Nvidia's CPU Doubling Is a System Play, Not a Chip Story

The market is reading Nvidia's CPU revenue projection all wrong. They see a chip company trying to steal server share from Intel and AMD. That is not what is happening. The ledger shows a different mechanism: Nvidia is not selling CPUs. It is selling a tighter coupling between compute and memory that makes the GPU faster. The CPU is the feed pipe. The GPU is the engine. And the pipe is about to double in revenue by fiscal 2028. That is not a product launch. That is a structural shift in how AI servers are built and priced. Let me start with the numbers, because the base is smaller than most people assume. Nvidia does not break out CPU revenue separately. Based on DGX and HGX system shipments, the value share of the Grace CPU within those systems, and supply chain data, I estimate the current run rate at $40-60 billion annually. That is roughly 3-5% of total revenue. Doubling that by FY2028 implies a target of $240-320 billion. The implied compound annual growth rate is 60-80%. That is not a mature business scaling. That is a new category being born. The context matters. The AI server market today is dominated by x86 CPUs from Intel and AMD, with Nvidia holding only 5-8% share. But the growth vector is not in the legacy rack-mount server space. It is in the new class of AI-optimized systems where the CPU is subordinated to the GPU. In that segment, Nvidia is not a challenger. It is the incumbent. The GB200 NVL72 system, which pairs Grace CPUs with Blackwell GPUs over NVLink-C2C, is the product that defines this category. And it is shipping in volume. The core insight is about interconnect, not cores. The Grace CPU is built on Arm's Neoverse V2 architecture, which is competitive but not superior to x86 on raw integer performance. The memory subsystem uses LPDDR5X, delivering over 480GB/s of bandwidth, which is 60-100% higher than standard DDR5. But the real differentiator is NVLink-C2C, which provides 900GB/s of CPU-to-GPU bandwidth. That is seven times the bandwidth of PCIe 5.0. In an AI workload, where the CPU is constantly feeding data to the GPU, that bandwidth advantage translates directly into higher utilization and lower latency. The CPU is not the bottleneck. That is the entire point. I have audited enough systems to know that the software stack is where the moat gets built. CUDA is the obvious layer, but DOCA and the Grace software stack extend the integration. When a customer buys a GB200 system, they are not buying a CPU and a GPU. They are buying a tightly coupled compute fabric where the software is pre-optimized for the hardware. The marginal cost of switching to an AMD EPYC or Intel Xeon in that context is not just the chip price. It is the cost of re-optimizing the entire software stack. That friction is the real barrier to entry. The contrarian angle is that this is not a zero-sum game against Intel and AMD. Nvidia is not trying to take over the general-purpose server market. The x86 fortress in enterprise legacy workloads is not under immediate threat. What is happening is more subtle. Nvidia is redefining the value distribution within the AI server. In a traditional server, the CPU is the master and the GPU is an accelerator. In a Grace-Blackwell system, the CPU is a data feeder and the GPU is the master. The value shifts from the CPU to the system integration. That is why the revenue can double without Nvidia taking significant share from Intel or AMD in their core markets. The growth is coming from a new category, not from cannibalizing an existing one. The financial impact is nuanced. The Grace CPU has a lower gross margin than the GPU, so the mix shift will dilute Nvidia's overall gross margin from around 75% to 70-73% by FY2028. But the system-level pricing power compensates for that. When a customer buys a GB200 NVL72, they are paying for the integrated system, not the individual components. The customer stickiness increases because the system is harder to replace than a single chip. The net effect on EPS is likely positive, even if the margin percentage declines. This is a classic systems play, not a component play. There are risks. The biggest one is a cyclical pullback in AI capital expenditure. If the hyperscalers cut their capex budgets, the CPU revenue projection falls apart. The second risk is AMD's MI400 series, which is a credible competitor in the AI accelerator space. If AMD can match Nvidia's system-level integration, the Grace CPU advantage narrows. The third risk is customer self-design. AWS Graviton and Google Axion are already in production, and if the hyperscalers decide to build their own CPU-GPU interconnects, the market shrinks. But the design cycle for a custom interconnect is long, and Nvidia's NVLink is already two generations ahead. The geopolitical layer adds another dimension. The US export controls on high-end AI chips to China have created a strange dynamic. Nvidia cannot sell its best systems to China, but neither can Intel or AMD sell their best x86 chips. The Chinese market is effectively frozen for all three. But in other regions, particularly the Middle East and Southeast Asia, the Arm architecture has a perceived neutrality that x86 lacks. Sovereign AI initiatives in those regions are more willing to adopt non-x86 solutions. That is a tailwind for Grace CPU adoption that is not captured in the standard competitive analysis. What I am watching is not the headline revenue number. I am watching whether Nvidia starts selling the Grace CPU as a standalone product, decoupled from the GPU bundle. If that happens, it signals a move into the general-purpose market. If it does not, the CPU remains a system component, and the revenue doubling is just a function of GPU system growth. The distinction matters for valuation. A system component is priced as part of a bundle. A standalone CPU is priced against Intel and AMD. The former is a monopoly position. The latter is a competitive market. The key signal to track is the hyperscaler direct procurement data. If AWS or Azure starts buying Grace CPUs in volume without the GPU bundle, that is a game changer. If they only buy the full GB200 system, the CPU is just a pass-through component. The second signal is the Vera CPU in the Rubin platform, which is expected to be a significant upgrade over Neoverse V2. If Vera delivers a 2x performance improvement, the standalone CPU argument becomes stronger. The third signal is the CoWoS packaging capacity. If TSMC can expand advanced packaging capacity fast enough, the supply constraint lifts and the system volume can grow. My base case is that the CPU revenue doubling happens, but it is not the story the market thinks it is. It is not a victory over Intel and AMD. It is a redefinition of the AI server architecture. The CPU becomes a component of a larger system, and the value accrues to the system integrator. Nvidia is the system integrator. The ledger remembers what the ego forgets: the revenue is real, but the competitive dynamics are more complex than a simple market share shift. Alpha hides in the friction of chaos, and the friction here is the software stack and the interconnect. Code does not lie, but it does obfuscate. The CPU revenue number is the obfuscation. The system-level margin is the truth. The question for the next two years is not whether Nvidia's CPU revenue doubles. It is whether the doubling creates a durable moat or just a temporary bundling advantage. If the software stack and the interconnect remain proprietary, the moat holds. If the industry standardizes on a common interconnect, the moat erodes. Silence in the order book is louder than noise. The order book for GB200 systems is loud. The question is what the order book looks like in 2027, when the Rubin platform ships and the competitive response from AMD and Intel is fully formed. That is the trade. That is the position. Everything else is noise.