Word Count: ~3,650
Hook: The Quiet Number Buried in Nvidia's Forecast
The headline number from Nvidia's latest investor guidance is not about GPUs. It's about CPUs. The company expects its CPU business revenue to more than double by fiscal year 2028, ending January 2028. That sounds like a modest target for a company already generating over $130 billion annually. But look closer.
The current CPU revenue base is roughly $4-6 billion per year. Doubling from a small base is easy math. The real signal is in the implied trajectory: reaching $24-32 billion by FY2028 requires a compound annual growth rate of 60-80%. That is not incremental growth. That is a strategic pivot from GPU vendor to full-stack AI compute platform.
I have spent the last 26 years watching hardware cycles, and I have learned one thing: when a component vendor starts bundling the motherboard, the memory subsystem, and the interconnect into a single SKU, they are not selling chips anymore. They are selling an architecture. Nvidia's CPU business is not about winning the benchmark wars against Intel Xeon or AMD EPYC. It is about redefining who controls the AI server.
The market has not priced this in. Most analysts still treat Nvidia as a GPU company with a CPU accessory. The data says otherwise.
Context: The Grace CPU and the Architecture Trap
Let me be explicit about what the Grace CPU is. It is an ARM-based processor, using the Neoverse V2 core design, manufactured on TSMC's 4N process node. It has 72 cores, LPDDR5X memory, and critically, an NVLink-C2C interconnect that provides over 900 GB/s of bandwidth between the CPU and the GPU.
Compare that to the standard PCIe 5.0 x16 interface, which delivers 128 GB/s. That is a sevenfold difference in bandwidth. In AI systems, memory bandwidth is the bottleneck. The CPU needs to feed data to the GPU as fast as possible; otherwise, the GPU sits idle waiting for data. NVLink-C2C turns the CPU from a general-purpose controller into a data feeder for the GPU. That is not a trivial difference. That is a new compute paradigm.
Tracing the noise floor to find the alpha signal.
In the current AI server market, the CPU is not the compute engine. The GPU is. The CPU's job is to manage data flow, orchestrate memory, and handle I/O. This is where Nvidia's Grace CPU has an inherent advantage. When you already have a system with 8 GPUs connected via NVLink, adding a Grace CPU with native NVLink-C2C means you can eliminate PCIe switches, reduce system power consumption, and save physical space. The marginal switching cost for a customer already using Nvidia GPUs is near zero.
Now, I want to be clear about something. The Grace CPU is not a general-purpose server CPU. It does not compete with Xeon in legacy workloads. But in the AI server market — the fastest-growing segment in the entire semiconductor industry — it is not competing on CPU performance. It is competing on system-level efficiency.
3. The System-Level Play: Why Nvidia Is Not Competing with Intel or AMD
Here is a critical insight: Nvidia is not trying to replace Intel's x86 server business. It is creating a new category — the AI server compute node.
Consider the total cost of ownership (TCO) for a GB200 NVL72 system. That is a rack containing 72 Blackwell GPUs and 36 Grace CPUs, all interconnected. When you compare this to a traditional x86-based AI server, the Nvidia system can deliver 30-50% better system-level performance per watt. That comes from the elimination of PCIe bottlenecks, tighter memory integration, and the ability to run the entire AI software stack — CUDA, DOCA, and the Grace CPU's software — as a single integrated unit.
But here is the counterintuitive part. Nvidia's CPU margin is lower than its GPU margin. Grace CPU is probably a 60% gross margin product, while the GPU is at 75% or above. If you look at the revenue mix, adding more CPU business will dilute Nvidia's overall gross margin. In the FY2028 scenario, with CPU revenue reaching $25-30 billion, that will be about 10% of total revenue. The margin dilution would be maybe 2-3 percentage points.
Now, why would Nvidia deliberately dilute its margins?
Because the alternative is worse. If Nvidia does not control the CPU, it has to integrate with Intel or AMD x86 CPUs, which brings PCIe bottlenecks, additional system complexity, and lower system-level performance. Nvidia is willing to sacrifice a few margin points on the CPU to create a sticky, system-level lock-in. This is the same playbook that Intel used in the 1990s with the Pentium Pro and the PCI bus.
The deeper logic is this: the customer is not buying a CPU. The customer is buying a system. The CPU is a cost, but the system-level performance is a differentiator. Nvidia is moving from a component vendor to a systems vendor. The company is now in the same category as Dell, HPE, and Lenovo, except with a 75% gross margin.
4. Competition: The Real Threat is AMD, Not Intel
Let's talk about the competitive landscape. In the AI server CPU market, Intel still has 40-50% share. AMD has 25-30%. Nvidia has 5-8%, but growing rapidly.
Intel is the incumbent, but the AI transition has been painful. The Xeon roadmap has not kept pace with AI workloads. The Gaudi accelerator has not created a cohesive ecosystem. Intel's x86 stronghold remains in the enterprise and general-purpose cloud, but the incremental AI server market is slipping away.
