's say you commit a billion
- 9:40dollars to procuring and installing a cutting edge GPU
- 9:43cluster in 2023. That investment is meant to have
- 9:46an economic lifespan that justifies that cost.
- 9:49However, because of architectural breakthroughs, the
- 9:52move from say Hopper to Blackwell or whatever comes
- 9:56next. That same computational power or
- 9:58better can often be acquired for half the price just two or three
- 10:02years later. Which means the original billion
- 10:05dollar investment suffers instant catastrophic
- 10:08technological depreciation. It's a race against time where
- 10:12the asset is literally shrinking in value faster than the
- 10:15training process can even be completed.
- 10:17And calls them fast depreciating assets if your billion dollar
- 10:21cluster is training a model that takes 18 months to complete and
- 10:25the technological benchmark shifts dramatically in month 15.
- 10:28The massive premium you paid for that 2023 performance is just
- 10:33it's incinerated capital. The hardware might be repurposed
- 10:35for inference later, sure, but the premium paid for that
- 10:38state-of-the-art training capability is lost forever.
- 10:41And that investment carries, as he says, minimal cash flow to
- 10:44show for it during that entire training phase.
- 10:46And compounding all of this is the competitive saturation, what
- 10:49he calls the foundation model race redundancy.
- 10:52Right now you have dozens of entities, big tech, well funded
- 10:57startups, sovereign funds, all building overlapping proprietary
- 11:02infrastructure. They're all chasing the same
- 11:04finite pool of demand for these general purpose LLM.
- 11:07'S exactly. But if the utility of AI is as
- 11:10vast as we believe it is, shouldn't the market support
- 11:13multiple foundation models? It should support some, yes, but
- 11:17probably not dozens of them, all commanding multibillion dollar
- 11:20premiums. That's a good point.
- 11:22And Eng points out the competitive threat from open
- 11:25source intelligence, particularly models like Meta's
- 11:28Llama series, which are, you know, rapidly closing the
- 11:31performance gap. So if open source models become
- 11:33good enough for say 80% of enterprise needs.
- 11:38Then the marginal cost of intelligence trends towards 0
- 11:41and when that happens, the pricing power of the proprietary
- 11:44models just collapses. And all that massive CapEx, it
- 11:47was spent building rival proprietary infrastructure.
- 11:50It can never achieve the predicted returns.
- 11:51It becomes a classic over investment cycle.
- 11:54And here's where the financial engineering comes in, which
- 11:57makes the bubble even harder to spot the circular risks that
- 12:01create synthetic demand. Oh, this is straight out of the
- 12:041990s dot com and telecom playbook.
- 12:06It absolutely is. I mean, think about the
- 12:08mechanics. A major cloud provider, a
- 12:11hyperscaler, needs to demonstrate revenue growth to
- 12:14justify its own enormous CapEx, right?
- 12:18So they invest billions into promising AI startups, but often
- 12:22it's not in liquid cash. They offer cloud credits.
- 12:26So Startup B gets $5 billion in funding, but crucially, that
- 12:30money is contractually obligated to be spent entirely on cloud
- 12:35provider A's compute services for training.
- 12:37Exactly. So cloud provider a the
- 12:38infrastructure owner books that subsequent cloud usage as
- 12:41revenue. Which looks great on their
- 12:43quarterly report. Of course, But it's not organic
- 12:45demand, it's subsidized demand. It's synthetic demand generated
- 12:49through vendor financing. So if the model builder startup
- 12:52B fails to secure real non subsidized revenue from the
- 12:56application layer, if their business model just collapses.
- 12:59That 5 billion in cloud revenue that cloud provider A booked
- 13:03just evaporates. And the expensive training
- 13:06clusters they built become stranded assets.
- 13:08This whole mechanism means that the current apparent demand for
- 13:12layer one infrastructure is potentially inflated, masking
- 13:16this underlying financial fragility.
- 13:18It's a powerful, self fulfilling cycle of investment that only
- 13:22works if that top layer, layer three, delivers real cash flow.
