The Verification Gap: Auditing the "Ten-Year Low" NVIDIA Thesis and Gavin Baker's All-In AI Bet

The claim propagated through blockchain media on July 19 with the velocity of a verified fact: NVIDIA was trading at its lowest forward price-to-earnings ratio in a decade. The attribution was precise. Gavin Baker, founder of Atreides Management, had presented his "all-in on AI infrastructure" positioning to investors, and the forward P/E statistic became the anchor for every subsequent commentary. At roughly 25-30 times forward earnings, the multiple did appear historically compressed for a company whose trailing earnings grew more than 130% year-over-year. The historical record contradicts the framing. In 2015, with NVIDIA stock trading near $20 and consensus earnings per share below $1, the forward P/E was approximately 22x. By mid-2016, it fell closer to 18x. The measurement basis was different — NVIDIA's market capitalization was a fraction of its current $3 trillion-plus valuation, and analyst coverage was thinner — but the multiple itself was lower. The statement "ten-year low forward P/E" is therefore not a straightforward financial fact. It is a selective comparison that requires specific measurement choices to function. This is the same pattern I encountered in 2018 while auditing ICO refund contracts on Ethereum. A statistic was reported as rigorous because it was derived from on-chain data. The chain showed successful refunds at a high rate. The full picture — including users structurally ineligible for refunds due to edge-case withdrawal logic — told a different story. The number was real. The conclusion was engineered. History verifies what speculation cannot. Baker's investment framework is public and well-documented. He has run Atreides Management since 2015, managing an estimated $10-15 billion in assets. His NVIDIA position dates to 2016, when the stock traded at $30-40 on a split-adjusted basis. He has described the AI infrastructure cycle as a multi-year total addressable market expansion, with GPU compute transitioning from optional accelerator to foundational infrastructure. The "all-in" statement, however, covers more than a single equity. It implies a portfolio positioned across the full infrastructure stack: chip design, advanced manufacturing, memory, networking, power generation, cooling, and data center construction. The macro backdrop supports the investment thesis. The four largest cloud providers — Microsoft, Amazon, Google, and Meta — guided combined capital expenditures above $300 billion for 2025, with AI-related allocation estimated at 55-60% of that total. This spending pipeline feeds directly into NVIDIA's data center segment, which generated roughly $115.2 billion in FY2025 revenue, up 93% year-over-year, and now constitutes approximately 89% of NVIDIA's total revenue. The company is no longer selling chips. It is selling AI factories. The structure of the claim, however, contains an internal tension. A forward P/E of 25-30x on a company with triple-digit earnings growth looks reasonable. But NVIDIA's current market capitalization requires a denominator — projected earnings — that is not yet realized. The forward P/E is not a statement about today. It is a statement about whether the next twelve months will deliver the earnings estimates currently embedded in the price. For that to happen, several structural conditions must hold. None are fully verified. NVIDIA must execute the Blackwell architectural transition without disruption. The global advanced packaging and memory supply chain must expand at pace with demand. AI inference demand must grow at a rate sufficient to absorb the capacity that training built. And the four largest customers must maintain capital expenditure discipline. Each condition carries its own failure modes. Structure outlasts sentiment. Forward price-to-earnings ratios are conceptually simple: current market capitalization divided by consensus earnings estimates for the next twelve months. The numerator is verifiable. The denominator is an aggregation of analyst predictions, corporate guidance, and embedded optimism. NVIDIA's forward P/E has been structurally compressed for three years because of the denominator's growth rate. When EPS grows at 50% or more annually, the forward multiple converges rapidly toward low levels despite an appreciating stock price. This is arithmetic, not anomaly. In FY2025, NVIDIA's EPS growth exceeded 130% on the strength of the AI buildout. At those rates, a $3 trillion company can trade at 30x forward earnings while appearing reasonably valued. The "ten-year low" framing fails for two reasons. First, it ignores that NVIDIA's forward P/E in 2015-2016 was equivalent or lower on a consistent measurement basis. Second, it treats the current forward estimate as a stable input when it is a moving target. In the months before the July selloff, analysts revised NVIDIA's forward EPS upward repeatedly, mechanically pulling the forward P/E down without any change in share price. A ratio that declines because the denominator is being recalibrated is not an indication of cheapness. It is an indication of momentum in estimates. The fundamental question is whether consensus forward EPS — widely estimated in the range needed to sustain a 25-30x multiple at current prices — is achievable. This requires interrogating the assumptions beneath the estimates. Based on my experience auditing financial protocols, the first thing to check