Latest / Deep Values: Value Investing Deep Dives / Are FANNG outdated? Let's talk about Capex and Earnings in 2026
Transcript
- 0:00Well, imagine taking the Hoover Dam, the entire US Interstate
- 0:04Highway System, and like the Apollo Space program.
- 0:08OK, wow. That is a lot of concrete and
- 0:10rocket fuel. Right, so roll all their
- 0:13inflation adjusted costs together into one massive global
- 0:16initiative and then hand the final bill to just four
- 0:20companies. Just four, yeah.
- 0:22Because that is exactly what we are witnessing right now in
- 0:24Silicon Valley. We are tracking a projected $630
- 0:28billion capital expenditure deployment by Amazon, Google,
- 0:32Meta and Microsoft, all set to be fully realized by 2026.
- 0:36It is staggering. I mean, this isn't just a
- 0:38standard upgrade cycle. It is an invisible, high stakes
- 0:42infrastructure war. It really is.
- 0:43It represents this, this violent secular shift in how capital
- 0:47flows through the modern economy because, you know, for a decade
- 0:50the tech sector was completely defined by capital, light high
- 0:53margin software scaling. Build it once, sell it a million
- 0:56times. Exactly.
- 0:57But now we are looking at heavy industry mechanics.
- 1:00I mean, if you track the CapEx to operating cash flow ratio for
- 1:03these 4 hyperscalers? Which is a crucial metric,
- 1:05right? Oh absolutely.
- 1:06It is the metric and the trajectory is just jarring.
- 1:10Like we saw it hovering around a historical norm of maybe 40%
- 1:13back in 2023, which is very healthy, but then by 2025 that
- 1:18metric vaulted to 65%. And it doesn't stop there.
- 1:22The forward models for 2026 suggest we are pushing an
- 1:25asymptote of nearly 90%. Wait, 90%.
- 1:28So almost every dollar of operating cash flow is just
- 1:30gone. Diverted directly into the
- 1:32furnace of artificial intelligence hardware.
- 1:35Which brings us to the core mission of our deep dive.
- 1:38Today, we're ripping apart the financial statements, reverse
- 1:41engineering the silicon strategies, and, you know,
- 1:44tracking the shifting power dynamics buried in our sources
- 1:47to find the actual truth behind this unprecedented spending
- 1:49spree. Because we really have to answer
- 1:511 defining question here. Exactly.
- 1:53Is this $630 billion outlay an act of collective defensive
- 1:58desperation by Big Tech? Or is it the bedrock of an
- 2:02entirely new macroeconomic order?
- 2:04Right. And by the time we wrap up, you
- 2:06will understand the mechanics of how this invisible silicon war
- 2:10dictates everything from the targeted ads on your mobile feed
- 2:13to the architecture of the enterprise software you rely on
- 2:17every day. So to understand the end game,
- 2:19we first have to look at the immediate panic this caused in
- 2:22the financial markets, right? Yeah, We have to ground this in
- 2:25the immediate reaction. You don't just deploy half a
- 2:27trillion dollars without collateral damage to your
- 2:29balance sheet, right? The initial cash flow crunch was
- 2:32severe enough to fundamentally re rate some of these equities.
- 2:36Amazon is like the quintessential case study here.
- 2:40Over a highly scrutinized trailing 12 month period,
- 2:43Amazon's free cash flow just evaporated.
- 2:46Literally disappeared. Completely.
- 2:48We tracked A collapse from a very robust $38 billion down to
- 2:52just $1.2 billion. That is, I mean, that is a 96%
- 2:56drawdown and free cash flow for a trillion dollar company.
- 2:59Yeah, that isn't just a minor capital reallocation.
- 3:02That's a financial Cliff that forces portfolio managers to hit
- 3:05the panic button. And they absolutely did hit the
- 3:07panic button. Beginning in July 2025, the
- 3:10market aggressively bifurcated. What does that mean for the
- 3:13listener? How did it split?
- 3:16Well, we witnessed this massive divergent in multiples.
- 3:19On one side you had these semiconductor designers, the
- 3:22pureplay infrastructure vendors selling the picks and shovels.
- 3:25Right, they saw massive gains. Exactly.
- 3:27But on the other side, the big tech hyperscalers who are
- 3:30actually absorbing these capital costs, they lagged way behind
- 3:34the market, punished the buyers and rewarded the suppliers.
