Hype fades; structure remains.
In late July 2025, Gavin Baker — founder of Atreides Management, one of the most vocal growth investors in this market cycle — was cited across financial media carrying a claim that instantly became a headline: NVIDIA Corporation is trading at its lowest forward price-to-earnings multiple in a decade. The associated position was even more marketable. Baker said he is “all-in” on AI infrastructure.
The story assembled itself with almost mechanical efficiency. A legendary technology investor. A market cap above three trillion dollars. A brutal summer selloff in which AI-related equities drew down between fifteen and forty percent from their peaks. The setup was the classic “blood in the streets” moment that growth narratives love to celebrate.
There is one problem. The math does not support the claim.
When a company grows earnings per share at 130% year-over-year, its forward P/E will always appear compressed. This is not a market insight. It is arithmetic. A forward multiple is a quotient: price divided by expected future earnings. When the denominator rises at a triple-digit rate, the quotient looks low regardless of the absolute value assigned to the numerator. Every holder of a hypergrowth stock is, by construction, staring at a forward P/E that appears more reasonable than the price history alone would justify.
This is not a valuation signal. It is a statistical artifact.
I have spent two decades testing this exact class of claim. In 2017, as a senior data analyst in Ho Chi Minh City, I manually audited 45 whitepapers from the ICO boom. Thirty-eight contained zero technical differentiation. The market disagreed with my analysis for six months. Then it agreed violently, and billions of dollars of speculative value evaporated. In 2020, I spent six months modeling yield farming strategies across Uniswap and Compound, and found that roughly 70% of the “yield” in DeFi was not economic return at all. It was inflationary token emission, priced as sustainable revenue. In 2024, I tracked institutional capital flows into Bitcoin exchange-traded funds and wrote “The Great Decoupling,” predicting that institutional adoption would sanitize crypto's rebel narrative. It did.
The AI infrastructure trade in 2025 is not identical to any of those episodes. But it rhymes — structurally, emotionally, and, most importantly, in its relationship between narrative and verified fact.
This article dissects the “decade-low forward P/E” claim against NVIDIA's actual trading history, examines the technical and commercial transitions embedded in Baker's infrastructure thesis, and advances a hypothesis that neither the bull case nor the bear case has explicitly articulated: the July 2025 selloff was not a failure of market understanding. It was a correct repricing of genuine uncertainty.
Context: The Man, the Thesis, the Backdrop
Gavin Baker is not a retail oracle. He is the founder and chief investment officer of Atreides Management, a technology-focused fund whose holdings have tracked some of the longest-running conviction positions in software and semiconductors. His public record includes persistent early exposure to NVIDIA during its transformation from a graphics vendor into the dominant supplier of AI computation. When Baker speaks about compute infrastructure, institutional capital listens.
The thesis he articulated in mid-2025 is broader than a single ticker. “All-in on AI infrastructure” is an assertion that an entire technological chain — accelerators, network fabrics, high-bandwidth memory, liquid cooling, power generation, data center construction — constitutes the defining capital buildout of this decade. The comparison to the interstate highway system or the global fiber backbone of the late 1990s is not hyperbolic. It is the scale of investment being discussed.
That thesis deserves respect. It may even be correct on a five-to-ten-year view. The AI infrastructure narrative is measurable in hundreds of billions of dollars of committed capital expenditures by the four largest technology companies in the world.
But respect for the thesis does not require accepting the valuation claim attached to it. And this is where a data-driven skepticism becomes necessary rather than merely temperamental.
The backdrop matters. NVIDIA's data center segment generated approximately $115 billion in revenue in fiscal 2025, up 93% year-over-year, representing roughly 89% of total company revenue. This is not a graphics card vendor. It is a national-scale compute utility in its adolescence. The July 2025 selloff occurred after a prolonged period of concentrated positioning, bullish options flow, and growing skepticism about AI's return profile. When a sector's expectations are this elevated, even moderate disappointment triggers outsized drawdowns. What happened in July was not anomalous. It is a recurring feature of infrastructure buildouts.