- 13:26Which makes the training layer the most unstable part of the
- 13:29entire stack. OK.
- 13:30Let's transition now to Layer 2 inference infrastructure if
- 13:34Layer 1 is that high risk volatile environment of
- 13:37discovery and incubation. Layer 2 is supposed to be the
- 13:40stable, predictable environment of execution.
- 13:43And we need to clearly separate the two because Ng claims that
- 13:46even amid this capital flood, inference infrastructure still
- 13:50needs significant investment. And that is the counterintuitive
- 13:53observation, isn't it? How can the market be saturated
- 13:55with chips yet still require more investment?
- 13:57Right? The answer lies in the technical
- 13:59specification and just the nature of the workload.
- 14:02Training is about solving 1 problem quickly.
- 14:05Inference is about solving billions of small problems
- 14:08efficiently. So training requires these
- 14:10massive synchronized GPU clusters with extremely high
- 14:14interconnect bandwidth. Like Envy Link, they need to
- 14:17coordinate computation across thousands of chips
- 14:20simultaneously. But inference is fundamentally
- 14:22different. It often runs on smaller
- 14:24isolated clusters or local servers, or even edge devices
- 14:28like phones or specialized industrial machinery on a
- 14:30factory floor. So what inference requires?
- 14:33Isn't that brute force synchronicity?
- 14:35No, it requires efficiency, low latency and energy optimization.
- 14:39You need the model to run cheap and fast, not just big.
- 14:42Which means that the expensive high performance training chips
- 14:45are often inefficient or just overkill for these
- 14:48run-of-the-mill inference tasks. Precisely.
- 14:51And the economic of demand are completely distinct.
- 14:53Training demand is periodic, it's bursty, and it's
- 14:56consolidated among a few dozen well known players.
- 14:59Inference demand, however, is continuous.
- 15:01It's real time, and it's distributed across millions of
- 15:04enterprises and billions of end users globally.
- 15:07The total potential scale of inference demand dwarfs training
- 15:10demand by orders of magnitude. But there is a utilization gap
- 15:14right now, isn't there? I mean, we have the chips, but
- 15:17the apps haven't scaled fast enough to use them all.
- 15:20We do, and that's the central caveat.
- 15:22The exponential growth Eng predicts for inference compute
- 15:25relies entirely on the successful deployment and
- 15:28scaling of the application layer layer 3.
- 15:31So if AI applications succeed in achieving labor substitution,
- 15:35generating vast amounts of value.
- 15:36Then inference demand will explode, making today's capacity
- 15:39seem totally insufficient and strong call for continued
- 15:43inference investment is therefore a direct calculated
- 15:46bet on the success of the applications above it.
- 15:49And if that bet pays off, this layer suggests A fundamental
- 15:52shift in the architecture of AI deployment.
- 15:55It moves away from the centralized winner take all
- 15:58dynamics of the training layer. Where only the Giants can play.
- 16:01Exactly. And it moves toward a more
- 16:03competitive, decentralized and ultimately commodity market for
- 16:06execution. Which is a crucial structural
- 16:08component for the long term health of the entire ecosystem.
- 16:12If intelligence itself becomes cheap to use, it removes a major
- 16:15barrier to adoption for small and medium sized enterprises.
- 16:18So if layer 2 becomes a commodity, it lowers the hurdle
- 16:22for layer three applications to scale rapidly.
- 16:24Effectively fueling the whole stack from the middle.
- 16:27So while Layer 1 might be a bubble driven by high CapEx and
- 16:30rapid depreciation, Layer 2 is more like an undervalued
- 16:33workhorse. Where efficiency and
- 16:35distribution will be the key economic differentiators going
- 16:38forward, rather than just pure horsepower.
- 16:41OK, so now we get to the core of Ang's thesis, the valuation
- 16:45inversion. We have massive capital in
- 16:47layers one and two, but Ang identifies layer 3, the
- 16:50application layer, as severely under invested.
