Why SpaceXAI Is Deploying Nvidia’s Vera CPU for Agentic AI

Nvidia says SpaceXAI will deploy its Vera CPUs to support the CPU-heavy work behind Grok’s next generation of agentic AI. That includes tool orchestration, code execution, data processing and simulations—tasks that coordinate what happens between GPU model calls rather than generating the model response itself. 1

The announcement also links Vera to SpaceXAI’s broader infrastructure plans, including Vera Rubin-based systems intended to scale toward gigawatts of capacity and an optimized Vera Rubin NVL72 system for a planned first-generation Starmind AI satellite. 1

Vera’s role in an agentic AI system

Traditional AI discussions often focus on the accelerator running the model. Agentic systems add a substantial control and execution layer: maintaining state, coordinating multiple model and tool calls, running code, processing intermediate data, simulating environments and repeatedly revising a plan.

That makes the CPU more consequential. A CPU serving an agentic workload may need to manage many concurrent processes, move large amounts of state and data, and keep GPU accelerators supplied with work. Vera is Nvidia’s attempt to design that CPU layer specifically around agentic AI, reinforcement learning and data-processing workloads rather than treating it as a generic host component. 33

The Vera specifications Nvidia is highlighting

Nvidia’s stated Vera specifications include:

• 88 custom Olympus CPU cores
• LPDDR5X memory with up to 1.2 TB/s of bandwidth
• Up to 1.5 TB of memory in configurations described by reporting on the chip 7 32
• Support for Vera servers, Vera Rubin systems and related AI infrastructure platforms 33

The design priorities map closely to the problems found in agentic workloads. More cores can support concurrent orchestration and simulation tasks. Greater memory bandwidth can reduce delays when agents repeatedly access state, intermediate results and data sets. In a Vera Rubin system, the CPU is part of a larger Nvidia platform that combines CPUs, Rubin GPUs, networking, DPUs and software rather than treating each component as an isolated purchase. 33 40

The 1.5 TB-per-socket figure should be treated carefully. It appears in secondary reporting and product descriptions, but the primary releases provided here do not independently confirm that exact configuration. It is safer to describe it as a reported or configurable figure, not as a universally established specification.

What the performance claims actually mean

At Vera’s March 16, 2026 launch, Nvidia claimed twice the efficiency and 50% faster performance than traditional rack-scale CPUs for the targeted workloads. 38 In the later SpaceXAI announcement, Nvidia cited up to 1.8× faster task completion than x86 CPUs across certain agentic AI, reinforcement-learning and data-processing workloads. 1

Those numbers are important signals about Nvidia’s positioning, but they are not equivalent to independent evidence that Vera is faster for every server workload. They are vendor claims tied to selected workload categories and comparisons. They do not establish general superiority over AMD EPYC or Intel Xeon in conventional enterprise computing.

For buyers, the more meaningful question is likely to be system-level performance: how quickly a complete CPU-GPU platform can execute an agent’s workflow, how much power it uses, and how effectively its software stack handles scheduling, memory movement and tool calls. Nvidia’s advertised CPU figures alone cannot answer that question.

From early evaluation to a deployment plan

The partnership developed in stages. Nvidia launched Vera on March 16, 2026, initially saying partner systems would become available in the second half of the year. 38 In May, Nvidia executive Ian Buck personally delivered early Vera systems to SpaceXAI, Anthropic, OpenAI and Oracle Cloud Infrastructure. Nvidia’s own account presents those deliveries as the first customer rollout, but the deliveries should not be confused with proof that every recipient had already completed a full-scale deployment. 11

The August 24 announcement moved SpaceXAI from early hardware recipient to a company with a publicly stated deployment plan. Nvidia says the processors will support Grok-related agentic workloads on Earth and form part of a planned space-optimized Vera Rubin system for Starmind. 1

The terrestrial portion is the nearer-term, easier-to-understand use case: pair CPU-heavy orchestration and data processing with Rubin GPUs in an AI data-center environment. The orbital portion is more speculative. Nvidia acknowledges that computing in orbit introduces different power, thermal, bandwidth, reliability and integration constraints. 1

Why the orbital plan is more than a chip benchmark

Putting an AI system in orbit would require solving problems that cannot be reduced to core count or memory bandwidth. A space system must account for power generation, heat rejection, radiation tolerance, launch mass, communications capacity and long-term serviceability.

That makes the planned Starmind payload a systems-engineering test as much as a Vera test. The announcement demonstrates an intended architecture and partnership; it does not yet demonstrate sustained, operational orbital computing at scale. Reporting described the launch target as as early as 2027, but that remains a future plan. 18

Nvidia’s move into the CPU layer

Vera broadens Nvidia’s AI infrastructure strategy. The company is no longer competing only for the accelerator in an AI server; it is also offering the CPU, interconnect, networking, storage and software components around it. Nvidia’s March announcement listed cloud providers and system makers planning to collaborate on Vera, while Oracle has said its next-generation OCI Supercluster will combine Vera CPUs with Rubin GPUs, BlueField DPUs, NVLink, SuperNICs and Ethernet switches. 38 40

Oracle Cloud Infrastructure is an especially important validation point because reporting says it plans to deploy hundreds of thousands of Vera CPUs beginning in 2026. 8 39 If those plans materialize, Vera would move beyond a flagship customer evaluation into a meaningful hyperscale deployment.

The competitive pressure falls most directly on AMD and Intel, whose server CPUs remain central to data-center infrastructure. Nvidia’s advantage is not necessarily a universal CPU lead; it is the possibility of selling a tightly integrated AI-factory platform. Customers may value that integration for agentic AI, while others may prefer heterogeneous fleets that preserve CPU choice and avoid dependence on one vendor.

The market reaction was not a verdict on Vera

Nvidia and SpaceX shares declined on the August 24 announcement day. One report put Nvidia down 2.91% and SpaceX down 1.44%. 18

That one-day movement does not show that investors rejected the partnership or that Vera failed technically. Share prices also reflect broader market conditions, earnings expectations, capital spending and positioning. The more useful test will be whether SpaceXAI and other customers turn announced plans into sustained deployments, and whether independent measurements confirm Nvidia’s efficiency and performance claims.

The bottom line

SpaceXAI’s Vera deployment is significant because it treats the CPU as a first-class component of agentic AI infrastructure. Nvidia says Vera’s 88 Olympus cores and high-bandwidth memory are designed for the coordination, execution, simulation and data movement that surround GPU inference and training. 1 33

The immediate commercial story is a high-profile AI customer adopting Nvidia’s CPU platform. The broader strategic story is Nvidia trying to capture more of the server stack, with Oracle’s planned hyperscale rollout providing another test of demand. The orbital Starmind system is the most ambitious extension of the plan—but until it launches and operates, it should be understood as a stated deployment objective rather than a demonstrated capability.