The Claim That Does Not Compile
The July 2025 AI selloff produced a familiar incantation, repeated across trading terminals, crypto newsletters, and Web3 aggregators: "NVIDIA is trading at its lowest forward P/E in a decade." The authority cited is Gavin Baker, founder of Atreides Management, a technology investor who has publicly declared an all-in position on AI infrastructure. The claim spreads like a smart contract with a reentrancy bug โ unchecked, unverified, and passed along because the narrative compiles.
It does not compile.
The code reveals what the pitch deck conceals. A forward P/E is not a fact. It is a quotient of two variables: a market price and a consensus guess about future earnings. When sell-side analysts revise earnings estimates upward by more than 130% in a single fiscal year โ as they did for NVIDIA in fiscal 2025 โ the forward P/E compresses mechanically. It is arithmetic, not signal. Calling that ratio "cheap" is structurally identical to calling a DeFi protocol "yield-bearing" because its emissions schedule advertises 40% APY. The yield is real. The source is a token printer. The claim fails under inspection, and reproducibility is the highest form of respect.
I have spent most of a decade auditing crypto protocols. The AI infrastructure narrative carries the same signature patterns I flagged in yield farms, algorithmic stablecoins, and NFT projects: unverified assumptions, maturity mismatches, and borrowed authority. The ticker is different. The architecture of the story is identical.
Context: The Substance Behind the Signal
Gavin Baker is not a retail influencer. He ran technology investing at Fidelity for more than a decade and later founded Atreides Management, a fund reportedly managing between $10 billion and $15 billion. He has held NVIDIA since 2016, through the crypto mining boom, the 2022 drawdown, and the current AI demand cycle. His public posture is not a meme. He is a long-duration growth investor, and his "all-in on AI infrastructure" framing extends beyond a single chip company to power generation, networking equipment, memory, and data center physical infrastructure.
The timing is deliberate. NVIDIA is mid-transition from the Hopper architecture to Blackwell. The GB200 NVL72 โ a rack-scale system interconnecting 72 Blackwell GPUs through NVLink โ entered volume shipment in the second half of 2025. This is not a product refresh. This is a business model mutation. NVIDIA no longer sells chips; it sells data centers. Individual customer orders now reach multiple billions of dollars. The data center segment produced approximately $115 billion in revenue in fiscal 2025, roughly 89% of total corporate revenue. Gross margins sit above 70%.
The macro environment is equally specific. Hyperscale cloud capital expenditures are projected to exceed $300 billion in 2025, up more than 35% year over year, with an estimated 55-60% of that sum attached to AI workloads. NVIDIA's data center revenue represents more than 30% of global cloud capital spending. This is a concentrated bet on a concentrated buyer base, routed through a concentrated supply chain. Smart contracts do not care about your narrative. Neither does the silicon supply chain.
Core: The Systematic Teardown
Finding One โ The forward P/E is a statistical artifact.
NVIDIA's earnings per share grew more than 130% in fiscal 2025. When the denominator of a valuation ratio grows at a triple-digit rate, the ratio appears compressed by definition. The "decade-low forward P/E" claim depends entirely on which earnings estimate is used and when the measurement is taken. Historical data shows NVIDIA traded at forward multiples of 15-25x in 2015-2016 โ a lower band than the approximately 25-30x forward multiple observed after the July 2025 correction. The "decade low" framing is a function of an aggressive analyst revision cycle, not a structural repricing of risk.
The uncomfortable math: at 30x forward earnings, a roughly $3 trillion market capitalization implies approximately $100 billion in expected annual earnings. To sustain that multiple, the market must believe EPS continues compounding at 40-50% annually for three to five years. If growth decelerates to 20%, the multiple compresses necessarily and permanently. A forward P/E at this scale is not an observation. It is a leveraged vote on unverified output.
This is precisely the analytical error I catalogued during DeFi summer. When I reverse-engineered Compound's governance and interest rate model in 2020, the TVL narrative was real and the mechanism was elegant. But the model carried a theoretical edge case: extreme volatility could destabilize the oracle feed. The core team initially ignored the finding. The 2022 correction validated it. The lesson translates directly: when the yield source is opaque, the displayed APY is a liability. NVIDIA's forward P/E assumes AI infrastructure return on investment materializes inside cloud providers' income statements within four to eight quarters. That assumption cannot be verified from the price chart.
Finding Two โ Customer concentration reads like an unaudited contract.
Four direct customers โ Microsoft, Amazon, Google, and Meta โ contribute more than 40% of NVIDIA's data center revenue. This is not a diversified industrial franchise. It is a smart contract with four whitelisted addresses controlling a majority of the treasury. A single cloud provider deferring procurement by one quarter creates visible volatility in NVIDIA's reported results.
The deeper risk is not that the four buyers collapse. It is that their incentives shift. All four are developing custom silicon: Google's TPU line, Amazon's Trainium and Inferentia, Microsoft's Maia, and Meta's MTIA. None currently threatens NVIDIA's dominance in general-purpose training. But each represents an exit option, and an exit option in the hands of your largest customers is structurally a call option against your gross margin.
I encountered the same dynamic during the 2021 NFT mania. I audited a high-profile PFP project whose contract inherited vulnerabilities from an outdated OpenZeppelin library. The project raised millions on aesthetic narrative; the code was the liability. The market eventually priced the difference between the image and the implementation. NVIDIA's moat โ CUDA's fifteen-year developer ecosystem and the NVL72 system-level lock-in โ is real. But customer-owned silicon is a slow, structural erosion that the "cheap P/E" narrative does not price, because it does not appear on a single quarter's income statement. It appears only in the trend lines.
