Why CPUs Are Now at the Center of the AI Race

[Daniel Nenni]

US chipmakers lead, but China players aim to increase local market share

The AI hardware race is shifting beyond GPUs. While graphics processors remain essential for dense tensor computation, CPUs are becoming strategic because modern AI systems are increasingly heterogeneous, distributed and agentic. Large language models do not only perform matrix multiplication; they retrieve data, call tools, manage context, schedule workflows, run containers, coordinate memory movement and orchestrate thousands of concurrent inference tasks. These control-plane functions depend heavily on CPU performance, memory bandwidth, cache hierarchy, I/O lanes and tight integration with accelerators.

This is why US chipmakers are repositioning CPUs as core AI infrastructure. Nvidia’s Vera CPU is designed for “agentic AI” and reinforcement-learning environments, extending the company beyond GPUs into full rack-scale systems that combine CPUs, GPUs, networking and DPUs. Reuters reported that Perplexity plans to use Nvidia’s Vera CPU, with the chip positioned for always-on AI agent workloads. Intel and AMD are also emphasizing server CPUs as AI platforms, pairing high core counts, DDR5/HBM memory support, chiplet packaging and accelerator interconnects with GPUs.

China’s opportunity lies in localization. US export controls have constrained access to advanced Nvidia GPUs, increasing pressure on Chinese cloud providers, AI labs and state-backed buyers to adopt domestic compute stacks. Reuters reported that Nvidia has pitched Vera CPUs to Chinese clients, partly to recover ground in a market hit by export restrictions and China’s self-reliance push. Meanwhile, Huawei, Alibaba, Baidu and DeepSeek-linked efforts are pushing local AI chips and server ecosystems. China still trails in advanced process nodes, high-bandwidth memory and software maturity, but domestic demand gives local CPU and accelerator vendors a path to increase market share.

Why CPUs are now at the center of the AI race

US chipmakers lead, but China players aim to increase local market share

asia.nikkei.com

 

[Daniel Nenni]

Why did we not see this coming? I bet ChatGPT knew.

 

[Jozo035]

China has long history of domestic CPUs since 90s and early 2000s with MIPS or SPARC, regularly being on top of TOP500 list since 2010s. They often don't rely on accelerators and use CPU performance for compute even if they have to install significantly more nodes (and deal with scalability issues).

 

[Paul2]

China has long history of domestic CPUs since 90s and early 2000s with MIPS or SPARC, regularly being on top of TOP500 list since 2010s. They often don't rely on accelerators and use CPU performance for compute even if they have to install significantly more nodes (and deal with scalability issues).

Highly dependent on the task. Many bulk computing tasks can't be turned into a lot of bulk numeric operations.

Picking SIMD resistant hashes, generating petabyte sized rainbow tables, NLP tokenizing all unencrypted wechat, http, and smtp traffic in China, detecting obfuscated/steganography VPN connections at multiple terabytes per second, automated vehicle tracking from space, submarine hunting from radar/hydrophone network data, etc

 

[Jozo035]

Highly dependent on the task. Many bulk computing tasks can't be turned into a lot of bulk numeric operations.

Picking SIMD resistant hashes, generating petabyte sized rainbow tables, NLP tokenizing all unencrypted wechat, http, and smtp traffic in China, detecting obfuscated/steganography VPN connections at multiple terabytes per second, automated vehicle tracking from space, submarine hunting from radar/hydrophone network data, etc

Sure, there are still applications like that, and companies/organizations are still building mainframes... although its interesting to put system like that on TOP500 list.

 

[Brady]

I work on agentic systems day in and day out.

How I see this is that while the massive training runs and the heavy inference tasks for generative content will continue to grow, it will be optimized to kingdom come and back.

On the other hand, “Agentic AI” is basically “software but with an LLM in there”, and the sheer amount of business logic and software out there I feel is so much bigger than the amount of content being created. For instance, imagine the number agent turns and conversation steps executed per piece of content created; it will be 100-to-1 in many cases.

So I generally agree that agentic is going to explode bigger than “GenAI content”, and so the amount of central logic orchestrating it will summarily explode.

Long live the CPU.