- 16:54And this is the paradox. The only layer that can generate
- 16:57the cash flow to justify the entire multitrillion dollar
- 17:01infrastructure is being systematically starved of
- 17:04capital. What's causing that?
- 17:05The cause is a deeply ingrained fear among venture capitalists
- 17:09and institutional money, the fear of the giants, or, as it's
- 17:12often called, sherlocking. They've been burned before.
- 17:15Yeah, OK, Explain Sherlocking. In this context.
- 17:17It implies that anything a startup builds can be just
- 17:19instantly copied. Precisely, it's the fear that if
- 17:23a startup builds a highly valuable, successful, but
- 17:26ultimately generic AI feature, let's say an automated customer
- 17:30service triage bot specialized for small businesses.
- 17:34The model giant, whether it's Google, Open AI, Microsoft.
- 17:37Will simply absorb that feature into its next native API update.
- 17:41The startup is instantly rendered obsolete or Sherlocked.
- 17:44And the VC investment is wiped out.
- 17:46Right. Why fund a company to build a
- 17:48feature that the platform provider can just give away for
- 17:50free next quarter? That fear, while it seems
- 17:53rational on the surface, it leads to paralysis, and N argues
- 17:58that this paralysis is based on a fundamental misunderstanding
- 18:01of how complex real world value is created.
- 18:04He contends that the recipe for building infrastructure you
- 18:07know, buy chips, pour concrete, secure power.
- 18:09It's conceptually simple and qualifiable.
- 18:12It's easy for a generalist investor to understand and
- 18:14model. But the recipe for building a
- 18:16truly successful high value AI application?
- 18:19Is complex. It's messy, and it's deeply
- 18:23specific to a vertical It requires proprietary domain data
- 18:27navigating these Byzantine workflow integration challenges
- 18:31within legacy enterprises and earning specialized user trust.
- 18:35And that complexity is precisely what generalist capital
- 18:38struggles to evaluate. But that's where the
- 18:39defensibility lies hidden in that complexity.
- 18:42And that leads to ENG's fundamental economic truism,
- 18:46which has to be correct for the entire AI economy to survive.
- 18:50The truism is just simple or. Really, it is almost by
- 18:54definition, applications have to be more valuable than the
- 18:57infrastructure, since we need them to be able to pay the
- 18:59infrastructure and technology providers.
- 19:01I mean, if the application layer doesn't generate more revenue
- 19:04than the layers beneath the cost, the entire structure is
- 19:07just a House of Cards. It is so if we are currently in
- 19:10what's called scenario A, where the total revenue generated by
- 19:13applications is less than the total cost of the
- 19:16infrastructure. And you have to include
- 19:18depreciation and debt service in that cost.
- 19:20Right then, the bubble burst. The market fails to sustain
- 19:23itself. The path to sustainability
- 19:25scenario B absolutely requires the application layer value to
- 19:30explode. It has to generate enough high
- 19:32margin revenue to service the debt taken on to build the data
- 19:35centers and the chips and still provide a profit margin for the
- 19:38application builder itself. Which?
- 19:40Requires applications to do things that generate massive,
- 19:43verifiable ROI, yes. So how then does a startup
- 19:47survive that threat of Sherlocking?
- 19:51ANG advocates for an almost aggressive defensibility through
- 19:55deep vertical integration. The key is moving away from the
- 19:59generic. A generic writing assistant is
- 20:01highly vulnerable. It's an easy feature for a
- 20:03foundation model to absorb. But consider something else A
- 20:06regulatory compliance agent for FDA filings in the
- 20:09Pharmaceutical industry. Now that is different.
- 20:11That requires proprietary access to internal document
- 20:14repositories and understanding of changing regulatory schemas,
- 20:18integration with specific legacy validation systems, and
- 20:21certified audit trail. And those elements act as a
- 20:23massive economic Moat. A Moat that a general purpose
- 20:26model simply cannot cross with a new API endpoint.
