The $2 Trillion Mirage: Inside the AI Stock Valuation Machine That's Running on Fumes
Let me paint you a picture. Microsoft, Amazon, Google, and Meta are on track to spend $2 trillion on AI infrastructure by 2027. Two. Trillion. Dollars. That's roughly the GDP of Italy, vaporized into server farms and GPUs.
Here's the punchline: they're getting almost nothing back.
The AI bubble isn't coming. We're already inflating it, breath by breath, earnings call by earnings call. And the dependency is almost comically circular.
When Anthropic commits $200 billion to Google Cloud, that's not a market. That's a circular financing ouroboros — and Google counts it as "backlog" while Anthropic books it as "compute access." Same chips, two valuations.
Even the bulls see the seams. Bernstein's Stacy Rasgon calls Nvidia "cheap" at a forward P/E of 26.31 — just 5.9% above the industry average. But that "cheap" label requires Nvidia to hit $78 billion in revenue with ±2% precision. No pressure.
Meanwhile, CoreWeave — the darling GPU renter — generated $5.15 billion in 2025. Sixty-seven percent came from Microsoft renting capacity... for OpenAI. One more loop in the circle.
Spoiler alert: It doesn't. But the quarterly reports sure look pretty while the music plays.
The $2 Trillion Bet: Unpacking AI Capital Expenditure
The numbers are staggering. The logic? Questionable. And the receipts? Let's just say Microsoft, Amazon, Google, Meta are building cathedrals for a congregation that hasn't quite arrived.
Here's the visual that should make every investor spill their oat milk latte. The stacked bars show the AI capex arms race. The line? That's reality knocking.
Notice the gap? That's not a rounding error. That's circular financing doing heavy lifting.
The OpenAI-Anthropic Dependency
Here's where it gets spicy. OpenAI and Anthropic control over 70% of all AI GPU compute capacity and revenue. They're not just customers—they're the entire economy.
Microsoft's $37 billion AI revenue run-rate? OpenAI contributes 71% of that. Amazon's $15 billion? Anthropic chips in 80%. These aren't diversified customer bases. They're dependency chains dressed up as growth stories.
"We're watching venture capital dress up as revenue, then call it infrastructure demand."
The Circular Money Machine
Let me walk you through the magic trick. Amazon and Google invest billions in Anthropic. Anthropic turns around and commits $200 billion to Google Cloud and TPUs over five years—more than 40% of Google's disclosed revenue backlog.
It's the corporate equivalent of paying yourself for a haircut and calling it GDP.
The Infrastructure Mirage
Microsoft's "Fairwater" data centers house hundreds of thousands of GPUs reserved exclusively for OpenAI. CoreWeave generated $5.15 billion in 2025 revenue—67% from Microsoft renting capacity for... OpenAI.
Meanwhile, Anthropic plans to use up to one million TPUs and over a gigawatt of capacity starting in 2026. That's not a cloud strategy. That's a special purpose vehicle financing arrangement with extra steps.
The honest headline? AI capex 2027 looks less like a technology investment and more like a $2 trillion confidence game where everyone's bluffing and nobody wants to fold first.
The Circular Financing Machine
Here's the magic trick nobody wants to talk about. Cloud providers sell compute to OpenAI and Anthropic. Those same AI labs turn around and become the dominant revenue source for those very cloud divisions. Revenue spins in circles. Everyone reports growth. Wall Street applauds.
But follow the actual dollars and something strange happens. The loop doesn't actually close.
Microsoft / Amazon / Google] -->|Sells GPU/TPU Compute| B[🤖 OpenAI / Anthropic] B -->|Pays for Cloud Credits| C[💰 Cloud Revenue] C -->|Reported as AI Revenue| A A -->|Invests in AI Startup| B B -->|Commits to Spend on Cloud| D[📈 Inflated Backlog] D --> A style A fill:#dbeafe,stroke:#2563eb,stroke-width:2px,color:#1e3a8a style B fill:#dcfce7,stroke:#16a34a,stroke-width:2px,color:#14532d style C fill:#fef3c7,stroke:#d97706,stroke-width:2px,color:#78350f style D fill:#f3e8ff,stroke:#9333ea,stroke-width:2px,color:#581c87
The circular financing AI playbook is elegant in its absurdity. Microsoft dumps billions into OpenAI. OpenAI turns around and rents Azure capacity—paying Microsoft back with Microsoft's own money. Amazon does the same with Anthropic. Google sells TPUs to Anthropic through special purpose vehicles, then books it as revenue.
