Groq builds the Language Processing Unit chip and operates GroqCloud, the inference cloud powering fast responses from open-weight models like OpenAI's gpt-oss and Qwen3.6. Since Nvidia's licensing deal in December 2025, Groq reorganized under CEO Adam Winter, raised $650 million in June 2026, and became a certified NVIDIA Cloud Partner in August 2026.
Groq designs the Language Processing Unit (LPU) inference chip and runs GroqCloud, an AI inference platform serving more than five million developers. In August 2026 Groq became a certified NVIDIA Cloud Partner and was valued at a reported $3.5 billion, down from $6.9 billion in September 2025, after its Nvidia licensing deal.
Founded: 2016 · HQ: Mountain View, CA, USA · Team: 300-400 · CEO: Adam Winter · Funding: $2.4B total raised across 7 rounds from 49 investors (Sep 2025: $750M led by Disruptive, Blackrock, Samsung, Cisco, Altimeter; Jun 2026: $650M led by Disruptive and Infinitum; Feb 2025: $1.5B Saudi Arabia infrastructure commitment) · Valuation: $6.9B (September 2025 Series D)
About Groq
Groq was founded in 2016 in Mountain View, California by Jonathan Ross, one of the designers of Google's original Tensor Processing Unit (TPU), and Douglas Wightman, a former Google X engineer. They left Google after concluding that GPU architecture, built for training, was a poor match for inference, the stage where trained models answer real user queries. Their thesis: purpose-built silicon could eliminate the latency that slows GPU-based inference, making AI responses feel instantaneous.
Groq's flagship product is the Language Processing Unit (LPU), a custom chip that stores model weights in on-chip SRAM rather than external high-bandwidth memory, eliminating the bandwidth bottleneck that limits GPU inference speed. GroqCloud, its cloud product, is a token-as-a-service API running third-party open-source models including Meta's Llama 4 and DeepSeek R1 on Groq's own hardware, accessed through an OpenAI-compatible interface with per-token pricing and a free developer tier.
In December 2025, Nvidia announced a non-exclusive licensing agreement for Groq's LPU inference architecture valued at approximately $20 billion. Groq retained its intellectual property and kept operating GroqCloud independently, while Nvidia gained rights to produce chips based on Groq's architecture. Nvidia unveiled the resulting Groq 3 LPU at GTC 2026 in March: a Samsung 4-nanometer chip targeting 1,500 tokens per second for agentic AI, with 150 terabytes per second of on-chip memory bandwidth, shipping in late 2026.
Groq has raised approximately $2.4 billion across seven rounds from 49 investors. Its September 2025 round raised $750 million at a $6.9 billion post-money valuation, led by Disruptive with Blackrock, Neuberger Berman, Samsung, Cisco, D1 Capital, Altimeter Capital, 1789 Capital, and Infinitum participating. In February 2025, Saudi Arabia committed $1.5 billion toward Groq's infrastructure, tied to a new data center in Dammam. In June 2026, Groq closed an additional $650 million led by Disruptive and Infinitum for GroqCloud buildout.
In August 2026, Groq became a certified NVIDIA Cloud Partner (NCP), deploying NVIDIA accelerated computing to NVIDIA's reference architecture inside its own data centers alongside its native LPU hardware (Groq newsroom, August 12, 2026). That same month, Bloomberg and Sacra reported Groq's valuation at $3.5 billion, roughly half its $6.9 billion valuation from September 2025, following the Nvidia licensing deal and the departure of senior staff. GroqCloud's model lineup also shifted through 2026: Groq retired llama-3.1-8b-instant and llama-3.3-70b-versatile on August 16 in favor of OpenAI's open-weight gpt-oss-20b and gpt-oss-120b models, and decommissioned the groq/compound and groq/compound-mini endpoints on September 21, 2026 (Groq's own deprecations documentation).
Groq earns revenue through IP licensing and cloud inference services. The Nvidia deal validated the hardware IP path at scale, while GroqCloud charges per million tokens consumed, with free access for individual developers and dedicated capacity for enterprise customers with SLA guarantees. Latka reported Groq's annual recurring revenue at approximately $172.5 million in 2025. The 2026 strategy centers on becoming the leading AI inference cloud rather than selling hardware to third-party operators.
Jonathan Ross left in December 2025 to join Nvidia as Chief Software Architect, with President Sunny Madra also moving to Nvidia as part of the deal. Adam Winter, formerly a Vice President at Groq, became CEO in late 2025 and is executing the pivot to an inference neocloud strategy. Claire Hart serves as Chief Operating Officer, Chief Legal Officer, and Board Member. As of mid-2026, Groq employs approximately 300 to 400 people, down from a peak of about 575 in late 2025, reflecting the leadership transition.
Groq's stated mission is to preserve human agency while building the AI economy by driving the cost of AI computation toward zero. Its research focus is inference hardware architecture, compiler technology, and systems software, not foundation model development. The LPU's defining feature is deterministic execution: memory access patterns are fixed at compile time, cutting the cache misses and latency variance that affect GPU inference under variable load.
Groq operates 13 data centers across North America, Europe, the Middle East, and Asia-Pacific, including Helsinki (Finland), Sydney (Australia), and Dammam (Saudi Arabia). It holds SOC 2 Type II certification and maintains a trust center at trust.groq.com. Main competitors include AWS Inferentia, Together AI, Fireworks AI, Cerebras Systems, and Nvidia's NIM inference service. Groq wins on raw token speed for streaming and real-time agent workloads but lacks fine-tuning and proprietary model capabilities enterprise buyers sometimes require. The June 2026 raise funds expansion to 200 megawatts of capacity by the end of 2027.