is whether the reported metric captures all relevant conditions or only the favorable ones. NVIDIA's architectural transition from Hopper to Blackwell represents the most consequential product cycle in its history. The B200 GPU and the GB200 NVL72 rack-scale system, announced in March 2024, moved into mass shipment in the second half of 2025. The NVL72 design integrates 72 Blackwell GPUs through NVLink 9 switching and is engineered for 100,000-GPU cluster training deployments. Commercially, this represents a transition from selling accelerators to selling data center infrastructure. A single customer order for NVL72 systems can reach tens of billions of dollars. The transition introduces supply chain exposure at three distinct layers. Advanced packaging is the first constraint. TSMC's CoWoS packaging substrate is the physical bottleneck for every high-end AI accelerator shipped by NVIDIA, AMD, and Google. 2025 monthly capacity is projected at 65,000-80,000 wafers, up from approximately 45,000-50,000 per month in 2024. Even at that growth rate, total capacity covers only about 80% of projected demand from major AI GPU makers. CoWoS is not a resource that can be added at will. It requires cleanroom construction, equipment qualification, and yield learning cycles that span quarters. The second constraint is high-bandwidth memory. SK Hynix, Samsung, and Micron have effectively pre-sold HBM3e production through 2026 to NVIDIA and other customers. The price for an 8-layer 24GB HBM3e stack exceeds $1,500, representing 40-50% of the B200's total material cost. Memory supply is not fungible. Each NVIDIA generation requires new memory specifications, creating a joint qualification cycle with memory vendors. Any HBM yield issue cascades directly into GPU production schedules. The third constraint is not in the chip supply chain at all. It is in the physical infrastructure layer. A 100,000-GPU cluster draws 80-120 megawatts of continuous power. Data center power density is moving from 5-10 kW per rack for conventional servers to 50-100 kW per rack for GPU systems. Existing electrical grids in data center hubs can require three to five years for new connection approvals. Liquid cooling has shifted from optional to mandatory. The physical buildout of AI infrastructure now depends more on electricians, grid interconnection queues, and cooling system manufacturers than on chip designers. Baker's "all-in" thesis, if it includes power utilities and cooling infrastructure providers such as Vertiv, is a bet on the pricing power of these physical constraints. The source article's framing of NVIDIA as "AI infrastructure" is correct but incomplete. NVIDIA's earnings are not a function of its own execution alone. They are a function of TSMC's capacity allocation decisions, memory vendors' yield curves, and utility companies' interconnection approval timelines. Complexity hides its own failures. The software layer provides the counterweight. CUDA has accumulated over 15 years of developer ecosystem lock-in. The major deep learning frameworks — PyTorch, JAX, TensorFlow — depend on CUDA for GPU execution. This dependency creates a switching cost that exceeds any hardware performance differential. Even if competitors achieve specification parity, software migration costs act as a protective barrier. However, the moat is being tested from three directions. Google's TPU v6 (Trillium) has begun serving external cloud customers, with demonstrated advantages in specific inference workloads. Amazon's Trainium and Inferentia ASIC families are being optimized internally for AWS workloads. And the developer ecosystem is diversifying: OpenAI's Triton language and JAX's XLA compiler route around CUDA's central role in certain workflows. The defensive position remains strong — NVIDIA's training market share holds at 80-95% across major analyst estimates, and AMD's MI300 series has not delivered the competitive threat that was projected — but the market structure is shifting from single-player dominance to a "one strong + multiple challengers" configuration. The forward P/E narrative does not price the gradual erosion of NVIDIA's pricing power in inference workloads, where ASIC alternatives are most competitive. This connects to NVIDIA's business model transformation. The company is moving beyond chip sales to system-level infrastructure: the NVL72 cabinet, the AI Enterprise software platform, DGX Cloud, and the networking stack comprising InfiniBand and Spectrum-X Ethernet. This vertical integration increases customer lock-in and raises average order value. It also increases delivery complexity and failure-mode exposure. A disruption in any layer — GPU, memory, NVLink switch, power delivery — delays the entire system, not just one component. NVIDIA's data center revenue concentration creates a structural dependency that the forward P/E does not capture. The top four direct customers — Microsoft, Amazon, Google, and Meta — account for more than 40% of data center revenue. The remaining share is distributed across enterprises and governments, but the demand signal originates from hyperscaler capital allocation decisions. Cloud capital expenditure is the macro variable that binds the entire AI infrastructure thesis. In 2025, hyperscaler capex guidance exceeded $300 billion. This figure has become a proxy for NVIDIA's revenue trajectory. However, capex guidance is discretionary and reversible. It is not a contract. It is a statement of intent