- 3:38Because investors were terrified about 2026, right?
- 3:41Driven entirely by the fear that these 2026 CapEx projections
- 3:45would structure and permanently impair their long term free cash
- 3:49flow yields. It feels like Big Tech is
- 3:51trapped in a high stakes game of musical chairs crossed with a
- 3:55prisoner's dilemma. That is a great way to put it.
- 3:57Because I mean no one dares to stop buying GPU's even if it
- 4:00bankrupts them right? Stopping means losing the AI war
- 4:03so they just keep signing the purchase orders.
- 4:05Did it look like a multi billion dollar staring contest?
- 4:08It did look like that, and you know that Prisoner's Dilemma
- 4:11framework is popular, but if we analyze the leading indicators
- 4:15and supply chain, it suggests a much more rational mechanism at
- 4:19play. Wait, really rational?
- 4:21Yeah, I pushed back on the panic narrative a bit.
- 4:24This isn't just preventative hoarding, it's actually a
- 4:26desperate attempt to catch up with compounding organic demand.
- 4:29OK. What's the evidence for that?
- 4:31We can verify this by tracking the leading times, so the delta
- 4:35between a hyperscaler executing a hardware contract and the
- 4:39actual deployment of those compute clusters.
- 4:42And those wait times are getting longer.
- 4:43Much longer. Those lead times are not
- 4:46compressing, they are stretching further out.
- 4:49But the most definitive signal is in the pricing mechanics of
- 4:53the compute itself. Let's dig into that pricing
- 4:55mechanism actually because historically compute is a
- 4:58depreciating asset. I mean, we are used to Moore's
- 5:01Law dynamics where the cost of a standardized compute instance
- 5:05steadily decays over time. Right, which was the standard
- 5:08operating procedure for a decade.
- 5:10Long term contract prices for raw compute were in a state of
- 5:13managed decline. However, crossing the threshold
- 5:16of December 2023, that vector totally inverted.
- 5:20We saw aggressive sustained spikes across the entire compute
- 5:24spectrum. So everything got more
- 5:25expensive. Everything.
- 5:27Spot prices, short term burst contracts and even long term
- 5:31reserved instances all moved vertically.
- 5:34But wait, if everyone is building out capacity
- 5:36simultaneously, expanding the total supply of available
- 5:39compute, A simultaneous spike across all contract durations
- 5:44implies a structural bottleneck. Where is the actual constraint?
- 5:48The constraint is the scale of the foundation model companies
- 5:51themselves. Like open AI and anthropic.
- 5:53Exactly. They are absorbing massive multi
- 5:56year tranches of compute for reinforcement learning and next
- 5:59generation multimodal training. And they are.
- 6:01Just eating up all. The supply they are locking up
- 6:03the reserved instances at such scale that the residual pool of
- 6:06compute available for the short term spot market is artificially
- 6:10constrained. Squeezing the short term supply.
- 6:12Right, The spot market is hyper elastic.
- 6:15When you drain the liquidity of available compute, the remaining
- 6:19sharp Herm instances command exorbitant premiums.
- 6:22It's like trying to get a last minute hotel booking on a
- 6:25holiday weekend. Oh, that makes sense.
- 6:26And it extends beyond just the logic layer.
- 6:29You know it's hitting the storage architecture as well.
- 6:31Right, because multimodal training requires ingesting
- 6:34exhabits of unstructured data like video and audio.
- 6:38Exactly. Which fundamentally alters the
- 6:40storage economics. Totally.
- 6:42And the financial friction here is that cloud providers
- 6:45structure their internal storage acquisitions on staggered multi
- 6:49year base rate contracts. OK.
- 6:51So they lock in a price, right? But when the spot price for
- 6:54enterprise flash and high bandwidth memory spikes during
- 6:57an AI hardware scramble, the cloud providers are forced to
- 7:00lock in those highly inflated storage costs across a
- 7:03three-year horizon. Ouch.
- 7:05Yeah, yeah. They aren't just absorbing a bad
- 7:08quarter of CapEx, they are cementing a higher fundamental
- 7:11floor for their infrastructure costs through the end of the
- 7:13decade. So if the underlying
- 7:15infrastructure costs are structurally elevated and the
- 7:18foundation model companies are the ones dictating the demand
- 7:21curve, it suggests a massive shift in industry power.
