The quiet numbers buried in Nvidia's earnings calls rarely make headlines. But when a company that sold roughly $130 billion worth of silicon in fiscal 2025 whispers about its CPU business doubling by fiscal 2028, the echo is seismic. I've been mapping the chaos in this industry since the Compound yield hunt days of 2020, and let me tell you โ this is not a story about a chip company selling more chips. This is a story about who gets to define the rules of AI infrastructure. And Nvidia is rewriting those rules faster than most people realize.
Let me start with the raw signal. Nvidia expects its CPU business revenue to more than double by fiscal 2028 (ending January 2028). That sounds like a headline. But the real signal is in the base: today, Nvidia's Grace CPU series contributes roughly $40-60 billion in system-attached revenue โ about 3-5% of total revenue. Doubling that means hitting $240-320 billion in CPU-attached revenue by 2028. That's not a rounding error. That's a paradigm shift hiding in plain sight.
The Old Guard Is Fighting a War They Didn't See Coming
Let me rewind. In the summer of 2020, I was dissecting Compound's interest rate models across five chains simultaneously, trying to find yield before the crowd did. I launched three Twitter threads before "yield farming" was even a term. And what I learned โ what stuck with me through the Bored Ape madness of 2021 and the Terra ashes of 2022 โ is that stories drive value, not just algorithms. Nvidia's CPU story is the same pattern: a narrative that starts as a technical footnote and becomes the dominant market force while the incumbents are still looking at their spreadsheets.
Here's the core architectural truth: Nvidia's Grace CPU is not trying to beat Intel or AMD at their own game. It's not even playing the same game. Grace is built on Arm's Neoverse V2 architecture, not x86. It's designed to feed GPUs, not to run general-purpose workloads. The entire chip is an architecture of the GPU โ a data feeder, a memory bandwidth juggler, a system-level orchestrator. Intel Xeon and AMD EPYC are general-purpose processors. Grace is a purpose-built companion.
The real numbers that matter: NVLink-C2C interconnect delivers 900GB/s+ between Grace CPU and Hopper/Blackwell GPU. That's a 7x bandwidth advantage over PCIe 5.0. When you're training a massive model, that bandwidth difference isn't a technicality โ it's the difference between a system that idles waiting for data and a system that runs at peak efficiency. Based on Nvidia's data and some third-party testing, the Grace+Blackwell combination delivers 30-50% better system-level performance-per-watt than x86+GPU alternatives. That's not a marginal advantage. That's a structural moat.
The Hybrid Cloud of AI Servers: From Card Vendor to System Integrator
Here's the narrative shift that most market observers are missing. Nvidia isn't competing in the traditional CPU market. They're redefining what an AI server even is. Think about the value chain: EDA/design โ manufacturing (TSMC) โ packaging โ system integration โ cloud/enterprise. Nvidia's core strength is at the system integration layer. By bundling Grace CPU with Blackwell GPU, they control the entire AI compute stack โ the hardware, the interconnect, and critically, the software stack with CUDA and DOCA.
Mapping the chaos to find the signal in the noise: the market share numbers tell a story of their own. Intel currently holds 40-50% of AI server CPU share, AMD 25-30%, and Nvidia only 5-8% but rising fast. Grace shipped in hundreds of thousands of units in 2024. With GB200 and GB300, Nvidia's CPU volumes will push into the millions. If the fiscal 2028 doubling happens, Nvidia could hold 20-25% of the AI server CPU market. That's not taking share from Intel or AMD โ that's creating a new category of server where the old players are already obsolete.
Here's what the incumbents are facing. Intel is trying to fight a two-front war: losing the AI CPU race to AMD on general performance and to Nvidia on system integration. Their Gaudi accelerators haven't coalesced into a coherent ecosystem. AMD's EPYC is strong on raw performance, and they're improving the Instinct integration, but they're still fighting the same architectural fight. Meanwhile, Nvidia's playbook is simple: when a customer already buys Nvidia GPUs, choosing Grace CPU saves them the PCIe switch costs, reduces system complexity, and lowers power consumption. The marginal switching cost to Grace is nearly zero โ the customer's already in the Nvidia ecosystem.
The Margin Dilution Question That Nobody Wants to Talk About
Here's the contrarian angle. The conventional wisdom is that Nvidia's CPU play is pure upside. But let me show you the financial friction. Nvidia's gross margin is around 75% โ absurd for hardware. That's software-like pricing in a silicon world. Grace CPU margins are lower โ maybe 55-65%. As CPU revenue grows to 10% of total revenue, the overall gross margin will slip to 70-73%. Operating margin could drop from 62% to 55-60%.
That's the dilution of the dream โ and yet, I see the net effect as positive. Because this isn't just about selling more chips. It's about raising the average order value and deepening the switching cost. When a hyperscaler buys a GB200 NVL72 system, they're not just buying a GPU. They're buying an entire computing architecture. The system-level integration creates a relationship that's much harder to unwind than a simple chip purchase. The stock market may initially see margin compression as a negative, but the customer stickiness and the higher ticket per unit outweigh the margin dilution. It's like when Amazon's margins dropped as they built AWS infrastructure โ the long-term annuity was worth more than the short-term margin.
