NVIDIA has agreed to spend approximately $20 billion to acquire assets from the artificial intelligence chip startup Groq. This is the largest acquisition in the company's history and continues its strategy of "acquiring" potential competitors before they become a threat to NVIDIA's leading position in the market.
This latest licensing agreement represents a repeat of the same approach taken just three months ago, reinforcing the view that decentralized AI infrastructure may be the only alternative to NVIDIA's expansion.
Prices tripled in three months thanks to an association with Trump Jr.
The deal was made just three months after Groq successfully Capital $750 million, valuing the company at $6.9 billion, with participation from BlackRock, Samsung, Cisco, and 1789 Capital, where Donald Trump Jr. was a partner. NVIDIA acquired almost all of Groq's assets, except for its cloud computing business. However, Groq called the deal a "non-exclusive licensing agreement."
Groq's CEO, Jonathan Ross – a former Google engineer and key figure in creating Google's Tensor Processing Unit – will join NVIDIA alongside president Sunny Madra and other senior executives. The startup will continue to operate independently under the leadership of its current CFO, Simon Edwards, who will now become its new CEO.
A recurring scenario
The deal with Groq repeats the pattern of NVIDIA's acquisition of Enfabrica three months earlier. In September 2023, NVIDIA spent over $900 million to attract Enfabrica's CEO and staff, and was granted the right to use the startup's technology. Both deals used a licensing structure instead of a outright purchase, thus avoiding antitrust scrutiny – similar to the $40 billion acquisition of Arm Holdings that was blocked in 2022.
The Kobeissi Letter succinctly summarizes NVIDIA's strategy: "We will buy you before you are strong enough to compete with us."
Technical advantages and competitive pressure
Groq's language processor uses on-chip SRAM instead of external DRAM, optimizing power consumption and claiming up to 10x energy savings. This architecture is powerful for real-time responsive tasks but limits model capacity – a limitation NVIDIA will continue to explore within its broader ecosystem.
This event coincided with Google's announcement of its 7th generation TPU, codenamed Ironwood, and the launch of Gemini 3 – an AI model trained entirely on TPU, currently topping many rankings. NVIDIA responded on X : “We are thrilled about Google’s success… NVIDIA remains a generation ahead of the industry – we are the only platform running all AI models.” As the “big players” begin making such reassuring statements, it’s clear that competitive pressure is increasing significantly.
Implications for Decentralized AI
While this deal doesn't directly impact the cryptocurrency market, it reinforces confidence in decentralized AI computing projects. Platforms like io.net are establishing themselves as alternatives to centralized AI infrastructure .
“People can Chia computing power across the network – whether from a data center or their own personal computer – contributing surplus GPU resources and receiving fair rewards through tokenomics,” Jack Collier, Director of Development at io.net, Chia BeInCrypto . The platform says enterprise clients like Leonardo.ai and UC Berkeley have significantly reduced operating costs by using the service.
However, the gap between expectations and reality remains large. NVIDIA's continued ownership of Groq's low-latency processing technology further helps them maintain their technological leadership, making it increasingly difficult for competitors to catch up with NVIDIA's performance.
This deal also raises serious questions about the independent development of other AI chip companies. Cerebras Systems – another NVIDIA competitor preparing for an IPO – may also face similar pressure. Whether Cerebras can maintain its independence or will eventually be drawn towards NVIDIA's "financial gravity" remains an open question.
