Silence is the first vote in a true consensus.
I wrote that sentence in 2017, four months into a post-mortem of The DAO โ tracing a reentrancy attack through transaction logs until I had mapped fourteen logical flaws in a system that had promised to distribute power. The most dangerous flaw, I concluded, was not in the clever code. It was in the story the code told, and no one audited that story until it was too late.
I am auditing a different story this week. It arrives from a blockchain and Web3 news outlet, which is the first curiosity. It concerns Gavin Baker, founder of Atreides Management, described as "all-in" on AI infrastructure. Its core claim is that NVIDIA is trading at its lowest forward price-to-earnings ratio in a decade, and that the July 2025 selloff in AI equities was a deviation from fundamentals so severe that patient capital was handed a gift.
Clean narratives deserve suspicion. Not because they are always wrong. Because they are always incomplete.
The Foundation of the Claim
Gavin Baker is a growth investor who has held NVIDIA since approximately 2016, when the stock traded near thirty to forty dollars. Atreides Management manages roughly ten to fifteen billion dollars per recent 13F filings. His discipline is long-duration conviction: identify a multiyear expansion of total addressable market, then sit through the noise. He is a credible source of directional conviction.
NVIDIA's fundamentals are, on the surface, extraordinary. Data center revenue for fiscal 2025 reached roughly $115.2 billion, up 93 percent year over year, and now accounts for approximately 89 percent of total revenue. The company is mid-transition from Hopper to Blackwell. The GB200 NVL72 โ a rack of seventy-two Blackwell GPUs networked through NVLink 9 โ is engineered for clusters of one hundred thousand GPUs. Order visibility, per company guidance, extends through calendar 2026. Gross margins exceed 70 percent.
The transition is not only architectural; it is commercial. NVIDIA has moved from selling chips to selling data centers. The NVL72 is a system commitment, lifting single-customer order values into the billions and extending lock-in across networking, software, and service. This is the "AI factory" model, and it explains why the company's gross margins and backlog have simultaneously expanded. But it also raises the cost of disappointment: a customer that delays a rack order is not delaying a component โ it is delaying an entire facility.
The moat is real. CUDA carries more than fifteen years of developer mindshare; PyTorch, JAX, and the core of modern deep learning are built atop it. Google's TPU, Amazon's Trainium, AMD's MI300X, and inference specialists like Cerebras and Groq press at the edges, but NVIDIA still commands an estimated 80 to 95 percent of the training market.
The article's pivot arrives two-thirds through: after the July selloff, NVIDIA's forward P/E โ roughly 25 to 30 times consensus estimates โ is claimed to be at a decade low, and Baker has responded with an all-in commitment to the AI infrastructure complex. The story treats this as an opportunity the crowd has overlooked. The July selloff, it should be said, was not an isolated event but part of a broader repricing of AI-related equities; some of it was technical, some of it a genuine reevaluation of how quickly AI investments would generate returns. The article treats the distinction as obvious. It is not.
I want to examine what that claim requires to be true.
Part One: The Arithmetic of the Artifact
Begin with the P/E, because the entire narrative hinges on one number, and numbers can be charmed into lying.
A forward P/E is not a fact. It is a quotient of a present price and a future earnings estimate โ and when that estimate is revised upward at a compound rate above 50 percent, the quotient is pulled downward by arithmetic rather than by value. NVIDIA's earnings per share grew roughly 130 percent in fiscal 2025. Analysts, watching the Blackwell order book, have lifted forward projections repeatedly. A rising denominator mechanically compresses the ratio.
I call this a statistical artifact. It resembles what I found in The DAO: a consensus can look sound because someone understated the worst-case execution path. Here, the forward ratio is only as trustworthy as the estimate embedded in it, and that estimate rests on three assumptions.
The first is that Blackwell's ramp is frictionless. That is demanding. CoWoS advanced packaging โ the process of joining a GPU die to high-bandwidth memory โ has constrained every major launch in recent years. Industry capacity is projected to rise from about 45,000 to 50,000 wafers per month in 2024 to 65,000 to 80,000 by the end of 2025, yet the market needs roughly 20 percent more than the fabs can supply. HBM3e from SK Hynix, Samsung, and Micron is effectively committed through 2026; a single HBM stack can cost over $1,500, as much as half the bill of materials of a Blackwell board. These are not procurement details. They are the walls of the corridor through which this thesis must walk.
The second assumption is that NVIDIA's anchor customers remain in a spending mood. Microsoft, Amazon, Google, and Meta together represent over 40 percent of data center revenue. Their combined capital expenditure guidance for 2025 exceeds $300 billion, roughly 55 to 60 percent AI-related. That is extraordinary tailwind. It is also a concentration of counterparty risk rarely priced. When I redesigned a DAO's governance tokenomics in 2020 โ quadratic voting, twelve virtual town halls โ I learned that a system with four dominant stakers is not a democracy. It is a committee with quorum problems. NVIDIA is four committee members away from a motion failing.
