Nvidia expects its CPU business revenue to more than double by fiscal 2028 (ending January 2028). That single sentence, buried in an earnings call, signals something far larger than a product roadmap update. This isn't about taking market share from Intel or AMD in the traditional server CPU market. It's about redefining what an AI server is — and who controls its economics.
The System-Level Play
Let me be clear about what Nvidia is actually doing here. The Grace CPU series isn't competing in the general-purpose computing market. It's designed to feed GPUs. The entire architecture — Arm Neoverse V2 cores, LPDDR5X memory optimized for bandwidth over capacity, and NVLink-C2C interconnect — is built around a single objective: maximize the throughput of the GPU next to it.
The market share numbers tell the story. In 2024, Nvidia held an estimated 5-8% of AI server CPU shipments. Intel still dominates at 40-50%, with AMD at 25-30%. But these figures are rearview mirror data. The GB200 superchip, which pairs Grace CPUs with Blackwell GPUs, is the real inflection point. When hyperscalers deploy GB200 NVL72 racks, they're not buying a CPU — they're buying a tightly coupled compute system where the CPU exists to serve the GPU.
The competitive logic is simple: when a customer has already committed to Nvidia GPUs, the marginal cost of switching to Grace CPUs is essentially zero. No additional PCIe switches, no system re-architecture, no software stack changes. The CUDA ecosystem already supports it. This is the lowest-friction path to CPU adoption the semiconductor industry has ever seen.
Bandwidth as the Battleground
The technical differentiation is striking. A standard PCIe 5.0 x16 link provides roughly 128 GB/s of bandwidth. NVLink-C2C delivers over 900 GB/s — a sevenfold advantage. In AI workloads, where the CPU is responsible for data feeding and tensor operations, this bandwidth differential is not a marginal improvement; it's a different category of performance.
In my experience analyzing AI infrastructure deployments, system-level performance per watt is the metric that matters most to cloud operators. Nvidia claims 30-50% system-level efficiency gains for Grace + Hopper/Blackwell combinations versus x86 alternatives. Third-party testing has largely confirmed these figures. This is the moat.
The technical roadmap follows a predictable trajectory. Grace Hopper (GH200) → Grace Blackwell (GB200/GB300) → the Rubin platform with Vera CPU + Rubin GPU and NVLink 6. Nvidia's CPU roadmap is now permanently fused to its GPU roadmap. The integration level between CPU and GPU becomes the core competitive dimension — not single-core performance.
The Financial Mechanics
The revenue numbers require some unpacking. Nvidia doesn't disclose CPU revenue separately, so estimates are based on DGX/HGX system shipments and the CPU's value share in those systems (roughly 15-20%). My calculations suggest the current base is $40-60 billion... correction — $4-6 billion. That's 3-5% of total revenue.
To reach $24-32 billion by FY2028 implies a compound annual growth rate of 60-80%. That's aggressive but not unrealistic if GB200/GB300 systems scale as planned. The bigger story is the margin structure. Grace CPU gross margins are structurally lower than GPUs. As CPU revenue grows to ~10% of total revenue, expect a gross margin dilution from 75% to roughly 70-73%.
The net effect on EPS, however, should be positive. System-level bundling increases average selling prices and customer stickiness. The integration complexity adds cost, but the pricing power is strong. This is the paradox: Nvidia's CPU business will lower its margin profile while strengthening its market position.
The Real Competition
AMD is the more immediate threat than Intel. EPYC processors lead in general-purpose performance, and AMD's Instinct GPU integration is improving. Intel's x86 fortress in enterprise servers remains defensible short-term, but the incremental AI market is slipping away.
The reality is that Nvidia is not fighting Intel or AMD for their existing markets. It's creating a new category — the AI-optimized system server — where the CPU is subservient to the GPU. In the AI server market, Nvidia doesn't need to beat Intel's Xeon on general-purpose benchmarks. It needs to make the question irrelevant.
Geopolitical Complexity
The export control situation adds layers of complexity. American restrictions on high-end AI chips to China limit Nvidia's CPU market in China — but they equally restrict Intel and AMD. The logic creates an interesting dynamic: non-x86 architectures gain geopolitical appeal in markets wanting to reduce dependence on American-dominated technology stacks. Arm's relative neutrality — despite SoftBank's ownership — provides Nvidia with a positioning advantage in Europe and the Middle East.
The Taiwan dependence remains the structural vulnerability. TSMC's 4N process and CoWoS advanced packaging are critical. A Taiwan strait crisis would be catastrophic for Nvidia's production. Samsung as a backup option remains a technical compromise.
What to Track
For the next 1-3 quarters, watch the data center revenue breakdown — specifically the share of DGX/HGX/GB200 systems. Pay attention to GB200 NVL72 adoption rates. Monitor AMD's MI400 series reception and Intel Gaudi 3's market traction. The real signal: whether any hyperscaler purchases Grace CPUs for non-GPU-bound workloads.
The longer-term question is more interesting. If Nvidia's CPU business doubles by 2028, it won't be because Nvidia stole 20-25% of the AI server CPU market. It'll be because the AI server itself has evolved into something fundamentally different — and Nvidia was the only vendor positioned to define its architecture. The question is whether Intel and AMD can adapt to a world where CPU performance is no longer the primary metric.