# Language Processing Unit Groq, Inc. is an American AI company that designs the Language Processing Unit (LPU), an application-specific integrated circuit (ASIC) optimized for fast inference of large language models and other AI workloads. Founded in 2016, it is headquartered in Mountain View, California. Groq, Inc. is an American artificial intelligence (AI) company that builds an AI accelerator application-specific integrated circuit (ASIC). The architecture was originally introduced as a Tensor Streaming Processor (TSP) but was later rebranded as a Language Processing Unit (LPU) following the widespread adoption of large language models after the breakthrough of ChatGPT. The company also develops related computer hardware and software to accelerate AI inference performance. Examples of the types of AI workloads that run on Groq's LPU are: large language models (LLMs), image classification, and predictive analysis. Groq is headquartered in Mountain View, California, and has offices in San Jose, California, Liberty Lake, Washington, Toronto, Canada, London, U.K. and remote employees throughout North America and Europe. In December 2025, Nvidia and Groq announced an agreement reportedly valued at approximately US$20 billion to license Groq's AI inference technology and to transfer several senior Groq executives to Nvidia. Groq stated that it would continue to operate as an independent company. ## History Groq was founded in 2016 by a group of former Google engineers, led by Jonathan Ross, one of the designers of the Tensor Processing Unit (TPU), an AI accelerator ASIC, and Douglas Wightman, an entrepreneur and former engineer at Google X (known as X Development), who served as the company’s first CEO. The founding team drew on experience from [DeepMind](https://www.wikiprompt.org/wiki/google-deepmind) and other AI research groups, aiming to build hardware specifically for the demands of [AI](https://www.wikiprompt.org/wiki/artificial-intelligence) workloads. Groq received seed funding from Social Capital's Chamath Palihapitiya, with a $10 million investment in 2017 and soon after secured additional funding. In April 2021, Groq raised $300 million in a series C round led by Tiger Global Management and D1 Capital Partners. Current investors include: Infinitum, The Spruce House Partnership, Addition, GCM Grosvenor, Xⁿ, Firebolt Ventures, General Global Capital, and Tru Arrow Partners, as well as follow-on investments from Infinitum, TDK Ventures, and XTX Ventures. After Groq’s series C funding round, it was valued at over $1 billion, making the startup a unicorn. On March 1, 2022, Groq acquired Maxeler Technologies, a company known for its dataflow systems technologies. It was decided Maxeler would maintain its brand due to the longstanding accomplishments of Dr. Oskar Mencer and team, publishing research in their field of expertise. On August 16, 2023, Groq selected Samsung Electronics' foundry in Taylor, Texas to manufacture its next-generation chips, on Samsung's 4-nanometer (nm) process node. This was the first order at this new Samsung chip factory. On February 19, 2024, Groq soft-launched a developer platform, GroqCloud™, to attract developers into using the Groq API and rent access to their chips. On March 1, 2024 Groq acquired Definitive Intelligence, a startup known for offering a range of business-oriented AI solutions, to help with its cloud platform. Groq raised $640 million in a series D round led by BlackRock Private Equity Partners in August 2024, valuing the company at $2.8 billion. On February 10, 2025, Groq announced that it had secured a US$1.5 billion commitment from the Kingdom of Saudi Arabia to expand delivery of its LPU-based AI inference infrastructure, tied to a new GroqCloud data center in Dammam, Saudi Arabia. As of 2025, Groq has established a dozen data centers across the U.S., Canada, the Middle East, and Europe, with its technology. In December 2025, Nvidia agreed to purchase assets from Groq for approximately US$20 billion, which is a record for Nvidia. Groq has described this as a non-exclusive licensing deal. As part of the deal, Groq founder Ross and Groq president Sunny Madra would join Nvidia. In May 2026, Groq was reported to be raising $650 million from existing investors to fund its transition to an AI inference cloud business, with the round structured as a pro-rata offering backstopped by investors Disruptive and Infinitum. ## Language Processing Unit Groq's initial name for their ASIC was the Tensor Streaming Processor (TSP), codenamed "Alan" but was later rebranded by Mark Heaps - VP of Brand, and Jonathan Ross - CEO/Founder, from the TSP to the Language Processing Unit (LPU) to make the nature of the processor more obvious. The LPU is designed specifically for the inference phase of [large language models](https://www.wikiprompt.org/wiki/large-language-model) and other AI models, focusing on low latency and high throughput. The LPU features a functionally sliced microarchitecture, where memory units are interleaved with vector and matrix computation units. This design facilitates the exploitation of dataflow locality in AI compute graphs, improving execution performance and efficiency. The LPU was designed off on two key observations: AI workloads exhibit substantial data parallelism, which can be mapped onto purpose-built hardware, leading to performance gains; and a deterministic processor design, coupled with a producer-consumer programming model, allows for precise control and reasoning over hardware components, allowing for optimized performance and energy efficiency. In addition to its functionally sliced microarchitecture, the LPU can also be characterized by its single-core, deterministic architecture. The LPU can achieve deterministic execution by avoiding the use of traditional reactive hardware components (branch predictors, arbiters, reordering buffers, caches) and by having all execution explicitly controlled by the compiler thereby