The Geopolitical Chessboard: De-x86 and the Arm Neutrality
This is where it gets interesting, and where my institutional-lens forecasting kicks in. The geopolitical dimension of Nvidia's CPU business is a double-edged sword โ but it's actually a tailwind in disguise. When the US restricts advanced AI chips to China, it hurts Nvidia's sales there. But it also cuts Intel and AMD off from China's market. So all three lose China. The hidden advantage is that Nvidia's CPU is Arm-based, not x86. In many parts of the world โ especially in Asia and the Middle East โ there's a growing desire to reduce x86 dependence because x86 is perceived as American dominance. Arm's relative neutrality makes Grace a more palatable option.
The "sovereign AI" movement โ countries building their own AI infrastructure โ is a real catalyst. Japan, the UAE, Saudi Arabia, even parts of Europe are pouring billions into homegrown AI compute. And when they want to build sovereign AI without checking American monopoly, the Arm-based Grace CPU with its flexible integration becomes more attractive. Nvidia is the one selling shovels in this gold rush, and the CPU piece is the shovel handle.
The Hidden Risk: The Hyperscaler's Own Silicon Ambitions
The contrarian angle that keeps me up at night is this: the biggest threat to Nvidia's CPU business isn't Intel or AMD โ it's Nvidia's own customers. AWS has Graviton. Google has Axion. Microsoft is working on custom silicon. The hyperscalers are the biggest buyers of AI compute, and they have both the incentive and the capability to build their own CPUs.
Why would they do it? Control over their own destiny. If you're AWS, you want to design the entire stack โ you don't want to be held hostage by a single vendor that controls both the GPU and the CPU. But here's the counterpoint: the custom silicon play is expensive. It takes 3-5 years to design and tape out a competitive CPU, and the AI hardware cycle is moving so fast that by the time you ship your custom CPU, the GPU generation has already moved forward. Nvidia's advantage is that they have the best GPU, and if you want the best GPU, you have to play in their system. The lock-in is real.
From the ashes of Terra, we learned to walk โ and in the Terra collapse, we learned that over-centralized architectures create fragility. The hyperscalers learned the same lesson from Nvidia: single-vendor dependence is a risk. That's why they're all investing in custom chips. But the counter-narrative is that the custom chip effort will struggle to keep pace with Nvidia's rapidly evolving architecture, and the cost is enormous. I see the self-design threat as medium-term, not existential. Nvidia's window is 2026-2028, when the Rubin platform with its next-gen Arm CPU comes out. If they execute on that roadmap, they're going to be hard to catch.
The Metrics That Matter: When Grace Goes Independent
The short-term signal that matters more than any benchmark is this: does Nvidia start selling Grace CPU as a standalone product, without the GPU bundle? Right now, Grace is tied to the system. But if they open it up as a standalone CPU for enterprise or inference use cases, that's a major expansion of the addressable market.
The second signal to track is the adoption of GB200 NVL72 systems โ the flagship rack-scale system. If hyperscalers adopt it at scale, the CPU revenue doubles automatically. Third, watch the evolution of TSMC's CoWoS packaging capacity. This is the physical bottleneck for Nvidia's entire roadmap. If the packaging capacity expands faster than expected, Nvidia ships more systems.
When I think about the net in the crowd, I'm watching for the moment when Nvidia's CPU business revenue grows at the same rate as its GPU business. That's the signal that the "system" strategy has worked. It's also the signal that Nvidia's margin will start to compress โ and the market will have to reprice the stock. The stock market is still pricing Nvidia as a GPU company. When the market realizes it's becoming a full-stack AI infrastructure company, the multiples will change.
The Takeaway: The CPU Is the New GPU Narrative
We're in the early innings of a game where the field is being redrawn. Nvidia's CPU business doubling is not just a revenue line item โ it's a strategic declaration that AI servers are no longer about raw GPU compute. They're about the integration between the CPU and the GPU. The CPU is no longer the general-purpose master control; it's the data feeder for the GPU. This inverts the entire value chain.
Here's my prediction: by 2028, the question won't be about Nvidia's CPU market share. It will be about how Intel and AMD are responding to a new competitor that doesn't even see them as the enemy. The real enemy is the old architecture โ the separation between CPU and GPU that's been the fundamental basis of computer design for 40 years. Nvidia is obliterating that separation.
When the crowd jumps, I look for the net. The crowd is still bullish on Nvidia's GPU growth. The net is the CPU โ the quiet, ugly, necessary piece of the system that holds the whole together. If Nvidia pulls this off, they're not just the biggest chip company in the world. They're the only company that owns the entire AI compute stack โ from the sand to the software. That's a narrative that will drive value for a decade.
Hunting for the next spark in the dry brush โ this CPU story is that spark. The question is: when the market realizes it, will you be already positioned?