AI search field leader Perplexity adopts Vera CPU! Amid the surge of intelligent entities, NVIDIA aims for a $200 billion general computing battlefield

According to the Zhitong Finance APP, Perplexity AI, an artificial intelligence startup focused on AI large model-driven search engines, confirmed on Tuesday local time that it plans to massively utilize NVIDIA's (NVDA.US) exclusively developed data center central processing unit (i.e., data center CPU - Vera CPU). Currently, this global company with the highest market value and AI chip super giant is striving to expand its market share and challenge Intel (INTC.US) and Advanced Micro Devices (AMD.US), two super giants in the x86 architecture CPU field with deep-rooted foundations in data center CPUs.

NVIDIA is leveraging the unprecedented demand for AI inference computing power and the CPU+GPU demand driven by the Agentic AI (i.e., AI agents) craze, actively upgrading itself from the "super dominator in the AI GPU field" to the "AI data center CPU+GPU+high-performance network infrastructure+rack-level AI computing power system" supplier in the AI era.

With the heavy launch of Claude Cowork by Anthropic and super AI agent tools like OpenClaw that can autonomously execute tasks expected to explode in 2026, the wave of AI agents is rapidly sweeping the globe. The bottleneck in AI computing architecture is shifting from GPUs centered on matrix multiplication throughput to data center CPUs focused on control flow, task orchestration, and memory/IO coordination, leading to a severe supply-demand imbalance for high-performance CPUs aimed at ultra-large-scale AI data centers.

As AI agents gain popularity worldwide, the main line of AI computing investment is shifting from "single-point computing power competition centered around GPUs" to "full-stack computing systems driven by AI agents." The next round of excess alpha returns will no longer solely belong to the strongest leaders in the AI GPU/AI ASIC fields but will systematically spread to the full-stack AI computing infrastructure layers, including high-performance CPUs for data centers, DRAM/NAND/HBM storage, AI PCBs, liquid cooling systems, data center optical interconnect systems, ABF substrates/glass substrates, MLCCs, electronic fabrics, and extensive wafer foundries. In this narrative shift of AI, data center CPUs, optical interconnects, and storage chips may emerge as the biggest winners.

From GPU Dominator to Data Center CPU Challenger, the Wave of AI Agents Reshapes Data Center CPU Demand

NVIDIA's management recently stated that it expects its "Vera" CPU products to generate approximately $20 billion in cumulative sales by the end of this fiscal year; compared to its AI-specific AI GPU series products, this is a more general-purpose computing chip. As AI companies like OpenAI, Anthropic, and DeepSeek focus on developing their own AI training/inference acceleration chips, the Vera CPU chip is an important part of NVIDIA's efforts to diversify its sales.

NVIDIA is entering a fiercely competitive data center CPU market, long dominated by Intel and AMD, which supply high-performance CPUs for various devices ranging from laptops to network server infrastructure However, many x86 architecture CPU chips were designed before the rise of so-called artificial intelligence "agents," which can autonomously execute extremely complex workflow tasks after receiving instructions from human users.

AI agents, capable of autonomously performing various tedious and complex tasks, are likely to be the ultimate trend in AI applications over the next decade. The emergence of AI agents signifies a shift in artificial intelligence from being an information assistance tool to a highly intelligent productivity tool, which is why Anthropic's valuation has surpassed $1 trillion, exceeding that of OpenAI.

Unlike human CPU users who take breaks between tasks, AI agents do not rest for a moment. Nate Kupp, Vice President of Computing Enterprises and Infrastructure at Perplexity, stated that NVIDIA's exclusively designed CPU executes AI agent coding tasks at a speed approximately 1.5 times that of traditional CPUs.

Kupp remarked in an interview, "Vera has truly impressed us because it aligns very precisely with many of our core AI inference workloads."

Founded in 2022, Perplexity has gained popularity among global developers and researchers for its real-time, efficient, and accurate AI outputs and inference results, along with a user interface similar to a search engine. The AI applications provided by the company can conduct in-depth research and can call upon AI large models beyond its own, including those from OpenAI and Anthropic PBC.