- 7:25A complete inversion of leverage in the tech ecosystem.
- 7:28I mean, if we look back at the Sauce era, roughly 2015 to 2021,
- 7:33the hyperscalers, AWS, Azure, Google cloud, they held all the
- 7:39power. They were the undisputed apex
- 7:41predators. They operated as toll roads.
- 7:44Their margin architecture was brilliant.
- 7:46Because they controlled the distribution bottleneck.
- 7:47Exactly. They would provision raw low
- 7:50margin server capacity, but then tightly couple it with their
- 7:53proprietary high margin software ecosystems.
- 7:56But they're relational databases and identity management tools.
- 7:59Right, they're specific telemetry tools.
- 8:01A workload that might theoretically generate a 40%
- 8:03gross margin on bare metal was magically transformed into a 65%
- 8:07margin business simply because the hyperscaler locked the
- 8:11customer into their proprietary software abstraction layer.
- 8:14But the sources indicate that the extraction layer has moved
- 8:17the foundation model companies are the new gatekeepers.
- 8:20The mechanism of this shift is entirely based on dynamic
- 8:23routing. How does that work?
- 8:25Well, when an enterprise Oregon, a user buys an API from
- 8:29Anthropic for instance, to power an internal application, that
- 8:33enterprise has 0 visibility into the underlying silicon executing
- 8:38the inference. They just know they're talking
- 8:40to clock. Exactly.
- 8:42Entropic acts as an abstraction engine.
- 8:44They dynamically route those API calls based on real time cost
- 8:48and latency metrics. 1 millisecond they might push a
- 8:51batch of requests through Amazon's Bedrock environment,
- 8:54and the next millisecond they might route overflowing traffic
- 8:57to Google's Vertex AI. It's the ultimate
- 9:00commoditization of the cloud. I mean think about it, when you
- 9:03stream a movie at home, do you care which specific data center
- 9:07is hosting the video file? No, of course not.
- 9:09You only care about the streaming service.
- 9:11You only care about Netflix or Disney.
- 9:13The best analogy here isn't digital, it's industrial.
- 9:16It's the electrical grid. Oh, I like that.
- 9:18Right. When an enterprise buys a
- 9:20massive industrial manufacturing robot, they care deeply about
- 9:24the brand of the robot, its capabilities, its precision, its
- 9:28software interface. That robot is the foundation
- 9:30model, but they absolutely do not care which specific power
- 9:33plant generated the electricity running through the wall socket
- 9:37to power. It which is terrifying for the
- 9:39hyperscalers. They are terrified of becoming
- 9:41the electrical grid, a capital intensive, heavily regulated
- 9:45utility that provides an interchangeable commodity.
- 9:48Because if they are just utilities, they lose their
- 9:50margins. Exactly.
- 9:51And that dynamic routing capability strips the
- 9:54hyperscalers of their pricing power.
- 9:56If AWS tries to raise the cost of their compute instances,
- 10:00Anthropic simply shifts the routing parameters to Google
- 10:03Cloud. The foundation model owns the
- 10:05customer relationship. The cloud provider just burns
- 10:08the electricity. Which provides the mathematical
- 10:10context for Open AI's staggering $90 billion valuation.
- 10:14Yeah, that valuation caused a lot of friction in the press.
- 10:17We saw the headlines balking at it, calling it a Texpic bubble,
- 10:21but if we unpack the numbers in the sources and assume they are
- 10:24tracking toward 10 billion in annual recurring revenue, we are
- 10:28looking at roughly A9 X forward ARR multiple.
- 10:32Which isn't actually fundamentally overpriced, no.
- 10:35If they maintain this gatekeeper status and eventually stabilize
- 10:39at 20 to 30% net margins, they are essentially modeling out to
- 10:43Microsoft's historical 35X price to earnings ratio.
- 10:48The valuation assumes they successfully and permanently
- 10:50relegate the hyperscalers to utility status.
- 10:53The math holds up, absolutely, but only if you assume the
- 10:56hyperscalers capitulate. Which they're not doing.
- 10:59No, and the $630 billion CapEx projection proves they're doing
- 11:04the exact opposite. They're fighting A2 front war.
- 11:06They are getting squeezed from the top by the foundation models
- 11:09demanding cheaper compute, and squeezed from the bottom by
- 11:12NVIDIA commanding massive premiums for their GPU's.