- 20:29So the insight here for you is that smart capital needs to
- 20:33execute a dramatic rotation. It needs to stop funding
- 20:36generic, easily replicable foundation models.
- 20:39And start funding the intensely complex vertical workflow agents
- 20:43that solve multi billion dollar problems for a specific
- 20:47industry. That's where N says the
- 20:49potential value is still much, much greater than the market
- 20:52currently recognizes. OK, so if the application layer
- 20:56is the necessary savior for this AI economy.
- 20:58Then the enabling technology that will finally drive that
- 21:01massive revenue jump is the shift from simple chat bots to
- 21:05complex agentic workflows. This is the difference between a
- 21:08minor efficiency gain and outright labor substitution.
- 21:12And this distinction is crucial for you to grasp.
- 21:14We're moving the conversation away from model size, which is
- 21:17mostly PR hype, and toward functional capability, which is
- 21:20business reality. So let's define the shift in the
- 21:23unit of work, the chatbot. That's the old paradigm.
- 21:26It's essentially A productivity aid.
- 21:28Right. It speeds up human work by
- 21:30performing information retrieval or some basic drafting.
- 21:33It competes with low value, low margin activities like a basic
- 21:37Google search or a junior employee summarizing a document.
- 21:40And you might see a modest, what, 10 to 20% productivity
- 21:43boost. At best and that is not going to
- 21:46close the trillion dollar revenue.
- 21:48OK, so then the agent, this is the new paradigm.
- 21:51This is a true labor substitute. It is.
- 21:54It's a system that can take a high level goal, plan the steps
- 21:57necessary, use external tools to gather data, execute the
- 22:01necessary code or interactions, and then iterate until that goal
- 22:05is met. And this competes directly with
- 22:07high value labor. A specialized analyst.
- 22:09A junior developer. A paralegal.
- 22:12And it's this agentic capability defined by Eng and others, that
- 22:15has to drive the superior economic performance needed to
- 22:18justify the infrastructure costs.
- 22:20He highlights 4 essential characteristics that empower
- 22:23agents to achieve this level of autonomous performance.
- 22:26The 1st, and maybe the most critical for reliability is
- 22:30reflection. Reflection is the agent's
- 22:32ability to critically assess and critique its own output before
- 22:35it delivers a final result. So like a coding agent, it
- 22:38doesn't just generate the Python code.
- 22:40No. It executes the code, it
- 22:42captures the runtime errors or failing tests, and then it uses
- 22:46that feedback to automatically debug and refactor its own code
- 22:50until the tests pass. Or a legal agent writing a
- 22:52brief. It reviews its own output
- 22:54against established case law, databases, internal style
- 22:57guides, identifies inconsistencies in tone or
- 23:00citation, and then fixes them autonomously.
- 23:03This self correction loop moves the software from being a
- 23:07suggestion engine to a truly reliable worker.
- 23:09And that drastically increases the trust a human supervisor can
- 23:13place in the system, which is absolutely essential for
- 23:15enterprise adoption. The second characteristic is
- 23:17tool use. This sounds fundamental for real
- 23:20world integration. It is.
- 23:22A chatbot is often confined to the data it was trained on.
- 23:25An agent has to step outside that training box.
- 23:27It needs robust, reliable access to extra external AP.
- 23:31Is like running a financial query via Bloomberg, or checking
- 23:35real time inventory levels through ASAP system, or
- 23:39interacting with a browser to navigate a third party website.
- 23:42So the agent becomes a coordinator of external
- 23:44services. Stitching together complex
- 23:46digital actions to achieve a business outcome rather than
- 23:49just generating text. OK.
- 23:51The third is planning. And this elevates the agent way
- 23:54beyond a simple prompt response. A human gives the agent a high
- 23:58level goal. Design and implement a marketing
- 24:01campaign targeting small businesses in the Midwest.
- 24:04The agent doesn't just start writing copy.
- 24:06No. First it breaks that high level
- 24:08goal down into subtasks. Analyze target demographics.
- 24:11Draft three ad copy variations. Identify relevant ad platforms.
