Gavin Baker Says NVIDIA Is at a Decade-Low Forward P/E. I Checked the Entire Chain.

Whispers before the ticker opens. That is how this story starts. Not with a headline. Not with a regulator. Not with a press release. It starts with a 13F filing, a Discord war room, and a spreadsheet full of supply-chain constraints. By the time the crypto-news wires lit up with 'Gavin Baker goes all-in on AI infrastructure' and 'NVIDIA is at its lowest forward P/E in a decade,' I had already spent six hours pulling NVIDIA's earnings revisions, TSMC's CoWoS guidance, HBM pricing, hyperscaler capex estimates, and power-grid interconnection queues. Here is the uncomfortable conclusion: Gavin Baker may be right about AI infrastructure. But the 'decade-low forward P/E' is not a fact. It is a narrative built on a statistical illusion, a very specific earnings estimate, and a very selective memory of NVIDIA's own history. The clock stops, but the chain doesn't. The chain runs from a fund manager's mouth to an SEC filing. From the filing to a quarterly earnings model. From the model to wafer starts at TSMC. From those wafers to HBM stacks from SK Hynix. From those stacks to 120-megawatt data centers that won't get grid interconnection until 2028. In a bull market, everyone wants to believe the headline. The chain is the only thing that separates a trade from a trap. I've seen this pattern before. During the Ethereum Merge, I found a 15% deviation in validator slashing rates hours before the mainstream outlets caught up. The skill that mattered was not prediction. It was knowing where the chain was likely to break. In 2025, the chain that matters runs through NVIDIA, TSMC, the BIS, and the electrical grid. If you are staring at a forward P/E, you are looking at the wrong end of the chain. Let me set the scene. Atreides Management is not a crypto hedge fund run by a teenager with a meme-coin terminal. Gavin Baker ran Fidelity's OTC fund. He has one of the best long-term growth track records of his generation. He started building an NVIDIA position in 2016, during the darkest days of the post-crypto-crash GPU glut. He understands semis, software and lock-in. When he says AI infrastructure is the biggest TAM expansion of his career, I don't roll my eyes. But I also don't treat his 13F like a prophecy. Atreides manages roughly $10 billion to $15 billion, based on the most recent SEC filings. NVIDIA's market cap is near $3 trillion. Gavin Baker's position is not a catalyst. It is a conviction. The market does not move because a $15 billion fund likes a $3 trillion stock. It moves because $10 trillion of passive capital and derivative flows follow the same narrative. That narrative is created by headlines. And this headline was created by a crypto-native media outlet, not by Baker's filing. What did Baker actually say? The phrase 'all-in' is doing a lot of work. In context, he appears to be making a broad thematic call: AI infrastructure is the trade of the decade. That does not necessarily mean his entire book is one ticker. It probably means a portfolio of AI-related positions. But the article narrowed it to NVIDIA because NVIDIA is the ticker that generates clicks. That is the first editing bias. Now let's attack the central claim. 'NVIDIA is trading at its lowest forward P/E in ten years.' Forward P/E is the price divided by the next twelve months of expected earnings. When a company's earnings are compounding above 100%, the forward P/E will always look lower than the trailing P/E. That is not a sign of cheapness. It is a function of a wildly growing denominator. Let me give you the numbers. In fiscal 2025, NVIDIA's data center segment generated roughly $115.2 billion in revenue, up 93% year over year. That segment is the overwhelming majority of the business, close to 89% of total revenue. EPS grew more than 130%. When you plug that into a P/E formula, the forward multiple compresses even while the stock price is hitting records. This is not a hidden discount. It is arithmetic. Now compare with NVIDIA's actual history. In 2015 and 2016, the forward P/E was roughly 15 to 25 times. The stock was considered a gaming and cryptocurrency speculation vehicle. In 2021, at the peak of the AI and crypto euphoria, forward P/E stretched to 60 to 80 times. In 2022, after the crypto crash and the Fed's rate hikes, it fell back to 25 to 40 times. After the July 2025 selloff, NVIDIA was probably sitting in the 25 to 30 times forward range. That is not the lowest in ten years. It is the same zone as 2022. And it is higher than 2015-2016. The 'decade-low' claim only works if you start the clock in 2023, ignore 2015-2016, and cherry-pick the consensus estimate that happens to be highest. That is not analysis. That is ad copy. Let me be even more precise. If you use forward one-year consensus EPS, there are periods in 2015 and 2016 when NVIDIA's forward P/E was around 15-20x. I can already hear the reply: 'That was a different company.' Different company? Same product architecture direction, same CUDA ecosystem, same founder-led culture. The business was smaller, but the forward multiple was lower. If you want to make the 'decade-low' claim, you have to define the starting date very carefully. Most people who repeat it have never looked at the old data. They are just passing along a narrative from a blog that was passing along a narrative from a tweet. But here is the deeper problem. Even if the forward P/E were the lowest in ten years, it would not mean the stock is cheap. At 25 to 30 times forward earnings and a $3 trillion market cap, the market is pricing in something like $100 billion