The third assumption is the one the article never articulates: that inference demand takes over from training demand as the next growth engine by 2026 and 2027. Training clusters are being built now; inference is where recurring revenue lives. But inference is a different economy. It is more price-competitive, more distributed, more aggressively optimized โ speculative sampling, FP8 and FP4 quantization, KV-cache reuse, continuous batching. Every optimization that lowers the marginal cost of a token compresses the hardware pricing power that justifies the forecast. The market senses what the narrative does not: the efficiency gains that make AI more accessible also slow the rate at which GPU capacity must grow.
There is also a question of historical baseline, which the phrase "decade-low" conveniently skips. NVIDIA's forward P/E in the mid-2010s, during the gaming and crypto-mining cycles, traded in the 15-to-25-times range. The 2021 peak saw forward multiples near 60 to 80 times. The 2022 trough settled at 25 to 40 times. By that measure, today's roughly 25-to-30-times forward multiple is not anomalous. It is ordinary โ the statistical center of a volatile series, not its historical floor. "Decade-low" is a narrative construction, not a measured fact.
Even if the forward multiple were genuinely low by NVIDIA's own history, the question remains: low relative to what? At a $3 trillion market capitalization, a 30-times multiple embeds an expectation of roughly $100 billion in annual earnings โ and for the multiple to hold, those earnings must keep growing at 40 percent-plus for years. This is not impossible. It is simply not the same as "cheap." Cheap is what you buy when no one else believes; this is what you buy when everyone believes and the only point of disagreement is the timing of the refund.
Part Two: The Governance of the Narrative
The arithmetic is the surface. The deeper problem is how the story travels, and this is terrain I know well.
The 13F form behind the article's premise is a quarterly snapshot of institutional holdings, not an interview. It does not convey an investment thesis. It does not distinguish a conviction position from a temporary allocation. "All-in on AI infrastructure" is an interpretation layered over a filing โ stripped of qualifiers by the time it appears in the article. A whisper, amplified, becomes a signal.
I call this narrative amplification through regulatory artifact, and it is a governance failure of a familiar kind. When a Web3 media outlet translates a single investor's disclosed position into a market signal, authority flows from the newscaster's reputation, not from audited reality. I have watched a whale's wallet movement narrated into a coordinated signal, an anonymous team's roadmap narrated into a roadmap of certainty. The volume of the retelling substitutes for the quality of the telling.
My own ecosystem has a special obligation to notice this. We spent a decade arguing that information should be verifiable at its source. Here we are, consuming a second-hand interpretation of a third-hand filing, through a channel whose editorial standards remain untested. The same pattern appeared in 2024, when I sat across from Geneva institutional investors and watched blockchain technology get repackaged as a "settlement layer" for funds that had no intention of supporting the values underlying it. The lexicon of conviction survives the migration; the conviction itself does not.
The cultural signal in this article is, for me, the most telling content of all. When the Web3 world โ built to distribute power โ becomes the vehicle through which Wall Street's most concentrated technology story is celebrated as a bargain, that is not ecosystem maturation. It is capital rotation. The Bitcoin ETF story plays the same melody: peer-to-peer electronic cash became a Wall Street toy, and within a year the audacity of that transformation was absorbed. Now the GPU, the engine of the most centralized infrastructure ever built, is being arrayed for the audience that was supposed to build something different.
Part Three: The Physical Oracle Problem
There is a third layer, connected directly to my daily work.
I have spent years writing about DeFi's oracle problem โ the uncomfortable fact that the price feeds keeping decentralized finance honest are delivered by centralized nodes. It is a joke of a solution, tolerated because the decentralized alternatives are not ready. The AI infrastructure trade contains the same structural weakness, at a larger scale and a different layer.
The true bottleneck is not silicon. It is the electrons reaching the silicon. A 100,000-GPU cluster draws 80 to 120 megawatts. Grid interconnection queues in the United States run three to five years. The companies that determine whether Blackwell's order book is fulfilled are not semiconductor companies at all: they are utilities, cooling specialists, transformer suppliers, producers of backup generation. "AI infrastructure" has been used as if it means GPUs. It increasingly means power plants.
This is a coordination failure awaiting its price. NVIDIA can manufacture the NVL72 as a unit of compute, but it cannot manufacture the megawatts that bring it to life. In 2026, as part of a protocol design project for Tallinn's AI startup hub, I spent four months integrating zero-knowledge proofs into AI-agent wallets โ an effort to prove an agent's origin without exposing proprietary data. The engineering was elegant, but the binding constraint was never cryptographic. It was the same physics that constrains every data center: power, land, and the patience of the grid operator. The AI buildout will be measured in megawatts before it is measured in models.
The second act of the AI story, if it arrives, will look nothing like the first. Training has been a centralized phenomenon, concentrated in frontier labs with near-infinite capital. Inference, by contrast, is diffuse; it happens in browsers and cars and factory floors. The infrastructure required to serve diffuse intelligence is not the same as the infrastructure required to build it. If Baker's thesis is correct, the second act will reward distributed deployment more than central computation โ and the valuation argument for NVIDIA will need to survive a shift in the mix toward lower-margin, higher-volume workloads.
