Nvidia's agent-tuned Vera CPU: 88 cores, 1.8x on agentic work
Olympus cores chase single-thread speed for latency-bound agent loops; the rack claims 6X CPU throughput, and Kimi K3's open weights reportedly drop July 27.
Copy markdown
Olympus cores, tuned for agent loops
Vera packs 88 custom Olympus cores and 176 threads with 164MB of L3 and up to 1.2 TB/s of LPDDR5X, and Nvidia claims up to 1.8x higher performance on agentic workloads versus x86. The bet: agent orchestration is single-thread- and latency-bound, so per-core IPC matters more to your loop times than raw core count.
A 256-chip rack, 6X the CPU throughput
The full Vera rack stacks 256 liquid-cooled CPUs for up to a 6X gain in CPU throughput over the prior generation, scaling dual-socket over NVLink-C2C with PCIe 6.4 and CXL 3.1. This is the host silicon your cloud-hosted Claude, GPT, and Gemini agents increasingly run on — more concurrent agent VMs packed per rack.
Per-VM encryption for agent sandboxes
Vera ships Confidential Computing with per-VM memory encryption, isolating each tenant or agent VM at the silicon level. After a week of sandbox escapes across Cursor, Codex, and Gemini CLI, that's a hardware-level answer to stopping one runaway agent from reading another's memory.
Kimi K3's open weights reportedly land July 27
Moonshot AI reportedly plans to release the weights of its 2.8-trillion-parameter K3 mixture-of-experts model on July 27, after demand outran its own servers. If it holds, self-hosters get zero-per-token inference of a model that's topped coding leaderboards — a real weekend project for anyone with the GPUs to run it.