One further observation belongs here. The article that triggered this analysis was published on a blockchain and Web3 news outlet. That is not an accident. The capital flows and attention dynamics of AI and crypto have converged. They now operate as overlapping expressions of the same narrative machinery — and the same narrative machinery produces “decade-low P/E” claims.
Core: Deconstructing the Decade-Low Claim
Part One — The Forward P/E Dissection
Let me be precise about the numbers.
NVIDIA's forward P/E after the July 2025 selloff, based on consensus next-twelve-month estimates, was approximately 25–30x. The “decade-low” characterization compares this figure against historical readings. The relevant history includes the 2015–2016 gaming and crypto-mining era at 15–25x forward; the 2021 speculative peak at 60–80x forward; the 2022 bear market trough at 25–40x forward; and the 2023–2024 AI euphoria phase at 30–60x forward.
The “decade-low” label fails on two levels. First, NVIDIA's forward P/E in 2015–2016 was lower than today's reading, meaning the claim is categorically false if measured consistently across a full decade. Second, if the claim is intended only relative to the 2021–2024 window, the correct description is “post-AI-boom low,” not “decade-low.” Precision matters because the entire narrative purchase of the claim depends on a comparison to 2015 valuations, which are not comparable to the current business.
Beyond semantics lies a structural issue.
Forward P/E is the ratio of current price to consensus expectations. When an estimate revision cycle is in its upward phase — as it has been for NVIDIA since Blackwell's March 2024 announcement — the denominator is being inflated every quarter. Analysts are not forecasting from the ground up. They are extrapolating from the most recent management guide and supply commentary. The result is a multiple that degrades mechanically as estimates rise, creating an illusion of increasing attractiveness.
A company growing at 130% per year is structurally incapable of sustaining the same forward multiple as a company growing at 20%. The ratio is compressed by the denominator. Investors who mistake this compression for cheapness are committing a category error.
The absolute valuation context compounds the problem. NVIDIA carries approximately three trillion dollars in market capitalization. At 30x forward earnings, the market is assigning roughly $100 billion in expected annual profit to this company. That price already assumes extraordinary execution across multiple years. To sustain the current multiple, NVIDIA must maintain compound annual EPS growth of 40% or more through fiscal 2027.
If that trajectory is interrupted, the consequence is not benign consolidation. It is a de-rating — a compression of the multiple driven downward by simultaneous reductions in both price and future earnings estimates. This is a “double compression,” and it occurs faster than the forward P/E narrative suggests.
Perhaps the most important detail in the NVIDIA valuation discussion is the distinction between forward P/E and trailing P/E. In a hypergrowth regime, the gap between these two metrics is enormous. NVIDIA's trailing P/E is meaningfully higher than its forward P/E because the denominator of the trailing calculation does not yet include the massive earnings contribution expected from Blackwell in fiscal 2026. The “low P/E” story only works if Blackwell's ramp is flawless.
Part Two — Blackwell and the Composite Moat
Under the valuation debate sits a technical transition that will determine whether the earnings estimates are even reachable.
NVIDIA's Blackwell platform began significant volume shipments in the second half of 2025. The flagship system, the GB200 NVL72, interconnects 72 Blackwell GPUs through an NVLink 9 fabric inside a single rack-scale unit, designed for training clusters at the 100,000-GPU scale. The commercial significance is not the chip itself but the unit of sale: NVIDIA is delivering turnkey compute pods with individual order values in the tens of billions.
This architecture is a direct response to the changing shape of frontier model training. Ten-thousand-GPU clusters were the norm in 2023. One-hundred-thousand-GPU clusters are the emerging frontier standard in 2025–2026. At that scale, conventional data center networking becomes the bottleneck, and the NVL72's integrated fabric is engineered to eliminate it.
