NVIDIA reports earnings on August 28. The headline number everyone will quote: forward PE of roughly 21x. That is a hardware company trading like a utility. The last time this name sat at these multiples, the market was convinced the AI trade was dead.
Let me state the obvious first. A 75% gross margin in hardware is not normal. TSMC runs at 55-60%. AMD lives at 50%. Intel is bleeding toward 40%. NVIDIA is sitting at software-company margins while selling physical silicon. That single number tells you more about their market position than any narrative about the AI trade.
Here is what the market is pricing in: growth deceleration. The 21x forward multiple is not the market being generous. It is the market saying the >100% year-over-year growth rates are about to hit a wall. Sara Alagic at M Science framed it exactly that way - investors are looking past this quarter to next year's slowdown. That view might be wrong.
The Blackwell story is not just a chip upgrade. It is a pricing event. The report states that NVIDIA is raising server prices by over 15% starting in early 2027. That is not passing through cost inflation. That is a demonstration that their pricing power has not eroded. When a company can raise prices 15% on a product line that already has a 36-52 week lead time, they are not at the mercy of the business cycle. They are the cycle.
Dig into the supply chain. Blackwell uses TSMC's 4NP process node. That's a customized version of a 5nm-class process. It is not the bleeding edge of Moore's law. The real bottleneck is not transistors. It is CoWoS packaging. TSMC's advanced packaging capacity is running at over 100% utilization. NVIDIA alone takes roughly 60% of that capacity. When you control 60% of the critical packaging resource in AI compute, your competitive moat is not just design architecture. It is physical infrastructure.
HBM3e memory costs are rising. SK Hynix is the primary supplier and their capacity is tight. NVIDIA's answer? Raise prices. They are not absorbing the cost. They are passing it downstream. In my years auditing token systems, I learned one universal truth: when a project can pass cost increases through without losing volume, that is a monopoly signal. NVIDIA has it.
The Numbers That Matter
Let me do the math that the market seems to be ignoring. The article states NVIDIA has a PEG ratio of around 0.5. For those who need a refresher: a PEG below 1 means the market expects growth to collapse. A PEG of 0.5 with a 75% gross margin is either a market inefficiency or a warning that the earnings model is about to break. I side with the former.
Free cash flow is strong. Operating cash flow at $28 billion against net income of $300 billion? That ratio is healthy. The balance sheet has over $30 billion in cash. They have the capacity for aggressive buybacks. The article indicates NVIDIA has authorization for roughly $15 billion in additional repurchases. That is a floor under the stock, not a ceiling.
The Contrarian View
Now let me argue against myself. There are three real risks that the 21x multiple is trying to price in.
First, the customer concentration. The top five customers - Microsoft, Meta, Amazon, Google, Oracle - represent over 40% of revenue. That is not an inherent problem when demand exceeds supply. It becomes a problem when those customers decide to vertically integrate. Google has TPUs. Amazon has Trainium. Microsoft has Maia. These are not jokes. They are legitimate ASICs designed for inference workloads. In 2026-2027, if inference becomes the dominant workload, NVIDIA's dominance in training chips does not automatically translate.
Second, the export control issue. China accounted for roughly 25% of NVIDIA revenue in FY2023. It has dropped to below 10%. The H20 chip is a hamstrung product. This is a structural loss that is not coming back. The market knows this. The question is whether the US/Europe AI demand can fully offset it. So far, the answer is yes. But the geopolitical risk premium is real.
Third, the supply chain dependency. TSMC for manufacturing, SK Hynix for HBM, CoWoS for packaging. Three external points of failure. If Taiwan gets complicated, NVIDIA faces a 6-12 month production gap. If HBM allocation shifts, same story. The article mentions that NVIDIA is working with Samsung and TSMC's Arizona fab for diversification. But those are years away from meaningful volume.
The Real Play
Here is the conclusion. The market is pricing NVIDIA like a hardware vendor that will eventually face commodity pricing. That has been the story for every hardware company since the semiconductor industry began. But NVIDIA is not selling chips. They are selling an AI infrastructure platform - hardware plus CUDA software plus NVLink interconnect plus system-level rack solutions.
I have audited enough code to know that the CUDA moat is real. It is not just a software library. It is a lock-in mechanism. Every AI developer who has been trained on CUDA is less likely to switch to AMD ROCm or a custom ASIC. That switching cost is not in the PE ratio.
The 21x multiple is a gift if the growth rate stays above 30% through 2025. The market is pricing in a slowdown to the low-teens. The data does not support that. CSP capex budgets in 2024 exceed $200B combined. The order book for Blackwell extends through the end of 2025. The inference market is just beginning to scale.
I do not need to be the smartest person in the room. I just need to read the data. The data says NVIDIA is an infrastructure monopoly at a cyclical low multiple. That is the definition of an asymmetric trade.
When the code bleeds, only the ledger survives. In this case, the code is Blackwell, and the ledger is TSMC's CoWoS line. Both are full.
The gas war taught me that speed is a tax. NVIDIA charges it.
Yield is the shadow cast by risk taken. The market is taking the risk of AI slowdown, and the yield is 21x earnings on a company growing at triple digits.
I do not trust whispers; I trust verified hashes. The hash here is the P&L statement. It is green.
Chaos is just data waiting for a ledger. This earnings call will be the ledger. Watch the Blackwell revenue line and the margin guide. If both are solid, the 21x multiple will look absurd by Q1 2025.
Migrations are just purgatory for lazy capital. Move early or get left at the 40x gate.
The Takeaway
The market has already priced in a slowdown. The question is whether that slowdown actually arrives. The evidence - the 15% server price increase, the 75% gross margin, the 60% CoWoS capacity lock, the CUDA ecosystem - all suggest the slowdown is not coming in 2025. The actual risk is 2026, when the CSP ASICs hit scale. Trade accordingly. The current valuation offers margin of safety for that risk.
Note that the data suggests that if NVIDIA's earnings beat expectations, the multiple could easily re-rate to 30x, implying 40% upside. The asymmetry is attractive.
The only question left is whether you can stomach the volatility. If the answer is yes, the ledger is in your favor.
