Nvidia Vera CPU Grok agents takes a step forward. At Hot Chips 2026 in Palo Alto on Monday, Nvidia disclosed that SpaceXAI has deployed the Vera CPU to power the next generation of Grok’s agentic AI infrastructure, making SpaceXAI the second hyperscaler after Meta to commit to the chip outside a full Vera Rubin rack configuration. The same day, Elon Musk posted on X to confirm a Q4 2027 launch target for Starmind, a space-optimized Vera Rubin NVL72 system developed jointly by SpaceX and Nvidia, with what Musk called “significant scale” slated to follow in 2028. The two announcements together indicate that the engineering thesis behind orbital AI compute is, at its foundation, a CPU story.
Nvidia Vera CPU Grok Agents: One Architecture Across Ground and Orbit
Public discussion of Starmind has largely centered on GPU counts, launch costs, and thermal management. Monday’s joint announcement reframes the rationale as a CPU-first problem. The constraint that Vera is engineered to relieve on the ground — agentic AI systems spending most of their runtime waiting on CPU-bound orchestration while GPUs sit idle — is the same constraint the Starmind architecture is intended to carry into low Earth orbit. The chip powering Grok’s AI factories on Earth is the same chip planned for the Starmind satellites, and that continuity is the substantive engineering claim of the program.
A standard chat completion places minimal load on a host CPU. An agent operates differently. It plans a task, assembles a context window from retrieval stores and conversation history, sends that material to the model, receives a response, parses tool calls from the output, executes those tools — often across code sandboxes, database queries, API calls, and file operations — collects the results, and schedules the next inference pass. Each step runs on the host CPU, not on the accelerator. The accelerator finishes its inference pass and then waits for the host to complete orchestration before the next token stream begins.
Why Agentic Workloads Force a CPU-First Design
An agentic AI fleet workload analysis from Nvidia’s technical team, drawing on telemetry from more than 163,000 agentic sessions, finds that over 97% of agent sessions exhibit unique execution profiles. That heterogeneity renders traditional CPU fleet strategies impractical. For a system completing a hundred agent steps in sequence, host-side delay compounds at every transition, and the gap between raw accelerator throughput and delivered agent throughput is determined almost entirely by how quickly the CPU can finish each orchestration cycle.
Nvidia has positioned Vera not as a replacement for GPU compute but as the layer that eliminates the wait between GPU compute cycles. The Vera product specification pairs 88 custom Olympus cores with LPDDR5X memory delivering up to 1.2 terabytes per second of bandwidth, a memory architecture tuned for the scatter-gather access patterns that agent orchestration demands. Nvidia states that Vera completes agentic AI, reinforcement learning, and data processing tasks up to 1.8 times faster than comparable x86 processors, a figure derived from Nvidia’s own internal testing rather than independent benchmarking.
“Vera gives us the CPU performance and memory bandwidth to run enormous amounts of orchestration, code and data processing while keeping GPUs doing what they do best,” said Mike Nicolls, president of SpaceXAI, in the joint release with Nvidia. “That means higher-performance AI agents and more useful work from every watt of compute.”
Vera’s Memory Bandwidth and Orchestration Profile
The bottleneck Vera targets is not raw FLOPS but the rate at which orchestration data can move between memory and the cores handling tool execution, retrieval, and scheduling. Scatter-gather access patterns dominate agent workloads because the host must reach across many small data structures — tool registries, conversation histories, vector indexes, sandbox states — in irregular sequences. LPDDR5X at 1.2 terabytes per second, paired with 88 Olympus cores, is configured to keep many such requests in flight simultaneously rather than serialize them through a narrower memory pipeline.
The same week as the SpaceXAI deployment, Nvidia confirmed that Grok’s agentic infrastructure is operating on Vera across multiple internal workloads, with telemetry showing orchestration cycle times shortened by a factor the company describes as consistent with its 1.8x figure. Nvidia’s framing in the Monday materials emphasizes that Vera is intended as the orchestration substrate for agent fleets, while Rubin-class GPUs handle the inference passes themselves. The two roles are kept distinct in the product positioning, and the NVLink fabric connecting them is sized to keep both sides fed.
Starmind’s Orbital Build Inherits Terrestrial Engineering
The Vera Rubin NVL72 rack around which Starmind’s AI1 satellite is built integrates 72 Rubin GPUs and 36 Vera CPUs in a single rack connected by NVLink 6’s 260-terabyte-per-second all-to-all fabric, with no internal copper cabling. The cable-free midplane cuts assembly time from roughly 100 minutes to about five minutes compared to the prior Grace Blackwell generation, a figure Nvidia published in connection with terrestrial data center deployments. In orbit, where there is no opportunity to route cables after launch and every mechanical connection that could vibration-fatigue during a Starship ascent is a candidate mission-ending failure, those properties become engineering requirements rather than convenience features.
The Monday announcement frames the deployment, the orbital system, and the broader roadmap as a single unified architecture. “Vera CPUs accelerating increasingly sophisticated AI agents, Vera Rubin powering the AI infrastructure behind Grok and gigawatt-scale AI factories on Earth, and NVIDIA accelerated computing extending into orbital AI infrastructure,” the statement reads. That language describes one hardware stack — CPU, GPU, NVLink fabric, software ecosystem — intended to execute the same workloads across terrestrial data centers and low Earth orbit. Software and tooling developed for Grok’s ground-based agents can, in principle, run on Starmind satellites without a complete rewrite, which is the continuity Musk referenced in his Aug. 24 X post describing the space-optimized design as “significantly simpler, lower cost, denser and lighter than a traditional rack.”
What Monday’s two announcements together establish is that the Nvidia Vera CPU Grok agents depend on is the same component SpaceX and Nvidia plan to launch in Q4 2027, with the terrestrial deployment already underway and the orbital variant targeting initial operations before scaled capacity arrives the year after.
Source: Nvidia Vera CPU Powers Grok Agents Now, Starmind Orbit in Q4 2027: One Architecture