- 11:15Right. The only escape velocity from
- 11:17that margin compression is vertical integration at the
- 11:19silicon level. The Silicon rebellion.
- 11:22This is where the engineering physics meet the financial
- 11:24statements. With model companies squeezing
- 11:27from the top and NVIDIA squeezing from the bottom, cloud
- 11:30providers are fighting back the only way they can by building
- 11:33their own custom silicon. The hyper scalers are
- 11:36aggressively migrating workloads to bespoke, internally developed
- 11:40ASICS application specific integrated circuits.
- 11:44Like Google's TPU, Amazon's tranium in Graviton.
- 11:47Meta is deploying the MTIA and Microsoft is attempting to scale
- 11:51Maya. The goal is to migrate inference
- 11:54workloads away from NVIDIA and keep those profit margins in
- 11:57house. Let's talk about Google's
- 11:59massive advantage here with ATPU, because they had a massive
- 12:02head start. Google is operating with a multi
- 12:04year structural advantage. Their tensor processing unit
- 12:07architecture is functionally distinct from a general purpose
- 12:09GPU. How much cheaper is it?
- 12:11It's incredible. By stripping away the legacy
- 12:14architecture required for FX rendering and generalized
- 12:17compute, and optimizing the silicon strictly for the matrix
- 12:21multiplication inherent in neural networks, they achieve
- 12:24massive thermal and cost efficiencies.
- 12:27Our sources indicate Google can train models on TPU's at roughly
- 12:3020 to 30% of the cost of open AI training on GPU.
- 12:34'S that is insane structural mode, 70% cost arbitrage.
- 12:39It means Google can theoretically build models two
- 12:41to three times larger than a competitor in the same
- 12:44generation utilizing the exact same capital footprint.
- 12:47Exactly. But the sources also highlight a
- 12:49fascinating conflict within Google because of this almost an
- 12:53internal civil war. Because if you only have a
- 12:55limited supply of these magical TPU's that print efficiency, who
- 12:59gets them? Right, it's an acute supply
- 13:01chain bottleneck. Do you give them to the internal
- 13:03Gemini team trying to build the world's best model, or the
- 13:06external Google Cloud clients who want to rent them?
- 13:08And that is the tension. Google only has fabrication
- 13:12capacity for roughly 4 million TPU's this year, scaling toward
- 13:16maybe 10 million next year. In the context of global AI
- 13:19demand 4 million units is severely constrained.
- 13:23So it's a 0 sum game internally. It creates a 0 sum conflict
- 13:26between Google's internal Gemini development unit which needs
- 13:29maximum compute density to achieve AGI, and Google Cloud
- 13:33Platform which desperately needs those Tpus to lure enterprise
- 13:37clients away from AWS and Azure. How do they untangle that
- 13:41tension? Because every TPU allocated to
- 13:44an external client is a direct tax on the speed of their
- 13:47internal foundational model development.
- 13:48They navigate it through strategic partnerships that
- 13:51offer dual value. Take Anthropic, for example.
- 13:54Anthropic training its models on Google's TPU infrastructure is a
- 13:57perfect strategic victory because.
- 13:58It's not just revenue. Exactly.
- 14:00From GCP's perspective, this isn't just revenue, it's a
- 14:04massive subsidized beta test of their entire hardware ecosystem.
- 14:09When a frontier model like Claude pushes your silicon to
- 14:11the absolute limit, it's stress test.
- 14:14The optical switching, the liquid cooling manifolds, The
- 14:17entire proprietary software stack required to orchestrate
- 14:2010s of thousands of chips. It proves to the broader
- 14:22enterprise market that the TPU ecosystem is a battle tested,
- 14:27hardened alternative to the NVIDIA monopoly.
- 14:29Precisely. Let's pivot to Amazon because
- 14:32while Google had a multi year head start on the silicon
- 14:34design, Amazon seems to be executing a master class in
- 14:38infrastructure deploying with Tranium.
- 14:40It's saving them billions in CapEx, but the sources point out
- 14:43this was born out of a severe physical bottleneck last year.
- 14:46It's a crucial point regarding the physical realities of AI
- 14:48CapEx. You cannot simply drop high
- 14:50density AI clusters into legacy cloud data centers.
- 14:53The thermal density of these new racks fundamentally breaks
- 14:56traditional forced air HBAC cooling architecture.