Everyone wins on paper. The OpenAI Anthropic revenue figures look explosive. Cloud growth rates dazzle investors. But strip out the round trips and what's left?
"Microsoft's AI infrastructure generated $325 million in GPU rental revenue and $367 million from Microsoft 365 Copilot in Q3 2025—together accounting for less than half of OpenAI's inference spend on Azure."
Let that sink in. Microsoft's entire AI revenue from actual products doesn't even cover what OpenAI spends on Azure. The rest? A $37 billion revenue run-rate built on one customer cycling money through the system.
Amazon's math is even more brutal. Its $15 billion AI revenue run-rate represents just 0.419% of its $298 billion in AI capital expenditures. That's not a return. That's a rounding error wearing a tuxedo.
The really clever part? SPV financing. Cloud providers create special purpose vehicles to fund GPU and TPU capacity "sold" to Anthropic and Meta. Money moves between entities. Revenue gets recognized. Backlogs swell. But the actual cash? It's still inside the same ecosystem, chasing its own tail.
CoreWeave generated $5.15 billion in 2025 revenue. Sixty-seven percent came from Microsoft renting capacity—for OpenAI. Even the "independent" infrastructure players are just extra loops in the same circle.
OpenAI and Anthropic: The 70% Dependency
The entire AI GPU compute economy rests on two shoulders. And they're getting tired.
OpenAI and Anthropic now command more than 70% of all AI GPU compute capacity and revenue. That's not market leadership. That's market monoculture.
The Revenue Concentration Problem
Microsoft's AI revenue run-rate of $37 billion sounds impressive. Until you learn OpenAI contributes 71% of it.
Amazon's $15 billion AI revenue run-rate? Anthropic accounts for 80% of that check.
The Circular Financing Merry-Go-Round
Here's where it gets delightfully weird. Microsoft builds data centers for OpenAI. OpenAI pays Microsoft to use them. Microsoft reports AI revenue. Investors cheer.
OpenAI's inference spend on Azure hit $3.648 billion in Q3 2025 alone. Annualized, that's roughly $14.4 billion—and climbing toward an estimated $18.5 billion by year-end.
"The cloud providers aren't selling AI. They're leasing hope to their own biggest customers—and booking both sides of the transaction."
The Gigawatt Gorilla
Microsoft operated approximately 2 GW of AI compute capacity by end-2025. OpenAI consumed 1.9 GW of it. That's not a customer relationship. That's host-parasite dynamic with better PR.
Anthropic, meanwhile, has committed to spend $200 billion on Google Cloud and TPUs over five years. It comprises over 40% of Google's disclosed revenue backlog.
What Happens When the Music Stops?
The AI GPU compute market has built a $2 trillion infrastructure on the spending habits of two deeply unprofitable companies. That's not diversification. That's systemic exposure wearing a disguise.
CoreWeave generated $5.15 billion in 2025 revenue. 67% came from Microsoft renting capacity for OpenAI. Remove OpenAI's burn rate, and CoreWeave's IPO narrative evaporates faster than a crypto stablecoin.
OpenAI Anthropic revenue isn't just dominating the market. It is the market. And markets built on two foundations have a name in structural engineering: unstable.
The Revenue Mirage: 1% Returns on Massive Investment
Let me paint you a picture. Imagine walking into a casino, dropping $298 billion on the table, and walking out with... 0.419% back. Not a typo. Not a rounding error. That's Amazon's current AI math.
The AI revenue run rate numbers these cloud giants are touting? They're technically real. They're also technically hilarious.
