Gavin Baker's NVIDIA "Decade-Low P/E" Is a Statistical Mirage — The All-In AI Trade Has a Blind Spot

Gavin Baker just doubled down on AI infrastructure with a public declaration. The Atreides Management founder's core claim: NVIDIA is trading at its lowest forward P/E in a decade, and July's selloff was a pure fundamental disconnect. The narrative is clean. Almost too clean. Let me be direct. I've spent a decade reading 13F filings, order books, and on-chain flows for a living. Baker's fund manages roughly $10–15 billion. NVIDIA's market cap is $3 trillion. Baker's entire position is a rounding error in NVIDIA's average daily volume. The market doesn't move because one growth fund manager says "buy." It moves because institutions see the same dataset and draw independent conclusions. The real signal isn't Baker's conviction — it's the timing of his disclosure. And here's where I sharpen my knife. The "ten-year low forward P/E" claim fails statistical scrutiny. Based on my audit experience pulling NVIDIA's historical valuation bands, the forward P/E in 2015–2016 was 15–25x. That's lower than today's estimated 25–30x forward multiple. The chart whispers before the market screams. Right now, the whisper says "expectation inflation," not "valuation compression." Baker didn't stumble into this position. He built his career riding mega-cap tech through drawdowns. He started accumulating NVIDIA in 2016, when the stock traded in the $30–40 range. His playbook: identify a 5-year, 10x TAM expansion, ignore short-term noise, add into fear. That strategy made him a fortune. The problem is he's now applying a 2016 playbook to a company that has already compounded 100x from those levels. The asymmetry has flipped. The AI infrastructure thesis itself is real. Cloud giants are guiding to a combined $300 billion+ in 2025 capital expenditures, with roughly 55–60% tied to AI. NVIDIA's data center business generated approximately $115.2 billion in FY2025 revenue — up 93% year over year — and now represents 89% of total revenue. This isn't a chip company anymore. It's an AI factory infrastructure provider, selling system-level racks that carry single-order values in the billions. The Hopper-to-Blackwell handoff defines this moment. Blackwell launched in March 2024. Mass shipments hit the second half of 2025. The GB200 NVL72 rack connects 72 Blackwell GPUs through NVLink 9 switching, creating a system-level solution designed for hundred-thousand-GPU training clusters. Each rack is a self-contained compute pod, with integrated liquid cooling, power delivery, and InfiniBand fabric — a far cry from the loose GPUs NVIDIA shipped during the Hopper generation. This generational shift raises per-customer order values and locks clients into longer procurement cycles. If Blackwell ramps cleanly, NVIDIA's 2026 fiscal year earnings guidance looks achievable. If the ramp stumbles, the entire "low P/E" narrative inverts. But a real transition carries real friction. TSMC's CoWoS advanced packaging remains the single tightest bottleneck in the AI supply chain. CoWoS capacity is expanding from roughly 45–50K wafers per month in 2024 to 65–80K per month in 2025 — still insufficient for every GPU order. HBM3e supply from SK Hynix, Samsung, and Micron is essentially pre-sold through 2026, with the memory alone consuming 40–50% of a B200's bill of materials. Every Blackwell rack demands 50–100kW of power density, and data center grid interconnects in the US and Europe face 3–5 year waitlists. The physical layer of AI infrastructure — not chip design — is now the binding constraint. I pulled NVIDIA's historical valuation bands across four distinct market regimes. Here's the scorecard. The 2015–2016 era: trailing P/E of 20–40x, forward P/E of 15–30x. Gaming and crypto mining demand drove cyclical swings. This was a pre-AI NVIDIA, still digesting the PC downturn and still building the CUDA ecosystem that would later become its fortress. The 2017–2018 period: trailing P/E of 30–50x, forward P/E of 25–40x. The AI/data center narrative was just emerging, but the crypto mining boom distorted demand signals badly. The 2019–2020 stretch: trailing P/E of 40–70x, forward P/E of 30–50x. Data center growth accelerated as cloud adoption went mainstream. The 2021 peak: trailing P/E of 80–110x, forward P/E of 60–80x. Crypto and AI hype overlapped into a full mania. That was the true valuation blow-off. The 2022 trough: trailing P/E of 30–50x, forward P/E of 25–40x. Crypto collapsed, rates surged, and NVIDIA's gaming revenue cratered. The stock got genuinely cheap because the earnings outlook was genuinely bleak. The 2023–2024 period: trailing P/E of 60–120x, forward P/E of 30–60x. AI mania pushed trailing multiples skyward while forward multiples stayed anchored to aggressive EPS growth assumptions. Post-July 2025 correction: trailing P/E of roughly 40–50x, forward P/E of approximately 25–30x. Now do the math. The 2015–2016 forward P/E range of 15–25x sits below today's 25–30x. So the "ten-year low" claim is only valid if you cherry-pick the denominator — for instance, using one-year forward consensus EPS that has been aggressively marked up following Blackwell guidance. A forward P/E is only meaningful when the earnings forecast is trustworthy. NVIDIA's EPS estimates have climbed at a dizzying clip because analysts are pulling forward Blackwell revenue projections. When the denominator gets inflated by 40–50% growth assumptions, the multiple looks mechanically low. That's not the market mispricing NVIDIA. That's the market pricing in a perfect execution scenario. The code is cold, but the hype is hot. Customer concentration is the hidden sword. Microsoft, Amazon, Google, and Meta — the four largest hyperscalers — contribute over 40% of NVIDIA's data center revenue. Their capex budgets are existential commitments to staying in the AI race. But every CFO is asking the same question: when does AI revenue catch up to AI capex? The investor calls right now are full of "short-term investment, long-term return" language. That's corporate speak for "we don't actually know when this pays off." If any hyperscaler guides capex lower in the next two quarters, NVIDIA's revenue trajectory gets repriced