Mission
To preserve human agency while building the AI economy by driving the cost of AI computation toward zero.
Products
- GroqCloud (AI inference cloud API): https://groq.com/groqcloud
- LPU (Language Processing Unit) (Custom inference chip): https://groq.com/technology
- GroqCloud Console (Developer portal): https://console.groq.com
- GroqCloud Enterprise (B2B SaaS with dedicated capacity): https://groq.com/enterprise
Compliance
SOC 2 Type II
Links
Frequently Asked Questions
What is Groq and what do they build?
Groq is an AI inference technology company founded in 2016 in Mountain View, California that designs the Language Processing Unit (LPU), a custom chip built specifically for AI inference rather than model training. The LPU stores model weights in on-chip SRAM rather than external memory, enabling token generation speeds that GPU-based cloud services cannot match at the same cost. Groq's cloud product, GroqCloud, runs third-party open-source models including Meta's Llama 4, DeepSeek R1, and Mistral models on its own LPU hardware, accessible via an OpenAI-compatible API. Developers can sign up at console.groq.com and start calling models within minutes using the same API format as OpenAI's SDK. In December 2025, Nvidia licensed Groq's LPU architecture for approximately $20 billion, establishing Groq's technology as one of the most commercially validated inference architectures in the industry. GroqCloud serves enterprise customers across financial services, technology, and media sectors, in addition to its five-million-plus developer community. As of 2026, Groq operates 13 data centers across North America, Europe, the Middle East, and Asia-Pacific.
Who founded Groq and who is the CEO?
Groq was co-founded in 2016 by Jonathan Ross and Douglas Wightman. Jonathan Ross was one of the designers of Google's Tensor Processing Unit (TPU), the AI accelerator chip Google built internally starting around 2013, and he led hardware architecture at Groq from founding through December 2025. Douglas Wightman was an entrepreneur and engineer who had worked at Google X before helping start the company. Ross served as CEO for the company's first seven years and reflected publicly in July 2026 that early management errors, including hiring engineers who needed direction and then giving them too much autonomy, cost the company three to four years of progress. In December 2025, Ross left to join Nvidia as Chief Software Architect following the LPU licensing deal, with President Sunny Madra also moving to Nvidia. Adam Winter, a former Vice President at Groq, became CEO in late 2025 and is leading the company's shift to an inference neocloud strategy. Claire Hart joined as Chief Operating Officer, Chief Legal Officer, and Board Member.
How much funding has Groq raised?
Groq has raised approximately $2.4 billion across seven rounds from 49 investors as of mid-2026. In September 2025, the company closed a $750 million round at a post-money valuation of $6.9 billion, led by Disruptive and backed by Blackrock, Neuberger Berman, Samsung, Cisco, D1 Capital, Altimeter Capital, 1789 Capital, and Infinitum. In June 2026, Groq announced $650 million in new growth capital, again led by Disruptive and Infinitum, to scale GroqCloud's data center footprint. Bloomberg and Sacra both reported Groq's valuation at $3.5 billion following that round (August 2026), roughly half its September 2025 valuation, reflecting the December 2025 Nvidia licensing deal and the departure of founder Jonathan Ross and other senior staff to Nvidia. That Nvidia agreement, a non-exclusive license for Groq's LPU inference architecture, was valued at approximately $20 billion and is an IP licensing transaction rather than an equity investment. Groq is not currently public and has not announced IPO plans.
What products does Groq make?
Groq makes two core products: the Language Processing Unit (LPU) chip and GroqCloud, the API platform that runs AI models on LPU hardware. The LPU is a dedicated inference accelerator that stores model weights in on-chip SRAM, delivering hundreds of tokens per second at lower cost than GPU cloud alternatives for streaming and real-time workloads. GroqCloud is an OpenAI-compatible API that lets developers call open-source models without managing hardware; through 2026 Groq shifted its default model lineup toward OpenAI's open-weight gpt-oss-20b and gpt-oss-120b models and Qwen3.6/3.8, retiring the older llama-3.1-8b-instant and llama-3.3-70b-versatile endpoints on August 16, 2026 and the groq/compound and groq/compound-mini endpoints on September 21, 2026 (Groq's own deprecations documentation). Pricing is per million tokens consumed, with a free tier for developers and enterprise plans featuring dedicated capacity and SLA guarantees. In August 2026, Groq became a certified NVIDIA Cloud Partner (NCP), deploying NVIDIA accelerated computing to NVIDIA's reference architecture inside its own data centers alongside its native LPU hardware. Groq does not build or train its own language models.
Who are Groq's main competitors?
Groq's main competitors in the AI inference cloud market are Together AI, Fireworks AI, AWS Inferentia via Amazon Bedrock, Nvidia's NIM inference service, and Cerebras Systems. Against Together AI and Fireworks AI, Groq's primary advantage is raw token generation speed on open-source models; Groq frequently appears in developer benchmarks as the fastest option for Llama and DeepSeek models, trading off a somewhat narrower model selection. Against AWS Inferentia and major cloud provider inference services, Groq wins on speed but loses on integration depth, enterprise support, and the ability to combine inference with cloud-native managed storage and vector databases. Cerebras Systems is the most architecturally similar rival, also building custom inference silicon with on-chip memory, and competes for the same developer segment that prioritizes latency above all else. Nvidia's NIM service, now incorporating Groq 3 LPU-derived chips, creates a unique dynamic where Groq's own architecture is available through a much larger distribution partner. Groq has no offering today for customers needing to fine-tune models, run proprietary models, or combine inference with cloud-native managed services. Tenstorrent is another emerging rival targeting low-latency inference silicon, representing the broader competitive threat from dedicated inference chip companies entering the market in 2025 and 2026.