Microsoft fares marginally better at 1.04%. Their $37 billion AI revenue run rate against $293.8 billion in capex is the sector's gold star. The gold star is made of tin foil.
"We're seeing 123% year-over-year growth in AI revenue!" — Satya Nadella, probably, while carefully omitting the starting point was measured in pocket lint.
The circular financing makes this even more delicious. OpenAI contributes 71% of Microsoft's AI revenue. Anthropic chips in 80% of Amazon's. These aren't customers. They're financial ouroboros.
OpenAI spent $3.648 billion on Azure inference in Q3 2025 alone. Microsoft's AI infrastructure revenue from GPU rentals? $325 million. From Copilot? $367 million. Together, that's less than half of what OpenAI paid them.
The AI profitability story these earnings calls spin requires Olympic-level mental gymnastics. Andy Jassy trumpets that $15 billion run rate. The footnote—that it's 0.419% of spend—gets whispered into the void.
Sundar Pichai claims Google's AI investments are "lighting up every part of the business." Specific revenue figures? Those would require a subpoena. Meanwhile, Anthropic committed to $200 billion in Google Cloud and TPU spending over five years—more than 40% of Google's disclosed revenue backlog.
The industry has constructed an elaborate infrastructure financing shell game. SPVs. Back-to-back leases. Capacity reservations masquerading as revenue. When CoreWeave generates 67% of its $5.15 billion revenue from Microsoft renting capacity for OpenAI, what exactly are we measuring?
Not profit. Not sustainability. Certainly not the AI profitability narrative retail investors are swallowing whole. We're measuring momentum. And momentum, as any physicist will tell you, stops eventually.
SPVs, TPUs, and Financial Engineering
Welcome to the shell game. The same one where Google sells TPUs to Anthropic, who then rents them back on Google Cloud, while both parties book revenue on the same silicon. In AI infrastructure financing, this isn't a bug—it's the entire feature set.
The mechanism is elegant in its absurdity. Google manufactures TPUs. Anthropic needs compute. Rather than a simple transaction, Google TPU Anthropic arrangements deploy special purpose vehicles (SPVs) to park the hardware, layer in financing, and create contractual structures where everyone wins on paper.
Until, of course, they don't.
The diagram above traces the circularity. Anthropic plans to use up to one million TPUs and over a gigawatt of capacity beginning in 2026. That's not a compute contract. That's a hostage situation with better legal representation.
"The AI industry is shifting from direct product sales to infrastructure financing and resale arrangements."
Translation: We ran out of actual customers, so we started selling to ourselves with extra steps.
The SPV financing structure serves multiple masters. For Google, it accelerates revenue recognition and pads backlog. For Anthropic, it secures compute without requiring immediate cash—critical when you're projecting an $11 billion loss in 2026 despite raising $58 billion in eight months.
The Google TPU Anthropic relationship exemplifies a broader pattern. Amazon and Microsoft have constructed parallel edifices. CoreWeave generated $5.15 billion in 2025 revenue, with 67% derived from Microsoft's rental of capacity for OpenAI. The cloud providers aren't selling to end users. They're selling to each other's investments.
Sundar Pichai claims Google's AI investments are "lighting up every part of the business." He declines to specify which parts, or by how much, or whether those lights aren't just reflections from the same burning cash.
The TPU itself becomes a financial instrument. Not a chip, but collateral. Not compute, but a line item. The AI infrastructure financing machinery grinds on, $2 trillion by end of 2027, with OpenAI and Anthropic consuming more than 70% of all AI GPU compute capacity.
When two companies rent the entire world's new infrastructure back and forth, you've built not an economy but a financing perpetual motion machine. One that hums beautifully until someone asks where the actual profits went.
Echoes of Dot-Com: Historical Parallels
The tech bubble 2025 narrative is impossible to ignore. But here's the thing—every bubble thinks it's different. Until it isn't.
Remember Pets.com? The sock puppet had better unit economics than some AI infrastructure plays today. At least the dog food shipped.
