Nvidia Groq 3 LPX Inference Rack Ships: 256 Accelerators and the Dedicated Inference Era

Nvidia Groq 3 LPX Inference Rack Ships: 256 Accelerators and the Dedicated Inference Era

Nvidia announced on August 24, 2026 that its Groq 3 LPX, the dedicated inference accelerator built from its $20B Groq acqui-hire, has entered full production and slots into the Vera Rubin platform with up to 256 LPX accelerators per rack. Nebius will be the first cloud customer, deploying LPX racks alongside Vera CPUs and Rubin GPUs. This marks the first time Nvidia has shipped a dedicated inference product separate from its training GPU line, reflecting the industry's recognition that inference workloads have fundamentally different hardware requirements than training.

The Groq 3 LPX is not a GPU — it is a purpose-built inference ASIC optimized for throughput and latency on transformer workloads. Unlike GPUs, which must support both forward and backward passes for training, the LPX is hardwired for the forward-pass-only inference path, allowing it to dedicate 100% of its silicon to token generation. Nvidia claims 3.2x inference throughput per watt versus the H100 GPU on Llama 4 405B workloads.

The Architecture Shift

The LPX rack architecture separates inference from training at the hardware level:

Previous: Training GPU (H100/B200) → also used for inference
New:      Training GPU (Rubin) → training only
          Inference ASIC (LPX)  → inference only
          CPU (Vera)            → orchestration & tool-calling

This separation matters because training and inference have opposite optimization profiles:

Performance Claims

Enterprise Impact

The LPX rack enables three new inference deployment patterns:

Dedicated inference clusters. Enterprises can deploy LPX-only racks for production inference without provisioning expensive training GPUs. A 256-LPX rack can serve approximately 50 million tokens per second — enough for 10,000 concurrent GPT-5.6-class agents.

Inference-as-a-Service pricing. Cloud providers like Nebius can offer inference at $0.14 per million tokens — 68% below current GPU-based pricing — making high-throughput agent deployments economically viable.

Edge inference. The LPX's 3.2x watts efficiency makes it suitable for edge deployments where power is constrained, enabling on-premise inference for regulated industries.

The Broader Context

Nvidia's LPX launch reflects a market that now spends more on inference than training. Gartner's August 2026 report confirms that inference spending has surpassed training for the first time, driven by the explosion of agent workloads that run inference continuously rather than in batch training runs. The LPX is Nvidia's bet that this shift is permanent.

The $20B Groq acqui-hire in 2025 was widely questioned at the time. With LPX in production and Nebius committed as the first customer, Nvidia has validated the thesis that inference deserves its own silicon.

Key Takeaways

• Nvidia's Groq 3 LPX enters full production with 256 accelerators per rack, delivering 3.2x inference throughput per watt versus H100 GPUs through purpose-built inference ASIC design
• The architectural separation of inference (LPX) from training (Rubin GPUs) reflects the market shift where inference spending now exceeds training for the first time
• Nebius deploys the first cloud LPX racks, enabling inference pricing at $0.14 per million tokens — 68% below current GPU-based inference costs

By Deepak Bagada, CEO at SaaSNext & Principal AI Architect.

Last tested: August 2026 with Python 3.12, Node v22, and latest framework releases.