AMD is the more credible threat. The EPYC server CPU has leading performance-per-watt in general-purpose compute. AMD also has the Instinct GPU (MI300/MI400 series), which is increasingly competitive. But AMD is still missing a tight CPU-GPU interconnect. The Infinity Fabric is a good inter-chip interconnect, but it does not provide the same bandwidth as NVLink-C2C. And the software stack is still not as mature as CUDA.
What about the custom silicon threat? AWS Graviton and Google Axion are Arm-based CPUs designed for specific workloads. These are a real competitive risk. But the design cycle is 3-5 years, and the economies of scale for a custom CPU only work if you have a massive, standardized workload. For most AI workloads — especially large language model inference — the GPU is the dominant compute element, and the CPU is peripheral.
Nvidia's advantage is the system integration. When a customer buys Nvidia GPU, adding the Grace CPU has the lowest marginal switching cost. When a customer buys an AMD GPU, adding an AMD CPU still requires PCIe and non-integrated memory. Nvidia's total system efficiency is the moat.
5. The Contrarian Angle: CPU Is Not the Business, the Interconnect Is
This is where I diverge from the mainstream analysis. Everyone is looking at the CPU revenue numbers. The real value of the Grace CPU is not the CPU itself. It is the interconnect lock-in.
NVLink-C2C is not just a CPU-to-GPU interconnect. It is a proprietary protocol that only works with Nvidia GPU. By integrating the CPU into the GPU ecosystem via NVLink, Nvidia is creating a closed loop. You cannot buy a Grace CPU and use it with AMD Instinct GPU. You cannot buy a Grace CPU and connect to a Google TPU. The Grace CPU is only useful when paired with Nvidia GPUs.
This has a strategic implication. Nvidia is not just selling a CPU. The company is using the CPU as a gateway drug for the entire ecosystem. If you buy Grace, you are committing to Nvidia's GPU roadmap, NVLink roadmap, and CUDA software stack. The system-level control is the real revenue driver, not the CPU.
But here is the risk. If customers start demanding open standards — and the open accelerator ecosystem continues to mature (e.g., UALink, PCIe 6.0, or CXL) — then the NVLink-C2C lock-in could become a liability. A customer that buys Grace CPU but later wants to switch to AMD GPU for a specific workload will face a forklift upgrade. This is a double-edged sword.
The second hidden risk is software. Grace CPU runs on Arm architecture. While the software ecosystem is mature enough for server use, the AI deployment stack is still not 100% portable. Many AI workloads are written for x86 (with AVX-512 instructions, for example) and do not run natively on Arm. Nvidia has done a lot of work with the CUDA and DOCA stacks, but the transition cost is non-trivial. The CPU roadmap for Grace is tied to the GPU roadmap. If the GPU roadmap slips — e.g., if the Rubin platform is delayed — then the CPU roadmap is also delayed. Nvidia is coupling its CPU and GPU roadmap, which means a single point of failure.
6. The Bear Case: Financial Dilution and the Hyperscaler Pushback
Let me walk through the financial implications. In FY2025, Nvidia's revenue is approximately $130 billion. The CPU-related revenue is $40-60 billion, which is about 3-5% of total revenue. If the CPU business doubles by FY2028, the revenue will be $25-32 billion. The total revenue will be $250-300 billion. The CPU will be about 10% of total revenue.
The gross margin impact is what matters. The GPU gross margin is around 75-80%. The CPU gross margin is around 60-65%. If you blend in 10% CPU revenue, the overall gross margin will drop from 75% to 72-73%. That is a 2-3% dilution.
But the operating margin will also decline. A system-level integration means more manufacturing complexity, more inventory, more logistics, and more service costs. This is a shift from a fabless chip company to a systems integrator. The operating margin will drop from ~62% to 55-60%.
Is this a good trade-off? It depends on the price elasticity of the demand. If the system-level performance gain — the 30-50% efficiency improvement — is real, then the customer will pay a premium. The net effect on EPS should be positive, as the customer is paying for the system value, not the CPU. The CPU is the cost of entry into the Nvidia ecosystem.
But there is a real risk in the hyperscaler pushback. Google has TPU, and Amazon has Trainium. Microsoft has Maia. Meta is exploring its custom chip. If the hyperscalers decide that the GPU+CPU system from Nvidia is too expensive and too closed, they can push for custom accelerators. The Grace CPU does not help in that case.
The market share projection is a 20-25% share in AI server CPUs by 2028. That is an aggressive target. But there is a path: if the GB200 and GB300 systems sell at volume, and the Rubin platform (Vera CPU + Rubin GPU) scales, then the CPU revenue could hit $20 billion. However, the uncertainty is high. The 55-60% probability for the base case is reasonable, but the tail risks are significant.