Competitors have not stood still. AMD's MI300X and MI325 have narrowed the raw specification gap. Google's TPU v6 Trillium has demonstrated competitive training performance and meaningful inference efficiency advantages, delivered as cloud services to external customers. Amazon's Trainium and Inferentia chips are deployed across substantial AWS capacity. Microsoft's Maia accelerator has entered internal deployment.
None of these alternatives eliminates NVIDIA's core advantage, which is the CUDA software ecosystem. Fifteen years of developer mindshare, integrated into PyTorch, JAX, and every major toolbox, creates a switching cost that no benchmark can capture. Even a rival accelerator with parity on price-performance must overcome a software stack that has accumulated operator kernels, debugging tooling, and deployment infrastructure for a decade and a half.
This is the “composite moat”: CUDA software, NVLink fabric, InfiniBand networking, rack-level engineering, and a fully integrated deployment stack. This is the real basis of Baker's conviction. A thesis built on hardware specifications invites comparison. A thesis built on a system-level standard operates on a different plane.
The hidden macro dynamic is the pushback. The customers being locked in are the same customers with the resources to build alternatives. Google, Amazon, Microsoft, and Meta are each investing billions to reduce dependency on NVIDIA. The top four direct customers contribute more than 40% of data center revenue. They are simultaneously NVIDIA's largest fans and its most motivated competitors. That tension does not appear in a forward P/E calculation. It appears — or fails to appear — in the 2026–2027 revenue guide.
Part Three — The Inference Transition and the Efficiency Trap
The Baker infrastructure thesis contains a second, implicit assumption: inference, not training, will become the dominant GPU demand engine between 2025 and 2027.
The assumption is directionally correct. Frontier model training exhausts a small pool of well-funded laboratories. Inference, by contrast, is expanding across enterprise software, consumer products, agentic workflows, and government deployments. The growth trajectory favors inference.
The complication is efficiency. The unit cost of inference is falling faster than almost any forecast. FP8 and FP4 quantization, speculative decoding, KV-cache retention, continuous batching, and model distillation are compressing cost per token in a nonlinear fashion. Compact models with remarkable task performance are running on commercial silicon at a fraction of the cost of frontier-scale systems.
Efficiency is not empathy. What that means here: every algorithmic efficiency gain reduces the number of GPUs required to serve a fixed volume of inference requests. The Jevons paradox argues that lower costs expand total demand. Historically, this argument has been correct in compute markets. But the dynamics operate with a lag, and the near-term causality runs in the opposite direction: efficiency improvement → reduced GPU procurement → lower forward demand.
If the inference efficiency curve continues to decline at 40–50% per year, the linear extrapolation of GPU demand from current hyperscaler capex guidance is materially overstated. The inference revenue projection for 2026–2027 may be the most exposed estimate in the entire AI equity complex.
Part Four — Physical Constraints: CoWoS, HBM, Power
The AI infrastructure trade has shifted its binding constraint from silicon to physics.
CoWoS advanced packaging remains the initial bottleneck. Taiwan Semiconductor Manufacturing Company's CoWoS capacity is projected to rise from roughly 45,000–50,000 wafers per month in 2024 to 65,000–80,000 per month in 2025. Nearly all of that capacity is allocated to NVIDIA, AMD, Google, and custom ASIC vendors. Packaging is a throughput limit, not a design problem.
High-bandwidth memory is the second constraint. SK Hynix, Samsung, and Micron have effectively sold out HBM3e production into 2026. HBM now represents 40–50% of a B200's bill of materials, meaning NVIDIA's margin profile is significantly dependent on memory pricing. NVIDIA has pricing power downstream, but upstream, it is competing with the entire industry for the same limited memory supply.
