Nvidia used the Hot Chips 2026 conference at Stanford University to roll out its most ambitious hardware lineup yet, anchored by a new CPU architecture built specifically for the kind of multi-step reasoning and tool-calling that defines agentic AI. The Vera CPU, Rubin GPU, BlueField-4 DPU, and Spectrum-X networking solutions were presented not as standalone products but as a co-designed stack, each layer optimized to work with the others.
The message from Nvidia’s August 24 presentation was clear: agentic workloads are different enough from traditional AI inference that they deserve their own silicon.
Vera CPU: 88 cores built for agents
The star of Nvidia’s Hot Chips showing was the Vera CPU, built around 88 custom Olympus cores. It pairs those cores with LPDDR5X memory capable of delivering up to 1.2 TB/s of bandwidth, a figure that matters because agentic workloads tend to be memory-hungry beasts. When an AI agent orchestrates multi-step reasoning, calls external tools, and processes data streams simultaneously, memory bandwidth often becomes the bottleneck before raw compute does.
Nvidia claims a 1.8x improvement in task completion time on agentic workloads compared to competing x86 CPUs. That’s a bold number, and it comes from internal benchmarks, so the usual caveats apply. But even the more modest comparison points are notable: Nvidia says the Vera architecture compiles the Linux kernel 14-22% faster than AMD’s 96-core EPYC 9655P.
The architectural secret sauce includes statically partitioned spatial multithreading and a 164 MB L3 cache. In practical terms, the spatial multithreading approach lets the chip dedicate specific hardware resources to specific threads rather than sharing them dynamically, reducing contention and improving predictability when an AI agent needs to juggle dozens of parallel operations.
The co-design philosophy
Nvidia didn’t present the Vera CPU in isolation. The conference sessions also covered the Rubin GPU, BlueField-4 DPU, and Spectrum-X networking, all framed as components of a unified platform where hardware decisions at one layer inform design choices at every other layer.
The BlueField-4 DPU handles data processing and security offloading at the network edge, freeing the CPU and GPU to focus on computation rather than data movement. Spectrum-X provides the networking layer that connects everything at scale.
SpaceXAI partnership and orbital ambitions
Perhaps the most eyebrow-raising announcement was a partnership with SpaceXAI to deploy Vera CPUs at scale for workloads related to Grok. The plan extends beyond terrestrial data centers: Nvidia and SpaceXAI are targeting an orbital version of the Vera Rubin NVL72 for deployment on the Starmind satellite by Q4 2027.
What this means for the competitive landscape
The Hot Chips presentation puts direct pressure on AMD, whose Venice CPU line is the explicit benchmark Nvidia is measuring against. A claimed 1.8x advantage on agentic workloads, if it holds up under independent testing, would represent a meaningful gap in a market segment that’s growing rapidly.
Hot Chips has historically been a venue for showcasing forward-looking architectures that take 12-18 months to reach volume production. The SpaceXAI deployment timeline of Q4 2027 suggests the Vera architecture is on roughly that schedule for at least some configurations.