Its trailing P/E sits around 34x. That is not a single stock problem. That is five of the loudest names in the AI infrastructure trade flashing the same warning at the same time.
Here is what matters: Nvidia is not just a chipmaker anymore. It is the public market's most liquid bet on whether AI infrastructure spending keeps compounding. When its valuation stretches, the stretch does not stay contained. It bleeds into networking, memory, power, cooling, and every data center REIT priced on the assumption that hyperscaler capex only goes up. The question is no longer whether AI demand is real. The question is whether the returns on that demand will arrive fast enough, and concentrate narrowly enough, to justify what investors are already paying.
I think we are watching the early stages of a repricing. Not of AI itself, but of who profits from building it.
The Capex Gravity Thesis
Here is the mental model. Call it the Capex Gravity Thesis.
Four pressure points inside Nvidia's infrastructure premium.
Every infrastructure boom follows the same arc. Capital floods in. Suppliers capture enormous margins. Then gravity kicks in. Customers find alternatives, margins compress, and the profits redistribute across a wider set of players. Railroads did it. Fiber optics did it. Cloud computing did it between 2015 and 2020.
The thesis is simple: the heavier the capex concentration in a single supplier, the stronger the gravitational pull toward diversification. Nvidia's gross margins expanded from 60.5% to 74.9% in a single year, according to its most recent quarterly filings. That is extraordinary. It is also a signal. Margins that high create a $50 billion annual incentive for every major customer to find a cheaper path.
Capex Gravity says the current moment is the peak of concentration, not the beginning of a permanent dynasty. The money is real. The demand is real. But the distribution of returns is about to shift.
The Fragile Chain Behind a $4.5 Trillion Valuation
The entire AI infrastructure trade rests on a chain with three links. Hyperscaler capex rises. Nvidia captures a dominant share of that spend. Adjacent infrastructure winners compound alongside it. Break any single link and the chain fails.
Start with the first link. Jensen Huang has cited "strong visibility of $1 trillion plus" in demand for Blackwell and Rubin platforms through end of 2027. That number is powerful. It is also a forward management statement, not realized revenue. Visibility is not the same as a purchase order. It is not the same as a durable profit pool. The gap between those concepts is where valuation risk lives.
Now look at the second link: Nvidia's share. Two unnamed customers account for 39% of Nvidia's revenue, according to analysis published on Investing.com in late 2025. Those same customers, almost certainly among Microsoft, Amazon, Google, and Meta, are building custom silicon. Amazon has Trainium. Google has TPUs. Meta is investing in internal accelerators. Every dollar they shift to in-house chips is a dollar that leaves Nvidia's revenue line.
The third link is the most fragile. Adjacent winners like AMD, Marvell, and Broadcom are priced as if Nvidia's dominance guarantees a rising tide for the whole sector. But if Nvidia's share compresses, the tide does not rise evenly. It recedes. AMD's MI300 family is gaining traction. Broadcom is building custom ASICs for the same hyperscalers that Nvidia depends on. The "picks and shovels" trade works only when one company owns the mine. Once the mine gets shared, the math changes for everyone.
Consider the valuation models circulating on Wall Street. One widely cited framework from TIKR assumes 30% revenue CAGR, 63% operating margins, and a 21x exit P/E to derive a fair value near $290 per share. If any single one of those assumptions slips by even a few percentage points, the implied upside evaporates. A 25% CAGR instead of 30%. A 58% operating margin instead of 63%. A 17x exit multiple instead of 21x. Each alone could cut the target by 20% or more.
The prediction markets reflect this uncertainty. On Polymarket, bettors give Nvidia only about a 45% chance of remaining the world's most valuable company by end of 2026. Alphabet sits at 29%. Apple at 16%. The crowd is not convinced this is permanent.