- 14:59Yeah, just melt. Basically, yeah.
- 15:01Amazon historically struggled intensely with the physical
- 15:04retrofitting required to Co locate power hungry GPU clusters
- 15:08alongside their bespoke tranium silicon.
- 15:11The power draw per rack escalated so violently that it
- 15:14required entirely new electrical substations and direct to chip
- 15:17liquid cooling manifolds. Which explains the lag.
- 15:21The capital was available, but the concrete and copper
- 15:24logistics literally couldn't move fast enough, which delayed
- 15:27their revenue acceleration last year.
- 15:28Yes, they hit a physical wall, but now that they've largely
- 15:32resolved those physical data center constraints, their
- 15:35ability to deploy Tranium and their ARM based Graviton CPU's
- 15:39is wildly accelerating AWS revenue.
- 15:42So they are finally hitting their stride.
- 15:44And the margin math is highly favorable.
- 15:46If AWS can offer an enterprise a 20% performance to cost
- 15:50improvement by shifting their inference workloads to Tranium,
- 15:53the enterprise happily refactors their code to accommodate the
- 15:56custom silicon. AWS pretends the remaining cost
- 15:59differential, effectively expanding their gross margins
- 16:01while capturing market share. Contrast that physical execution
- 16:05with Microsoft's current silicon posture, because Maya seems
- 16:09functionally restricted compared to what Google and Amazon are
- 16:12deploying. Microsoft is definitely exposed
- 16:15on the silicon front. Maya is currently relegated to
- 16:17handling smaller specialized models.
- 16:20The architecture is just not yet capable of orchestrating the
- 16:23massive multimodal workloads demanded by GPT 4 class
- 16:27foundation models. So they are stuck renting.
- 16:30This forces Microsoft to remain highly dependent on standard
- 16:33high margin external hardware for the most critical revenue
- 16:37generating workloads, limiting their ability to shield their
- 16:40margins from hardware vendors. Which creates a perfect bridge
- 16:43to examine the actual revenue generation.
- 16:45We've mapped the $630 billion capital drain, the data center
- 16:49physics, and the silicon strategies.
- 16:51But where is the actual money coming from?
- 16:53We need to categorize the mysterious AI revenue everyone
- 16:56claims to have. We really do, because everyone
- 16:58in the Fortune 500 is reporting AI revenue to appease analysts.
- 17:02But we need to isolate the actual net new commercial
- 17:05behavior. Let's look at the varying
- 17:07definitions. Right, like Google integrates it
- 17:09directly into Google Cloud and Workspace, accelerating GCP to
- 17:1363% growth. AWS focuses strictly on Bedrock
- 17:17revenue, hitting 28% growth at 15 quarter high.
- 17:20But let's scrutinize Microsoft's headline number, $37 billion in
- 17:24annualized AI revenue. That $37 billion figure requires
- 17:28intense forensic accounting. It is a highly blended metric.
- 17:32It aggregates the raw infrastructure revenue from
- 17:35renting Azure GPU's, the API revenue share they capture from
- 17:38Open AI, and the seat based software revenue generated by
- 17:42injecting Copilot into the Office 365 ecosystem.
- 17:45I want to push back aggressively on the software component of
- 17:48that metric. I mean, it's easy to slap an AI
- 17:51label on an existing Office 365 subscription.
- 17:54Slapping A Copilot button into an Excel spreadsheet and raising
- 17:57the enterprise licensing fee by 20% is a brilliant pricing
- 18:00lever, but how much of this is genuinely new behavior?
- 18:03Or is it just Microsoft extracting rent from their
- 18:05existing distribution monopoly? It's a very valid skepticism,
- 18:08and it highlights a deep structural vulnerability in
- 18:11Microsoft's software strategy, driven largely by their
- 18:13responsible AI architecture. What does that mean in practice?
- 18:17Well, Microsoft's mandate is enterprise compliance.
- 18:20They force their products through this layer.
- 18:22They cannot risk a rogue output from an AI agent hallucinating
- 18:25financial data or violating a corporate brand standard.
- 18:29Consequently, they route their AI outputs through heavily
- 18:33restrictive, latency inducing filtering layers.
- 18:36So the corporate armor is weighing down the application.
- 18:38Exactly. While it's necessary for
- 18:40enterprise compliance, it artificially throttles the speed
- 18:43and lateral creativity of the models.
