Here is what happened over the past few quarters, and I almost missed it. I was so focused on the GPU order flow—the Blackwell shipments, the CoWoS packaging bottlenecks—that I ignored the quiet number buried in Nvidia's latest guidance. The one that says CPU revenue could more than double by FY2028. That is not a product update. That is a structural re-rating of who captures value in the AI server market.
We have seen this movie before in crypto. A project pivots from a single narrative to a full-stack ecosystem. The market initially discounts the expansion, then prices it as a moat. Nvidia is not just selling chips anymore. It is selling the server itself, the interconnect, and the software that makes the whole thing sing. The question is: do we treat this like the DeFi yield trap of 2020, where the complexity hid the risk, or like the early institutional integration of 2025, where the complexity was the security?
Context: The Quiet System Integration
Let's establish the base. Nvidia's CPU business, the Grace series, sits in the fabless design layer. But the strategic play is system-level integration. They are not trying to beat Intel on single-core speed. They are redefining the CPU's role as the GPU's data feeder. In an AI server, the CPU is not the master anymore. It is a high-speed orchestrator for the GPU.
This is a fundamental shift from the traditional x86 world. My audit of the Golem network back in 2017 taught me the same lesson: the narrative is always ahead of the technical reality. Nvidia's narrative is that a tightly coupled CPU+GPU system beats a generic CPU with a discrete GPU. And the technical reality, based on the data, is that they are right—on system-level performance-per-watt.
Today, Nvidia's CPU share in AI servers is roughly 5-8%, but it is rising fast. The GB200 and GB300 platforms are forcing Grace CPUs into the mainstream because they are not optional. It is the necessary glue. This is not about stealing Intel's enterprise base; it is about owning the incremental AI server market. The current context is that the AI hardware market is splitting into two: the general-purpose legacy and the accelerated-computing future. Nvidia is building the infrastructure for the latter.
Core: The System-Level Margin Analysis
Let me walk you through the financials, using my own modeling framework. Nvidia does not disclose CPU revenue separately, so I have to triangulate. In FY2025 (ending January 2025), the CPU-related revenue—mostly from DGX/HGX systems—was likely in the $40-60 billion range? No, that's the whole data center revenue. Wait, let me correct. The CPU value within those systems is about 15-20%. So, the CPU-specific contribution is around $6-10 billion. But the guidance says the business doubles. That implies a base of $40-60 billion in CPU systems, or a standalone base of $240-320 billion by FY2028? That math is wrong.
Let me re-frame. The report states that if the revenue doubles, the CPU business could hit $240-320 billion. That implies the current base is substantial, not the $40-60 billion in CPU-related revenue. Actually, the report is clear: the base is $40-60 billion in CPU-related revenue, and doubling that means the system-level revenue. The key metric is that the CPU contributes about 10% of total revenue by FY2028. That is the core insight. They are not a CPU company; they are a systems company.
The real technical analysis is on the memory and interconnect. LPDDR5X memory offers 480GB/s of bandwidth, while the standard DDR5 is under 300GB/s. That is a 60-100% advantage. But the game-changer is NVLink-C2C, which delivers 900GB/s+ compared to PCIe 5.0's 128GB/s. That is a 7x bandwidth advantage. In my copy trading community, I would call this a 'high leverage order flow.' The system is not just faster; it is structurally more efficient at the system level.
The core insight: Nvidia is not selling a CPU. They are selling a memory architecture that makes the GPU never starve. This is the biggest data point from my analysis. The financial impact is that the gross margin will dilute, dropping from ~75% to ~70-73%, because system integration is more expensive than a single chip. But the EPS effect is positive because the bundle pricing locks in the customer.
Contrarian: The 'Self-Cannibalization' Myth
Everyone is focused on how Nvidia is attacking Intel and AMD. But the contrarian angle is that Nvidia is eating its own ecosystem. By pushing the Grace CPU, they are cannibalizing the traditional x86 server market they relied on for the host CPUs. But here's the blind spot: the real threat is not Intel or AMD. It is the customer's self-sufficiency.
AWS has Graviton. Google has Axion. These are in-house ARM-based CPUs. If they scale, they will squeeze Nvidia's system-level margin. But based on my 2020 DeFi yield trap experience, I know that speed and simplicity often beat complexity. The cloud vendors are realizing that building a custom CPU is expensive and slow. Nvidia's Grace is the 'ready-made' solution. It has lower switching costs because it is already paired with the GPU. The contrarian view is that the hyperscalers will abandon their custom chips and just buy the Nvidia bundle.
But I think the opposite is true. The 'sovereign AI' trend will drive more hyperscaler diversification. Yet, Nvidia's advantage is that they are not just selling hardware. They are selling 'trust' in the system. As I always say, Trust is the only asset that survives the crash. And in the AI hardware space, the trust in the integration is the asset.
Takeaway: The New Value Distribution
What does this mean for the market? The AI server's value is shifting from the 'chip' to the 'system.' The CPU is not a substitute for the GPU; it is the amplifier. The 2028 target is not a sales forecast. It is a declaration that Nvidia wants to own the entire AI rack, from the power delivery to the interconnect to the memory.
We walk away from greed, we stay for trust. The profit and the margins are in the integration, not the raw components. The message to the flock is simple: watch the system-level efficiency, not the single-core speed. The real competition will be on the 'integrated efficiency' and the software stack. The future is a system that is not just faster, but more transparent about its own efficiency. Transparency is the shield against the next bubble. We don't need a new GPU. We need a new way to value the GPU. That is the new rule.