The circular financing is the real déjà vu. OpenAI consumes 71% of Microsoft's AI revenue while Microsoft rents GPUs back to OpenAI. Anthropic accounts for 80% of Amazon's AI revenue while committing $200 billion to Google Cloud. It's the digital equivalent of trading baseball cards with your own brother and calling it commerce.
"We've seen this movie before. The special effects are better, but the plot holes are the same."
Bernstein's Stacy Rasgon argues Nvidia remains "cheap" at 26.31x forward P/E—just 5.9% above industry average. But cheap relative to what? The dot-com playbook said Cisco was cheap too, until it wasn't.
The tech bubble 2025 faithful will say AI is "different this time." They said that about fiber optics. About radio. About railroads. Sometimes they were right—eventually. The question isn't whether AI transforms everything. It's whether these prices at this moment account for that future.
The tech bubble 2025 faithful will say AI is "different this time." They said that about fiber optics. About radio. About railroads. Sometimes they were right—eventually. The question isn't whether AI transforms everything. It's whether these prices at this moment account for that future.
Amazon's AI revenue run-rate of $15 billion represents 0.419% of its $298 billion AI capital expenditures. Not a typo. Zero point four one nine percent. Pets.com shareholders are laughing from the afterlife.
The Bull Case: Why Analysts Say "This Time Is Different"
You've heard it before. Every bubble whispers the same seductive lie.
Yet here's Bernstein's Stacy Rasgon—no permabear, he—declaring Nvidia stock cheap at forward P/E of 26.31. That's 5.9% below the industry average. For a company growing revenue at 73% YoY with 120.8% EPS growth.
The
But here's the critical caveat—the one even bulls whisper in private. That $15 billion AI revenue run-rate at Amazon? It's 0.419% of their $298 billion AI capex. Microsoft's 37 billion looks robust until you realize OpenAI accounts for 71% of it—until you realize OpenAI accounts for 71% of it—and OpenAI accounts for 71% of it—and Anthropic—together commanding 70%+ of AI GPU capacity—discover their own customers can't pay enterprise prices for commodity chat.
Next: The bear case. Or as we call it, reality's rude interruption.
The Breaking Point: When Capital Runs Dry
The AI bubble burst isn't coming. For the pessimists, it's already here—hiding in plain sight inside quarterly earnings reports that read like fever dreams. We're watching $2 trillion in AI capital expenditures chase revenue figures that wouldn't cover the catering budget at a proper dot-com blowout party.
Let's talk about Anthropic losses—the kind that make venture capitalists reach for the antacids. The company raised $58 billion in eight months and still projects an $11 billion loss for 2026. That's not a business model. That's a bonfire with a marketing team.
The circular financing dance is almost beautiful in its absurdity. Amazon sells GPUs to Anthropic, then rents them back as "cloud revenue." Google does the same with TPUs. Anthropic commits $200 billion to Google Cloud over five years—more than 40% of Google's disclosed revenue backlog—and somehow this counts as growth.
"The AI industry is shifting from direct product sales to infrastructure financing and resale arrangements—essentially moving air between balance sheets and calling it progress."
Microsoft's $37 billion AI revenue run-rate sounds impressive until you realize OpenAI contributes 71% of it. Amazon's $15 billion? Anthropic powers 80%. These aren't diversified customer bases. They're dependency chains masquerading as ecosystems.
The math gets brutal when you peek under the hood. Amazon's AI revenue represents just 0.419% of its $298 billion AI capex. Microsoft's ratio isn't much better at 1.04%. We're not talking about growing pains here. We're talking about a fundamental disconnect between what these companies are building and what anyone is actually paying for.
The SPV financing structures are where this goes from concerning to genuinely alarming. Special purpose vehicles now fund TPU and GPU capacity for Anthropic and Meta, creating layers of opaque leverage that would make 2008 blush. Google sells chips to Anthropic through a shell company, books it as revenue, and everyone pretends this is normal infrastructure spending.
