One chipmaker to rule them all

TL;DR: During its GPU Technology Conference keynote yesterday, Nvidia announced the Vera CPU (yes, you read that right). It’s a new chip designed not for the training work that made it a $4.5 trillion company, but for the kind of inference work that AI is moving toward. This marks Nvidia’s most direct challenge to Intel and AMD yet, and it’s a sign that the AI chip race may no longer be winnable on GPUs alone.

What happened: Clad in his trademark leather jacket, Jensen Huang walked onstage at GTC yesterday to declare: The AI “inference inflection” is here. As the industry shifts away from training and towards practical use—ChatGPT alone runs about 2.5 billion prompts a day—the GPU titan is answering with a new 88-core CPU that’s built to excel at inference.

By data center standards, 88 cores is modest—compare it to the 288 in Intel’s forthcoming Clearwater Forest and the 256 in AMD’s EPYC Venice. But Vera is promising to eke out more AI performance per core, and it can communicate with Nvidia's GPUs through a proprietary, high-speed link that other chips can't tap into, accelerating the overall workload. Vera is also tailored specifically for agentic AI, which is a highly compute-intensive slice of inference. The CPU is part of a broader seven-chip platform called Vera Rubin, which also includes next-gen GPUs, networking chips, and the Groq 3 LPU, which is designed for extremely fast inference. All together, Nvidia is signaling that it’s diversifying its offerings beyond the beefiest AI chips that reign supreme for training.

The CPU pivot: GPUs can handle a lot of tasks simultaneously, which is ideal for training on massive datasets. Agentic AI workloads, though, benefit from a processor that can order tasks logically—putting on shoes after pants, essentially—and that’s where CPUs shine. Agentic workloads are also far more token-heavy, and that high demand is outpacing what current CPUs can handle.

The year of the AI agent: AI is entering its errand-running era, shifting from chatbots that answer questions like “what’s the cheapest flight to London” to agents that can book the trip themselves. Since last year, OpenAI folded its Operator agent into ChatGPT; Anthropic shipped Claude Code and Cowork, and Microsoft launched Copilot Cowork across Microsoft 365—while smartphone makers are increasingly integrating on-device AI agents. Gemini’s agentic features have been rolling out to new Samsung phones, and Apple's long-awaited Siri overhaul is expected later this year.

Nvidia’s competition heats up: Nvidia's GPUs still dominate training, but inference is a far more crowded fight. Intel holds about 60% of data center CPU market share. AMD has roughly 24%, and Nvidia sits at just 6%. Google, Amazon, and Meta are all developing custom chips to cut their Nvidia dependence—Meta's latest, a family of four MTIA processors, is aimed squarely at inference. Nvidia has taken notice: It spent $20 billion last December to license inference tech from specialized AI chipmaker Groq (one of its founders helped design Google’s TPUs).

Bottom line: Nvidia’s CPU expansion is a notable about-face from a company that has long insisted GPUs could handle all of AI's needs—though it's still betting big on them, projecting $1 trillion in orders across its GPU-led Blackwell and Vera Rubin platforms through 2027. The play for Nvidia, it seems, is to become the one-stop AI shop: CPUs, GPUs, networking, and software—no other chipmakers required. —WK