Groq LPU is a custom inference chip with on-chip SRAM and deterministic scheduling delivering low-latency AI inference for LLMs.
The Groq LPU (Language Processing Unit) is a tensor streaming processor that combines compute, SRAM, and networking on a single die. It uses a software-scheduled, deterministic execution model designed for predictable latency and high throughput on generative AI inference workloads.
In December 2025, Nvidia entered a non-exclusive licensing agreement for Groq's inference technology, validating the LPU architecture. Groq continues to operate independently and in June 2026 raised $650M to scale its inference cloud business.
Cerebras Systems
Cerebras Cloud (Inference & Training Platform), Wafer-Scale Engine 3 (WSE-3)
cerebras.ai
d-Matrix
Corsair AI Compute Platform
d-matrix.ai
SambaNova Systems
DataScale, RDU
sambanova.ai
NVIDIA
DGX
nvidia.com
Majestic Labs
Prometheus
majestic-labs.ai
Etched.ai
Sohu
etched.com
| Attribute | LPU | Cerebras Cloud (Inference & Training Platform), Wafer-Scale Engine 3 (WSE-3) | Corsair AI Compute Platform | DataScale, RDU | DGX | Prometheus | Sohu | |
|---|---|---|---|---|---|---|---|---|
| Provider |  | | | | | | | |
| Founded | 2016 | 2016 | 2019 | 2017 | 1993 | 2023 | 2022 | |
| Sells To | Enterprise | — | Enterprise | Enterprise | Consumers, Enterprise | Enterprises | B2B, Enterprise | |
| Pricing Model | Recurring | — | — | Recurring, Software | Licensing, Product Sales, Subscription | Direct Sales | Product Sales, Recurring | |
| Ownership | Venture-backed | — | Private | Venture-backed | Public | Venture-backed | Venture-backed |
Provider
Founded 2016
Sells To Enterprise
Pricing Model Recurring
Ownership Venture-backed
Provider
Founded 2016
Sells To —
Pricing Model —
Ownership —
Provider
Founded 2019
Sells To Enterprise
Pricing Model —
Ownership Private
Provider
Founded 2017
Sells To Enterprise
Pricing Model Recurring, Software
Ownership Venture-backed
Provider
Founded 1993
Sells To Consumers, Enterprise
Pricing Model Licensing, Product Sales, Subscription
Ownership Public
Cerebras Wafer-Scale Engine is a family of wafer-scale AI processors.
Analog compute-in-memory processor delivering 25 TOPS for edge AI inference.
Analog in-memory inference platform collocating deep sub-threshold compute inside multi-level non-volatile memory.
Maverick-2 is NextSilicon's runtime-reconfigurable dataflow accelerator for high-performance computing and AI workloads.
Licensable AI processor IP for on-device inference across edge silicon designs.
Cerebras Wafer-Scale Engine is a family of wafer-scale AI processors.
Analog compute-in-memory processor delivering 25 TOPS for edge AI inference.
Analog in-memory inference platform collocating deep sub-threshold compute inside multi-level non-volatile memory.
Maverick-2 is NextSilicon's runtime-reconfigurable dataflow accelerator for high-performance computing and AI workloads.
Training API for fine-tuning open-source language models.