OpenAI taking a 10% stake in AMD while Nvidia invests $100 billion in OpenAI. Microsoft accounting for nearly 20% of Nvidia's annualized revenue while also being a major customer of CoreWeave, in which Nvidia holds a significant equity stake. The lines between revenue and equity are blurring among a small group of companies that keep showing up on both sides of every transaction. That is not what a healthy, competitive market looks like. That is a system optimizing for capex velocity rather than capital efficiency.
Whether this web of cross-ownership creates systemic risk or just looks messy is genuinely unclear. But the pattern rhymes with late-cycle behavior in previous infrastructure booms. When the builders start financing each other, the question shifts from "Is demand real?" to "Who is actually the end customer paying with real revenue?"
The data center buildout is also increasingly constrained by physics, not just chips. Power, land, water for cooling, and networking bandwidth all bind before GPU supply does. Once those bottlenecks hit, incremental GPU demand grows more slowly than current 85% year-over-year revenue growth implies. Nvidia's own segmentation tells the story. Its edge computing segment, covering PCs, workstations, consoles, robotics, and automotive, now represents less than 8% of total revenue. The company is almost entirely a data center play. That magnifies cyclicality. If AI data center spend normalizes, there is no diversified base to cushion the fall.
My read on this: the market is pricing Nvidia as an infrastructure platform with utility-like durability. But its customer concentration, margin profile, and competitive dynamics look more like a cyclical supplier at peak pricing power. Those two frameworks produce very different valuations. The gap between them is the reckoning.
2031
Three signals inside the same shift
Nvidia's gross margins are creating their own gravitational pull toward alternatives.
Gross margins expanded from 60.5% to 74.9% in a single year. That level of extraction creates a $50 billion annual incentive for hyperscalers to accelerate custom silicon programs like Amazon's Trainium and Google's TPUs. Every quarter at these margins funds the competition.
Two customers control 39% of revenue while the builders finance each other.
OpenAI takes a 10% stake in AMD while Nvidia invests $100 billion in OpenAI. Microsoft accounts for nearly 20% of Nvidia's annualized revenue while also being a major CoreWeave customer. When builders start financing each other, the question shifts from demand reality to who is the actual end customer paying with real revenue.
Microsoft 365's agentic layer signals value is migrating from infrastructure to apps.
Microsoft 365 is deploying a new agentic layer that breaks complex requests into delegated tasks, enabled through the admin center. This is the application layer monetizing AI workloads directly. By 2031, the compute market will likely stratify into frontier training, large-scale inference, and edge inference, with GPU suppliers capturing a shrinking share of total value.
Pull back six years. The question that matters is not whether Nvidia stays dominant in 2026. It is whether the AI infrastructure stack in 2031 looks anything like today's.
History offers a pattern. In 2005, Cisco was the undisputed infrastructure winner of the internet buildout. Its routers and switches powered every data center. By 2011, its market cap had stagnated while cloud-native companies captured the real value. The infrastructure was essential. The infrastructure vendor's premium was not.
Nvidia's CUDA ecosystem is its deepest moat. Over 4 million developers build on it. Switching costs are real and measured in years, not months. But switching costs create switching incentives. Every quarter that hyperscalers pay 75% gross margins to Nvidia is another quarter they invest more aggressively in alternatives. The asymmetric bet is not on whether CUDA lock-in persists. It is on whether the rate of defection exceeds the rate of new demand.
By 2031, the AI compute market will likely have stratified into three tiers. Frontier training, where Nvidia probably retains significant share because performance matters most. Large-scale inference, where custom silicon from cloud providers captures an increasing portion because cost per token matters most. And edge inference, where ARM-based chips, specialized accelerators, and mobile processors dominate because power efficiency matters most.
The compounding question is not "Will AI demand grow?" It almost certainly will. The global AI market is projected to reach $1 to $2 trillion annually by 2030 across software, services, and hardware. The compounding question is "Will returns on AI infrastructure concentrate in GPU suppliers, or redistribute to application-layer companies that actually monetize the workloads?"
