August 24, 2026 was a pivotal day for AI infrastructure. At Hot Chips, NVIDIA announced two developments that signal a fundamental shift in how enterprise AI agents will be built and run. First, NVIDIA Vera CPU, the first processor designed specifically for the orchestration demands of AI agents, is being adopted by SpaceXAI for gigawatt-scale agentic AI operations. Second, NVIDIA Groq 3 LPX, a new interactive inference accelerator extending the Vera Rubin platform, has entered full production with Nebius as its first cloud adopter.
These are not incremental hardware announcements. Together, they mark the emergence of a purpose-built hardware stack for AI agents: specialized silicon at every layer of the agentic loop.
NVIDIA Vera CPU: The First CPU Built for AI Agents
For years, CPU optimization in the AI era meant making x86 systems better at moving data to GPUs. NVIDIA Vera takes a different approach entirely.
Vera is designed around the specific demands of agentic AI workloads. When an AI agent is running, a large share of its compute budget is not spent on model inference at all. The agent is orchestrating tools, executing code, parsing API responses, running simulations, and deciding what to do next. This CPU-intensive orchestration layer is the bottleneck that Vera targets directly.
The technical specifications reflect that focus: 88 NVIDIA-designed Olympus cores, NVIDIA Spatial Multithreading, and 1.2TB/s of LPDDR5X memory bandwidth. Across agentic AI, reinforcement learning, and data-processing workloads, NVIDIA says Vera delivers up to 1.8x faster task completion compared with x86 CPUs.
“Agentic AI requires a new kind of computing system, one built not only to generate answers, but to take action,” said Ian Buck, NVIDIA’s vice president of hyperscale and high-performance computing, in the official announcement. “Vera gives AI agents the CPU performance to act in real time, executing code, processing data and coordinating complex tasks.”
The framing is significant. This is not NVIDIA pitching a faster general-purpose processor. It is a CPU built for a specific, new workload: the orchestration layer of an autonomous AI agent.
| Specification | NVIDIA Vera CPU | Typical x86 (reference) |
|---|---|---|
| Core architecture | 88 Olympus cores (NVIDIA-designed) | Intel/AMD multi-core |
| Memory bandwidth | 1.2 TB/s (LPDDR5X) | 100-300 GB/s (DDR5) |
| Agentic task speed | Up to 1.8x faster than x86 | Baseline |
| Primary focus | Agent orchestration, code execution, RL | General purpose |
| Platform integration | Vera Rubin NVL72 co-design | Standalone or HGX pairing |
SpaceXAI Commits to Gigawatt-Scale Agentic AI with Vera
The significance of the Vera CPU is amplified by who is adopting it first. SpaceXAI, the company behind Grok, is deploying Vera CPUs to accelerate the orchestration layer behind its next generation of agentic AI workloads.
SpaceXAI’s scale is not a typical enterprise reference. The company is scaling toward gigawatts of computing capacity. At that level, even modest efficiency gains in CPU orchestration translate into enormous savings in power and cost. “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. “That means higher-performance AI agents and more useful work from every watt of compute.”
SpaceXAI’s full infrastructure plan around this announcement spans three layers:
- NVIDIA Vera CPUs for agentic orchestration at massive scale
- NVIDIA Vera Rubin GPU platform for Grok model training and inference
- Starmind AI satellite, an orbital deployment built on an optimized NVIDIA Vera Rubin NVL72 system
The combination represents one of the most ambitious AI infrastructure deployments announced to date. Elon Musk earlier this year committed exclusively to NVIDIA GPUs for SpaceXAI’s AI operations, describing them as “the best.” The Vera CPU and Groq 3 LPX adoption extends that architecture commitment across the full agentic stack.
NVIDIA Groq 3 LPX: Ultrafast Token Generation Goes to Production
Alongside the Vera CPU announcement, NVIDIA confirmed that NVIDIA Groq 3 LPX is now in full production. This is a complementary product: an interactive AI inference accelerator that extends the Vera Rubin platform by dramatically increasing token generation rates for latency-sensitive workloads.
In Artificial Analysis benchmarking, Groq 3 LPX demonstrated world-class speed specifically for agentic coding and other time-sensitive applications. Nebius is the first AI cloud to adopt it, deploying it through Nebius Token Factory.
“Generation is the phase of inference that determines how responsive an AI system actually is, and that is exactly what NVIDIA Groq 3 LPX is built to accelerate,” said Danila Shtan, CTO of Nebius. “As the first AI cloud bringing it to production, we are making sure every step of an agent’s loop feels instant, through the same API developers are already using, with no migration to a new stack.”