Power is the third constraint, and the one that binds most visibly in 2025. A 100,000-GPU cluster requires 80–120 megawatts of continuous electrical capacity. Hyperscalers face grid interconnection queues of three to five years in multiple US and European jurisdictions. Data center power density has moved from 5–10 kilowatts per rack to 50–100 kilowatts per rack, making liquid cooling a strategic requirement rather than an engineering preference.
The supply chain for AI infrastructure now extends deep into the energy economy. Microsoft, Google, and Amazon have signed power purchase agreements for nuclear, solar, and geothermal capacity at unprecedented scale. Natural gas turbines are being repurposed for data center load. Small modular nuclear reactors, once a speculative blueprint, are entering procurement pipelines.
Baker's “all-in” framing may implicitly include these adjacent plays. Atreides may well hold positions across the full stack: accelerators, networking, memory, power, cooling. That would be a structurally more robust trade than a single-stock NVIDIA bet. Public 13F filings are insufficient to confirm the full scope. But the media narrative narrowed the story to NVIDIA, which is exactly how narratives narrow — toward the most recognizable ticker.
Part Five — The Web3 Mirror
Let me return to the meta-observation.
Why does a Web3 publication carry a story about NVIDIA's forward P/E? Because the audiences have merged. In 2025, AI and crypto are not separate stories. They are parallel expressions of the same speculative impulse — the search for critical infrastructure, hard assets, and high-conviction narratives.
In 2021, the crypto industry built dozens of Layer 2 rollup networks, dedicated data availability layers, and cross-chain protocols before transaction demand justified the throughput. Most of that infrastructure sits underutilized relative to capacity. The buildout was funded on the expectation that applications would arrive. Some did. Most value accrued to the infrastructure builders early and to applications later — and much was destroyed in between.
AI infrastructure is executing the same playbook. Hyperscalers are committing capital ahead of verified application-layer revenue because the cost of being late is perceived as higher than the cost of being early. That might be correct. It is also a faith claim, not a measurement. Structurally, AI capex resembles the 2021 Layer 2 expansion and the 2017 ICO pre-sales more than it resembles any mature capital cycle.
The difference is scale. The sums are an order of magnitude larger, the builders are the world's largest companies, and the infrastructure is physical rather than virtual. But the narrative shape is identical: build the pipes, price them like cathedrals, and hope the applications arrive before the cost of capital rises.
The data availability layer debate in crypto offers a precise analogy. The narrative said every rollup needs a dedicated DA layer. The reality is that 99% of rollups generate insufficient data to justify dedicated DA infrastructure. AI infrastructure investment may face a similar dynamic: massive compute buildout ahead of sufficient, margin-accretive demand.
Part Six — Geopolitics and the Parallel Universe
The AI infrastructure thesis has a geographic ceiling that neither the bull nor the bear case has priced with discipline.
NVIDIA's China revenue has declined from approximately 17% of total revenue in fiscal 2024 to roughly 13% in fiscal 2025, and the trend is accelerating. The Bureau of Industry and Security continues to tighten export controls, and proposed frameworks contemplate country-level caps on aggregate compute shipments.
The result is a bifurcated world. In China, Huawei's Ascend processors and Cambricon's accelerators are scaling under a state-backed domestic substitution program. Chinese domestic chip content in the domestic AI market is projected to rise from roughly 20% in 2024 to 50% by 2027. The technical gap remains real, but the policy tailwind is enormous, and the software ecosystem — MindSpore, CANN, and a growing set of domestic framework integrations — is maturing precisely because NVIDIA alternatives are disappearing from the market.
This creates the possibility of two parallel AI compute ecosystems, each with independent software stacks, supply chains, and strategic logics. For Baker's infrastructure thesis, this is not necessarily negative. Sovereign AI procurement across the Gulf states, Europe, and parts of Asia is creating demand independent of US hyperscaler capital expenditure. National labs, defense establishments, and state-backed enterprises are a new customer class with long procurement cycles and low price sensitivity.
