Here's a paradox that sums up the AI chip market right now: Nvidia is launching a brand-new processor specifically designed to compete with Intel, yet it is handing Intel a lucrative contract for its flagship data centre servers. The contradiction reveals something important about how the technology landscape is shifting.
Intel announced at Nvidia GTC 2026 in San Jose that its Xeon 6 processor will serve as the host CPU in Nvidia's DGX Rubin NVL8 systems, extending the x86 pairing the two companies established with the Xeon 6776P in current DGX B300 Blackwell-based platforms. It is a win for Intel at a critical moment. But the real story is why Nvidia felt compelled to build its own CPU at all.
For years, Nvidia dominated artificial intelligence through its graphics processing units. Chips like the H100 and its successor Blackwell became the industrial backbone of the AI boom, with companies burning through cash to acquire them. Yet something has shifted. According to Nvidia's head of AI infrastructure, CPUs are becoming the bottleneck in terms of growing out AI and agentic workflow. This is not a marginal problem. The emergence of agentic AI, which requires systems to reason, plan, and execute actions across multiple steps, demands something GPUs alone cannot deliver: rapid CPU performance to orchestrate, schedule, and manage the expensive GPU resources.
The host CPU is responsible for task orchestration, memory management, scheduling, and data movement to GPU accelerators. With inference workloads shifting toward agentic AI and reasoning systems, those functions place increasingly heavy demands on per-core performance and memory bandwidth.
This is where Nvidia's new Vera CPU enters the picture. Nvidia announced 88-core Vera data centre CPUs at GTC 2026, claiming impressive 50 per cent performance gains over standard CPUs, fuelled by a 1.5x increase in IPC from its Olympus cores. The company also unveiled a CPU-only rack containing 256 Vera processors, claiming a 6x gain in CPU throughput and twice the performance in agentic AI workloads.
The financial stakes are substantial. Mercury Research estimates the server CPU market share in the last quarter of 2025 was dominated by Intel at 60 per cent, AMD at 24.3 per cent, and Nvidia at 6.2 per cent, with the remaining share split among in-house Arm-based CPUs from hyperscalers like Amazon, Microsoft and Google. Nvidia is not starting from a position of strength in the traditional CPU market. Yet hyperscalers are already committing. Amazon Web Services, Anthropic, Black Forest Labs, Cisco, Cohere, CoreWeave, Cursor, Dell Technologies, Google, Harvey, HPE, Lambda, Lenovo, Meta, Microsoft, Mistral AI, Nebius, Nscale, OpenAI, OpenEvidence, Oracle Cloud Infrastructure, Perplexity, Runway, Supermicro, Thinking Machines Lab and xAI are expected to adopt Rubin.
Yet why select Intel for the NVL8 system when Nvidia is building its own CPU? The answer reveals institutional inertia and practical constraints. Intel Xeon processors are used as the host CPU for DGX Rubin NVL8 systems due to their capability to support fast memory speeds, balanced performance across a range of workloads, lower long-term total cost of ownership, and their mature, enterprise-proven software ecosystem. In other words, Intel won because its architecture is entrenched. Existing AI software stacks expect x86 instruction sets. Retraining workloads to run on Arm-based Vera introduces risk and cost.
Still, Nvidia is playing the long game. The evolution of the Vera CPU and its integration into deployable rack-scale systems marks Nvidia's entry into direct CPU sales, positioning itself as a competitor to Intel and AMD in the traditional CPU market. The company is essentially hedging its bets. For customers who need to maintain architectural compatibility with legacy systems, the Intel partnership persists. For hyperscalers building custom data centres from scratch, Nvidia's Vera offers a chance to optimise the entire stack around AI workloads rather than accept compromises baked into decades-old x86 design.
Nvidia designed its CPU specifically to help its star GPUs run AI workloads. Single-threaded performance becomes much more important than dollars per core because the goal is to ensure that very expensive GPU resources do not sit idle. This represents a fundamentally different philosophy from Intel and AMD, which optimise for general-purpose computing across many use cases.
The market will decide which approach wins. Intel maintains the advantage of software compatibility, cost certainty, and proven reliability. Nvidia offers optimisation for a very specific (and increasingly lucrative) use case. The real competition, however, may come from the custom in-house CPUs that Amazon, Google, Microsoft, and Meta are already building. Those chips, designed for internal use and optimised for proprietary workloads, face no pressure to satisfy external customers and no legacy constraints. They may prove faster, cheaper, and more efficient than anything Nvidia, Intel, or AMD can offer.
For now, the CPU renaissance in AI infrastructure is a win for everyone building chips. The question is whether that market can sustain multiple serious competitors, or whether, like the GPU market before it, consolidation will eventually favour one or two dominant players.