And when a cloud provider answers investor questions about return on capital with "short-term investment, long-term return," I hear the exact language of a DAO treasury that spoke that way three months before a governance crisis. Vague speech about future returns is the mask of an unmeasured current risk โ a lesson my time auditing DAOs taught me to take seriously.
Part Four: The Parallel Ledger
The article also ignores the envelope in which any thesis must operate: geopolitical segmentation.
NVIDIA's China revenue has fallen from roughly 17 percent of total in fiscal 2024 to perhaps 13 percent in fiscal 2025 under export controls. The AI diffusion rules under discussion in Washington would impose country-level quotas on advanced chips, splitting the global compute market into regions. Meanwhile, Huawei's Ascend line, supported by an explicit localization drive, is expected to raise domestic Chinese chip adoption from about 20 percent to 50 percent by 2027.
This matters for valuation. A "decade-low P/E" computed for a company serving the entire world has a different meaning than one for a company serving part of it. Two AI compute platforms are quietly being built โ one around CUDA, one around a domestic Chinese stack โ and bifurcation will eventually adjust the growth assumptions beneath the forward estimate. Markets price the full-addressable-world case until they do not.
There is also a competitive undercurrent the article waves away. The system-level lock-in represented by the NVL72 โ a customer buying a data center in a box, with switching fabric and software pre-integrated โ is the source of NVIDIA's pricing power. But it is also the source of customer resentment. Google, Amazon, and Microsoft are all investing heavily in custom silicon; Meta is not far behind. The Open Compute Project is pressuring the economics of proprietary interconnect. NVIDIA may hold 80 to 95 percent of training today, but the one-superpower model is evolving into a one-superpower-plus-challengers model, and that evolution is priced, if at all, as a slow bleed rather than a sudden break.
What the article also misses is the possibility that the real beneficiaries of the AI-infrastructure cycle are not chipmakers at all. The power utilities โ Constellation, Vistra, the regional grids with available interconnection capacity โ are becoming the rarest asset in the chain. In the same way that oracle latency is DeFi's hidden vulnerability, power latency is the hidden vulnerability of the AI factory. If I were asked to design a portfolio around the physical constraints of this buildout, I would start with the providers of electrons, not the fabricators of silicon.
The Counterweight
Discipline requires me to argue against myself.
I am not making a bear case for NVIDIA. The demand for compute is not fictitious. The bottlenecks I have described are the kind of physical constraints that historically favor the incumbent with the deepest supplier relationships, the most committed order book, and the most inertial software ecosystem. CUDA is not a moat because of its specifications; it is a moat because the world's frameworks, research papers, and brightest minds have internalized CUDA as the environment they inhabit. Switching costs are generational.
Baker's reported thesis โ first act cloud training, second act enterprise adoption, sovereign AI, and inference demand โ is more coherent than most narratives I audit. If his position is genuinely an infrastructure-complex position, incorporating Broadcom, Micron, TSMC, Vertiv, and the power producers, then "NVIDIA at a decade-low P/E" is a headline, not a thesis. The thesis is that the physical constraints of the buildout will command better prices across the entire chain, and the GPU is the most liquid proxy for that expectation.
The blind spot I confess is that my instinct to demand accountability from concentrated power can curdle into reflexive skepticism of concentrated success. The market's caution about AI ROI is partly rational โ it is the hedge of an industry burned by every hype cycle and unwilling to be caught again. Three hundred billion dollars of annual capital expenditure, concentrated in five institutions, with AI-attributable revenue still a fraction of the base โ that is not a fraud. It is a categorical wager. History will record it either as one of the most intelligent capital allocations ever made, or as one of the most expensive consensus errors. The probability is not 90/10. It is closer to 60/40 that this buildout is real. A 25-to-30-times forward multiple on a 60/40 coin is not a gift.
That is my disagreement with the article. "Low P/E" collapses vision and price into a single statement. A cheap multiple is not a conclusion; it is an invitation to inspect the assumptions beneath it โ packaging capacity, customer concentration, the inference demand curve, regional segmentation โ and to ask of the system the question any auditor asks: what does it look like under stress?
The Choice Before Us
I have spent a decade inside one of history's most ambitious experiments in distributed trust, and I have watched my own industry smuggle the old concentration back in, under the label of scale. Bitcoin became an ETF. Ethereum became a securities debate. Now the blockchain press celebrates a three-trillion-dollar chipmaker as a bargain. None of this means NVIDIA is a bad business. It means the ethical architecture we built to inspect concentrated power is being replaced by the architecture of convenience.
The question for 2026 is not whether NVIDIA is cheap. It is whether the infrastructure being built around it can be governed by anyone other than its builders. We designed on-chain governance for treasuries measured in billions. We are not applying those lessons to hundreds of billions of dollars in annual expenditure, allocated by five institutions, shielded behind forward guidance and phrases like "long-term return."
Silence is the first vote in a true consensus. The consensus around AI infrastructure is loud. The silence at its center is the absence of an audit trail. A system that cannot be audited is a system that does not want to be trusted.
The ledger of silicon will be balanced eventually, either by transparency or by consequence. We still get to choose which arrives first. That choice does not have to wait for the next bear market. It can be made in the next quarterly report โ if we are sober enough to read it that way.
