Nvidia's CPU Gambit: Redefining the AI Server Value Stack

Most analysts are reading Nvidia's CPU revenue projection backwards. They see a market share grab against Intel and AMD. That's the wrong frame. The real story is about who gets to define the value distribution rules inside an AI server. And Nvidia just moved to write those rules itself. Nvidia expects its CPU business revenue to more than double by fiscal 2028, ending January 2028. The market treats this as a side note to the GPU juggernaut. It's not. This is a structural play to cement dominance over the entire AI compute stack, not just the accelerator socket. My framework for this is simple: ignore the marketing, trace the order flow. Where does the value accrue in an AI server? Currently, the GPU takes the lion's share. But the CPU is the traffic cop. It feeds data, manages the system, and increasingly, runs the inference logic. Whoever controls that intersection controls the system's efficiency ceiling. Let's break down the actual position. Nvidia's Grace CPU is an Arm-based design, fabbed on TSMC's 4N process. It's not a general-purpose x86 killer. It's a data feeder for Hopper and Blackwell GPUs. The key differentiator is NVLink-C2C, the chip-to-chip interconnect. It delivers over 900 GB/s of bandwidth between CPU and GPU. Compare that to a standard PCIe 5.0 x16 slot at 128 GB/s. That's a 7x advantage. It's not measured yet in the context of total cost of ownership, but the math is brutal. This changes the system calculus. A Grace+Blackwell pair eliminates the need for separate PCIe switches, reduces system power draw, and cuts latency. From a quant perspective, this is about capital efficiency. You're paying for a tightly coupled system that delivers more usable compute per watt and per square foot of data center space. For hyperscalers, that's the only metric that matters at scale. Now, the market context. Intel still holds roughly 40-50% of AI server CPU sockets, leveraging its enterprise legacy. AMD is at 25-30% with strong EPYC performance. Nvidia sits at 5-8%, but it's growing fast. The GB200 superchip, pairing two Blackwell GPUs with a Grace CPU, is the vehicle for this growth. Based on my analysis of supply chain data and system teardowns, the Grace CPU represents roughly 15-20% of the bill of materials for these systems. If Nvidia's CPU revenue doubles to a range of $240-320 billion by FY2028, that implies a massive ramp in unit shipments, from hundreds of thousands to millions of units. Here's the contrarian angle. The market fears AMD's MI400 series or Intel's Gaudi as competitive threats. That's the wrong battleground. The real threat to Nvidia's CPU ambitions isn't AMD or Intel. It's the hyperscalers' own silicon. AWS has Graviton. Google has Axion. These are Arm-based CPUs designed for their specific workloads. If these in-house chips prove sufficient for AI inference and data management, they could displace Grace in the very systems where Nvidia currently bundles it. This is the classic vertical integration risk. Your customer can become your competitor if you don't maintain a system-level efficiency gap that they cannot replicate in-house. But here's what the bears miss. Designing a CPU is one thing. Integrating it with a GPU over a proprietary high-bandwidth interconnect, and having the software stack (CUDA, DOCA) tuned for that specific pairing, is an entirely different beast. The hyperscalers can build a Graviton, but they can't build an NVLink-C2C equivalent overnight. That interconnect and the software integration form the real moat. It's not about the chip. It's about the system. The financial impact is nuanced. Nvidia's gross margins are around 75%. Grace CPU, with its higher cost of system integration, will dilute that slightly. Expect operating margins to compress from 62% to the high 50s by FY2028. But this is a deliberate trade. The CPU+GPU bundle increases average selling price and customer lock-in. The net effect on earnings per share should be positive because the system-level pricing power more than compensates for the margin dilution. It's a volume-for-margin swap that makes sense if you believe AI infrastructure spending continues to compound. Geopolitics adds another layer. Export controls on China affect all three players. Nvidia can't sell its high-end chips there. Neither can Intel or AMD. But the Arm architecture's perceived neutrality, relative to US-dominant x86, gives Nvidia an edge in some non-US markets pursuing sovereign AI initiatives. It's a minor tailwind, but worth noting for the 2026-2028 horizon. The risk matrix is clear. The biggest risk is an AI capex cyclical downturn. If hyperscalers cut spending, the CPU revenue projection evaporates. Second is the AMD threat, specifically if their MI400 series with integrated Infinity Fabric gains real traction. Third is the self-designed chip movement. Supply chain risk, specifically over-reliance on TSMC for 4N process and CoWoS advanced packaging, is a medium-level concern that could cause delivery delays. So, what are the signals to track? For the next two quarters, watch Nvidia's data center revenue mix. Specifically, the percentage coming from DGX/HGX systems and GB200 NVL72 rack-scale solutions. If that mix shifts meaningfully toward the full systems, the CPU narrative is playing out. Also, monitor hyperscaler announcements. If AWS or Google starts talking about Grace adoption in their fleet, that's a massive bullish signal. If they talk about their own silicon replacing it, that's the bear case. By 2027, the Rubin platform, with its Vera CPU, will tell us a lot. If Nvidia decouples Grace from GPU bundles and sells it as a standalone CPU, that's a signal they're confident in the Arm ecosystem for general-purpose workloads. That would be a direct assault on Intel's last fortress. Until then, the strategy is defensive and opportunistic. Nvidia's CPU business is not about stealing market share. It's about redefining the competitive dimension of AI servers. The new battleground isn't single-core performance. It's CPU-GPU integration density. And on that axis, Nvidia has no peer. The question isn't whether Nvidia's CPU revenue doubles. It's whether the hyperscalers let them keep the keys to the system. That's the trade to watch.