- 18:46It forces the output into a safe generic corridor.
- 18:49Leaving them vulnerable. Yes, this leaves them completely
- 18:52vulnerable to faster AI native startups.
- 18:55Leaner competitors unburdened by legacy compliance frameworks are
- 18:59shipping specialized, highly iterative tools and code weekly
- 19:02that functionally outperform Microsoft's bundled offerings.
- 19:05So if the seat based model is vulnerable to specialized
- 19:08competitors, how does Microsoft defend the revenue multiple?
- 19:12Where is the true software revenue ceiling?
- 19:14The defense lies in the transition from seat based
- 19:17pricing to consumption based pricing, specifically through
- 19:21platforms like Copilot Studio. Right now, Copilot penetration
- 19:24is sitting at roughly 5% of their 400 million user base,
- 19:28primarily because CF OS are gating the $30 monthly per seat
- 19:32cost. It's a linear, easily capped
- 19:34expense. But Copilot Studio changes the
- 19:37unit economics entirely. Completely Studio allows an
- 19:41enterprise to build autonomous internal AI agents.
- 19:45Imagine a bespoke HR agent designed to cross reference
- 19:48thousands of employee records, compliance documents and payroll
- 19:51databases simultaneously. The Enterprise doesn't pay a
- 19:54flat $30 fee for that agent. No, they pay for token
- 19:56consumption. Every time that agent executes a
- 19:59multi step query, it burns thousands of tokens.
- 20:01The sources indicate that under this consumption model, a single
- 20:04enterprise deployment can easily scale to 10,000 or $20,000 a
- 20:07month in compute utilization, rather than a flat $30 fee.
- 20:12The transition from human gated software to autonomous token
- 20:15consumption represents an exponential revenue curve.
- 20:17That is a fascinating long term play, but it still requires A
- 20:21lengthy enterprise sales cycle and deep behavioral changes
- 20:24within the Fortune 500. While everyone watches the
- 20:27dramatic enterprise software battles, the most immediate and
- 20:30tangible AI monetization is quietly happening right in front
- 20:33of us in the ads on our phones. Yeah, the stealth AI winners,
- 20:37the advertising algorithms. Let's focus on Beta's recent
- 20:41quarterly performance. It presented a paradox that
- 20:44completely subverts traditional advertising economics.
- 20:47Metas financials are the cleanest current proof of
- 20:49concept for immediate AIROI. They reported a 19% growth in
- 20:54total. Ad impressions delivered.
- 20:56Let's pause on that. In any traditional commodity
- 20:58market, if you artificially flood the ecosystem with a 19%
- 21:02increase in supply, the unit price ice craters.
- 21:05Increasing supply usually tanks the price.
- 21:08Except their price per ad actually increased by 12%
- 21:11simultaneously. They expanded inventory and
- 21:13expanded yield concurrently. Mechanistically, how did they
- 21:16pull that off without degrading the user experience to the point
- 21:19of churn? By utilizing foundation models
- 21:21to fundamentally alter the nature of the ad inventory
- 21:24itself, they aggressively scaled overlay ads or partnership ads.
- 21:29Taking a page from TikTok, Meta uses AI to seamlessly match
- 21:32creator content with relevant sponsor overlays.
- 21:36Historically, Meta's recommendation engines relied on
- 21:39rigid matrix factorization machine learning.
- 21:41If a user was watching a high fidelity sports highlight, the
- 21:45system could only safely serve a sports adjacent ad with matching
- 21:48metadata tags. A rigid tagging system restricts
- 21:51your usable inventory. You can't put a disjointed used
- 21:54car ad over highly specific sports content without annoying
- 21:57the user. Exactly.
- 21:59But by transitioning their core recommendation engine to an
- 22:01architecture based on their Llama Foundation model, the LTM,
- 22:05they moved from rigid tagging to semantic understanding.
- 22:08The AI understands the nuanced context of the video and the
- 22:11intent of the user in real time. And in an industry where a 2%
- 22:15algorithm improvement is considered groundbreaking,
- 22:17introducing foundation models into ad matching is
- 22:20revolutionizing conversion rates.
- 22:22Yes, it can dynamically inject a soft, highly relevant sponsor
- 22:27overlay that feels native to the content rather than an
- 22:30interruption, like matching a sports creator with a sneaker
- 22:33ad. Perfectly integrated.