to $120 billion of net income over the next twelve months. That is not a discount. That is a contract. The contract says NVIDIA must execute a flawless architecture transition, keep gross margins above 70%, convince hyperscalers to keep spending, and avoid a geopolitical tornado that is already spinning. What would actually make NVIDIA cheap? Earnings growth of 40% to 50% per year for four or five years, with no serious supply-chain break, no export-control shock, and no return-on-investment reckoning from the hyperscalers. That is a lot of ifs. The market is not stupid to trade at 25 times forward earnings. It is pricing an execution risk premium. Let's talk about what NVIDIA is selling. It is not selling chips. It is selling AI factories. The GB200 NVL72 is a rack with 72 Blackwell GPUs connected through NVLink 9 switches. It is a system-level product designed for 100,000-GPU clusters. When a cloud provider buys a GB200 NVL72, it is not buying a component. It is buying a data-center skeleton. Some of those orders are reportedly worth tens of billions of dollars and are locked through calendar 2026. That transformation changes the revenue model. NVIDIA used to sell a GPU and walk away. Now it sells integrated racks, networking through InfiniBand and Spectrum-X, software through NVIDIA AI Enterprise and DGX Cloud, and services. The ASP per customer has gone from thousands of dollars to tens of billions. This is why NVIDIA's gross margin can stay above 70%. It is also why the stock is an execution contract: if any component of the rack slips, the entire rack slips. The order book is a promise. The supply chain is the reality. And the supply chain is not a calm river. CoWoS advanced packaging capacity at TSMC was roughly 45,000 to 50,000 wafers per month in 2024. It is expected to reach 65,000 to 80,000 in 2025. That sounds like a big expansion, but demand is drinking almost all of it. HBM3e is even tighter. SK Hynix, Samsung and Micron are basically sold out through 2026. A single HBM3e stack costs over $1,500 and can represent 40-50% of the bill of materials for a B200. This is not an accessory. It is the spine of the product. And power is the hidden killer. A 100,000-GPU cluster needs 80 to 120 megawatts. In many parts of the United States and Europe, grid interconnection queues are three to five years long. Microsoft and Google are signing power purchase agreements with nuclear, geothermal and gas turbine providers. The AI infrastructure build-out is no longer just a semiconductor story. It is an energy story. If the power isn't there, the GPU doesn't run. And the GPU doesn't generate revenue. That is the part of the chain that no forward P/E captures. Let me also add the demand-side reality. Microsoft, Amazon, Alphabet and Meta are expected to spend over $300 billion in combined capital expenditures in 2025, up around 35% year over year. Depending on the estimate, 55% to 60% of that total is tied to AI. That is the demand engine behind NVIDIA. But the returns from that spending are still opaque. Listen to any hyperscaler earnings call and you will hear the same phrase: 'We are investing ahead of demand.' This is polite CFO-speak for 'we do not have a visible payback period yet.' It is not a sign of fraud. It is a sign of uncertainty. And uncertainty is not a free option. It is a discount factor. If one hyperscaler blinks and cuts AI capex guidance, NVIDIA's forward P/E will re-rate higher at the same time that EPS estimates get cut lower. That is the double-whammy that kills momentum stocks. The July 2025 selloff was not a panic. It was the first honest repricing of that uncertainty. The competitive picture is more complex than the article's 'NVIDIA is unbeatable' framing. NVIDIA still controls 80-95% of the AI training market. The CUDA ecosystem is a 15-year moat. Every serious deep-learning framework in the world is wired into CUDA. Switching costs are massive. But AI is moving from training to inference, and inference is a different battlefield. Google's TPU v6 Trillium is already available to external cloud customers. Amazon's Trainium and Inferentia chips are deployed at scale in its own data centers. Microsoft has Maia. Cerebras and Groq are attacking niche inference workloads with architectural advantages. None of these will kill NVIDIA in the next twelve months. But they do not need to. They only need to soften NVIDIA's pricing power at the margin. Every percentage point of share that goes to a custom ASIC is a percentage point of high-margin revenue that NVIDIA cannot get back. And there is a second, slower-moving threat: China. US export controls are already reshaping the addressable market. NVIDIA's China revenue has fallen from roughly 17% of total revenue in fiscal 2024 to around 13% in fiscal 2025. If the BIS tightens further, that number can go to zero. China is building a parallel AI ecosystem with Huawei Ascend, Cambricon and a state-driven push for domestic chips. By 2027, domestic Chinese AI chips could hold more than 50% of the domestic market. That is not a niche. That is a parallel universe. The 'global AI infrastructure' narrative is not global anymore. It is a fortress with walls on both sides. I have spent a career reading exchange balance sheets. The crypto world taught me that the phrase 'proof of reserves' is theater unless it is continuous. A snapshot proves that a balance sheet existed at one instant. It does not prove that the reserves will survive a bank run. The forward P/E argument for NVIDIA works the same way. It is a single snapshot of analyst estimates, and those estimates are not continuously audited. They