- 24:15Allocate budget setup tracking. And then it manages the
- 24:18sequential execution of those tasks.
- 24:20Requiring minimal human intervention to move from Step 1
- 24:23to completion. And finally, the most advanced
- 24:25characteristic multi agent collaboration.
- 24:28This mimics A specialized decentralized human team.
- 24:32So instead of 1 monolithic AI, you have several specialized
- 24:35agents working together, each with a distinct persona or skill
- 24:39set. So you could have a data
- 24:40scientist agent analyzing customer trends.
- 24:43And a copywriter generating ad text based on the scientists
- 24:46findings. And a chief marketing officer
- 24:49agent reviewing the strategy and approving the final budget.
- 24:52Exactly. They communicate and negotiate
- 24:55internally to produce a robust, multifaceted solution.
- 24:58So this combination reflection tool use planning and
- 25:01collaboration. This is what enables that
- 25:04massive productivity jump required to close the revenue
- 25:07gap. Let's use that coding example
- 25:09again to show the economic impact Ang is pointing toward.
- 25:11OK, if you have a developer using a basic chatbot, they
- 25:15might get a 20% speed boost. They still have to review, debug
- 25:19and plan everything. But if that same developer is
- 25:22managing a swarm of highly capable collaborating agents, A
- 25:25coder agent, A reviewer agent, A debugging agent.
- 25:29The human's role changes from hands on keyboard to pure
- 25:32supervisor and product owner. And that human can become what,
- 25:36500% more productive? Or even manage multiple projects
- 25:39at once. That level of efficiency
- 25:42generates application value, the ROI that is fondly commensurate
- 25:45with the billions poured into the infrastructure.
- 25:48That 500% jump is the only thing that transitions us from
- 25:51scenario A to scenario B. And N is betting that these
- 25:55agentic workflows, the green shoots, he sees in specialized
- 25:58areas like legal contracts, drug discovery, research, automated
- 26:02finance. Will grow into a dense forest of
- 26:04ROI, generating the necessary billions to pay the high cost of
- 26:09the underlying compute. To properly measure the risk
- 26:12inherent in ENG's tripartite stack, we have to look at the
- 26:15historical precedentthe.com mirror of the late 90s and early
- 26:192000s. Right, there are clear
- 26:21structural parallels, but they're also crucial divergences
- 26:24that suggest this boom is fundamentally more stable, as
- 26:27the sources suggest. Let's analyze the primary
- 26:29capital source difference first. In 2000, the capital flooding
- 26:33the system was largely driven by speculative IP OS, highly
- 26:36leveraged startups, and, you know, exuberant retail investors
- 26:39chasing easy money. It was shaky capital based on
- 26:43perceived promise and momentum. Today, the capital is vastly
- 26:47different. It's primarily sourced from deep
- 26:49corporate balance sheets, Big tech companies that are highly
- 26:53profitable holding massive cash reserves.
- 26:55And powerful sovereign wealth funds making long term strategic
- 26:59bets. Which is why Eng assesses the
- 27:012025 boom as more stable. The key players, the ones
- 27:06building the infrastructure, they have the financial
- 27:08resilience to absorb short term volatility and appreciation
- 27:12without immediate default. They can weather a severe market
- 27:15correction, unlike the unprofitable startups of two
- 27:18decades ago. But the utility of the
- 27:20technology is maybe the most convincing divergence.
- 27:23In 2000, utility was low. We had dial up speeds,
- 27:26fragmented Internet access. And business models that were
- 27:29often more theoretical than practical, like pets.com.
- 27:32Today, ANG assesses the utility is high and real AI is not a
- 27:36future promise. It's deeply integrated into
- 27:38production grade systems right now.
- 27:40Solving complex problems in code generation.
- 27:42Medical diagnostics. Sophisticated supply chain
- 27:45optimization. The technology is genuinely
- 27:47transformative, which makes the underlying demand fundamentally
- 27:50more durable than the speculative concepts of the
- 27:53latest. 90s However, the risk in layer one infrastructure still
- 27:56carries that historical echo of the dark fiber fallacy reborn,
- 28:01which Eng warns about specifically.