- 22:35And because that bridging results in higher conversion
- 22:37rates, the advertisers are willingly bidding up the price
- 22:40of the newly created inventory. It's like AI gave Meta the
- 22:45ability to instantly build a customized, highly targeted
- 22:48billboard in the middle of a desert where no one thought you
- 22:51could sell anything. And it's working instantly.
- 22:53It really is. When shifting to Google, they
- 22:56are executing a very similar mechanism but implying it to the
- 22:59search interface to monetize the UN monetizable.
- 23:02The classic long tail query. Exactly Consider previously
- 23:05worthless long tail queries. Before, if you asked a three
- 23:08paragraph question detailing your budget, your airline
- 23:11preferences, and your exact dietary restrictions for a trip
- 23:13to Tokyo, traditional search would panic.
- 23:16It yielded no ads. Advertisers bid on exact match
- 23:19keywords like Tokyo flights, not on conversational essays.
- 23:22But now AI allows Google to capture it.
- 23:25The generative model parses the complex intent of that 200 word
- 23:29query, initiates A dynamic multi turn conversation, and
- 23:33dynamically inserts high value AI ad slots exactly when the
- 23:36user is ready to buy. All without cannibalizing
- 23:39traditional search ad space at the top of the page.
- 23:41It is ruthless efficiency, but we have to pivot to the
- 23:44macroeconomic consequences of that efficiency because this
- 23:48efficiency has a dark side. Meta is making a fortune on ads,
- 23:52yet their stock plummeted recently.
- 23:53Why? Because the shift to AI is
- 23:55fundamentally altering the human workforce, replacing operating
- 23:59expenses OpEx with capital expenses CapEx.
- 24:02Right, when a corporation unlocks massive productivity
- 24:05gains through automation, that capital doesn't sit idle, it is
- 24:08redirected. And we saw exactly how it's
- 24:11redirected in Meta's financial panic.
- 24:12They added ten billion dollars to their CapEx guidance,
- 24:15triggering fears about massive depreciation costs hitting their
- 24:18bottom line by 2026 or 2027. The market panicked, assuming
- 24:22Meadow would absorb the margin compression.
- 24:24But Meta offset this massive increase in silicon CapEx by
- 24:27executing severe structural cuts to their OpEx, primarily through
- 24:31heavy human layoffs. They are directly substituting
- 24:35the ongoing recurring cost of software engineering salaries
- 24:38with the upfront depreciable cost of GPU infrastructure.
- 24:42Let's isolate the mechanics of how software engineering is
- 24:44actually changing, because this isn't just an efficiency drive
- 24:48to cut underperforming staff. The fundamental unit of Labor in
- 24:51Silicon Valley is transforming. It's a sobering look at the
- 24:55reality of AI in coding. We have crossed a Rubicon in
- 24:58code generation. We moved incredibly rapidly from
- 25:01AI writing 50% of boilerplate syntax to specialized agents
- 25:05capable of writing 100% of code for complex functions
- 25:08autonomously. So what does the engineer do
- 25:10now? The role of the human engineer
- 25:12has been completely altered. They are transitioning from
- 25:15being the primary typists of logic to acting as managers,
- 25:18orchestrating dozens of autonomous AI coding agents
- 25:21simultaneously. An engineer is no longer writing
- 25:23a single specific API integration.
- 25:26They're writing the prompt that instructs 50 agents to write 50
- 25:29integration simultaneously, and then they act as the QA layer
- 25:33reviewing the output. And the trajectory is steep.
- 25:36Our sources project that AGI capabilities within the specific
- 25:40domain of software engineering will approach 90% capability by
- 25:44the 2028 to 2029 window. If a single engineer's output
- 25:48has multiplied exponentially, corporate leadership faces a
- 25:51binary choice. They can produce 10 times more
- 25:54products, hoping the market can absorb that volume.
- 25:57Or they can maintain their current product road map and lay
- 26:00off a massive percentage of the workforce, dropping those OpEx
- 26:03savings directly to the bottom line.
- 26:05History suggests corporations will heavily favor the latter.
- 26:08Margin expansion always wins, and this is forcing
- 26:11macroeconomic theories that were historically relegated to
- 26:13academic white papers into the center of corporate strategy.
- 26:17I want to touch on that and partially because the sources
- 26:19mentioned the rising discourse around AI layoff taxes and
- 26:22transfer payments as systems like universal basic income.