If you believe returns redistribute, then the 2025 to 2026 period is the peak of the infrastructure premium, not the midpoint. The flywheel still spins, but the value accrues to a different set of players. That is not a bearish view on AI. It is a bearish view on paying 34x earnings for a cyclical supplier at peak margins during peak capex concentration.
There is a contrarian case worth taking seriously. If Nvidia successfully transitions from chip vendor to full-stack AI platform, with software revenue, networking, and systems-level integration creating recurring revenue streams, then the utility framework might be correct. The company has the talent and the installed base to attempt it. But attempting it and achieving it are different things. The 70% rule for decision velocity applies here: you make the call with 70% of the information and adjust. Right now, 70% of the evidence points toward margin compression and share redistribution by the end of the decade.
What to Build This Weekend
You do not need to predict Nvidia's stock price to act on the Capex Gravity Thesis. You need to stress-test your own exposure.
Step one: open your brokerage account and calculate your effective Nvidia concentration. If you hold an S&P 500 index fund, Nvidia represented 7.99% of the index as of late 2025 and 12.18% of the Nasdaq. If you also hold individual semiconductor or AI ETF positions, your true exposure might be 15% or higher. Write that number down.
Step two: use a free tool like Microsoft Copilot Cowork, which just launched as a background agent inside Microsoft 365, to build a simple scenario model. Ask it to create a spreadsheet comparing three cases for your portfolio: Nvidia grows at 30% CAGR through 2030, Nvidia grows at 15% CAGR as margins compress, and Nvidia revenue flattens by 2028 as custom silicon captures inference workloads. You do not need a finance degree. You need three columns and honest assumptions.
Step three: map the AI infrastructure names in your portfolio against the three-tier framework. Which ones depend on frontier training demand? Which ones benefit from inference scaling? Which ones have exposure to edge AI? The Oracle-to-PostgreSQL AI migration tool that Microsoft just shipped inside VS Code is a small example of where value is migrating: away from proprietary lock-in, toward open ecosystems. Look for that pattern across your holdings.
Step four: set a calendar reminder for Q1 2027. That is when Nvidia's $1 trillion demand visibility window closes. If revenue guidance for the following quarter shows deceleration, the Capex Gravity Thesis is playing out. If it accelerates, update your priors. Either way, you will have a framework for interpreting the data instead of reacting to headlines.
The reps matter more than the prediction. Build the model. Test the assumptions. Break things on a spreadsheet before the market breaks them in your portfolio.
Stress-test your AI infrastructure exposure in three moves.
- Map your concentration risk. List every holding in your portfolio that depends on hyperscaler capex growth. Include Nvidia, AMD, Marvell, Broadcom, and data center REITs. Calculate what percentage of your total equity exposure traces back to a single assumption: that AI infrastructure spend only goes up.
- Model the margin compression scenario. Take the TIKR framework (30% CAGR, 63% operating margins, 21x exit P/E) and run three downcases: 25% CAGR, 58% margins, 17x exit. Each alone cuts the target by 20% or more. If your thesis requires all three assumptions to hold simultaneously, you are underpricing risk.
- Explore the Microsoft 365 agentic layer. Enable it through your Microsoft 365 admin center to start delegating complex requests. Understanding how the application layer monetizes AI workloads will sharpen your view on where long-term value actually accrues, infrastructure or software.
The demand is real. The distribution of returns is about to shift.
Nvidia at 34x trailing earnings with 74.9% gross margins is not a platform with utility-like durability. It is a cyclical supplier at peak pricing power during peak capex concentration. History says the infrastructure is essential but the infrastructure vendor's premium is not permanent. The Capex Gravity Thesis does not require AI to fail. It only requires returns to redistribute from GPU suppliers to the application-layer companies that actually monetize the workloads. Seventy percent of the evidence points toward margin compression and share redistribution by the end of the decade.