Perplexity declined to disclose the scale of NVIDIA CPUs it plans to purchase. NVIDIA previously revealed that OpenAI, Anthropic, and Oracle plan to make large-scale purchases of its Vera CPUs. Reportedly, the main buyers of the Vera CPU AI labs and cloud vendors listed by NVIDIA include Anthropic, OpenAI, SpaceXAI, CoreWeave, Oracle Cloud Infrastructure, Lambda, Nebius, and Nscale.

NVIDIA, the AI chip leader, enters the data center CPU territory! Intel and AMD face a new front in the data center.

The core advantage of NVIDIA's Vera CPU compared to traditional x86 server CPUs from Intel and AMD is not simply that "general-purpose CPU benchmarks are stronger," but rather that it redesigns the role of CPUs in AI factories for AI agent workloads. Traditional x86 server CPUs have long served general tasks such as databases, virtualization, web services, and enterprise applications; Vera, on the other hand, is aimed at sandbox code execution, tool invocation, retrieval, data processing, task scheduling, and GPU orchestration within the agent loop.

NVIDIA claims that Vera can achieve task completion 1.8 times faster than x86 across various AI agent workloads; Perplexity also states that Vera executes AI agent coding tasks about 1.5 times faster than traditional CPUs. This indicates that Vera truly penetrates the "GPU generates the next action, CPU executes the action and feeds back the result" agent closed loop, rather than the traditional enterprise server replacement market In the intelligent agent link, a large amount of workload is not only consumed in token generation on GPUs but also in CPU-dominated processes such as Python interpretation, web scraping, database retrieval, RAG index access, lexical processing, task queue scheduling, RPC/IPC communication, and KV state updates. This means that what determines user experience is increasingly not the peak computing power of a single GPU, but whether the CPU has sufficient core count, thread concurrency, cache hierarchy, memory bandwidth, and PCIe/CXL/interconnect scheduling capabilities to support high-frequency tool calls and high-density task switching. Once the CPU cores, memory subsystem, or I/O scheduling are insufficient, even if the GPU has nominal computing power, utilization will collapse due to data preparation, task coordination, and system waiting.

Therefore, it is undeniable that the bottleneck of AI computing architecture is shifting from GPUs centered on matrix multiplication throughput to data center CPUs focused on control flow, task orchestration, and memory/IO coordination. The root of this change lies in the essential migration of workload paradigms. The CPU is no longer just a general-purpose computing chip but has become the control plane processor, system orchestration engine, and resource scheduling hub of the intelligent agent era. The statement that "underestimated CPUs have become the new bottleneck for AI" is not an emotional judgment but an inevitable result of AI workloads evolving from "inference computing problems" to "complex system engineering problems."

From the perspective of hardware platforms, Vera's advantages are concentrated in high bandwidth, low latency, GPU collaboration, and energy efficiency density. NVIDIA disclosed that the Vera Rubin NVL72 integrates 72 Rubin GPUs, 36 Vera CPUs, ConnectX-9 network cards, and BlueField-4 DPUs into a rack-level AI supercomputing system. A single Vera Rubin Superchip contains 88 self-developed Olympus Arm-compatible CPU cores, 1.5TB LPDDR5X CPU memory, and provides 1.8TB/s bandwidth through NVLink-C2C.

Compared to traditional server architectures where x86 CPUs collaborate with GPUs via PCIe, Vera's value lies in packaging CPUs, GPUs, DPUs, networks, and memory into a unified AI factory system, reducing data transport bottlenecks and increasing token output per megawatt. NVIDIA claims that the Vera Rubin NVL72 can reduce the cost of highly interactive deep inference intelligent agent AI to one-tenth per million tokens compared to the GB200 NVL72, achieving up to 10 times the token output per megawatt. Jensen Huang recently stated in an interview that Vera allows NVIDIA to enter a new serviceable market for CPUs worth approximately $200 billion.

Among Wall Street analysts, the most optimistic target price for NVIDIA reaches $500, set by Baird senior analyst Tristan Gerra, maintaining an "outperform" rating. If the stock price reaches $500, it corresponds to a market value of approximately $12.2 trillion (current market value is about $4.8 trillion), implying about 154% upside potential from the current stock price. Baird's core bullish logic is not merely betting on a surge in GPU demand but believes that NVIDIA is expanding from an "AI GPU supplier" to a full-stack AI computing infrastructure platform provider at the AI factory level It is emphasized that the Vera CPU opens up a new $200 billion growth opportunity in the CPU market