Groq raises at $3.5bn, half its September mark, with Nvidia inside the round

Groq announced on Monday that it raised $350m at a $3.5bn valuation, down from the $6.9bn it carried last September ^[1][9]. The round was led by Disruptive, whose founder Alex Davis is now Groq's executive chairman, with participation from Nvidia ^[2][3] - the same Nvidia that late last year licensed Groq's language processing unit technology and hired founder and chief executive Jonathan Ross along with much of the senior team ^[4][5].

That sequence is the story, not the discount. Nvidia never bought the company; it took non-exclusive rights to the LPU technology in a deal widely reported at around $20bn and described just as widely as a "not-acqui-hire" ^[4]. Ross, a former Google engineer who worked on that company's tensor chips, went with it ^[5]. The reported price of the licence is roughly 5.7 times what the remaining business is now worth on paper ^[20], which is a reasonable measure of where investors think the value sat.

What was left rebuilt itself around a different business. Co-founder Doug Wightman took over as chief executive and repositioned Groq from chip designer to data-centre operator selling inference by the token ^[6]. Per TechCrunch, that company now operates Nvidia systems, which makes the surviving Groq an Nvidia customer ^[7], with new capital aimed at customers wanting medium and larger clusters of Nvidia accelerated computing for training and inference ^[11]. It runs 13 data centres across North America, Europe, the Middle East and Asia Pacific, serving more than 6 million developers, enterprises and AI-native companies ^[8]. The inference cloud, processing trillions of tokens a week, came with the pivot rather than with the departed chip team ^[10].

The company rejects the down-round framing. A spokesperson told TechCrunch that Groq sees this as establishing a new valuation for the "post-Nvidia-licensing-deal version of Groq" ^[14]. Arithmetically it is a markdown of about 49 percent in under a year ^[13]. Davis said in a statement that the plan is to build "the world's leading AI inference cloud" and that inference will become the largest layer of AI infrastructure ^[12].

The demand case is not the weak part. The weak part is position. Nvidia already supplies the GPUs behind CoreWeave, Lambda and Nebius while investing billions into some of them ^[15], so a friendly, dependent inference supplier part-funded by Nvidia is a familiar shape rather than a distinctive one. The economics of that shape are unsettled: CoreWeave posted strong second-quarter revenue growth and landed contracts including Meta and Anthropic, yet investors stayed focused on heavy capital expenditure, debt reliance, fast-depreciating hardware and the path to free cash flow ^[16]. Whether neoclouds return enough on that investment remains open ^[17]. Groq's own financials are private ^[18].

Meanwhile the thesis Groq abandoned still prices well elsewhere. TNW notes that rival inference-chip startups such as Fractile have raised at buoyant valuations, and that one London challenger recently tripled its worth to $3.3bn while betting openly against Nvidia ^[19]. The repricing here was specific: it followed the loss of founders and engineers, not a collapse in interest in inference silicon.

Three things to watch. First, capacity: Groq intends to go from 54 megawatts to more than 200 by 2027 ^[21], nearly a fourfold build ^[24], while TNW describes the ambition as passing 200 megawatts within a year ^[22]. Second, whether $1bn across June's $650m and this $350m funds that build or is a bridge to debt ^[23][25]. Third, what Nvidia's stake buys the company in practice, since GPU allocation is the variable that decides whether a neocloud grows on schedule.