NVIDIA shipped its first Vera CPUs to three leading AI research labs—Anthropic, OpenAI, and SpaceX AI—in late January 2025, followed by Oracle Cloud Infrastructure, marking the chipmaker's first foray into custom CPU design and a significant departure from its quarter-century GPU-centric strategy. The Vera initiative represents NVIDIA's direct answer to the emerging market for agentic AI inference, where autonomous systems must perform repeated reasoning tasks, decision trees, and memory lookups at vastly lower cost per token than traditional large language model serving. According to NVIDIA CEO Jensen Huang at Dell Technologies World, Vera-based inference delivers costs at roughly one-tenth the per-token expense of GPU alternatives, while agent sandbox workloads run 50% faster on Vera compared to traditional CPUs and enterprise data queries execute 3x faster. The arrival at Anthropic and OpenAI—both building frontier agentic systems—signals these organizations view custom silicon as essential infrastructure for next-generation AI deployment.
Vera's technical specifications and market positioning remain partially obscured by NVIDIA's limited public disclosure. The company has not released detailed die size, core count, memory bandwidth, TDP, or instruction set architecture (ISA) details, leaving independent verification of performance claims to third-party analysis. Agentic AI inference differs fundamentally from batched LLM token generation: agents require rapid context switching, sparse memory access patterns, and low-latency decision branching rather than the sustained compute throughput GPUs excel at. This workload mismatch explains NVIDIA's pivot—GPUs' massive parallelism wastes power and transistors on sequential agent reasoning. However, competitive pressures loom: AMD's MI325X and custom silicon efforts from Meta, Google, and others are targeting identical inference markets. Intel's Gaudi accelerators and emerging CPU alternatives from startups add further pressure. Industry analysts note NVIDIA must demonstrate meaningful volume production, competitive pricing, and software ecosystem support (particularly CUDA compatibility or parity) to defend market share in this nascent segment.
Broader timing and logistics remain unclear. NVIDIA has not announced Vera's general availability date, production capacity, or enterprise pricing, making it difficult to assess whether this represents a genuine second product line or a limited-run experiment. The shipments to three AI labs suggest engineering validation rather than commercial launch. NVIDIA's parallel strategy—maintaining aggressive GPU scaling with Blackwell and beyond while simultaneously launching Vera—indicates the company views agentic and generative workloads as distinct markets requiring different silicon. Industry observers question whether NVIDIA can manage two competing architectures and ecosystems simultaneously, particularly given CUDA's entrenched position. The moves announced at GTC Taipei and Google I/O underscore NVIDIA's determination to own every layer of AI infrastructure: from data center training (Hopper, Blackwell GPUs) through foundational inference (GPUs) to specialized agentic reasoning (Vera). Success depends on delivering Vera to market within months, not years, and proving cost and performance advantages that justify developer migration away from GPU-centric workflows.