Two years ago, if you asked Silicon Valley who was most likely to challenge Nvidia in the AI chip business, Groq would have been a top answer. Now, Groq has hitched its wagon to Nvidia.
Nvidia's $20B 'Not-Acquisition'
In December, Nvidia signed a $20 billion deal with Groq. But it's not a traditional acquisition. Structured as a "licensing + hiring" arrangement, Nvidia gained rights to Groq's LPU (Language Processing Unit) chip technology and brought in Groq's founder and CEO Jonathan Ross, along with most of the core engineering team. Silicon Valley calls this a "not-an-acquisition."
Why the roundabout approach? A full acquisition would face antitrust scrutiny, while licensing technology and hiring a few people is much cleaner. Nvidia got the technology and talent it wanted while avoiding regulatory hurdles.
Existing shareholders didn't lose out. In February, Groq distributed the first tranche of the deal—$7.6 billion—to shareholders, roughly $64 per share.
What Remains: Groq2 and the 'Inference Cloud'
With the core team gone, the Groq shell remains with a few employees and the brand. They've relaunched as Groq2, with interim CEO Adam Winter and CFO Matt Eng. The new direction is no longer chip design but an AI "inference cloud" (neocloud).
In simple terms, large language models operate in two phases: training and inference. Inference is when you type a query and the model generates a response—the part where compute demand is growing fastest. Groq2 aims to specialize in real-time inference applications, offering an optimized cloud service for developers and enterprises to offload inference workloads.
For this new venture, Groq is raising $650 million. According to Axios, existing investors Disruptive and Infinitum have said they will cover the entire amount even if no one else participates.
From Challenger to Dependent
The most nuanced aspect is this shift:
- Two years ago: Groq designed its own chips to compete head-on with Nvidia.
- Now: Groq2 operates an inference cloud that relies on others' compute.
- Position: From Nvidia challenger to a layer on top of Nvidia hardware.
- Revenue model: From proprietary LPU architecture to optimization and services.
The most valuable chip technology was sold to Nvidia, the founder joined Nvidia, and the remaining company pivots to a business that must run on Nvidia hardware. A company that once aimed to disrupt the table has chosen to sit at it.
This isn't necessarily a bad outcome—the inference cloud is a real business, and $7.6 billion is real money. But the story of "the one most likely to challenge Nvidia" has officially ended.
Sources: After Nvidia's $20B not-acqui-hire, CocoLoop; AI chip startup Groq reportedly raising $650M (TechCrunch); Groq lines up $650m for 'neocloud' spin-out after $20b Nvidia deal (Axios).