- 26:25Are these emerging as serious systemic conversations among the
- 26:28institutional capital deploying this $630 billion?
- 26:32They are emerging as serious macroeconomic topics to manage
- 26:35the transition. Yes, because it becomes an issue
- 26:38if it threatens the broader economic demand curve.
- 26:40Consider the baseline metrics. Traditional enterprise IT
- 26:44spending hovers around 3 to 4% of total revenue.
- 26:47Human operating expenses salaries often represent 30% or
- 26:51more. The financial incentive to
- 26:52replace that 30% OpEx layer with an expanded I key CapEx layer is
- 26:57overwhelming. But if that transition happens
- 26:59simultaneously across the global economy, the discussion around
- 27:02layoff taxes or Ubi isn't driven by altruism.
- 27:05It's an economic debate about preventing systemic demand
- 27:08destruction during the transition phase, ensuring there
- 27:11are still human consumers to buy the software.
- 27:14It is the ultimate economic paradox which forces us to
- 27:18synthesize everything we've discussed and deliver a verdict
- 27:21on the core question. We've seen the hardware costs,
- 27:23the power struggles, the defensive silicon deployments,
- 27:26and the revenue streams. Does the math ultimately justify
- 27:30the $630 billion investment? Let's compare the numbers.
- 27:33The projected total CapEx is over $630 billion.
- 27:38Against that, the currently trackable annualized AI revenue
- 27:41across cloud, APIs and subscriptions is roughly 100
- 27:45billion to $150 billion. At first glance, this is a
- 27:48massive gap, $150 billion revenue stream trying to justify
- 27:53a $630 billion outlay. It looks like a catastrophic
- 27:57capital misallocation. It does if you assume the
- 27:59revenue is static and the CapEx is an annual recurring cost.
- 28:02Yeah, but the market is beginning to calm down because
- 28:04that fundamentally misunderstands the nature of the
- 28:07capital cycle. CapEx is a multi year
- 28:09forward-looking investment. The infrastructure being
- 28:11commissioned today is designed to cover the next four years of
- 28:14infrastructure needs, right? The 100 to $150 billion ARR
- 28:17figure is merely a snapshot of monetization today.
- 28:21So we have to factor in the compounding growth rate of the
- 28:24ARR. Exactly.
- 28:26If the model companies ARR accelerates to $400 billion by
- 28:29next year, the ratio of incremental cost to incremental
- 28:33revenue tightens dramatically. The market is realizing that the
- 28:37underlying business model is actually solidifying.
- 28:40So essentially the panic happened because people were
- 28:42looking at the price of a four year mortgage and comparing it
- 28:45to a single month's salary. But now that salary is doubling
- 28:49every few months. The math over a four year
- 28:52horizon is actually going to underwrite the debt.
- 28:54The velocity of the revenue acceleration is proving capable
- 28:57of catching the scale of the capital deployment.
- 28:59It is a massive gamble, but a mathematically sound one.
- 29:03To bring this all full circle, we started by trying to
- 29:05comprehend the sheer scale of a $630 billion gamble.
- 29:09We explored how power shifted from the clouds to the models,
- 29:12how custom silicon like Google's TPU and Amazon's Tranium are
- 29:15fighting back, how hidden ad revenues are funding the war,
- 29:19and how human OpEx is being transformed into silicon CapEx.
- 29:22We've mapped the transition, but as we close, there's a final
- 29:26provocative thought I want to leave the listener to Mull over.
- 29:28Go for it. We've established that Big Tech
- 29:30is successfully replacing human labor with machine
- 29:33infrastructure to justify the $630 billion spend.
- 29:38But what happens in a few years when these highly advanced
- 29:41models start optimizing their own code?
- 29:43Oh well. And not just that, what happens
- 29:45when they start negotiating their own computing costs across
- 29:48different clouds autonomously? As the infrastructure becomes
- 29:52completely autonomous, Where exactly does the human consumer
- 29:55fit into an economy driven by AI selling to AI?
- 29:59That is a staggering macroeconomic reality to leave
- 30:02hanging in the air, an entirely closed loop autonomous economy.
- 30:07Thank you for joining us on this deep dive.
- 30:08Keep questioning the fundamental mechanics behind the headlines
- 30:11and we will catch you next time.