Karl Freund

The rise of Agentic AI, capable of setting goals and executing multi-step tasks, is fundamentally altering AI infrastructure requirements. Unlike traditional generative AI, these proactive systems demand ultra-low latency and consume far more tokens, pushing the CPU from a mere support role to a critical orchestrator of complex workflows. This shift is expected to quadruple CPU core demand in data centers, driving a need for highly efficient, hybrid computing across devices, edge, and cloud. Qualcomm is strategically positioning itself to capitalize on this, leveraging its mobile and edge strengths. Its "DragonFly" platform, an "efficiency-first" inference solution for data centers, promises superior performance per watt through advanced memory and liquid cooling, aiming to make agentic AI affordable and scalable despite Qualcomm's late entry into the data center market.

Qualcomm is gearing up to transform itself into an Agentic AI Infrastructure company. We look into what that means, and its upcoming DragonFly AI Server chip

The Dragonfly has its first test flight at the Qualcomm Investor DayQualcomm

Cadence is rapidly deploying "Super Agents" to offer agentic AI support, now with AuraStack, for Printed Circuit Board and advanced multi-chip packaging design.

AI stock prices are surprisingly falling, despite analysts consistently raising earnings and price targets, leading to over $2.3 trillion lost in chip stocks since late June. Analysts remain bullish, citing robust revenue growth and conservative valuation metrics, exemplified by KeyBanc's $1750 target for Micron based on 2027 earnings. Conversely, the market is fearful, with investors protecting gains amid concerns over demand peaking, competitive threats, and potential CAPEX declines. Historical data reveals similar divergences in the past, often resulting in significant stock declines or prolonged recovery periods. This creates a "wall of worry," forcing investors to weigh fear against the industry's exciting potential.

Nvidia's new Vera CPU is strategically positioned to capture the $200 billion server CPU market, extending Nvidia's leadership beyond GPUs. Designed for agentic AI, Vera boasts re-engineered, fast single-threaded cores and LPDDR5 memory, significantly boosting efficiency and keeping GPUs busy. Its monolithic architecture provides exceptional memory bandwidth and single-core performance, leading to surprising gains in diverse applications. Perplexity sees 1.5-1.9X improvement in agentic AI, while NYSE achieves a six-fold P99 latency reduction, and Los Alamos reports 3-7X better scientific simulation performance. This broad appeal projects Vera to generate $20 billion in revenue in 2026, solidifying Nvidia's comprehensive data center compute dominance.

Nvidia's Vera CPU, now in volume production, is strategically positioned to capture a significant share of the $200 billion server CPU market. Re-engineered from its Grace predecessor, Vera excels in agentic AI workloads with its fast single-threaded cores and LPDDR5 memory, efficiently keeping GPUs busy. Beyond AI, its exceptional performance, memory bandwidth, and power efficiency are attracting diverse applications. Perplexity reports 1.5-1.9X gains, NYSE achieves a six-fold P99 latency reduction, and Los Alamos sees 3-7X better scientific simulation performance. Analysts project Vera could generate $20 billion in revenue for Nvidia in 2026, reinforcing its comprehensive data center compute dominance.

Infleqtion will deploy one of the world's first fault-tolerant, NVIDIA NVQLink-integrated neutral-atom quantum computer in Illinois by 2027.

The AI industry is witnessing a profound shift in inference processing, moving beyond a single accelerator model. "Disaggregated inference" is emerging, driven by the realization that prompt ingestion (prefill) and token generation (decode) are distinct workloads with different physics and bottlenecks. This specialization, crucial for production AI, allows for optimized hardware, leading to better utilization, lower costs, and faster response times. NVIDIA is pairing GPUs with LPUs, while AMD and Cerebras combine high-throughput systems with wafer-scale decode. Intel focuses on orchestration, and Google optimizes its TPUs. This new era demands proprietary silicon and comprehensive system design, making inference a "systems business" where owning the full stack, not just a chip, ensures competitive advantage.

While EDA incumbents like Cadence and Synopsys integrate AI, three startups are vying for market adoption, which could also portend a new pricing model