Finding Three โ The supply chain is the real smart contract, and it has a reentrancy problem.
Advanced packaging capacity โ TSMC's CoWoS โ is the binding constraint. Global monthly CoWoS capacity is expected to rise from roughly 45,000 to 50,000 wafers per month in 2024 to approximately 65,000 to 80,000 in 2025, with NVIDIA, AMD, and Google absorbing most of the output. HBM3e supply from SK Hynix, Samsung, and Micron is largely pre-sold through 2026, with HBM accounting for 40-50% of a B200's bill of materials. If either link slips, Blackwell shipments miss, consensus earnings miss, and the "low P/E" quietly becomes the "value trap P/E."
A bug in the contract is a feature in the exploit. In this case, the bug is confirmation bias around order books. Orders are not immutable transactions. They are reschedulable promises, subject to the customer's own revenue reality. When a cloud provider's AI monetization lags its capex commitment, the provider delays the next tranche. The market treats this as a negative surprise, even though the mechanism was visible in advance. Nothing in the forward P/E calculation includes this optionality, because the option is not priced until it is exercised.
Then there is power. A 100,000-GPU cluster demands 80 to 120 megawatts. Cloud providers face grid interconnection waits of three to five years in parts of the United States and Europe. Data center power density is migrating from 5-10 kW per rack toward 50-100 kW. Liquid cooling is no longer optional. These physical constraints shift pricing power away from chip architects toward energy, thermal, and grid infrastructure providers. If the thesis is genuinely "all-in on AI infrastructure," the rational expression is a basket, not a single ticker. NVIDIA is the most visible asset in the trade. It is not the most resilient one.
Finding Four โ The maturity mismatch is hidden inside the infrastructure trade.
This is where stablecoin auditing becomes directly relevant. Ethena's sUSDe โ and every yield-bearing stablecoin before it โ operates on stacked assumptions: funding rates stay positive, basis positions hold, and bull markets persist. The product functions without disruption in expansion. It unwinds violently when the market turns. The maturity mismatch between the stablecoin's liabilities and its yield source is the structural flaw.
AI infrastructure exhibits the same geometry. Cloud providers are signing five-year leases and multi-billion-dollar GPU purchase commitments on the cost side. On the revenue side, AI inference monetization and application-layer adoption are unproven. The capex is contractual. The off-take is not. When a hyperscaler reports an AI revenue growth rate that diverges from its capex growth slope, the market will reprice the entire chain. The temporary absence of evidence is not evidence of absence. But in a market that has already assigned $3 trillion to the top of the chain, the burden of proof sits with the revenue line, not with the price chart.
Finding Five โ The "smart money" signal has a sample size problem.
Atreides Management's $10-15 billion under management is roughly 0.3-0.5% of NVIDIA's market capitalization. A fund manager's conviction โ however genuine, however well-earned through a decade of correct positioning โ cannot move a market of this size. The "smart money is buying" framing is an appeal to authority, and the data does not support the inference. SEC 13F filings are lagging indicators. By the time the public observes the position, the position is history.
I am not asserting that Baker is wrong. The code of this cycle is not yet written. I am asserting that the claim "NVIDIA is at a ten-year low forward P/E" fails reproducibility. A decade of trading history includes materially lower forward multiples at materially lower earnings bases. The current ratio is a mathematical consequence of the fastest earnings ramp in the history of semiconductors. The narrative discards that mechanical truth because the mechanical truth does not support a compelling headline.
Contrarian: What the Bulls Got Right
A competent audit documents what a project does right. The bulls have three defensible points.
First, the CUDA moat is genuine. Fifteen years of developer ecosystem accumulation creates a barrier that does not appear on a spec sheet. PyTorch, JAX, and the deep learning toolchain compile through CUDA. Hardware competitors can match the silicon; they cannot match the migration cost. That switching cost is a tax that protects the 70%+ gross margin.
Second, the system-level pivot is a lock-in mechanism. The NVL72 rack, the NVLink fabric, the networking stack, and the software platform transform the procurement unit from a chip into a data center. Abandoning NVIDIA after committing to rack-scale deployment means re-plumbing the entire infrastructure stack. Customers will not make that decision inside a single product cycle.
Third, the reasoning demand curve is the second S-curve. Training was the first wave; inference is the second. If inference workloads inflect nonlinearly โ and available usage data from major AI application providers suggests they are โ the demand horizon extends through 2027 and beyond. The bear case systematically undervalues the installed base currently moving AI from experimentation into production.
We audited the soul, and it was hollow. The shell, however, is extremely well constructed.
Takeaway: The Verification Queue
The "decade-low P/E" thesis contains one sentence of value buried inside it: the July selloff was partially emotional. Everything else is a statistical artifact that will not survive contact with a quarterly earnings call.
The verification points are knowable. NVIDIA's next earnings guidance. Cloud provider capital expenditure revisions. The slope of AI revenue growth relative to the slope of AI capex growth. BIS export rule updates. TSMC's CoWoS capacity expansion schedule. These are the transaction inputs. Run the audit each quarter and hold the narrative to the same standard you would hold a yield claim. Position size is a statement of faith. Verification cadence is a statement of discipline.
Logic is the only currency that never inflates. Spend it on the numbers that can be verified, not the narratives that cannot. When EPS growth decelerates toward 20%, the forward P/E will not look cheap. It will look like what it always was: a leveraged contract on unverified output. Know your position. Audit the assumptions. The market will deliver its verdict on its own schedule, and the verdict will not care about your cost basis.