Skip to content

Back to the catalog

Photon Capital (@PhotonCap) on X
x.comthread

Photon Capital (@PhotonCap) on X

Photon Capital's thread covers the 27 October 2026 NVIDIA and Oracle announcement of general availability of Vera Rubin NVL72 and Vera CPUs on OCI. It lists specs (Spectrum-6 at 102.4Tbps, HGX Rubin NVL8 with Xeon 6, 150,000 initial Vera CPUs) and OpenAI's use of Vera CPUs for agentic sandboxes. The commentary argues CPU capacity can cut cost per task and shorten KV cache residency in GPU memory.

TL;DR Thread summarises OCI's Vera Rubin NVL72 and Vera CPU launch and argues CPU capacity for agent tool execution can lower cost per task.
Open original ↗
Document

What it claims

On 27 October 2026, NVIDIA and Oracle announced general availability of Vera Rubin NVL72 racks and standalone Vera CPUs on Oracle Cloud Infrastructure (OCI). The plan is to grow Vera Rubin-based OCI Superclusters to multi-gigawatt scale. Oracle gave an initial supply of 150,000 Vera CPUs, available in select regions this quarter. OpenAI says it is moving Vera CPUs on OCI into production to run code and tools inside agentic sandboxes.

How it works

NVL72 links 72 GPUs per rack over NVLink. Between racks, OCI uses Spectrum-X built on Spectrum-6, plus BlueField-4 DPUs. Spectrum-6 switches 102.4Tbps and supports both pluggable optics and co-packaged optics (CPO).

OCI also plans an x86 option, HGX Rubin NVL8, which links eight Rubin GPUs over NVLink. Its planned host is two 64-core Intel Xeon 6 CPUs with 4TB of DDR5 and 6.4Tbps of RDMA networking. OCI Kubernetes Engine (OKE) scales CPU and GPU resources independently, so the CPU-to-GPU ratio can be set per workload.

Why CPU capacity matters for agents

Photon Capital argues that when an agent's CPU work (tool calls, code execution) blocks its next model call, adding CPU can lower cost per task more than adding GPUs. It also argues that shorter tool waits shorten how long a KV cache sits in GPU memory, which ties CPU speed to inference memory efficiency. Photon Capital says both benefits still need to be checked against real GPU utilization and agent completion times.

Limits

The NVL8 is a planned product, and its specifications are planned, not shipped. Its NVLink domain stops at eight GPUs, so jobs spanning more GPUs depend on the 6.4Tbps RDMA fabric rather than NVLink. Spectrum-6 supports CPO, but the announcement does not confirm that OCI uses CPO or any specific optics supplier. The announcement gave no total expansion capacity, CAPEX figure or completion timeline for the multi-gigawatt build-out. Regional availability is described only as "select regions," with no regions named. The thread gives no pricing, benchmarks or measured agent-workload results.

Organisations named here

  • NVIDIA
  • Oracle
  • Oracle Cloud Infrastructure
  • OpenAI
  • Intel
  • Photon Capital
  • Spectrum-6
  • Xeon 6

Technologies named here

  • NVLink
  • Vera Rubin NVL72
  • Vera CPU
  • HGX Rubin NVL8
  • Spectrum-X
  • BlueField-4 DPU
  • OCI Kubernetes Engine
  • OCI Superclusters
  • Kubernetes
  • BlueField-4
  • RDMA
  • co-packaged optics
  • Oracle Cloud Infrastructure

Techniques named here

  • Independent CPU and GPU scaling
  • CPU offload for agentic tool execution

Also named here

  • KV cache

Related in inference