violently. We trade the panic, not the price — but the panic starts with a single CFO whisper on an earnings call. Competition is a slow bleed, not a sudden break. NVIDIA still commands 80–95% of the AI training accelerator market — that's not hyperbole, that's Mercury Research and Omdia territory. But the edges are eroding. Google's TPU v6 Trillium is now commercially available to external cloud customers, targeting inference workloads with aggressive price-performance. Amazon's Trainium and Inferentia keep improving. AMD's MI300X shipped below expectations, but the MI400 series lurks. And the custom ASIC players — Cerebras, Groq — are carving out inference niches where the full CUDA stack is overkill. None of these threaten NVIDIA's training dominance today. But they're building alternative supply for the exact workloads the bull case needs to grow: inference at scale. If NVIDIA's inference share erodes to 60% by 2028, the "AI infrastructure forever" multiple breaks. The "AI factory" lock-in is real — and it's a double-edged sword. NVIDIA's business model has shifted from selling chips to selling entire data center systems. The NVL72 rack plus NVLink plus CUDA plus networking (InfiniBand, Spectrum-X Ethernet) creates a compounding lock-in: a customer buying a rack is structurally committed to NVIDIA's ecosystem for the next 3–5 years. This is why NVIDIA maintains gross margins above 70%. This is also why the "low forward P/E" argument has a structural supporter: the company's revenue quality is genuinely higher today than it was in 2016. Recurring software (AI Enterprise, DGX Cloud), networking gear, and system integration create revenue stickiness that didn't exist a decade ago. But lock-in cuts both ways. Hyperscalers hate dependency. Google, Amazon, Microsoft, and Meta are all accelerating their custom silicon programs precisely to reduce NVIDIA exposure. The same "AI factory" strategy that creates NVIDIA's margin strength is incubating the customer-side rebellion that will eventually compress those margins. The inference demand assumption needs a stress test. Baker's thesis implicitly assumes inference replaces training as the next GPU demand wave in 2025–2027. That assumption deserves more skepticism than it's getting. Inference efficiency is improving faster than the bulls realize. FP8/FP4 quantization, speculative sampling, KV cache reuse, continuous batching — these optimizations are slashing per-token cost. Small models like DeepSeek's open-source family and Llama-3 mid-size variants are handling vertical workloads without flagship Blackwell GPUs. The training market is safe. But the inference upside may be dramatically smaller than the linear projections suggest. If inference demand grows at 30% instead of 80%, the "AI infrastructure forever" narrative cracks — and the forward P/E math breaks with it. Now the angle nobody's covering. Baker's "all-in AI infrastructure" framing — if taken literally — isn't a single-stock bet. It's a basket. The intelligent configuration includes Broadcom for networking, TSMC for foundry, Micron and SK Hynix for HBM, Vertiv for power and cooling, and utilities like Constellation Energy or Vistra for electricity. The AI infrastructure trade is not NVIDIA. NVIDIA is one crowded position within a much larger physical buildout. And the highest-elasticity position isn't the GPU manufacturer — it's the power provider. Rack density is going from 5–10kW to 50–100kW. The companies selling liquid cooling and electrical distribution are seeing unit economics improve faster than any chip vendor. Pixels hold value when code forgets — but nothing runs without electrons. Second blind spot: sovereign AI. Export controls have already cut NVIDIA's China revenue share from roughly 17% in FY2024 to an estimated 13% in FY2025. But the bigger trend is the parallel universe forming. China's domestic AI chip ecosystem — Huawei Ascend, Cambricon — is scaling under a national substitution policy. By 2027, domestic chips could capture 50%+ of China's AI accelerator market. That's not just lost revenue for NVIDIA. That's a second, independent AI ecosystem growing outside CUDA's orbit. The world is splitting into two compute spheres. Baker's all-in thesis doesn't touch this because it pretends geopolitical segmentation doesn't exist. It does — and it compounds every quarter. Third blind spot: the ROI verification gap. Nobody in this trade wants to acknowledge that AI capex returns are unproven. In my work building trading signals, I've learned that when a narrative requires a 3–5 year validation window, the market trades the narrative, not the validation. NVIDIA at 30x forward earnings sounds reasonable at 40% EPS CAGR. But if EPS growth decelerates to 20%, the multiple de-rates — and a de-rating coinciding with estimate cuts is the classic double-kill. The 30x forward P/E isn't an entry point. It's a priced-for-perfection multiple. The "ten-year low" claim is the most dangerous narrative element in this trade. It implies NVIDIA has never been cheaper. It invites investors to stop doing math and start feeling opportunistic. July's selloff wasn't "fundamentals disconnected." It was the market repricing uncertainty around AI ROI. Dismissing that as noise is exactly the kind of overconfidence that gets portfolios destroyed. What am I watching now? NVIDIA's next earnings report for Blackwell revenue recognition pace — actual units shipped, not guided. Hyperscaler capex commentary over the next two quarters. BIS export rule updates. And the single most important chart in the market: the slope of AI revenue growth versus AI capex growth at Microsoft, Google, and Amazon. See the pattern before it prints. If AI revenue acceleration flattens against a still-rising capex curve, the AI trade flips from "buildout" to "overbuild" in exactly one quarter. Gavin Baker is a great investor. That doesn't make this call right. NVIDIA is a magnificent company. That doesn't make it cheap. The forward P/E looks low for one reason — because the machine is expected to keep compounding at unprecedented rates. Speed is the new currency of trust, but trust isn't analysis. And liquidity is the only truth that bleeds. Watch the ROI slope, not the P/E ticker.