- 28:02The dark fiber analogy is so powerful because it illustrates
- 28:06the timing mismatch perfectly. In the late 90s telecom boom,
- 28:10billions were spent laying fiber optic cable, the dark fiber.
- 28:15Anticipating massive Internet traffic growth that took years
- 28:19longer to materialize than expected.
- 28:21And the companies laying the fiber went bankrupt because they
- 28:23couldn't service the debt, even though the fiber itself was
- 28:26eventually used. So the training infrastructure
- 28:28today faces the exact same solvency crisis if the demand
- 28:31for their train models arrive slower than expected.
- 28:34But we established that GP US are even riskier than fiber.
- 28:37Much riskier. It's because of the asset
- 28:39liability mismatch. Dark fiber last 20 years.
- 28:42If the exponential traffic growth takes 10 years to arrive,
- 28:45the asset is still valuable and eventually close off.
- 28:48But GPU's operating under Moore's Law acceleration are
- 28:52often functionally obsolete within three years.
- 28:54So if companies take on a billion dollars in debt to build
- 28:57a cluster today, anticipating a specific ROI in five years, but
- 29:01the asset loses 75% of its value in three years, they face a
- 29:05crippling solvency crisis. That fast depreciating debt
- 29:09liability is the specific cratch mechanism Eng identifies in
- 29:13layer one. Beyond the purely financial
- 29:15risks, Eng also points to several non financial market
- 29:18briefs that could slow the rate at which the application layer
- 29:21scales, keeping that revenue gap wide open.
- 29:24Geopolitics is a massive 1. Huge the US China technology
- 29:27race and the strict export controls on advanced chips are
- 29:31forcing the creation of parallel, often less efficient
- 29:34markets globally. So China, for instance, has to
- 29:37invest heavily in domestic chip fabrication that may be several
- 29:40generations behind. Which leads to distorted global
- 29:43demand figures and potentially creates artificial scarcity or
- 29:46inflated valuation bubbles in local markets.
- 29:48And then? There's a very practical break,
- 29:50one that Eng is deeply invested in addressing himself, and
- 29:53that's the skills gap. The application layer simply
- 29:56cannot grow fast enough if there aren't enough skilled human
- 29:59engineers to build the agentic orchestrations we just talked
- 30:02about. It takes time to teach engineers
- 30:04how to look at incredibly complex enterprise workflows and
- 30:08translate them into robust multi step AI agents.
- 30:12Ng notes just as human work is complicated, it's taking
- 30:16everyone a while to look at the very complex workflows and to
- 30:19build AI agentic workflows. This human capital shortage acts
- 30:23as a governor ensuring that the application layer growth is
- 30:26inherently slower than the capital pouring into the
- 30:29training layer. Which increases the risk of
- 30:31oversupply. And finally, we have the most
- 30:33fundamental constraint of all physical limits, specifically
- 30:37the energy bottleneck. This is where finance and
- 30:40physics collide. The computational demands are
- 30:43reaching such a level that growth is being constrained by
- 30:45power grid capacity, not just chip availability.
- 30:49So data centers can't secure sufficient megawatts and many
- 30:52are struggling with this. Growth just stops instantly.
- 30:55It creates the strange scarcity bubble where existing power data
- 31:00center racks are incredibly valuable, but billions of
- 31:04dollars of brand new unpowered chips sitting in warehouses
- 31:09become functionally worthless until a new power plant comes
- 31:12online. The.
- 31:13AI bubble, paradoxically might pop due to a lack of
- 31:17electricity. It's a real possibility.
- 31:19I mean, it's just an overwhelming volume of data to
- 31:21process, right? The market complexity is it's
- 31:23just unprecedented. It really is, and it's hard to
- 31:26separate the signal from all that noise.