are revised in batches. When supply chains break, the estimates move slower than the stock price. That is why a 'cheap' stock can keep getting cheaper. The figures are stale. The narrative is fresh. Trust no one, verify everything, move fast. I do not trust the phrase 'lowest forward P/E in a decade.' I want to see the actual chain. I want to see the wafer starts. I want to see the HBM allocation. I want to see the power contracts. I want to see the hyperscaler capex guidance. That is the only way to know if the discount is real or just a function of an over-exuberant EPS forecast. Here is the contrarian angle the headlines missed: the July selloff was not irrational, and Gavin Baker's confidence is not a signal to abandon discipline. The market's fear was not about NVIDIA's product position. It was about an unverified ROI curve. Hyperscalers are spending enormous amounts of money on infrastructure that has not yet produced a visible, quantifiable return. The efficiency of AI is improving fast. Quantization, speculative sampling, KV-cache reuse, model distillation, and smaller models are all reducing the number of GPU-seconds required for the same output. If that efficiency curve accelerates, the demand forecast that justifies the 'low forward P/E' will be too high. You will not need a recession to hurt NVIDIA. You will just need the inference-efficiency curve to bend faster than the capex curve. There is also a structural disconnect in the phrase 'all-in.' If Baker is truly all-in on AI infrastructure, he should own much more than NVIDIA. The best expression of the trade may be a basket: NVIDIA for training, Broadcom for networking and custom ASIC design, TSMC for manufacturing, SK Hynix and Micron for memory, Vertiv for cooling and power infrastructure, and utilities such as Constellation Energy or Vistra for the electrons. The scarcity that matters in 2026 is not just GPU die space. It is electrical power, liquid cooling, and the ability to interconnect racks without waiting three years for a grid permit. If power is the true bottleneck, the highest-beta AI trade is not NVIDIA. It is the company selling the power and the cooling. If I had Gavin Baker on a call, I would ask him one question: which part of the chain would he own first if he could only own one position? His answer would tell us more than the 'all-in' phrase. Does he pick NVIDIA, with its CUDA lock-in? Does he pick TSMC, with its manufacturing monopoly? Does he pick Vertiv, with its power and cooling exposure? Or does he pick a utility, because power is the true bottleneck? That question is more important than any P/E ratio. It forces a real thesis. So what do I watch now? First, NVIDIA's next quarterly report. I ignore the headline EPS. I want the Blackwell revenue recognition and the gross margin. If the GB200 NVL72 is shipping in volume and gross margins are holding above 70%, the near-term story stays intact. If the margin dips because CoWoS or HBM costs are higher than expected, that is the first crack. Second, hyperscaler capex language. I listen for a change in vocabulary. When a CFO says 'opportunity' instead of 'return,' they are telling you they do not have a payback formula. When they start saying 'return on AI investment,' the market will breathe a little easier. Until then, the uncertainty is priced into the multiple. Third, BIS rule changes. Export controls are the regulatory equivalent of a stablecoin losing its peg. The damage happens first in the charts, not in the press release. Any new rule on China-bound AI chips will change NVIDIA's total addressable market overnight. Fourth, TSMC's CoWoS capacity guidance for 2026. If TSMC announces another massive expansion, the supply constraint is easing. If it stays flat, the bottleneck is permanent for the cycle. Fifth, the inference demand curve. This is the most important variable for the 2026-2027 thesis. If training demand simply shifts to inference without another S-curve of new demand, NVIDIA will still be the leader, but the growth rate will slow. A slowdown in growth is enough to break the 'cheap forward P/E' argument, because the multiple and the earnings will both move down. And one more thing: the power meter. I spent years tracking validator profitability in crypto. Now I spend time looking at data-center power deals. The next bottleneck is not a secret. It is the electrical grid. Watch the utilities. Watch the liquid cooling names. Watch the companies that build the factories that hold the GPUs. That is where the AI infrastructure trade starts to look like a real portfolio. The takeaway is not 'sell NVIDIA' and it is not 'buy NVIDIA.' The takeaway is that 'decade-low forward P/E' is a slogan, not a thesis. The clock stops, but the chain doesn't. Gavin Baker may be early to the most important infrastructure cycle of his career. But the same was true for many smart people who bought the wrong part of the chain at the wrong time. NVIDIA is a great company. It is not a cheap company. It is an execution contract at an execution-friendly price. The difference matters. Liquidity flows where trust is liquid. In crypto, trust evaporates when a reserve snapshot turns out to be a lie. In AI infrastructure, trust evaporates when a hyperscaler blinks on capex or a supply chain slips by one quarter. That is the moment when the 'lowest forward P/E in a decade' becomes the 'highest forward P/E on a cut estimate.' Speed is the only currency that matters. Move with the facts, not the headlines. Verify the chain. And if you do build the AI infrastructure basket, do not forget the power grid. The GPUs can't run on vibes.