NVIDIA hand-delivers first Vera CPUs to Anthropic, OpenAI, SpaceX, Oracle — first standalone NVIDIA CPU lands at frontier labs

What was delivered

The reporting from Bloomberg, NVIDIA’s official blog, wccftech, TweakTown, TipRanks, and TNW confirms:

• Product: NVIDIA Vera — first NVIDIA standalone data-center CPU
• Architecture: ARM-based, purpose-built for agentic AI workloads
• Production status: full production since March 2026 (per Bloomberg / TipRanks)
• First customers: Anthropic, OpenAI, SpaceX, Oracle Cloud Infrastructure
• Delivery: hand-delivered by Ian Buck (NVIDIA’s GM of Hyperscale and HPC)
• Sites: Anthropic San Francisco, OpenAI Mission Bay, SpaceX Palo Alto (delivered Friday late May 2026), Oracle Cloud Infrastructure Santa Clara (delivered the following Monday)
• Successor to: NVIDIA Grace (the previous-generation CPU released alongside Hopper GPUs)
• Competitive position: directly competes with Intel Xeon and AMD EPYC

Why Vera is structurally important

Three reads matter.

1. NVIDIA is becoming a full-stack compute supplier. For most of the AI compute era, NVIDIA sold GPUs while customers paired them with Intel Xeon or AMD EPYC CPUs. Vera changes that. NVIDIA now sells:

• GPUs (H100 → B100 → B200 → Rubin)
• CPUs (Vera, formerly Grace)
• Networking (Spectrum-X Ethernet, NVLink, InfiniBand)
• Coherent memory architecture (NVLink C2C ties Vera CPUs directly to NVIDIA GPUs at memory-coherent latency)
• Reference platforms (DGX systems, MGX server designs)
• Software (CUDA, NIM microservices, AI Enterprise)

This is the same vertical-integration pattern Apple ran with Apple Silicon — NVIDIA designing every layer of the compute stack for AI-specific workloads, leaving Intel and AMD with the workloads where the integration premium doesn’t apply.

2. Agentic AI workloads have distinct compute requirements. Traditional ML training is GPU-dominated, CPU-light. Agentic AI is different — agents spend significant compute on orchestration logic, tool calls, planning, and context management that runs on CPU rather than GPU. As Dynamic Workflows, Project Polaris, and similar parallel-subagent architectures become the dominant Claude Code / Copilot deployment pattern through H2 2026, CPU optimization for agentic workloads matters more. Vera is NVIDIA’s bet that those workloads should run on NVIDIA-designed silicon, not Intel or AMD.

3. The compute supply chain story for frontier labs keeps getting more complex. For each customer:

Compare to twelve months ago when “AI compute supply chain” meant “what NVIDIA GPUs can you get and when.” The 2026 picture is multi-vendor across CPU, GPU, custom AI accelerator (TPU / Trainium / Maia), and increasingly multi-cloud (AWS + Google + Microsoft + Oracle). For Claude, ChatGPT, and other API consumers, the practical implication is more reliable serving (multi-vendor diversification) and continued downward pressure on per-token costs (cross-vendor pricing competition).

What it means for the four customers

For Anthropic: Vera is the fifth distinct compute-source layer in the Anthropic supply chain (alongside NVIDIA GPUs, AWS Trainium, Google TPU via the $40B investment, and the early-stage Microsoft Maia 200 talks). For Anthropic’s drive toward sustainable operating profit and the October 2026 IPO trajectory, having a multi-vendor compute portfolio is procurement leverage, not just resilience.

For OpenAI: Vera is the bridge between NVIDIA’s current GPU stack and OpenAI’s reported in-house silicon ambitions. Running Vera alongside Microsoft Maia 200 (which is also serving OpenAI’s GPT-5.2 in production) creates a CPU-GPU compute portfolio that doesn’t depend on Intel or AMD for any layer.

For SpaceX: Vera adds to the Colossus 1 and forthcoming Colossus 2 data-center buildouts. The Anthropic-SpaceX 300+ MW capacity deal reaches a milestone with Vera deployment — SpaceX’s data centers can now offer NVIDIA-full-stack capacity to Anthropic and other AI tenants.

For Oracle Cloud: Vera delivery positions OCI as a serious AI compute hyperscaler. Combined with Oracle’s Stargate partnership work, this is OCI’s move from “AI compute also-ran” to “credible AI compute supplier.”

What it means for Claude and ChatGPT users

Practically: nothing changes operationally. Vera-powered serving is invisible to end users — your Claude API call lands on whatever hardware is closest and cheapest. Anthropic and OpenAI just gained another lever in the routing decision.

Structurally: every additional compute supplier in the stack improves serving reliability (lower probability of region-wide outage) and continues the downward pressure on per-token economics. For Claude Code at $2.5B+ ARR and ChatGPT on confidential S-1 filing, every cent of compute cost reduction at the unit level translates into material P&L improvement.

The honest caveats

Three caveats:

Vera benchmarks vs Xeon / EPYC aren’t yet independently verified at production scale. NVIDIA’s claims about Vera CPU performance for agentic workloads come from internal testing. Independent third-party benchmarking will surface over the next 30-90 days as customers deploy Vera in production.

“First delivery” is symbolic, not volume-shipping. Hand-delivered first systems are press-event-scale. Full deployment at Anthropic / OpenAI / SpaceX / Oracle scale will take quarters, not weeks. The structural impact on per-token costs is a 2026-2027 story, not a Q3 2026 one.

ARM-based CPUs require software-stack work. Vera being ARM-based (vs Intel/AMD’s x86) means some workloads need recompilation. Most modern AI infrastructure (Linux + Docker + Python + CUDA) supports ARM cleanly, but edge cases (proprietary x86 tools, legacy enterprise software) will require migration effort.

What it changes for Pick Right readers tomorrow

If you’re a Claude Pro subscriber, nothing changes immediately. If you’re a Claude Code heavy user, expect continued capacity expansion and stable-to-improving rate limits through H2 2026 as Vera-powered serving scales.

For broader context, see the Claude review, the ChatGPT review, the Anthropic-Microsoft Maia 200 chip talks, the SpaceX Colossus capacity unlock, the Google + Broadcom multi-gigawatt TPU partnership, and the Anthropic S-1 filing for the broader compute-supply-chain and IPO-pipeline picture.

Update (26 August 2026): two developments since have changed the economics around Vera rather than the hardware itself. Systems shipping in early 2027 are reported to cost more than 15% extra on memory pass-through, with Vera racks specifically shipping at reduced memory capacity. And OpenAI has published first benchmarks for its in-house Jalapeño inference chip, claiming better throughput per watt than Nvidia’s GB200 and GB300 — the clearest sign yet that Nvidia’s largest inference customers intend to serve a share of their own traffic on custom silicon.

Update, 25 August 2026 — Vera’s memory configuration has been cut. This article called the per-token cost impact of Vera “a 2026-2027 story.” It is, but not in the direction implied. Bloomberg reported on 22 August that Vera Rubin and Grace Blackwell systems shipping in early 2027 are expected to cost more than 15% extra, and that Vera CPU racks are now shipping with SOCAMM memory reduced from roughly 55TB to 28TB per rack — because DRAM, LPDDR and HBM4 have repriced sharply. Nvidia did not comment and the sources are unnamed, so treat the figures as reporting rather than policy. For buyers the capacity cut matters more than the price: memory capacity is what binds long-context inference, so the shortage is likelier to surface as context limits and concurrency caps than as a new rate card. What the memory squeeze means for AI pricing.

Questions or corrections? Email Pick Right. Want the full list? See all news.