subject to quarterly revision, board review, and — critically — evidence of return on investment. The ROI evidence is not yet visible. In recent investor communications, cloud providers have described their AI investments in temporal abstractions: "long-term," "multi-year," "early innings." The absence of quantitative ROI disclosures for AI capex is itself a data point. In my 2020 work auditing DeFi lending protocols, I observed a parallel pattern. Protocols reported total value locked as a growth metric while the underlying deposits were incentive-driven and dependent on continuous reward emissions. The metric was real. The economic durability was not. When emissions ended, TVL contracted sharply. NVIDIA's revenue is currently backstopped by capex commitments that are similarly temporary in nature. The next growth leg introduces an additional variable. Training demand drove the 2023-2024 expansion. Inference is expected to drive the 2026-2027 cycle. But inference workloads are more price-sensitive than training. Optimization techniques — FP8 and FP4 quantization, speculative decoding, KV cache reuse, continuous batching — are reducing the compute cost per token at a steady rate. If inference efficiency improves faster than application-level demand grows, the number of GPUs required per unit of AI-generated value declines. This dynamic could compress NVIDIA's incremental pricing power in its highest-potential growth segment. The source article's thesis implicitly assumes that first-generation AI deployments will generate second-generation demand. That assumption has not been tested in the market. Export controls compound the market risk. NVIDIA's China revenue has declined from approximately 17% of total revenue in FY2024 to roughly 13% in FY2025, driven by successive rounds of U.S. export restrictions. The trend is directionally clear and likely to continue. China is simultaneously developing domestic alternatives — Huawei's Ascend line and Cambricon chips — under a national substitution policy. By 2027, domestic chips could account for more than half of China's AI accelerator market. NVIDIA's long-term addressable market is not as large as its decade-long growth narrative implies. The forward P/E figure does not isolate this geopolitical discount. Additional structural risk comes from the "sovereign AI" trend. Governments in the Middle East, Europe, and Southeast Asia are building national AI compute infrastructure. This demand is real — Saudi Arabia, the UAE, and several European nations have announced multi-billion-dollar AI compute programs. But sovereign AI procurement is slow, politically conditioned, and subject to export approval processes. It cannot be relied upon to backfill any shortfall in hyperscaler spending. The July AI selloff has been framed as an emotional overreaction. The evidence supports a different reading: the market repriced AI infrastructure assets in response to ambiguous ROI disclosures and supply chain risks. This was not panic. It was probability adjustment. The distinction is visible in the structure of the selloff. NVIDIA's stock declined while forward earnings estimates remained elevated. A purely emotional selloff would typically coincide with downward revision of forward estimates — the market pricing in lower future earnings. Instead, the multiple compression reflected a discount applied to uncertain earnings rather than a repricing of consensus fundamentals. The market was not rejecting NVIDIA's forecast. It was lowering confidence in the forecast's execution probability. Pressure reveals the cracks in logic. The "all-in AI infrastructure" narrative arriving through blockchain media is itself a signal worth scrutiny. Web3 outlets have a structural incentive to translate traditional market narratives into crypto-native terms. The same mechanism produced the "liquidity fragmentation" storyline that circulated to promote cross-chain products, and the "decentralized sequencing" roadmap that has remained a presentation slide for two years. These narratives are not analytical frameworks. They are distribution vehicles. A fund manager's 13F filing or investor letter becomes a news headline, which becomes a buy signal for GPU-related tokens, without any intermediate verification step. Evidence does not negotiate. Baker's thesis contains a legitimate insight: AI infrastructure is expanding, and NVIDIA is the highest-quality exposure to that expansion. But the "ten-year low forward P/E" claim is an unverified state variable. It depends on consensus estimates that assume Blackwell yield, CoWoS capacity, HBM supply, power availability, hyperscaler capital discipline, and sustained inference demand growth. Any one of these dependencies failing shifts the denominator downward and re-rates the multiple upward. The verification window is measurable. The upcoming earnings report will reveal Blackwell pricing, margin trajectory, and forward order visibility. The next round of hyperscaler earnings calls will disclose whether capex guidance holds. The September-December period will expose whether CoWoS expansion keeps pace with demand. Each data point either validates or invalidates the claim. Silence is the strongest proof of truth. The market will confirm or reject Baker's thesis through price action over the next two quarters. Investors who accept the "ten-year low" framing as a completed analysis — rather than a conditional forecast — are betting on narrative. The numbers will decide.