7. Geopolitics: The x86 Erosion and the Arm Advantage
There is a geopolitical dimension that most financial analysts ignore. The AI server CPU market is not just a technology battle. It is also a geopolitical battle.
Nvidia's Arm-based CPU has an inherent advantage over x86 in certain regions. For example, in China, the US export controls have restricted Nvidia's advanced GPU sales. But the Grace CPU, when bundled with the GPU, is also subject to the same restrictions. The China market is effectively closed for Nvidia's top-tier AI systems. That is a headwind.
However, the Arm architecture provides a political neutrality that x86 does not. In the Middle East, Southeast Asia, and Europe, there is a growing demand for "sovereign AI" — the ability to build AI infrastructure without relying on a single US-based architecture. Arm-based CPUs, in this case, are seen as a neutral alternative. Nvidia's Grace CPU can be framed as an "Arm-based solution" that is not subject to the x86 dominance of the US. This is a subtle but real advantage.
But there is a counter-risk. Arm is now controlled by SoftBank, and there are geopolitical risks around Arm's licensing. If the US were to apply export controls on Arm architecture, Nvidia's CPU business would be vulnerable. In this scenario, RISC-V would be a backup, but the ecosystem is not mature enough for high-performance server CPUs.
The Taiwan issue is also critical. Grace CPU is manufactured on TSMC's 4N process. If the Taiwan Strait conflict escalates, the supply chain is severely disrupted. Nvidia's fallback options (Samsung or Intel foundry) are technically inferior for this level of integration.
8. The Forward-Looking Judgment: The System is the Product, and the CPU is the cost of entry
So what is the bottom line?
Nvidia's CPU revenue doubling is not a hardware story. It is a system strategy. The Grace CPU is the glue that binds the GPU, the interconnect, and the software stack. The CPU is the entry ticket to the Nvidia system. The customer is not buying a CPU. The customer is buying the highest-performance AI system on the market.
The implication for the industry is significant. For Intel and AMD, the threat is not in the CPU core count or performance-per-watt. The threat is the system-level integration. When the GPU and the CPU are fused into a single NVLink domain, the CPU becomes a commodity. The system becomes the product. This is the same pattern that Intel used in the 1990s to win the PC market: integrate the chipset, the CPU, and the motherboard, and you control the architecture.
The question is whether this strategy will work in the AI era. The market is evolving fast, and the custom-silicon trend is real. But the GPU is still the bottleneck. The GPU is the rare compute element. If Nvidia controls the GPU, the CPU is just an additional line of defense.
The real signal to track is not the CPU revenue. It is the attach rate of the Grace CPU to Nvidia GPU. If the attach rate is above 60%, Nvidia's CPU is a core part of the system. If the attach rate is below 20%, the CPU is an optional accessory, and the CPU business is not a real moat.
For the next 3 quarters, I am watching the GB200 NVL72 shipment volumes. If the volumes hit 100,000 units in the next 12 months, the CPU revenue will be around $10 billion. The doubling is on track. If the volumes miss, the CPU business will not reach the FY2028 target.
The biggest risk to the bull case is not AMD or Intel. It is the hyperscalers. If Google, AWS, and Microsoft all commit to custom accelerators and custom CPUs, the system-level lock-in will not be broken. But the AI inference market is growing so fast that the demand for general-purpose AI servers will be high enough for Nvidia to capture.
Takeaway
The Nvidia CPU business is a system-level power grab. The CPU is the bait, the interconnect is the hook, and the system is the catch. The company is moving from a GPU vendor to a full-stack AI platform. The CPU revenue is the signal, but the system control is the value.
The market is still pricing Nvidia as a GPU company. The transition to a system-level vendor will be a slow, gradual process, but the implications are profound. If Nvidia controls the CPU-GPU interconnect, it controls the AI server architecture. This is the new battleground for AI hardware.
Will Intel or AMD be able to respond? They have the CPU expertise, but they do not have the GPU. The AI server is a GPU-centric system. The CPU is a supporting role. If Nvidia keeps the CPU-GPU interconnect, the company will own the architecture.
The message is clear: build first, ask questions later. The AI server is the new system. Nvidia is the new system vendor.
The data is in the interconnect. The profit is in the system.
Analyst Notes
- The FY2028 CPU revenue target of $25-30B is plausible but requires a successful GB200/GB300 ramp.
- The GPU attach rate is the key indicator to monitor.
- The threat from AMD is real but requires a better CPU-GPU interconnect.
- The geopolitical factors are not the primary drivers in the near-term.
- The base case probability is 55-60%. The bear case is 25%. The bull case is 15-20%.
Code does not lie, but it does hide. The CPU is the hidden code in the Nvidia system.
Volatility is the price of entry, not the exit. The AI market is volatile, but the system-level architecture is the stable foundation.
Tracing the noise floor to find the alpha signal. The signal is not the CPU. It is the system.