AI chipmaker Groq raises $650M and re-staffs: what Nvidia's $20B 'not-acqui-hire' deal means

AI inference (model-serving) chip specialist Groq has confirmed it has closed a $650 million funding round. According to TechCrunch, the round was led by BlackRock's tech growth fund; Founders Fund, Tiger Global and Samsung Catalyst Fund also participated. The company's valuation rose to $7.5 billion.

The investment came after a dramatic personnel reshuffle. Earlier this year, Nvidia extracted about 80 senior engineers — including Groq founder Jonathan Ross — and a number of core IP licences from Groq in what has been described as a $20 billion «not-acqui-hire». The structure was designed to avoid a classic acquisition and so sidestep antitrust scrutiny by the FTC and the EU.

The deal's structure is as follows: Nvidia paid Groq $20 billion in cash; in return it received an unlimited cross-licence to Groq's Language Processing Unit (LPU) architecture and brought the engineers, including Ross, onto Nvidia's payroll. Groq carries on as an independent company — but has lost half of its strongest technical team.

Groq's new chief executive, Stuart Pann, told TechCrunch: «This investment funds a re-staffing from the ground up and a next-generation LPU architecture; within three years, we aim to recover the performance advantage our founding team created».

Groq's advantage rests on deterministic inference speed, based on permanently allocated SRAM cache. Compared with Nvidia's H200/B200, it serves tokens 3 to 5 times faster, but it does not have the broad software-stack advantage of Nvidia's CUDA ecosystem.

The term «not-acqui-hire» has emerged in the sector over the past two years. Microsoft's deal with Inflection AI (the Mustafa Suleyman team), Amazon's with Adept and Google's with Character.AI follow a similar pattern, designed to avoid antitrust review. The US FTC and the UK CMA are working to tighten the rules so these structures do not fall outside their scope.

The FTC, in a report published in May, said: «These transactions place de facto mergers beyond the reach of contemporary antitrust oversight». The European Commission classified the Microsoft-Inflection deal in February as «a concentration» and opened a formal investigation.

Groq's new team will focus on three areas: 1) the next generation of the LPU architecture; 2) the software stack (PyTorch integration and a custom compiler); 3) the inference-as-a-service cloud platform. The company already provides API access to more than 12,000 developers.

Jonathan Hofer, head of technology at lead investor BlackRock, said: «Groq's IP and customer base are solid; we are investing in a three-year roadmap under new technical leadership». The investor's read: even with the loss of its founders, Groq can keep its place as one of the most important alternative architectures in the inference market.

Groq's situation reflects a broader debate about how start-ups should compete with dominant incumbents in the AI era. Y Combinator chief Garry Tan posted on X: «Not-acqui-hires monetise founding teams but hollow out the original company; it could make the sector less inviting for the next generation of founders».