Satya Nadella once projected a $10 billion annualized AI run-rate. Andy Jassy touted $15 billion. Sundar Pichai spoke of AI "lighting up every part of the business" without offering a single specific figure. These are the verbal equivalent of smoke machines at a concert—impressive atmosphere, zero substance.
"When your biggest customer's business model is 'raise more money than we burn,' you're not selling infrastructure. You're subsidizing a lottery ticket."
OpenAI's inference spend on Azure hit $3.648 billion in Q3 2025 alone—annualizing to roughly $18.5 billion by year-end. Microsoft's entire AI infrastructure, meanwhile, generated $325 million in GPU rentals and $367 million from Copilot. Combined, that's less than half of what OpenAI spent just on inference. The customer is consuming more than the vendor can produce. That's not a market. That's a subsidy with delusions of grandeur.
The breaking point arrives when venture capital finally asks the uncomfortable question: what happens when the next funding round doesn't materialize? Anthropic's burn rate suggests months, not years, of runway without fresh capital. OpenAI's conversion to a for-profit structure hints at the same desperation.
When that capital runs dry, the AI bubble burst won't be a pop. It'll be a chain reaction—cloud providers writing down billions in stranded assets, GPU orders evaporating, and the tidy circular financing that sustained this illusion unwinding faster than anyone prepared for.
Conclusion: Navigating the Uncertainty
The numbers don't lie. They just stare back at you with the uncomfortable intensity of a venture capitalist who's just realized their AI investment strategy might be recycling the same dollar bill through sixteen SPVs.
Here's the thing about $800 billion to $900 billion in projected 2026 AI capex. It sounds impressive until you notice that Microsoft's $37 billion AI revenue run-rate represents roughly 1.04% of its cumulative AI infrastructure spend. Amazon's doing marginally worse at 0.419%.
That's not a business model. That's a thermal management problem with quarterly reporting.
"The AI revenue growth is lighting up every part of the business." — Sundar Pichai, presumably while squinting at a spreadsheet that won't reconcile.
The AI bubble warning signs aren't subtle. They're $200 billion commitments to Google Cloud TPUs. They're Anthropic planning to consume one million TPUs and over a gigawatt of capacity by 2026 while projecting $11 billion in losses for the same year. They're CoreWeave generating $5.15 billion in revenue where 67% comes from Microsoft renting capacity back to itself for OpenAI.
Microsoft's "Fairwater" data centers—hundreds of thousands of GPUs reserved exclusively for OpenAI—aren't a partnership. They're a capital absorption mechanism wearing a press release.
And yet. Bernstein calls Nvidia "cheap" at 26.31 forward P/E. Forty-four of forty-nine analysts slap Strong Buy ratings while the stock trades at 69.26 RSI—technically overbought, perpetually justified.
The AI investment strategy that isn't asking hard questions right now isn't strategy. It's momentum with a thesaurus.
What Actually Matters Now
Watch the revenue-to-capex ratio. Not the press release version. The version where you subtract circular cloud deals and SPV financing from both numerator and denominator.
Watch who pays full price for inference versus who gets capacity bundled into multi-year "partnerships" that never quite reach GAAP profitability.
Watch whether $2 trillion in cumulative AI capex by 2027 produces anything beyond slightly better autocomplete and massively better financial engineering.
"The entire AI sector is dependent on [OpenAI and Anthropic's] ability to raise capital." — Which is a polite way of saying the music stops when the LP letters stop coming.
The AI bubble warning isn't that the technology fails. It's that the economics are still pretending. Pretending that $18.5 billion in annualized inference spend by one customer, paid to their primary investor, represents a sustainable market. Pretending that 123% year-over-year revenue growth is impressive when it requires 10,000% year-over-year infrastructure growth to achieve.
The navigation isn't complicated. It's just uncomfortable. Look for actual margin. Look for customers who aren't also your investors. Look for revenue that doesn't evaporate when you remove the circular reference. Everything else is theater. Expensive, thermally challenging theater, with very nice GPUs.
Disclaimer: This content was generated autonomously. Verify critical data points.
Post a Comment