Nvidia's $20 billion Groq deal is really about blocking Google's TPU momentum

Nvidia is paying around $20 billion for a non-exclusive license to AI chip startup Groq's technology. The deal positions the company against Google's growing TPU threat.

Groq builds specialized processors designed to run large language models. The company's Language Processing Units (LPUs) are optimized specifically for inference tasks—running AI models rather than training them.

As part of the agreement, Groq CEO Jonathan Ross, President Sunny Madra, and other team members will join Nvidia to continue developing the licensed technology. Groq confirmed the deal in a press release but didn't share financial details. Ross co-founded Groq in 2016 after working on TPU chip development at Google.

The $20 billion price tag is roughly three times the $6.9 billion valuation Groq reached during a September funding round. According to CNBC, the deal came together fast. Alex Davis of Disruptive, the lead investor in that last round, confirmed the $20 billion cash payment. Nvidia had $60 billion in cash and equivalents at the end of October, though the company hasn't confirmed the amount or what exactly the payment covers.

Simon Edwards, who became CFO in September, will take over as CEO. Groq will remain independent, and its GroqCloud service will keep running, but the deal looks a lot like an acquisition in disguise. The structure mirrors transactions Microsoft, Google, and Amazon have used over the past two years to bring in talent and technology from AI startups without triggering regulatory scrutiny through formal acquisitions. Investor and Groq co-founder Chamath Palihapitiya wrote on X: "Today we close this almost decade chapter and Jonathan starts a new one with Nvidia."

Google's TPUs are gaining ground on Nvidia's dominance

The transaction targets the growing inference chip market, the processors needed to run AI applications like chatbots. While Nvidia's GPUs remain the go-to choice for training new AI models, cheaper and faster alternatives may work just fine for running them.

"We plan to integrate Groq’s low-latency processors into the NVIDIA AI factory architecture, extending the platform to serve an even broader range of AI inference and real-time workloads," Nvidia CEO Jensen Huang wrote in an internal email to employees, according to CNBC.

Nvidia faces increasing competition from Google's Tensor Processing Units (TPUs), which handle both training and inference. Apple has trained its largest AI models on TPUs rather than Nvidia GPUs, and Anthropic is also leaning more heavily on Google's chips. Meta and OpenAI are building their own specialized chips to reduce their Nvidia dependence.

According to SemiAnalysis, Google's TPUv7 "Ironwood" nearly matches Nvidia's Blackwell generation in theoretical computing power (FLOPs) and memory bandwidth. The real advantage, though, is cost. SemiAnalysis reports that OpenAI negotiated a roughly 30 percent discount on its Nvidia fleet simply by making a credible threat to switch to TPUs or alternatives.

Groq's revenue forecasts had dropped sharply before the deal

The chip startup had been struggling, cutting its 2025 revenue projections by about three-quarters, The Information reports. Back in July, Groq had forecast cloud revenue above $40 million and total revenue exceeding $500 million for the year.

A Groq spokesperson blamed the shortfall on limited data center capacity in a region where the company planned to install more chips. Despite raising around $1.8 billion from investors including BlackRock, Tiger Global Management, Samsung, and Cisco, Groq couldn't crack Nvidia's market dominance.

Nvidia holds its lead partly through its proprietary CUDA programming language, which developers have built their workflows around. Other chip startups are also seeking buyers: Intel is in talks to acquire AI chip startup SambaNova, according to The Information, with an announcement possible as early as next month. Meta bought startup Rivos in October, and AMD acquired the Untether AI team in June.