guaranteeing determinism in execution of an LPU program. This deterministic approach contrasts with the speculative execution used in many [CPU](https://www.wikiprompt.org/wiki/intel) designs, which can introduce variability in performance. The first generation of the LPU (TSP) yields a computational density of more than 1TeraOp/s per square mm of silicon for its 25×29 mm 14nm chip operating at a nominal clock frequency of 900 MHz. The second generation of the LPU (LPU v2) will be manufactured on Samsung's 4nm process node. The LPU's architecture is particularly suited for [transformer](https://www.wikiprompt.org/wiki/transformer)-based models, which dominate modern [generative AI](https://www.wikiprompt.org/wiki/generative-ai) applications. Groq hosts open-source large language models running on its LPUs for public access. Access to these demos is available through Groq's website and its playground for Developers. The LPU's performance has been benchmarked against [Nvidia](https://www.wikiprompt.org/wiki/nvidia) GPUs, showing significant speedups in token generation for models like Llama and Mistral, though independent verification is ongoing. ## Architecture and Design The LPU's design philosophy centers on eliminating the overhead of general-purpose processors. By using a single core with a large, software-managed memory, the LPU avoids the need for complex cache hierarchies and coherence protocols. Each functional slice includes its own memory and compute units, allowing data to flow directly between them without passing through a shared bus. This is similar in spirit to the dataflow architectures explored by earlier projects like [Xerox PARC](https://www.wikiprompt.org/wiki/xerox-parc) and [MIT CSAIL](https://www.wikiprompt.org/wiki/mit-csail), but applied to modern AI workloads. The compiler plays a crucial role in the LPU's operation. It schedules all instructions and data movements ahead of time, ensuring that the hardware never stalls waiting for data. This static scheduling is possible because AI inference graphs are typically known in advance, unlike interactive applications. The result is a processor that achieves high utilization and predictable performance, which is critical for real-time AI services. The LPU also supports a range of numerical precisions, including FP32, BF16, and INT8, allowing developers to trade off accuracy for speed. This flexibility is important for deploying models across different hardware constraints, from edge devices to cloud data centers. ## Software and Ecosystem Groq provides a software stack that includes a compiler, runtime, and APIs for popular frameworks like PyTorch and TensorFlow. The GroqWare suite includes tools for model optimization, profiling, and deployment. The company has also partnered with [Hugging Face](https://www.wikiprompt.org/wiki/hugging-face) to offer a wide range of open-source models on its platform, making it easy for developers to experiment with LPU acceleration. GroqCloud, launched in 2024, allows users to rent LPU capacity on demand, similar to cloud services from [AWS](https://www.wikiprompt.org/wiki/amazon-web-services), [Azure](https://www.wikiprompt.org/wiki/azure), and [Google Cloud](https://www.wikiprompt.org/wiki/google-cloud). The platform supports both interactive inference and batch processing, with a pay-as-you-go pricing model. As of 2025, GroqCloud hosts over 100,000 developers and has processed billions of requests. The company also maintains a research division that collaborates with academic institutions like [Stanford AI Lab](https://www.wikiprompt.org/wiki/stanford-ai-lab) and [Berkeley AI Research](https://www.wikiprompt.org/wiki/berkeley-ai-research) to explore new architectures for AI. Groq's technology has been used in applications ranging from [chess engines](https://www.wikiprompt.org/wiki/chess-computer) to [autonomous vehicles](https://www.wikiprompt.org/wiki/waymo), though the primary focus remains on LLM inference. ## Market Position and Competition Groq competes with other AI chip startups such as [SambaNova](https://www.wikiprompt.org/wiki/samba-nova) and [Graphcore](https://www.wikiprompt.org/wiki/graphcore), as well as established players like [AMD](https://www.wikiprompt.org/wiki/amd) and [Intel](https://www.wikiprompt.org/wiki/intel). The LPU's deterministic design gives it an edge in latency-sensitive applications, but it lacks the flexibility of GPUs for training tasks. Groq has positioned itself as a specialist in inference, a market that is growing rapidly with the proliferation of AI applications. In 2025, the company's valuation reached $2.8 billion after its Series D round, and the Nvidia deal in December 2025 further validated its technology. The licensing agreement allows Nvidia to incorporate Groq's inference technology into its own products, potentially expanding the reach of LPU-based solutions. ## Future Directions Looking ahead, Groq plans to release the LPU v2, manufactured on Samsung's 4nm process, which is expected to offer significant performance improvements over the first generation. The company is also exploring new memory technologies, such as [HBM](https://www.wikiprompt.org/wiki/hbm), to increase bandwidth and reduce latency further. Groq's transition to an AI inference cloud business, as indicated by the 2026 funding round, suggests a strategic shift from hardware sales to service provision. This aligns with industry trends where compute is increasingly delivered as a utility, similar to the evolution of cloud computing. ## See Also - Central processing unit - Graphics processing unit - Tensor processing unit - AI accelerator ## References - Groq, Inc. official website - Press releases and announcements from Groq - Industry reports on AI hardware ## External Links - [Groq official website](https://groq.com) --- Source: https://www.wikiprompt.org/wiki/groq-lpu License: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/) Last updated: 2026-09-09T01:54:55.479078+00:00