But NVIDIA cannot serve significant portions of that demand under current restrictions. The “all-in AI infrastructure” narrative requires an asterisk: it depends on market accessibility, and that accessibility is being constrained by policy on an almost quarterly basis.
Contrarian: What the Bull Case Refuses to See
The contrarian position is not that NVIDIA is doomed. The contrarian position is that the July 2025 selloff was functionally efficient.
Look at what the market knew at that moment. A multi-hundred-billion-dollar capex cycle was underway. The transition from Hopper to Blackwell carried real execution risk — initial yields, packaging constraints, memory allocation, and power supply all carried uncertainty. The customers funding the buildout were simultaneously developing internal silicon, with timelines repeatedly accelerated. The inference cost curve was bending downward, undermining linear GPU demand projections. And the standard hyperscaler response to ROI questions — “we are in the early innings” — is not an ROI statement.
A rational investor presented with these conditions makes a choice: extrapolate the capex trend and buy every dip, or require a margin of safety. In July 2025, a meaningful segment of the market chose the second option. That is risk management, not mispricing.
The framing of NVIDIA's “decade-low” forward P/E as a gift presumes the market is wrong and the optimist is right. But the forward P/E is “low” because the consensus estimate embeds assumptions about sustained 40%+ growth that the market judges — reasonably, given the evidence — to be less than fully certain. The multiple is a risk premium. It is the market charging for the possibility of disappointment in a system where disappointment has been systematically under-priced for three years.
There is also a lesson from my own experience. In 2022, after the LUNA and FTX collapses, I spent three months in deliberate withdrawal from public markets. When I returned, I focused exclusively on infrastructure projects with sustainable economic models. I analyzed Polygon's ZK-rollup roadmap with a small group of developers in Vietnam, assessing technical resilience rather than price action. That discipline — technical verification over narrative resonance — is the correct framework for evaluating NVIDIA today.
The infrastructure is real. The question is whether the revenues will arrive fast enough to justify the prices paid before they were visible. In 2017 and 2021, the answer was no. This cycle may differ — the builders are larger, the capital is more patient, and the demand signals are more concrete. But a different answer requires different evidence.
Code doesn't feel. Markets eventually do.
Takeaway: Structural Signals, Not Quotients
The “decade-low P/E” claim should be handled with the same skepticism I applied to ICO whitepapers and DeFi yield formulas. It is a narrative device, not a data point. The analysis must be built on the chain of verification events that will confirm or falsify the AI infrastructure thesis.
Here is what to watch over the next four to eight quarters.
Hyperscaler capital expenditure guidance, and the slope of AI revenue growth relative to capex growth. NVIDIA fiscal Q3 and Q4 2026 results for Blackwell revenue ramp and gross margin trajectory. Taiwan Semiconductor's CoWoS capacity positioning. BIS rulemaking activity on export controls. The first hyperscaler to reduce capex guidance while describing it as “efficiency optimization.” And the AI revenue disclosures of Microsoft, Amazon, and Google, which will resolve whether the gap between AI spending and AI revenue is closing or widening.
The market is not about to discover whether AI is real. AI is real. It is about to discover whether the return on AI investment justifies the prices paid for infrastructure before the returns were measurable. That discovery is the next narrative inflection.
When that discovery arrives, the narrative will not be “NVIDIA is cheap.” It will be “AI efficiency is finally real,” and the winners will be the infrastructure layers that can sustain margins through an efficiency correction. The losers will be the assets priced on the assumption that efficiency improvements would never arrive.
Hype fades; structure remains. The structure of AI infrastructure is formidable. But structure and valuation are separate things. Baker's thesis may be correct at the structural level while the market price of NVIDIA remains fair.
The investment question is not whether AI infrastructure will be built. It will be. The question is whether current prices have already paid for the building, and whether “decade-low” forward multiples are measuring the future or repeating the oldest mistake in technology investing — confusing arithmetic with insight.
Position according to structural signals. The quotients will follow.
