Nvidia expects CPU business revenue to more than double by fiscal 2028, ending January 2028. That is the headline. The problem: Nvidia does not disclose CPU revenue separately. The base is an estimate. The multiplier is a projection. The ledger bleeds where code is silent — and in this case, the code is the financial reporting itself.
This is not a story about Nvidia stealing CPU sockets from Intel and AMD. It is a story about redefining what a CPU is for in an AI server. The market is pricing this as a share-shift narrative when it is actually a value-chain redefinition.
Context: The AI Server CPU Chessboard
Current AI server CPU share sits at Intel 40-50%, AMD 25-30%, Nvidia 5-8%. Nvidia's Grace family — ARM-based, Neoverse V2 cores, fabricated on TSMC 4N — is not a general-purpose CPU play. It is a GPU companion, designed to feed data to Hopper and Blackwell accelerators through NVLink-C2C interconnect.
The architecture tells you everything. Grace pairs with LPDDR5X memory pushing 480GB/s of bandwidth, versus DDR5's sub-300GB/s on x86 platforms. The NVLink-C2C interconnect delivers 900GB/s+, roughly seven times PCIe 5.0 x16's 128GB/s. This is not a CPU competing on core count. It is a data pipeline component engineered for one job: keeping GPUs fed.
Nvidia's positioning is system-level integration. Fabless design, TSMC manufacturing, but the strategic leverage sits in the software stack — CUDA, DOCA, and the tightly coupled Grace-GPU ecosystem. The company is not competing for the general-purpose server market. It is defining a new category where the CPU is a peripheral to the accelerator. Intel and AMD sell CPUs that happen to sit next to GPUs. Nvidia sells a system where the CPU exists to serve the GPU.
Core: What "Doubling" Actually Means
Based on DGX/HGX system shipments and Grace's estimated 15-20% value share within those systems, CPU-related revenue lands around $40-60 billion for FY2025. Doubling implies $240-320 billion by FY2028E. That is a 60-80% compound annual growth rate, but the base is small relative to Nvidia's ~$130 billion total revenue. CPU is roughly 3-5% of the top line today. This is not a new business line. It is an augmentation of an existing monopoly.
The growth drivers are structural. GB200 and GB300 systems require Grace as a mandatory component — there is no x86 option in Nvidia's flagship rack-scale offerings. Inference workloads demand higher CPU throughput per GPU than training, and inference is where the market is heading. Sovereign AI infrastructure programs in Europe, the Middle East, and Southeast Asia are creating policy-driven demand that favors non-x86 architectures. In my experience auditing hardware supply chains, these government-driven procurement cycles are slower to materialize but stickier once locked in.
The financial mechanics deserve scrutiny. Grace CPU gross margins run below GPU margins — likely 50-60% versus the 75%+ GPU range. System integration complexity raises operating costs. Nvidia's overall gross margin will face structural pressure as CPU revenue scales. The net effect on EPS is likely positive because bundling raises average selling prices and customer stickiness. But the margin dilution narrative will resurface every earnings cycle. Volatility is the price of admission.
The technology roadmap reinforces the bundling thesis. Grace Hopper gives way to Grace Blackwell, then GB300, then the Rubin platform with Vera CPU paired with Rubin GPU over NVLink 6. Each generation tightens the CPU-GPU coupling. By 2028, the integration is likely to be a single silicon package, not two chips on a board. That trajectory makes Intel and AMD's discrete CPU approach structurally disadvantaged in AI workloads. The system-level performance-per-watt advantage of the integrated approach is estimated at 30-50% over x86-plus-GPU configurations, based on Nvidia's published data and select third-party benchmarks. Those numbers warrant skepticism — vendor-published benchmarks always do — but the architectural logic is sound.
The competitive math favors Nvidia in one specific scenario: when a customer has already committed to Nvidia GPUs, the marginal cost of adopting Grace is minimal. No PCIe switches, no system-level re-architecture, lower power and space requirements. The switching cost is lowest for the customer already locked into the CUDA ecosystem. That lock-in is the moat. It is not the chip. It never was.
Against Intel, the threat is asymmetric. Intel's Xeon Max and Gaudi accelerators have not formed a coherent AI ecosystem, and its enterprise x86 fortress is a shrinking island. AMD is the more credible threat — EPYC leads on performance-per-watt in general-purpose workloads, and the MI400 series integration with Instinct GPUs is improving. But AMD is playing the same game Nvidia defined: GPU-CPU co-design. It is a follower in a race where Nvidia sets the pace. Chaos is just unquantified variance.
Contrarian: The Real Threat Is Not Intel or AMD
The conventional reading frames this as Nvidia versus the x86 duopoly. The blind spot is customer self-designed silicon. AWS Graviton and Google Axion are already deployed at hyperscale. If the largest AI buyers conclude that custom ARM CPUs integrated with their own accelerators beat a vendor-locked stack, Nvidia's system-level bundling loses its gravitational pull. The customer who builds their own CPU is the customer Nvidia cannot lock in.
Export controls cut both ways. U.S. restrictions on advanced AI chips to China limit Nvidia's addressable market there, but they equally constrain Intel and AMD. All three lose China. The difference: ARM's perceived neutrality versus x86's American provenance gives Grace a geopolitical edge in markets seeking supply-chain diversification. In my audit experience, procurement decisions in sovereign AI programs increasingly weigh architecture provenance alongside raw performance.
The deeper contrarian read: Nvidia's CPU doubling is not about market share. It is about redefining the value allocation rules inside an AI server. If CPU+GPU integration becomes the primary competitive dimension — rather than CPU core count or single-thread performance — then Intel and AMD are competing in a game Nvidia has already rewritten.
The scenario distribution for FY2028E CPU revenue deserves explicit modeling. Bear case: $150-200 billion, triggered by AI demand cyclicality and accelerated customer self-design. Base case: $240-320 billion, assuming GB/Rubin platforms ship on schedule. Bull case: $350-400 billion, driven by inference explosion and sovereign AI procurement. My probability assignment: 25% bear, 55-60% base, 15-20% bull. That distribution is not a prediction. It is a risk framework. Trust no one, verify everything, compute always.
Takeaway: Watch the Signals, Not the Headlines
The question is not whether Nvidia doubles CPU revenue by 2028 — the base is small enough that execution risk is manageable. The question is whether Grace moves from mandatory GPU companion to standalone product. If Nvidia starts selling Grace CPUs outside GPU bundles, the competitive dynamic shifts entirely. Watch hyperscaler direct procurement, watch the Vera CPU in the Rubin platform, watch whether x86 vendors mount an effective counter in AI-specific silicon. Manual audits save what algorithms miss.
Skepticism is the only viable alpha. The ledger will tell you what the press release hides — but only if you read it with the right tools.
