The market is staring at the wrong number.
Over the past seven days, Nvidia's equity has been trading like a mature utility, not the monopoly supplier of the AI gold rush. The stock sits at roughly 15x forward EV/EBITDA. That is a 45% discount to its own five-year average of 27x. It is a 50% discount to AMD's 32x multiple. The market is pricing Nvidia as if it is a cyclical hardware vendor with a peak quarter behind it.
The market is wrong.
This is not a call on the price of silicon. This is a call on the structure of the balance sheet. Behind the clean income statement, Nvidia has accumulated roughly $150 to $200 billion in off-balance-sheet purchase commitments. The bears call this a liability. I call it a moat. And I have spent the last four years watching liquidity flow through this exact kind of arrangement.
Charts lie. Liquidity speaks.
Context: The Architecture of the Commitment
Let me break down what these commitments actually are.
First, there is the OpenAI deal. Nvidia has committed to invest approximately $100 billion in compute infrastructure, specifically targeting 10 gigawatts of AI compute capacity by 2030. This is not a donation. This is a "compute-for-equity" arrangement. Nvidia is converting its chip supply into an ownership stake in the most important AI frontier lab on the planet.
Second, there are the supply agreements with Taiwan Semiconductor Manufacturing Company (TSMC). Nvidia has locked in CoWoS advanced packaging capacity and 3nm wafer starts through 2026 and 2027. This is not a simple purchase order. This is a pre-payment for the physical bottleneck of the entire AI supply chain. TSMC's CoWoS capacity is currently running above 95% utilization. Nvidia consumes roughly 60% of that output.
Third, there are the HBM4 memory agreements with SK Hynix and Samsung. High Bandwidth Memory is the second bottleneck. Nvidia has pre-paid to secure supply for the Vera Rubin platform, slated for 2026.
These commitments are structured as operating leases or purchase obligations. They do not show up on the balance sheet as debt. But they are as real as debt. They are fixed costs. They must be paid regardless of whether the AI demand cycle turns.
This is the core of the bear thesis. If AI capital expenditure peaks in 2026, Nvidia is left holding the bag on billions of dollars of unused wafer starts and packaging capacity. The bank's own analysis suggests a worst-case scenario of $50 billion in stranded costs. That is roughly 10% of the company's current enterprise value.
The market is discounting this scenario heavily. That is why the multiple is at 15x.
But the market is missing the other side of the ledger.
Core: The On-Chain Truth of the AI Order Flow
The most important data point in this entire story is not Nvidia's P&L. It is the capital expenditure guidance from the four largest cloud service providers: Microsoft, Amazon, Google, and Meta.
These four companies collectively account for roughly 40-50% of Nvidia's revenue. Their AI CapEx-to-revenue ratio currently sits between 15% and 20%. This is the tell. It is not at a cyclical peak. It is in the early innings of a structural build-out.
Let me put this in historical context. In the 1990s, telecom companies spent heavily on fiber optic infrastructure. That build-out took over a decade. It created massive overcapacity in the short term, but it also created the physical layer for the internet economy. We are in the equivalent phase for AI compute. The CapEx cycle for AI infrastructure is projected to last 5 to 7 years, through roughly 2030.
Nvidia is not selling a product. It is selling the pickaxes for a gold rush that is still expanding.
Here is the second data point that the market is misreading: the inference market. The bear case relies heavily on the assumption that CSP self-designed chips (Google TPU, AWS Trainium, Microsoft Maia) will erode Nvidia's market share in inference.
This is true. It is happening. But it is happening at the low end of the market.
The high-end inference workload—the kind that requires the lowest latency and the highest throughput for frontier models like GPT-5 or Claude 4—still runs on Nvidia. The CUDA ecosystem is not a feature. It is a gravitational force. There are over 4 million developers writing code for CUDA. They are not migrating to a proprietary TPU compiler just because a cloud provider wants to save 20% on unit cost.
The switching cost is not the chip. It is the software stack. And that stack is 15 years deep.