- 31:28Exactly. So for you, as you're trying to
- 31:30track all this volatility, remember that, you know,
- 31:32objective data-driven analysis is really key.
- 31:35And if you're looking for ways to track that total market
- 31:38temperature based on valuation metrics that kind of filter out
- 31:41the day-to-day craziness, we'd actually encourage you to search
- 31:44and download the iOS app Buffett Indicator from the Apple App
- 31:47Store, right? It gives you those objective
- 31:49metrics, helps you gauge if the market's, you know, overheating
- 31:51or maybe starting to stabilize. OK, So to synthesize Andrew
- 31:55ENG's comprehensive framework, we have to accept that his
- 31:58answer to the core question is there an AI bubble is
- 32:01sophisticated, it's necessarily nuanced.
- 32:04It is not a simple yes or no. Structurally, the answer is a
- 32:08qualified yes. There is likely a bubble in the
- 32:11training infrastructure layer. This is characterized by capital
- 32:14chasing redundant model creation fueled by that synthetic demand,
- 32:19and facing catastrophic rapid depreciation risks due to
- 32:22technological obsolescence. Conversely, the answer is a firm
- 32:26no. There is not a bubble in the
- 32:27application layer. That layer is fundamentally
- 32:30undercapitalized relative to its potential.
- 32:32It holds the latent value necessary to sustain the entire
- 32:36technological stack. The opportunity for growth here
- 32:38is still vast, largely untapped and structurally shielded if
- 32:42capital is deployed correctly. And the burst of the training
- 32:45bubble, if it happens, may not be the cataclysm of the
- 32:47Pessimosphere, as ANG suggests. It might be a necessary, almost
- 32:52healthy event. It would consolidate redundant
- 32:54market efforts, eliminate excess capacity, and, crucially, it
- 32:57would lower the marginal cost of compute for the application
- 33:00layer. Which would finally trigger the
- 33:02true golden age of AI deployment and adoption.
- 33:05The soft landing scenario depends entirely on a dramatic
- 33:08and efficient rotation of capital.
- 33:10We have to stop funding the 50th General Purpose Foundation model
- 33:13and instead start funding the 5000 TH Agentic workflow that
- 33:18integrates proprietary dietary data and solves a billion dollar
- 33:21problem for the manufacturing or healthcare or finance industry.
- 33:26The genuine profound utility of AI means the value gap will
- 33:29eventually close, but it's a high stakes race against the
- 33:33debt and depreciation cycles of that infrastructure layer.
- 33:36It is. And this brings us to our final
- 33:38provocative thought for you to carry forward.
- 33:40Shift your attention away from the model leaderboards, that
- 33:43obsession with who has the largest parameter count or the
- 33:46smartest chatbot. That's primarily PR hype.
- 33:49It's driven by the training layers need to justify its
- 33:51massive CapEx. Instead, start watching the ROI
- 33:54leaderboards. Who is saving their clients the
- 33:56most verifiable money with highly integrated vertical
- 33:59agentic workflows? That is the true business
- 34:01reality that will close the revenue gap and dictate the long
- 34:04term winners in this economy. The potential for genuine labor
- 34:08substitution and economic transformation remains immense.
- 34:12And by focusing on that utility and recognizing the structural
- 34:16differences within the AI stack, we can move past speculative
- 34:20volatility toward verifiable durable growth.
- 34:24Thank you for joining us on this deep dive into Andrew Mang's
- 34:27tripartite thesis. We'd love for you to continue
- 34:30this exploration with us if you found this level of structural
- 34:33analysis helpful and want to access more advanced content and
- 34:36market insights. Welcome to join our Deep Value
- 34:39Club membership. You can also subscribe to us on
- 34:41Apple and Spotify Podcasts channel Deep Value investing.
- 34:45And a final reminder, all analysis in this discussion is
- 34:48based on publicly available sources for educational purposes
- 34:51only and does not constitute investment advice.
- 34:54Please conduct your own due diligence before making any
- 34:57investment decisions. We'll see you next time on the
- 34:59Deep Dive.