NVIDIA Vera CPU Secured by Perplexity, Targeting $200 Billion Market
0xBroomberg
AI search platform Perplexity confirmed plans to purchase Nvidia's Vera CPU at scale, becoming the sixth known bulk buyer; Nvidia is using the chip to push beyond GPU supply into a full-stack AI compute platform, entering the ~$200 billion CPU market long dominated by Intel and AMD.
What is Vera, and why isn't it just another CPU?
Traditional server CPUs — Intel Xeon, AMD EPYC — handle databases, virtualization, and web services. Vera is redesigned from scratch for AI-agent workflows: sandbox code execution, tool calls, retrieval, task scheduling, and GPU orchestration.
This means → Vera is not "a faster general-purpose chip." It is a CPU optimized exclusively for AI workloads, on a fundamentally different design path from x86.
Perplexity VP Nate Kupp said Vera runs AI-agent coding tasks roughly 1.5× faster than conventional CPUs. Nvidia's own benchmarks show 1.8× faster task completion across multiple AI-agent workloads versus x86.
Why are AI agents changing the role of the CPU?
Historically, GPUs handled the heavy math — training and inference — while CPUs just dispatched tasks. But as AI agents grow more complex, Python interpretation, database retrieval, RAG index access, and task-queue scheduling are consuming a rising share of total compute.
In plain terms = no matter how fast the GPU is, the system stalls if the CPU can't keep up — the CPU is shifting from supporting actor to system bottleneck.
This reflects a broader industry pivot from "stack more GPU power" toward GPU-CPU co-optimization — and Vera is Nvidia's direct answer to that shift.
Where is Nvidia's "rack-level integration" moat?
Nvidia packages Vera inside the Vera Rubin NVL72 — a single rack integrating 72 Rubin GPUs, 36 Vera CPUs, plus ConnectX-9 NICs and BlueField-4 DPUs, forming a complete AI supercompute unit.
Each Vera Rubin superchip carries 88 custom Arm-compatible CPU cores and 1.5 TB of LPDDR5X memory, linked via NVLink-C2C — Nvidia's proprietary chip-to-chip interconnect — at 1.8 TB/s bandwidth.
This means → Intel and AMD sell individual CPUs. Nvidia sells an entire system — GPU, CPU, networking, and memory all designed in-house. That level of integration is a moat competitors cannot replicate quickly.
How large is the cost advantage?
Nvidia says the Vera Rubin NVL72 cuts per-million-token cost for highly interactive, deep-reasoning AI agents to one-tenth of the prior-generation GB200 NVL72.
It also delivers up to 10× tokens per megawatt — cheaper and greener.
In plain terms = for data-center buyers, the same AI inference job could cost dramatically less in both electricity and compute — a hard metric that directly drives procurement decisions.
How big is the customer roster, and what comes next?
Publicly confirmed Vera CPU buyers now include Anthropic, OpenAI, SpaceX AI, CoreWeave, Oracle, Lambda, Nebius, and Nscale. Perplexity's addition brings the list to at least nine.
Nvidia management expects Vera CPU to generate roughly $20 billion in cumulative sales by fiscal year-end.
This means → the customer base spans the full AI value chain — from foundation-model training (OpenAI, Anthropic) to cloud infrastructure (Oracle, CoreWeave) to AI applications (Perplexity). Whether Nvidia delivers on the $20 billion target is the key test of whether its CPU strategy can genuinely displace Intel and AMD.
Content is for reference only, not financial advice.