Let me give you a personal example. In my previous role, I led a team that developed a mean-reversion strategy for Layer 2 tokens. We ran our backtesting on AWS. We tried to optimize the compute cost by using AWS Trainium instances. The hardware was fine. The performance was acceptable. But we spent three weeks rewriting our PyTorch code to be compatible with the custom compiler. We lost a month of research time. We went back to Nvidia A100s. The cost of the hardware was irrelevant. The cost of the engineering time was everything.
This is the hidden dynamic that the market does not price. The unit economics of the chip matter less than the total cost of the ecosystem. Nvidia understands this. They are not selling a chip. They are selling time-to-market.
The Contrarian Angle: The Off-Balance-Sheet Moat
The market views the $150 to $200 billion in commitments as a risk. I view it as the single most underappreciated competitive barrier in the semiconductor industry.
Here is the logic.
AMD cannot get CoWoS capacity. They are competing for the same TSMC packaging lines. But Nvidia has already pre-paid for the next two years of output. This means AMD's MI400 series, despite being a competitive chip on paper, will be supply-constrained at launch. You cannot sell what you cannot build.
The same dynamic applies to HBM4. SK Hynix and Samsung are ramping production. But Nvidia has locked in the first wave of supply. Every other AI chip designer is fighting for the scraps.
This is the essence of what I call the "liquidity speak" principle. The commitment is not just a promise to buy. It is a promise to buy that prevents competitors from buying. It is an offensive weapon disguised as a financial liability.
The second contrarian angle is the business model transformation. The OpenAI deal is not just a supply agreement. It is a signal of vertical integration. Nvidia is moving from being a chip vendor to being an AI infrastructure operator.
This is the same playbook that Amazon used with AWS. Amazon realized that its internal infrastructure could be monetized as a service. Nvidia is doing the same thing with compute. The DGX Cloud offering, combined with the OpenAI commitment, positions Nvidia to capture the operating margin of the AI cloud, not just the hardware margin.
If the market starts to value Nvidia as an infrastructure operator—at 25 to 30x EV/EBITDA—rather than a hardware vendor at 15x, the re-rating potential is significant. We are talking about a 60-100% upside to the current valuation.
Now, I need to be honest about the risks. I am not a perma-bull. The risk of an AI CapEx cycle correction is real. If the CSPs see a slowdown in AI monetization, they will cut CapEx. And Nvidia's off-balance-sheet commitments will become a drag. That is the nature of leverage. It cuts both ways.
The probability of a cyclical downturn in 2026-2027 is roughly 30-40%. That is not a negligible risk. But the market is pricing this risk as if it is a certainty.
FOMO is a tax on the unobservant. But so is excessive pessimism.
Takeaway: The Signals That Matter
Forget the price target. Forget the quarterly revenue beat. Focus on the following signals.
First, watch the TSMC monthly revenue reports. If CoWoS capacity expands from 45,000 wafers per month to 60,000 wafers per month by the end of 2025, as planned, the supply constraint eases. This is a bullish signal for Nvidia's revenue trajectory.
Second, watch the CSP CapEx guidance. Microsoft, Amazon, Google, and Meta all report quarterly. If they maintain or increase their AI CapEx intensity, the demand cycle is intact.
Third, watch for the Vera Rubin launch in 2026. If the platform hits its performance targets on TSMC's 3nm process, the technological moat extends for another two years.
The market is a voting machine in the short term and a weighing machine in the long term. Right now, it is voting for a cyclical decline. The weight of the evidence suggests a structural expansion.
Don't marry the bag, respect the chart. But also respect the order flow. The money is still moving. The commitments are still being signed. The liquidity is still speaking.
The question is not whether Nvidia is a good company. It is whether the market is willing to see the balance sheet for what it really is: a war chest of locked-up supply that your competitors cannot touch.
I have seen this pattern before. In 2020, during DeFi Summer, the teams that locked up liquidity early won the market. The ones that waited for certainty were left holding the bag. The same logic applies to silicon. The supply is the alpha. And Nvidia has bought it all.
Trust the data, ignore the discord. The data says the moat is widening.
The next 18 months will tell us if the market is willing to re-rate this reality. I am watching the order flow. You should too.