Following Nebius, the inference cloud Groq plans to be among the earliest adopters of the platform. The product name reflects an integration of Groq’s LPU (Language Processing Unit) technology into NVIDIA’s Vera Rubin system architecture, combining NVIDIA’s rack-scale infrastructure with Groq’s ultrafast token generation capabilities.
For enterprise AI teams, Groq 3 LPX addresses a specific production challenge: agent response latency. When an AI agent completes a reasoning step and needs to generate the next output or tool call, token generation speed determines how quickly the loop continues. Faster generation means more responsive agents, shorter wait times per step, and better utilization of the orchestration layer running on Vera CPUs.
Starmind: AI in Orbit
The orbital dimension of the SpaceXAI announcement deserves attention. SpaceXAI’s planned first-generation Starmind AI satellite will be based on an optimized NVIDIA Vera Rubin NVL72 rack-scale system. The satellite is designed for the requirements of orbital computing: the same accelerated computing architecture as terrestrial AI factories, adapted for space and powered by solar energy.
NVIDIA and SpaceXAI are already working to adapt the Vera Rubin architecture to orbital requirements while preserving a common software ecosystem. This means that models and agent harnesses developed for terrestrial deployment could eventually run on the same architecture in orbit.
For enterprise builders, Starmind is not a near-term product consideration. However, it signals something important: the compute infrastructure behind frontier AI agents is being designed, from the outset, to be architecture-portable across environments. The same system that runs Grok in a terrestrial AI factory will eventually run AI in orbit.
What This Means for Enterprise AI Builders in 2026
The immediate implications for enterprise teams building agentic AI products are concrete.
Orchestration is now a first-class infrastructure concern. The emergence of purpose-built CPU architecture for agent orchestration confirms what practitioners have discovered in production: model inference is only part of the compute cost of an agentic system. Tool use, code execution, and state management between model calls are CPU-bound workloads with real latency and cost implications. Teams designing multi-step agent architectures should profile their orchestration costs separately from inference costs.
Token generation speed is an agent quality metric. Fast token generation, which Groq 3 LPX accelerates, is not just a user experience concern. In multi-agent systems and real-time agentic workflows, slow generation creates cascading delays across the entire pipeline. As Groq 3 LPX infrastructure reaches cloud providers via Nebius Token Factory and the Groq inference cloud, enterprise teams will gain access to faster generation at competitive cost without changing their existing API integrations.
Hardware-software co-design is coming to enterprise AI. The SSI and NVIDIA Vera Rubin deal in July (Safe Superintelligence and NVIDIA: What Their $5B Compute Partnership Means for Enterprise AI) focused on GPU compute for frontier model training. This week’s announcements add CPU compute for agent orchestration and LPU-style accelerators for token generation. The full hardware stack for enterprise AI agents is being purpose-built layer by layer.
SpaceXAI’s infrastructure choices are an enterprise leading indicator. The AI models behind Grok run on SpaceXAI infrastructure before they reach enterprise deployments via Cursor, the SpaceXAI API, and Grok Bot for enterprise teams. As covered when Grok Bot launched, the performance characteristics of SpaceXAI’s underlying infrastructure directly shape what enterprise teams can build on top of it. Vera CPU-accelerated orchestration will show up in the products enterprise teams are already using.
The Hardware Stack for Enterprise AI Agents Is Complete
The two announcements from Hot Chips 2026 complete a picture that has been forming throughout this year. AI agent infrastructure is not a GPU-scaling problem. It is a systems problem requiring purpose-built components at every layer of the agentic loop:
- Model inference: GPUs (Vera Rubin NVL72, Blackwell, Hopper)
- Token generation: LPU-style accelerators (Groq 3 LPX, Groq LPU)
- Agent orchestration: Purpose-built CPUs (NVIDIA Vera)
For teams building production AI agents today, this architecture is mostly invisible behind cloud abstractions. But as purpose-built agent compute reaches providers through Nebius Token Factory, Groq’s inference cloud, and eventually through major hyperscalers adopting the Vera Rubin platform, the latency and cost characteristics of agentic workloads will improve meaningfully.
The teams that understand the full compute model of an agentic system, not just the model inference cost but the orchestration cost and the generation latency, will be better positioned to architect systems that remain economically viable as they scale.
If your enterprise is architecting agentic AI systems and you want to think through the infrastructure, cost, and deployment strategy, talk to the Enera team.