It Was Jensen Huang's 'Vera' That Solved OpenAI's CPU Shortage
The solution turned out to be the CPU. Since the arrival of OpenAI's next-generation agentic AI 'GPT-6 Astra,' attention across the semiconductor industry has shifted beyond graphics processing units (GPUs) to the central processing units (CPUs) that handle control — and Nvidia CEO Jensen Huang's proprietary CPU 'Vera' has emerged as the answer. With AI moving beyond simple conversation to an 'agent' model (AI that performs tasks on the user's behalf) that directly executes computer operations, the burden on the CPU, which governs the entire system, has surged far beyond that on the GPU, which handles computation.
On the 7th, Huang said on X that 'GPT-6 Astra was trained on a fleet of roughly 100,000 Nvidia Grace Blackwell NVLink72 systems,' adding, 'In just four years, we have gone from ChatGPT through o1 to Astra,' before declaring, 'AGI has finally arrived,' and congratulating the OpenAI team. The heart of running large language models (LLMs) has long been the GPU, which handles massive matrix multiplications, but Astra is a different animal. It is an autonomous agent that carries out strings of computer-control tasks — managing operating system environments, driving web browsers, compiling code inside sandboxes (virtual execution environments), and invoking tools across multiple steps. Because such input/output control and serial system processes fall to the CPU rather than the GPU, the longer an agent runs, the more exponentially its CPU workload grows.
The rise of agentic AI has lifted the strategic value of Nvidia's next-generation superchip platform, 'Vera Rubin.' When CPU bottlenecks emerge, cloud service providers (CSPs) running data centers have typically bought additional x86 server CPUs from Intel or AMD and paired them with their GPU servers — a route that runs through the PCIe pathway, a serial interface for shuttling data between components, producing data latency and wasted power. Nvidia bypasses this by packaging its own Vera CPU and the next-generation Rubin GPU into a single package linked by an ultra-high-speed interconnect (NVLink-C2C). Nvidia stresses that this yields advantages in data transfer speed and power efficiency over the previous approach of combining CPUs from outside manufacturers.
Because AI agent infrastructure demands enormous memory bandwidth and capacity on both the GPU and CPU sides, this architectural shift also ripples through the domestic memory chip industry. The Rubin GPU will carry HBM4 for large-scale parameter inference, while the Vera CPU will come with LPDDR5X supporting real-time virtual machine operation and context caching. Samsung Electronics and SK hynix hold the lead in HBM4 supply, and they are also armed with mass-production competitiveness in the LPDDR5X market — which is expanding from mobile into servers — as well as the LPCAMM2 module market.
Lee Young-jin, a global AI/software researcher at Samsung Securities, noted that 'researchers used to face a cap on the number of experiments they could manage at once, but with agents taking over coding and monitoring, the human time bottleneck is disappearing.' He added, 'As the constraint that determines research speed shifts from humans to hardware, the importance of computing resources will only grow.' He cited OpenAI's GPU allocation practices as evidence: even when GPU allocation for a particular model was cut by 59.2% under a temporary research restriction, about 85% of the reduced computing was quickly reallocated to experiments on other model families. 'Given the robust latent demand for AI infrastructure, hardware investment will continue to expand beyond simple token processing toward reliably handling long-running agent workloads,' the researcher said.
